Version 3.

0
2/20/2010

NVIDIA CUDA™
Programming Guide






ii CUDA Programming Guide Version 3.0




CUDA Programming Guide Version 3.0 iii

Table of Contents
Chapter 1. Introduction ..................................................................................... 1
1.1 From Graphics Processing to General-Purpose Parallel Computing ................... 1
1.2 CUDA™: a General-Purpose Parallel Computing Architecture ........................... 3
1.3 A Scalable Programming Model ..................................................................... 4
1.4 Document’s Structure ................................................................................... 5
Chapter 2. Programming Model ......................................................................... 7
2.1 Kernels ........................................................................................................ 7
2.2 Thread Hierarchy .......................................................................................... 8
2.3 Memory Hierarchy ...................................................................................... 10
2.4 Heterogeneous Programming ...................................................................... 11
2.5 Compute Capability .................................................................................... 14
Chapter 3. Programming Interface ................................................................. 15
3.1 Compilation with NVCC ............................................................................... 15
3.1.1 Compilation Workflow .......................................................................... 16
3.1.2 Binary Compatibility ............................................................................. 16
3.1.3 PTX Compatibility ................................................................................ 16
3.1.4 Application Compatibility ...................................................................... 17
3.1.5 C/C++ Compatibility ............................................................................ 18
3.2 CUDA C ..................................................................................................... 18
3.2.1 Device Memory .................................................................................... 18
3.2.2 Shared Memory ................................................................................... 21
3.2.3 Multiple Devices ................................................................................... 27
3.2.4 Texture Memory .................................................................................. 28
3.2.4.1 Texture Reference Declaration ....................................................... 29
3.2.4.2 Runtime Texture Reference Attributes ............................................ 29
3.2.4.3 Texture Binding ............................................................................ 30
3.2.5 Page-Locked Host Memory ................................................................... 33
3.2.5.1 Portable Memory ........................................................................... 33



iv CUDA Programming Guide Version 3.0

3.2.5.2 Write-Combining Memory .............................................................. 33
3.2.5.3 Mapped Memory ........................................................................... 33
3.2.6 Asynchronous Concurrent Execution ..................................................... 34
3.2.6.1 Concurrent Execution between Host and Device .............................. 34
3.2.6.2 Overlap of Data Transfer and Kernel Execution ............................... 35
3.2.6.3 Concurrent Kernel Execution .......................................................... 35
3.2.6.4 Concurrent Data Transfers ............................................................. 35
3.2.6.5 Stream ......................................................................................... 35
3.2.6.6 Event ........................................................................................... 37
3.2.6.7 Synchronous Calls ......................................................................... 37
3.2.7 Graphics Interoperability ...................................................................... 37
3.2.7.1 OpenGL Interoperability ................................................................ 38
3.2.7.2 Direct3D Interoperability ............................................................... 40
3.2.8 Error Handling ..................................................................................... 46
3.2.9 Debugging using the Device Emulation Mode ........................................ 47
3.3 Driver API .................................................................................................. 49
3.3.1 Context ............................................................................................... 51
3.3.2 Module ................................................................................................ 52
3.3.3 Kernel Execution .................................................................................. 52
3.3.4 Device Memory .................................................................................... 54
3.3.5 Shared Memory ................................................................................... 57
3.3.6 Multiple Devices ................................................................................... 58
3.3.7 Texture Memory .................................................................................. 58
3.3.8 Page-Locked Host Memory ................................................................... 60
3.3.9 Asynchronous Concurrent Execution ..................................................... 61
3.3.9.1 Stream ......................................................................................... 61
3.3.9.2 Event Management ....................................................................... 62
3.3.9.3 Synchronous Calls ......................................................................... 63
3.3.10 Graphics Interoperability ...................................................................... 63
3.3.10.1 OpenGL Interoperability ................................................................ 63
3.3.10.2 Direct3D Interoperability ............................................................... 65
3.3.11 Error Handling ..................................................................................... 72
3.4 Interoperability between Runtime and Driver APIs ........................................ 72



CUDA Programming Guide Version 3.0 v

3.5 Versioning and Compatibility ....................................................................... 73
3.6 Compute Modes ......................................................................................... 74
3.7 Mode Switches ........................................................................................... 74
Chapter 4. Hardware Implementation ............................................................ 77
4.1 SIMT Architecture ....................................................................................... 77
4.2 Hardware Multithreading ............................................................................. 78
4.3 Multiple Devices ......................................................................................... 79
Chapter 5. Performance Guidelines ................................................................. 81
5.1 Overall Performance Optimization Strategies ................................................ 81
5.2 Maximize Utilization .................................................................................... 81
5.2.1 Application Level .................................................................................. 81
5.2.2 Device Level ........................................................................................ 82
5.2.3 Multiprocessor Level ............................................................................ 82
5.3 Maximize Memory Throughput .................................................................... 84
5.3.1 Data Transfer between Host and Device ............................................... 85
5.3.2 Device Memory Accesses ...................................................................... 85
5.3.2.1 Global Memory .............................................................................. 86
5.3.2.2 Local Memory ............................................................................... 87
5.3.2.3 Shared Memory ............................................................................ 88
5.3.2.4 Constant Memory .......................................................................... 88
5.3.2.5 Texture Memory ........................................................................... 89
5.4 Maximize Instruction Throughput ................................................................ 89
5.4.1 Arithmetic Instructions ......................................................................... 90
5.4.2 Control Flow Instructions ..................................................................... 92
5.4.3 Synchronization Instruction .................................................................. 93
Appendix A. CUDA-Enabled GPUs .................................................................... 95
Appendix B. C Language Extensions ................................................................ 97
B.1 Function Type Qualifiers ............................................................................. 97
B.1.1 __device__ .......................................................................................... 97
B.1.2 __global__ .......................................................................................... 97
B.1.3 __host__ ............................................................................................. 97
B.1.4 Restrictions ......................................................................................... 98
B.2 Variable Type Qualifiers .............................................................................. 98



vi CUDA Programming Guide Version 3.0

B.2.1 __device__ .......................................................................................... 98
B.2.2 __constant__ ...................................................................................... 98
B.2.3 __shared__ ......................................................................................... 99
B.2.4 volatile ................................................................................................ 99
B.2.5 Restrictions ....................................................................................... 100
B.3 Built-in Vector Types ................................................................................ 101
B.3.1 char1, uchar1, char2, uchar2, char3, uchar3, char4, uchar4, short1,
ushort1, short2, ushort2, short3, ushort3, short4, ushort4, int1, uint1, int2, uint2,
int3, uint3, int4, uint4, long1, ulong1, long2, ulong2, long3, ulong3, long4, ulong4,
longlong1, longlong2, float1, float2, float3, float4, double1, double2 .................. 101
B.3.2 dim3 ................................................................................................. 102
B.4 Built-in Variables ...................................................................................... 102
B.4.1 gridDim ............................................................................................. 102
B.4.2 blockIdx ............................................................................................ 102
B.4.3 blockDim ........................................................................................... 102
B.4.4 threadIdx .......................................................................................... 102
B.4.5 warpSize ........................................................................................... 103
B.4.6 Restrictions ....................................................................................... 103
B.5 Memory Fence Functions .......................................................................... 103
B.6 Synchronization Functions ......................................................................... 104
B.7 Mathematical Functions ............................................................................ 105
B.8 Texture Functions ..................................................................................... 106
B.8.1 tex1Dfetch() ...................................................................................... 106
B.8.2 tex1D() ............................................................................................. 107
B.8.3 tex2D() ............................................................................................. 107
B.8.4 tex3D() ............................................................................................. 107
B.9 Time Function .......................................................................................... 107
B.10 Atomic Functions ...................................................................................... 108
B.10.1 Arithmetic Functions .......................................................................... 108
B.10.1.1 atomicAdd() ................................................................................ 108
B.10.1.2 atomicSub() ................................................................................ 108
B.10.1.3 atomicExch() .............................................................................. 109
B.10.1.4 atomicMin() ................................................................................ 109
B.10.1.5 atomicMax() ............................................................................... 109



CUDA Programming Guide Version 3.0 vii

B.10.1.6 atomicInc() ................................................................................. 109
B.10.1.7 atomicDec() ................................................................................ 109
B.10.1.8 atomicCAS() ............................................................................... 109
B.10.2 Bitwise Functions ............................................................................... 110
B.10.2.1 atomicAnd() ................................................................................ 110
B.10.2.2 atomicOr() .................................................................................. 110
B.10.2.3 atomicXor() ................................................................................ 110
B.11 Warp Vote Functions ................................................................................ 110
B.12 Profiler Counter Function .......................................................................... 111
B.13 Execution Configuration ............................................................................ 111
B.14 Launch Bounds ......................................................................................... 112
Appendix C. Mathematical Functions ............................................................ 115
C.1 Standard Functions ................................................................................... 115
C.1.1 Single-Precision Floating-Point Functions ............................................. 115
C.1.2 Double-Precision Floating-Point Functions ........................................... 118
C.1.3 Integer Functions .............................................................................. 120
C.2 Intrinsic Functions .................................................................................... 120
C.2.1 Single-Precision Floating-Point Functions ............................................. 120
C.2.2 Double-Precision Floating-Point Functions ........................................... 122
C.2.3 Integer Functions .............................................................................. 123
Appendix D. C++ Language Constructs ........................................................ 125
D.1 Polymorphism .......................................................................................... 125
D.2 Default Parameters ................................................................................... 126
D.3 Operator Overloading ............................................................................... 126
D.4 Namespaces ............................................................................................. 127
D.5 Function Templates .................................................................................. 127
D.6 Classes .................................................................................................... 128
D.6.1 Example 1 Pixel Data Type ................................................................. 128
D.6.2 Example 2 Functor Class .................................................................... 129
Appendix E. NVCC Specifics ........................................................................... 131
E.1 __noinline__ ............................................................................................ 131
E.2 #pragma unroll ........................................................................................ 131
E.3 __restrict__ ............................................................................................. 132



viii CUDA Programming Guide Version 3.0

Appendix F. Texture Fetching ........................................................................ 135
F.1 Nearest-Point Sampling............................................................................. 136
F.2 Linear Filtering ......................................................................................... 136
F.3 Table Lookup ........................................................................................... 138
Appendix G. Compute Capabilities ................................................................ 139
G.1 Features and Technical Specifications ........................................................ 139
G.2 Floating-Point Standard ............................................................................ 140
G.3 Compute Capability 1.x ............................................................................. 142
G.3.1 Architecture ....................................................................................... 142
G.3.2 Global Memory .................................................................................. 143
G.3.2.1 Devices of Compute Capability 1.0 and 1.1 ................................... 143
G.3.2.2 Devices of Compute Capability 1.2 and 1.3 ................................... 144
G.3.3 Shared Memory ................................................................................. 144
G.3.3.1 32-Bit Strided Access ................................................................... 145
G.3.3.2 32-Bit Broadcast Access ............................................................... 145
G.3.3.3 8-Bit and 16-Bit Access ................................................................ 145
G.3.3.4 Larger Than 32-Bit Access ........................................................... 146
G.4 Compute Capability 2.0 ............................................................................. 146
G.4.1 Architecture ....................................................................................... 146
G.4.2 Global Memory .................................................................................. 148
G.4.3 Shared Memory ................................................................................. 150
G.4.3.1 32-Bit Strided Access ................................................................... 150
G.4.3.2 Larger Than 32-Bit Access ........................................................... 150
G.4.4 Constant Memory .............................................................................. 151




CUDA Programming Guide Version 3.0 ix

List of Figures
Figure 1-1. Floating-Point Operations per Second and Memory Bandwidth for the CPU
and GPU 2
Figure 1-2. The GPU Devotes More Transistors to Data Processing ............................ 3
Figure 1-3. CUDA is Designed to Support Various Languages or Application
Programming Interfaces ...................................................................................... 4
Figure 1-4. Automatic Scalability .............................................................................. 5
Figure 2-1. Grid of Thread Blocks ............................................................................. 9
Figure 2-2. Memory Hierarchy ............................................................................... 11
Figure 2-3. Heterogeneous Programming ............................................................... 13
Figure 3-1. Matrix Multiplication without Shared Memory ......................................... 23
Figure 3-2. Matrix Multiplication with Shared Memory.............................................. 27
Figure 3-3. Library Context Management ................................................................ 52
Figure 3-4. The Driver API is Backward, but Not Forward Compatible ....................... 74

Figure E-1. Nearest-Point Sampling of a One-Dimensional Texture of Four Texels ... 136
Figure E-2. Linear Filtering of a One-Dimensional Texture of Four Texels in Clamp
Addressing Mode ............................................................................................ 137
Figure E-3. One-Dimensional Table Lookup Using Linear Filtering .......................... 138
Figure G-1. Examples of Global Memory Accesses by a Warp, 4-Byte Word per Thread,
and Associated Memory Transactions Based on Compute Capability ................... 149
Figure G-2 Examples of Strided Shared Memory Accesses for Devices of Compute
Capability 2.0 .................................................................................................. 152
Figure G-3 Examples of Irregular and Broadcast Shared Memory Accesses for Devices
of Compute Capability 2.0................................................................................ 153







CUDA Programming Guide Version 3.0 1

Chapter 1.
Introduction
1.1 From Graphics Processing to
General-Purpose Parallel Computing
Driven by the insatiable market demand for realtime, high-definition 3D graphics,
the programmable Graphic Processor Unit or GPU has evolved into a highly
parallel, multithreaded, manycore processor with tremendous computational
horsepower and very high memory bandwidth, as illustrated by Figure 1-1.
Chapter 1. Introduction


2 CUDA Programming Guide Version 3.0



Figure 1-1. Floating-Point Operations per Second and
Memory Bandwidth for the CPU and GPU

The reason behind the discrepancy in floating-point capability between the CPU and
the GPU is that the GPU is specialized for compute-intensive, highly parallel
computation – exactly what graphics rendering is about – and therefore designed
such that more transistors are devoted to data processing rather than data caching
and flow control, as schematically illustrated by Figure 1-2.
NV30
Jan
2003
Jun Apr
2004 2005 2007 2006 2008
NV35
NV40
G70
G71
G80
G92
GT200
Jun Nov Mar May Jun
GT200 = GeForce GTX 280
G92 = GeForce 9800 GTX
G80 = GeForce 8800 GTX
G71 = GeForce 7900 GTX
G70 = GeForce 7800 GTX
NV40 = GeForce 6800 Ultra
NV35 = GeForce FX 5950 Ultra
NV30 = GeForce FX 5800
3.0 GHz
Core2 Duo
3.2 GHz
Harpertown
NV30
NV40
G71
G80
G80
Ultra
G80
Ultra
Northwood
Prescott EE
Woodcrest
Harpertown
Chapter 1. Introduction


CUDA Programming Guide Version 3.0 3



Figure 1-2. The GPU Devotes More Transistors to Data
Processing

More specifically, the GPU is especially well-suited to address problems that can be
expressed as data-parallel computations – the same program is executed on many
data elements in parallel – with high arithmetic intensity – the ratio of arithmetic
operations to memory operations. Because the same program is executed for each
data element, there is a lower requirement for sophisticated flow control, and
because it is executed on many data elements and has high arithmetic intensity, the
memory access latency can be hidden with calculations instead of big data caches.
Data-parallel processing maps data elements to parallel processing threads. Many
applications that process large data sets can use a data-parallel programming model
to speed up the computations. In 3D rendering, large sets of pixels and vertices are
mapped to parallel threads. Similarly, image and media processing applications such
as post-processing of rendered images, video encoding and decoding, image scaling,
stereo vision, and pattern recognition can map image blocks and pixels to parallel
processing threads. In fact, many algorithms outside the field of image rendering
and processing are accelerated by data-parallel processing, from general signal
processing or physics simulation to computational finance or computational biology.
1.2 CUDA™: a General-Purpose Parallel
Computing Architecture
In November 2006, NVIDIA introduced CUDA™, a general purpose parallel
computing architecture – with a new parallel programming model and instruction
set architecture – that leverages the parallel compute engine in NVIDIA GPUs to
solve many complex computational problems in a more efficient way than on a
CPU.
CUDA comes with a software environment that allows developers to use C as a
high-level programming language. As illustrated by Figure 1-3, other languages or
application programming interfaces are supported, such as CUDA FORTRAN,
OpenCL, and DirectCompute.
Cache
ALU
Control
ALU
ALU
ALU
DRAM
CPU
DRAM








GPU
Chapter 1. Introduction


4 CUDA Programming Guide Version 3.0


Figure 1-3. CUDA is Designed to Support Various Languages
or Application Programming Interfaces
1.3 A Scalable Programming Model
The advent of multicore CPUs and manycore GPUs means that mainstream
processor chips are now parallel systems. Furthermore, their parallelism continues
to scale with Moore‟s law. The challenge is to develop application software that
transparently scales its parallelism to leverage the increasing number of processor
cores, much as 3D graphics applications transparently scale their parallelism to
manycore GPUs with widely varying numbers of cores.
The CUDA parallel programming model is designed to overcome this challenge
while maintaining a low learning curve for programmers familiar with standard
programming languages such as C.
At its core are three key abstractions – a hierarchy of thread groups, shared
memories, and barrier synchronization – that are simply exposed to the programmer
as a minimal set of language extensions.
These abstractions provide fine-grained data parallelism and thread parallelism,
nested within coarse-grained data parallelism and task parallelism. They guide the
programmer to partition the problem into coarse sub-problems that can be solved
independently in parallel by blocks of threads, and each sub-problem into finer
pieces that can be solved cooperatively in parallel by all threads within the block.
This decomposition preserves language expressivity by allowing threads to
cooperate when solving each sub-problem, and at the same time enables automatic
scalability. Indeed, each block of threads can be scheduled on any of the available
processor cores, in any order, concurrently or sequentially, so that a compiled
CUDA program can execute on any number of processor cores as illustrated by
Figure 1-4, and only the runtime system needs to know the physical processor
count.
Chapter 1. Introduction


CUDA Programming Guide Version 3.0 5

This scalable programming model allows the CUDA architecture to span a wide
market range by simply scaling the number of processors and memory partitions:
from the high-performance enthusiast GeForce GPUs and professional Quadro and
Tesla computing products to a variety of inexpensive, mainstream GeForce GPUs
(see Appendix A for a list of all CUDA-enabled GPUs).


A multithreaded program is partitioned into blocks of threads that execute independently from each
other, so that a GPU with more cores will automatically execute the program in less time than a GPU
with fewer cores.
Figure 1-4. Automatic Scalability
1.4 Document’s Structure
This document is organized into the following chapters:
 Chapter 1 is a general introduction to CUDA.
 Chapter 2 outlines the CUDA programming model.
 Chapter 3 describes the programming interface.
 Chapter 4 describes the hardware implementation.
 Chapter 5 gives some guidance on how to achieve maximum performance.
 Appendix A lists all CUDA-enabled devices.
GPU with 2 Cores

Core 1 Core 0
GPU with 4 Cores

Core 1 Core 0 Core 3 Core 2
Block 5 Block 6
Multithreaded CUDA Program
Block 0 Block 1 Block 2 Block 3
Block 4 Block 5 Block 6 Block 7

Block 1 Block 0

Block 3 Block 2

Block 5 Block 4

Block 7 Block 6

Block 0 Block 1 Block 2 Block 3

Block 4 Block 5 Block 6 Block 7
Chapter 1. Introduction


6 CUDA Programming Guide Version 3.0

 Appendix B is a detailed description of all extensions to the C language.
 Appendix C lists the mathematical functions supported in CUDA.
 Appendix D lists the C++ constructs supported in device code.
 Appendix E lists the specific keywords and directives supported by nvcc.
 Appendix F gives more details on texture fetching.
 Appendix G gives the technical specifications of various devices, as well as
more architectural details.







CUDA Programming Guide Version 3.0 7

Chapter 2.
Programming Model
This chapter introduces the main concepts behind the CUDA programming model
by outlining how they are exposed in C. An extensive description of CUDA C is
given in Section 3.2.
Full code for the vector addition example used in this chapter and the next can be
found in the vectorAdd SDK code sample.
2.1 Kernels
CUDA C extends C by allowing the programmer to define C functions, called
kernels, that, when called, are executed N times in parallel by N different CUDA
threads, as opposed to only once like regular C functions.
A kernel is defined using the __global__ declaration specifier and the number of
CUDA threads that execute that kernel for a given kernel call is specified using a
new <<<…>>> execution configuration syntax (see Appendix B.13). Each thread that
executes the kernel is given a unique thread ID that is accessible within the kernel
through the built-in threadIdx variable.
As an illustration, the following sample code adds two vectors A and B of size N
and stores the result into vector C:
// Kernel definition
__global__ void VecAdd(float* A, float* B, float* C)
{
int i = threadIdx.x;
C[i] = A[i] + B[i];
}

int main()
{
...
// Kernel invocation with N threads
VecAdd<<<1, N>>>(A, B, C);
}
Here, each of the N threads that execute VecAdd() performs one pair-wise
addition.
Chapter 2. Programming Model


8 CUDA Programming Guide Version 3.0

2.2 Thread Hierarchy
For convenience, threadIdx is a 3-component vector, so that threads can be
identified using a one-dimensional, two-dimensional, or three-dimensional thread
index, forming a one-dimensional, two-dimensional, or three-dimensional thread
block. This provides a natural way to invoke computation across the elements in a
domain such as a vector, matrix, or volume.
The index of a thread and its thread ID relate to each other in a straightforward
way: For a one-dimensional block, they are the same; for a two-dimensional block
of size (D
x
, D
y
), the thread ID of a thread of index (x, y) is (x + y D
x
); for a three-
dimensional block of size (D
x
, D
y
, D
z
), the thread ID of a thread of index (x, y, z) is
(x + y D
x
+ z D
x
D
y
).
As an example, the following code adds two matrices A and B of size NxN and
stores the result into matrix C:
// Kernel definition
__global__ void MatAdd(float A[N][N], float B[N][N],
float C[N][N])
{
int i = threadIdx.x;
int j = threadIdx.y;
C[i][j] = A[i][j] + B[i][j];
}

int main()
{
...
// Kernel invocation with one block of N * N * 1 threads
int numBlocks = 1;
dim3 threadsPerBlock(N, N);
MatAdd<<<numBlocks, threadsPerBlock>>>(A, B, C);
}
There is a limit to the number of threads per block, since all threads of a block are
expected to reside on the same processor core and must share the limited memory
resources of that core. On current GPUs, a thread block may contain up to 512
threads.
However, a kernel can be executed by multiple equally-shaped thread blocks, so that
the total number of threads is equal to the number of threads per block times the
number of blocks.
Blocks are organized into a one-dimensional or two-dimensional grid of thread
blocks as illustrated by Figure 2-1. The number of thread blocks in a grid is usually
dictated by the size of the data being processed or the number of processors in the
system, which it can greatly exceed.

Chapter 2: Programming Model


CUDA Programming Guide Version 3.0 9


Figure 2-1. Grid of Thread Blocks

The number of threads per block and the number of blocks per grid specified in the
<<<…>>> syntax can be of type int or dim3. Two-dimensional blocks or grids can
be specified as in the example above.
Each block within the grid can be identified by a one-dimensional or two-
dimensional index accessible within the kernel through the built-in blockIdx
variable. The dimension of the thread block is accessible within the kernel through
the built-in blockDim variable.
Extending the previous MatAdd() example to handle multiple blocks, the code
becomes as follows.
// Kernel definition
__global__ void MatAdd(float A[N][N], float B[N][N],
float C[N][N])
{
int i = blockIdx.x * blockDim.x + threadIdx.x;
int j = blockIdx.y * blockDim.y + threadIdx.y;
if (i < N && j < N)
C[i][j] = A[i][j] + B[i][j];
Grid
Block (1, 1)
Thread (0, 0) Thread (1, 0) Thread (2, 0) Thread (3, 0)
Thread (0, 1) Thread (1, 1) Thread (2, 1) Thread (3, 1)
Thread (0, 2) Thread (1, 2) Thread (2, 2) Thread (3, 2)
Block (2, 1) Block (1, 1) Block (0, 1)
Block (2, 0) Block (1, 0) Block (0, 0)
Chapter 2. Programming Model


10 CUDA Programming Guide Version 3.0

}

int main()
{
...
// Kernel invocation
dim3 threadsPerBlock(16, 16);
dim3 numBlocks(N / threadsPerBlock.x, N / threadsPerBlock.y);
MatAdd<<<numBlocks, threadsPerBlock>>>(A, B, C);
}
A thread block size of 16x16 (256 threads), although arbitrary in this case, is a
common choice. The grid is created with enough blocks to have one thread per
matrix element as before. For simplicity, this example assumes that the number of
threads per grid in each dimension is evenly divisible by the number of threads per
block in that dimension, although that need not be the case.
Thread blocks are required to execute independently: It must be possible to execute
them in any order, in parallel or in series. This independence requirement allows
thread blocks to be scheduled in any order across any number of cores as illustrated
by Figure 1-4, enabling programmers to write code that scales with the number of
cores.
Threads within a block can cooperate by sharing data through some shared memory
and by synchronizing their execution to coordinate memory accesses. More
precisely, one can specify synchronization points in the kernel by calling the
__syncthreads() intrinsic function; __syncthreads() acts as a barrier at
which all threads in the block must wait before any is allowed to proceed.
Section 3.2.2 gives an example of using shared memory.
For efficient cooperation, the shared memory is expected to be a low-latency
memory near each processor core (much like an L1 cache) and __syncthreads()
is expected to be lightweight.
2.3 Memory Hierarchy
CUDA threads may access data from multiple memory spaces during their
execution as illustrated by Figure 2-2. Each thread has private local memory. Each
thread block has shared memory visible to all threads of the block and with the
same lifetime as the block. All threads have access to the same global memory.
There are also two additional read-only memory spaces accessible by all threads: the
constant and texture memory spaces. The global, constant, and texture memory
spaces are optimized for different memory usages (see Sections 5.3.2.1, 5.3.2.4, and
5.3.2.5). Texture memory also offers different addressing modes, as well as data
filtering, for some specific data formats (see Section 3.2.4).
The global, constant, and texture memory spaces are persistent across kernel
launches by the same application.
Chapter 2: Programming Model


CUDA Programming Guide Version 3.0 11


Figure 2-2. Memory Hierarchy
2.4 Heterogeneous Programming
As illustrated by Figure 2-3, the CUDA programming model assumes that the
CUDA threads execute on a physically separate device that operates as a coprocessor
to the host running the C program. This is the case, for example, when the kernels
execute on a GPU and the rest of the C program executes on a CPU.














Global memory
Grid 0
Block (2, 1) Block (1, 1) Block (0, 1)
Block (2, 0) Block (1, 0) Block (0, 0)
Grid 1
Block (1, 1)
Block (1, 0)
Block (1, 2)
Block (0, 1)
Block (0, 0)
Block (0, 2)
Thread Block

Per-block shared
memory
Thread
Per-thread local
memory
Chapter 2. Programming Model


12 CUDA Programming Guide Version 3.0

The CUDA programming model also assumes that both the host and the device
maintain their own separate memory spaces in DRAM, referred to as host memory and
device memory, respectively. Therefore, a program manages the global, constant, and
texture memory spaces visible to kernels through calls to the CUDA runtime
(described in Chapter 3). This includes device memory allocation and deallocation as
well as data transfer between host and device memory.
Chapter 2: Programming Model


CUDA Programming Guide Version 3.0 13


Serial code executes on the host while parallel code executes on the device.
Figure 2-3. Heterogeneous Programming
Device
Grid 0
Block (2, 1) Block (1, 1) Block (0, 1)
Block (2, 0) Block (1, 0) Block (0, 0)
Host
C Program
Sequential
Execution

Serial code


Parallel kernel
Kernel0<<<>>>()








Serial code




Parallel kernel
Kernel1<<<>>>()



Host
Device
Grid 1
Block (1, 1)
Block (1, 0)
Block (1, 2)
Block (0, 1)
Block (0, 0)
Block (0, 2)
Chapter 2. Programming Model


14 CUDA Programming Guide Version 3.0

2.5 Compute Capability
The compute capability of a device is defined by a major revision number and a minor
revision number.
Devices with the same major revision number are of the same core architecture. The
major revision number of devices based on the Fermi architecture is 2. Prior devices
are all of compute capability 1.x (Their major revision number is 1).
The minor revision number corresponds to an incremental improvement to the core
architecture, possibly including new features.
Appendix A lists of all CUDA-enabled devices along with their compute capability.
Appendix G gives the technical specifications of each compute capability.







CUDA Programming Guide Version 3.0 15

Chapter 3.
Programming Interface
Two interfaces are currently supported to write CUDA programs: CUDA C and the
CUDA driver API. An application typically uses either one or the other, but it can
use both under the limitations described in Section 3.4.
CUDA C exposes the CUDA programming model as a minimal set of extensions to
the C language. Any source file that contains some of these extensions must be
compiled with nvcc as outlined in Section 3.1. These extensions allow
programmers to define a kernel as a C function and use some new syntax to specify
the grid and block dimension each time the function is called.
The CUDA driver API is a lower-level C API that provides functions to load
kernels as modules of CUDA binary or assembly code, to inspect their parameters,
and to launch them. Binary and assembly codes are usually obtained by compiling
kernels written in C.
CUDA C comes with a runtime API and both the runtime API and the driver API
provide functions to allocate and deallocate device memory, transfer data between
host memory and device memory, manage systems with multiple devices, etc.
The runtime API is built on top of the CUDA driver API. Initialization, context,
and module management are all implicit and resulting code is more concise. CUDA
C also supports device emulation, which facilitates debugging (see Section 3.2.8).
In contrast, the CUDA driver API requires more code, is harder to program and
debug, but offers a better level of control and is language-independent since it
handles binary or assembly code.
Section 3.2 continues the description of CUDA C started in Chapter 2. It also
introduces concepts that are common to both CUDA C and the driver API: linear
memory, CUDA arrays, shared memory, texture memory, page-locked host
memory, device enumeration, asynchronous execution, interoperability with
graphics APIs. Section 3.3 assumes knowledge of these concepts and describes how
they are exposed by the driver API.
3.1 Compilation with NVCC
Kernels can be written using the CUDA instruction set architecture, called PTX,
which is described in the PTX reference manual. It is however usually more
Chapter 3. Programming Interface


16 CUDA Programming Guide Version 3.0

effective to use a high-level programming language such as C. In both cases, kernels
must be compiled into binary code by nvcc to execute on the device.
nvcc is a compiler driver that simplifies the process of compiling C or PTX code: It
provides simple and familiar command line options and executes them by invoking
the collection of tools that implement the different compilation stages. This section
gives an overview of nvcc workflow and command options. A complete
description can be found in the nvcc user manual.
3.1.1 Compilation Workflow
Source files compiled with nvcc can include a mix of host code (i.e. code that
executes on the host) and device code (i.e. code that executes on the device). nvcc‟s
basic workflow consists in separating device code from host code and compiling the
device code into an assembly form (PTX code) and/or binary form (cubin object).
The generated host code is output either as C code that is left to be compiled using
another tool or as object code directly by letting nvcc invoke the host compiler
during the last compilation stage.
Applications can then:
 Either load and execute the PTX code or cubin object on the device using the
CUDA driver API (see Section 3.3) and ignore the generated host code (if any);
 Or link to the generated host code; the generated host code includes the PTX
code and/or cubin object as a global initialized data array and a translation of the
<<<…>>> syntax introduced in Section 2.1 (and described in more details in
Section B.13) into the necessary CUDA C runtime function calls to load and
launch each compiled kernel.
Any PTX code loaded by an application at runtime is compiled further to binary
code by the device driver. This is called just-in-time compilation. Just-in-time
compilation increases application load time, but allow applications to benefit from
latest compiler improvements. It is also the only way for applications to run on
devices that did not exist at the time the application was compiled, as detailed in
Section 3.1.4.
3.1.2 Binary Compatibility
Binary code is architecture-specific. A cubin object is generated using the compiler
option –code that specifies the targeted architecture: For example, compiling with
–code=sm_13 produces binary code for devices of compute capability 1.3. Binary
compatibility is guaranteed from one minor revision to the next one, but not from
one minor revision to the previous one or across major revisions. In other words, a
cubin object generated for compute capability X.y is only guaranteed to execute on
devices of compute capability X.z where z≥y.
3.1.3 PTX Compatibility
Some PTX instructions are only supported on devices of higher compute
capabilities. For example, atomic instructions on global memory are only supported
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 17

on devices of compute capability 1.1 and above; double-precision instructions are
only supported on devices of compute capability 1.3 and above. The –arch
compiler option specifies the compute capability that is assumed when compiling C
to PTX code. So, code that contains double-precision arithmetic, for example, must
be compiled with “-arch=sm_13” (or higher compute capability), otherwise
double-precision arithmetic will get demoted to single-precision arithmetic.
PTX code produced for some specific compute capability can always be compiled to
binary code of greater or equal compute capability.
3.1.4 Application Compatibility
To execute code on devices of specific compute capability, an application must load
binary or PTX code that is compatible with this compute capability as described in
Sections 3.1.2 and 3.1.3. In particular, to be able to execute code on future
architectures with higher compute capability – for which no binary code can be
generated yet –, an application must load PTX code that will be compiled just-in-
time for these devices.
Which PTX and binary code gets embedded in a CUDA C application is controlled
by the –arch and –code compiler options or the –gencode compiler option as
detailed in the nvcc user manual. For example,
nvcc x.cu
–gencode arch=compute_10,code=sm_10
–gencode arch=compute_11,code=\’compute_11,sm_11\’
embeds binary code compatible with compute capability 1.0 (first –gencode
option) and PTX and binary code compatible with compute capability 1.1 (second
-gencode option).
Host code is generated to automatically select at runtime the most appropriate code
to load and execute, which, in the above example, will be:
 1.0 binary code for devices with compute capability 1.0,
 1.1 binary code for devices with compute capability 1.1, 1.2, 1.3,
 binary code obtained by compiling 1.1 PTX code for devices with compute
capabilities 2.0 or higher.
x.cu can have an optimized code path that uses atomic operations, for example,
which are only supported in devices of compute capability 1.1 and higher. The
__CUDA_ARCH__ macro can be used to differentiate various code paths based on
compute capability. It is only defined for device code. When compiling with
“arch=compute_11” for example, __CUDA_ARCH__ is equal to 110.
Applications using the driver API must compile code to separate files and explicitly
load and execute the most appropriate file at runtime.
The nvcc user manual lists various shorthands for the –arch, –code, and
÷gencode compiler options. For example, “÷arch=sm_13” is a shorthand for
“÷arch=compute_13 ÷code=compute_13,sm_13” (which is the same as
“÷gencode arch=compute_13,code=\’compute_13,sm_13\’”).
Chapter 3. Programming Interface


18 CUDA Programming Guide Version 3.0

3.1.5 C/C++ Compatibility
The front end of the compiler processes CUDA source files according to C++
syntax rules. Full C++ is supported for the host code. However, only a subset of
C++ is fully supported for the device code as described in detail in Appendix D. As
a consequence of the use of C++ syntax rules, void pointers (e.g., returned by
malloc()) cannot be assigned to non-void pointers without a typecast.
nvcc also support specific keywords and directives detailed in Appendix E.
3.2 CUDA C
CUDA C provides a simple path for users familiar with the C programming
language to easily write programs for execution by the device.
It consists of a minimal set of extensions to the C language and a runtime library.
The core language extensions have been introduced in Chapter 2. This section
continues with an introduction to the runtime. A complete description of all
extensions can be found in Appendix B and a complete description of the runtime
in the CUDA reference manual.
The runtime is implemented in the cudart dynamic library and all its entry points
are prefixed with cuda.
There is no explicit initialization function for the runtime; it initializes the first time
a runtime function is called (more specifically any function other than functions
from the device and version management sections of the reference manual). One
needs to keep this in mind when timing runtime function calls and when
interpreting the error code from the first call into the runtime.
Once the runtime has been initialized in a host thread, any resource (memory,
stream, event, etc.) allocated via some runtime function call in the host thread is
only valid within the context of the host thread. Therefore only runtime functions
calls made by the host thread (memory copies, kernel launches, …) can operate on
these resources. This is because a CUDA context (see Section 3.3.1) is created under
the hood as part of initialization and made current to the host thread, and it cannot
be made current to any other host thread.
On system with multiple devices, kernels are executed on device 0 by default as
detailed in Section 3.2.3.
3.2.1 Device Memory
As mentioned in Section 2.4, the CUDA programming model assumes a system
composed of a host and a device, each with their own separate memory. Kernels
can only operate out of device memory, so the runtime provides functions to
allocate, deallocate, and copy device memory, as well as transfer data between host
memory and device memory.
Device memory can be allocated either as linear memory or as CUDA arrays.
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 19

CUDA arrays are opaque memory layouts optimized for texture fetching. They are
described in Section 3.2.4.
Linear memory exists on the device in a 32-bit address space for devices of compute
capability 1.x and 40-bit address space of devices of compute capability 2.0, so
separately allocated entities can reference one another via pointers, for example, in a
binary tree.
Linear memory is typically allocated using cudaMalloc() and freed using
cudaFree() and data transfer between host memory and device memory are
typically done using cudaMemcpy(). In the vector addition code sample of
Section 2.1, the vectors need to be copied from host memory to device memory:
// Device code
__global__ void VecAdd(float* A, float* B, float* C, int N)
{
int i = blockDim.x * blockIdx.x + threadIdx.x;
if (i < N)
C[i] = A[i] + B[i];
}

// Host code
int main()
{
int N = ...;
size_t size = N * sizeof(float);

// Allocate input vectors h_A and h_B in host memory
float* h_A = (float*)malloc(size);
float* h_B = (float*)malloc(size);

// Initialize input vectors
...

// Allocate vectors in device memory
float* d_A;
cudaMalloc((void**)&d_A, size);
float* d_B;
cudaMalloc((void**)&d_B, size);
float* d_C;
cudaMalloc((void**)&d_C, size);

// Copy vectors from host memory to device memory
cudaMemcpy(d_A, h_A, size, cudaMemcpyHostToDevice);
cudaMemcpy(d_B, h_B, size, cudaMemcpyHostToDevice);

// Invoke kernel
int threadsPerBlock = 256;
int blocksPerGrid =
(N + threadsPerBlock – 1) / threadsPerBlock;
VecAdd<<<blocksPerGrid, threadsPerBlock>>>(d_A, d_B, d_C, N);

// Copy result from device memory to host memory
// h_C contains the result in host memory
cudaMemcpy(h_C, d_C, size, cudaMemcpyDeviceToHost);

// Free device memory
Chapter 3. Programming Interface


20 CUDA Programming Guide Version 3.0

cudaFree(d_A);
cudaFree(d_B);
cudaFree(d_C);

// Free host memory
...
}
Linear memory can also be allocated through cudaMallocPitch() and
cudaMalloc3D(). These functions are recommended for allocations of 2D or 3D
arrays as it makes sure that the allocation is appropriately padded to meet the
alignment requirements described in Section 5.3.2.1, therefore ensuring best
performance when accessing the row addresses or performing copies between 2D
arrays and other regions of device memory (using the cudaMemcpy2D() and
cudaMemcpy3D() functions). The returned pitch (or stride) must be used to access
array elements. The following code sample allocates a width×height 2D array of
floating-point values and shows how to loop over the array elements in device code:
// Host code
float* devPtr;
int pitch;
cudaMallocPitch((void**)&devPtr, &pitch,
width * sizeof(float), height);
MyKernel<<<100, 512>>>(devPtr, pitch);

// Device code
__global__ void MyKernel(float* devPtr, int pitch)
{
for (int r = 0; r < height; ++r) {
float* row = (float*)((char*)devPtr + r * pitch);
for (int c = 0; c < width; ++c) {
float element = row[c];
}
}
}
The following code sample allocates a width×height×depth 3D array of
floating-point values and shows how to loop over the array elements in device code:
// Host code
cudaPitchedPtr devPitchedPtr;
cudaExtent extent = make_cudaExtent(64, 64, 64);
cudaMalloc3D(&devPitchedPtr, extent);
MyKernel<<<100, 512>>>(devPitchedPtr, extent);

// Device code
__global__ void MyKernel(cudaPitchedPtr devPitchedPtr,
cudaExtent extent)
{
char* devPtr = devPitchedPtr.ptr;
size_t pitch = devPitchedPtr.pitch;
size_t slicePitch = pitch * extent.height;
for (int z = 0; z < extent.depth; ++z) {
char* slice = devPtr + z * slicePitch;
for (int y = 0; y < extent.height; ++y) {
float* row = (float*)(slice + y * pitch);
for (int x = 0; x < extent.width; ++x) {
float element = row[x];
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 21

}
}
}
}
The reference manual lists all the various functions used to copy memory between
linear memory allocated with cudaMalloc(), linear memory allocated with
cudaMallocPitch() or cudaMalloc3D(), CUDA arrays, and memory
allocated for variables declared in global or constant memory space.
As an example, the following code sample copies some host memory array to
constant memory:
__constant__ float constData[256];
float data[256];
cudaMemcpyToSymbol(constData, data, sizeof(data));
cudaGetSymbolAddress() is used to retrieve the address pointing to the
memory allocated for a variable declared in global memory space. The size of the
allocated memory is obtained through cudaGetSymbolSize().
3.2.2 Shared Memory
As detailed in Section B.2 shared memory is allocated using the __shared__
qualifier.
Shared memory is expected to be much faster than global memory as mentioned in
Section 2.2 and detailed in Section 5.3.2.3. Any opportunity to replace global
memory accesses by shared memory accesses should therefore be exploited as
illustrated by the following matrix multiplication example.
The following code sample is a straightforward implementation of matrix
multiplication that does not take advantage of shared memory. Each thread reads
one row of A and one column of B and computes the corresponding element of C
as illustrated in Figure 3-1. A is therefore read B.width times from global memory
and B is read A.height times.
// Matrices are stored in row-major order:
// M(row, col) = *(M.elements + row * M.width + col)
typedef struct {
int width;
int height;
float* elements;
} Matrix;

// Thread block size
#define BLOCK_SIZE 16

// Forward declaration of the matrix multiplication kernel
__global__ void MatMulKernel(const Matrix, const Matrix, Matrix);

// Matrix multiplication - Host code
// Matrix dimensions are assumed to be multiples of BLOCK_SIZE
void MatMul(const Matrix A, const Matrix B, Matrix C)
{
// Load A and B to device memory
Matrix d_A;
Chapter 3. Programming Interface


22 CUDA Programming Guide Version 3.0

d_A.width = A.width; d_A.height = A.height;
size_t size = A.width * A.height * sizeof(float);
cudaMalloc((void**)&d_A.elements, size);
cudaMemcpy(d_A.elements, A.elements, size,
cudaMemcpyHostToDevice);
Matrix d_B;
d_B.width = B.width; d_B.height = B.height;
size = B.width * B.height * sizeof(float);
cudaMalloc((void**)&d_B.elements, size);
cudaMemcpy(d_B.elements, B.elements, size,
cudaMemcpyHostToDevice);

// Allocate C in device memory
Matrix d_C;
d_C.width = C.width; d_C.height = C.height;
size = C.width * C.height * sizeof(float);
cudaMalloc((void**)&d_C.elements, size);

// Invoke kernel
dim3 dimBlock(BLOCK_SIZE, BLOCK_SIZE);
dim3 dimGrid(B.width / dimBlock.x, A.height / dimBlock.y);
MatMulKernel<<<dimGrid, dimBlock>>>(d_A, d_B, d_C);

// Read C from device memory
cudaMemcpy(C.elements, Cd.elements, size,
cudaMemcpyDeviceToHost);

// Free device memory
cudaFree(d_A.elements);
cudaFree(d_B.elements);
cudaFree(d_C.elements);
}

// Matrix multiplication kernel called by MatMul()
__global__ void MatMulKernel(Matrix A, Matrix B, Matrix C)
{
// Each thread computes one element of C
// by accumulating results into Cvalue
float Cvalue = 0;
int row = blockIdx.y * blockDim.y + threadIdx.y;
int col = blockIdx.x * blockDim.x + threadIdx.x;
for (int e = 0; e < A.width; ++e)
Cvalue += A.elements[row * A.width + e]
* B.elements[e * B.width + col];
C.elements[row * C.width + col] = Cvalue;
}

Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 23


Figure 3-1. Matrix Multiplication without Shared Memory
The following code sample is an implementation of matrix multiplication that does
take advantage of shared memory. In this implementation, each thread block is
responsible for computing one square sub-matrix C
sub
of C and each thread within
the block is responsible for computing one element of C
sub
. As illustrated in Figure
3-2, C
sub
is equal to the product of two rectangular matrices: the sub-matrix of A of
dimension (A.width, block_size) that has the same line indices as C
sub
, and the sub-
matrix of B of dimension (block_size, A.width) that has the same column indices as
C
sub
. In order to fit into the device‟s resources, these two rectangular matrices are
divided into as many square matrices of dimension block_size as necessary and C
sub
is
computed as the sum of the products of these square matrices. Each of these
products is performed by first loading the two corresponding square matrices from
global memory to shared memory with one thread loading one element of each
matrix, and then by having each thread compute one element of the product. Each
thread accumulates the result of each of these products into a register and once
done writes the result to global memory.
A
B
C

B.width A.width
0
col
A
.
h
e
i
g
h
t

B
.
h
e
i
g
h
t

B
.
w
i
d
t
h
-
1

row
0
A.height-1
Chapter 3. Programming Interface


24 CUDA Programming Guide Version 3.0

By blocking the computation this way, we take advantage of fast shared memory
and save a lot of global memory bandwidth since A is only read (B.width / block_size)
times from global memory and B is read (A.height / block_size) times.
The Matrix type from the previous code sample is augmented with a stride field, so
that sub-matrices can be efficiently represented with the same type. __device__
functions (see Section B.1.1) are used to get and set elements and build any sub-
matrix from a matrix.
// Matrices are stored in row-major order:
// M(row, col) = *(M.elements + row * M.stride + col)
typedef struct {
int width;
int height;
int stride;
float* elements;
} Matrix;

// Get a matrix element
__device__ float GetElement(const Matrix A, int row, int col)
{
return A.elements[row * A.stride + col];
}

// Set a matrix element
__device__ void SetElement(Matrix A, int row, int col,
float value)
{
A.elements[row * A.stride + col] = value;
}

// Get the BLOCK_SIZExBLOCK_SIZE sub-matrix Asub of A that is
// located col sub-matrices to the right and row sub-matrices down
// from the upper-left corner of A
__device__ Matrix GetSubMatrix(Matrix A, int row, int col)
{
Matrix Asub;
Asub.width = BLOCK_SIZE;
Asub.height = BLOCK_SIZE;
Asub.stride = A.stride;
Asub.elements = &A.elements[A.stride * BLOCK_SIZE * row
+ BLOCK_SIZE * col];
return Asub;
}

// Thread block size
#define BLOCK_SIZE 16

// Forward declaration of the matrix multiplication kernel
__global__ void MatMulKernel(const Matrix, const Matrix, Matrix);

// Matrix multiplication - Host code
// Matrix dimensions are assumed to be multiples of BLOCK_SIZE
void MatMul(const Matrix A, const Matrix B, Matrix C)
{
// Load A and B to device memory
Matrix d_A;
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 25

d_A.width = d_A.stride = A.width; d_A.height = A.height;
size_t size = A.width * A.height * sizeof(float);
cudaMalloc((void**)&d_A.elements, size);
cudaMemcpy(d_A.elements, A.elements, size,
cudaMemcpyHostToDevice);
Matrix d_B;
d_B.width = d_B.stride = B.width; d_B.height = B.height;
size = B.width * B.height * sizeof(float);
cudaMalloc((void**)&d_B.elements, size);
cudaMemcpy(d_B.elements, B.elements, size,
cudaMemcpyHostToDevice);

// Allocate C in device memory
Matrix d_C;
d_C.width = d_C.stride = C.width; d_C.height = C.height;
size = C.width * C.height * sizeof(float);
cudaMalloc((void**)&d_C.elements, size);

// Invoke kernel
dim3 dimBlock(BLOCK_SIZE, BLOCK_SIZE);
dim3 dimGrid(B.width / dimBlock.x, A.height / dimBlock.y);
MatMulKernel<<<dimGrid, dimBlock>>>(d_A, d_B, d_C);

// Read C from device memory
cudaMemcpy(C.elements, d_C.elements, size,
cudaMemcpyDeviceToHost);

// Free device memory
cudaFree(d_A.elements);
cudaFree(d_B.elements);
cudaFree(d_C.elements);
}

// Matrix multiplication kernel called by MatMul()
__global__ void MatMulKernel(Matrix A, Matrix B, Matrix C)
{
// Block row and column
int blockRow = blockIdx.y;
int blockCol = blockIdx.x;

// Each thread block computes one sub-matrix Csub of C
Matrix Csub = GetSubMatrix(C, blockRow, blockCol);

// Each thread computes one element of Csub
// by accumulating results into Cvalue
float Cvalue = 0;

// Thread row and column within Csub
int row = threadIdx.y;
int col = threadIdx.x;

// Loop over all the sub-matrices of A and B that are
// required to compute Csub
// Multiply each pair of sub-matrices together
// and accumulate the results
for (int m = 0; m < (A.width / BLOCK_SIZE); ++m) {

Chapter 3. Programming Interface


26 CUDA Programming Guide Version 3.0

// Get sub-matrix Asub of A
Matrix Asub = GetSubMatrix(A, blockRow, m);

// Get sub-matrix Bsub of B
Matrix Bsub = GetSubMatrix(B, m, blockCol);

// Shared memory used to store Asub and Bsub respectively
__shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
__shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];

// Load Asub and Bsub from device memory to shared memory
// Each thread loads one element of each sub-matrix
As[row][col] = GetElement(Asub, row, col);
Bs[row][col] = GetElement(Bsub, row, col);

// Synchronize to make sure the sub-matrices are loaded
// before starting the computation
__syncthreads();

// Multiply Asub and Bsub together
for (int e = 0; e < BLOCK_SIZE; ++e)
Cvalue += As[row][e] * Bs[e][col];

// Synchronize to make sure that the preceding
// computation is done before loading two new
// sub-matrices of A and B in the next iteration
__syncthreads();
}

// Write Csub to device memory
// Each thread writes one element
SetElement(Csub, row, col, Cvalue);
}
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 27


Figure 3-2. Matrix Multiplication with Shared Memory
3.2.3 Multiple Devices
A host system can have multiple devices. These devices can be enumerated, their
properties can be queried, and one of them can be selected for kernel executions.
Several host threads can execute device code on the same device, but by design, a
host thread can execute device code on only one device at any given time. As a
consequence, multiple host threads are required to execute device code on multiple
devices. Also, any CUDA resources created through the runtime in one host thread
cannot be used by the runtime from another host thread.
The following code sample enumerates all devices in the system and retrieves their
properties. It also determines the number of CUDA-enabled devices.
int deviceCount;
cudaGetDeviceCount(&deviceCount);
int device;
for (device = 0; device < deviceCount; ++device) {
cudaDeviceProp deviceProp;
cudaGetDeviceProperties(&deviceProp, device);
if (dev == 0) {
A

B
C
C
sub


BLOCK_SIZE
B.width A.width
BLOCK_SIZE BLOCK_SIZE
B
L
O
C
K
_
S
I
Z
E

B
L
O
C
K
_
S
I
Z
E

B
L
O
C
K
_
S
I
Z
E

b
l
o
c
k
R
o
w

row
0
BLOCK_SIZE-1
B
L
O
C
K
_
S
I
Z
E
-
1

0 col
blockCol
A
.
h
e
i
g
h
t

B
.
h
e
i
g
h
t

Chapter 3. Programming Interface


28 CUDA Programming Guide Version 3.0

if (deviceProp.major == 9999 && deviceProp.minor == 9999)
printf("There is no device supporting CUDA.\n");
else if (deviceCount == 1)
printf("There is 1 device supporting CUDA\n");
else
printf("There are %d devices supporting CUDA\n",
deviceCount);
}
}
By default, the device associated to the host thread is implicitly selected as device 0
as soon as a non-device management runtime function is called (see Section 3.6 for
exceptions). Any other device can be selected by calling cudaSetDevice() first.
After a device has been selected, either implicitly or explicitly, any subsequent
explicit call to cudaSetDevice() will fail up until cudaThreadExit() is called.
cudaThreadExit() cleans up all runtime-related resources associated with the
calling host thread. Any subsequent API call reinitializes the runtime.
3.2.4 Texture Memory
CUDA supports a subset of the texturing hardware that the GPU uses for graphics
to access texture memory. Reading data from texture memory instead of global
memory can have several performance benefits as described in Section 5.3.2.5.
Texture memory is read from kernels using device functions called texture fetches,
described in Section B.8. The first parameter of a texture fetch specifies an object
called a texture reference.
A texture reference defines which part of texture memory is fetched. As detailed in
Section 3.2.4.3, it must be bound through runtime functions to some region of
memory, called a texture, before it can be used by a kernel. Several distinct texture
references might be bound to the same texture or to textures that overlap in
memory.
A texture reference has several attributes. One of them is its dimensionality that
specifies whether the texture is addressed as a one-dimensional array using one
texture coordinate, a two-dimensional array using two texture coordinates, or a three-
dimensional array using three texture coordinates. Elements of the array are called
texels, short for “texture elements.”
Other attributes define the input and output data types of the texture fetch, as well
as how the input coordinates are interpreted and what processing should be done.
A texture can be any region of linear memory or a CUDA array.
CUDA arrays are opaque memory layouts optimized for texture fetching. They are
one-dimensional, two-dimensional, or three-dimensional and composed of
elements, each of which has 1, 2 or 4 components that may be signed or unsigned
8-, 16- or 32-bit integers, 16-bit floats (currently only supported through the driver
API), or 32-bit floats. CUDA arrays are only readable by kernels through texture
fetching and may only be bound to texture references with the same number of
packed components.
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 29

3.2.4.1 Texture Reference Declaration
Some of the attributes of a texture reference are immutable and must be known at
compile time; they are specified when declaring the texture reference. A texture
reference is declared at file scope as a variable of type texture:
texture<Type, Dim, ReadMode> texRef;
where:
 Type specifies the type of data that is returned when fetching the texture; Type
is restricted to the basic integer and single-precision floating-point types and any
of the 1-, 2-, and 4-component vector types defined in Section B.3.1;
 Dim specifies the dimensionality of the texture reference and is equal to 1, 2, or
3; Dim is an optional argument which defaults to 1;
 ReadMode is equal to cudaReadModeNormalizedFloat or
cudaReadModeElementType; if it is cudaReadModeNormalizedFloat
and Type is a 16-bit or 8-bit integer type, the value is actually returned as
floating-point type and the full range of the integer type is mapped to [0.0, 1.0]
for unsigned integer type and [-1.0, 1.0] for signed integer type; for example, an
unsigned 8-bit texture element with the value 0xff reads as 1; if it is
cudaReadModeElementType, no conversion is performed; ReadMode is an
optional argument which defaults to cudaReadModeElementType.
3.2.4.2 Runtime Texture Reference Attributes
The other attributes of a texture reference are mutable and can be changed at
runtime through the host runtime. They specify whether texture coordinates are
normalized or not, the addressing mode, and texture filtering, as detailed below.
By default, textures are referenced using floating-point coordinates in the range
[0, N) where N is the size of the texture in the dimension corresponding to the
coordinate. For example, a texture that is 64×32 in size will be referenced with
coordinates in the range [0, 63] and [0, 31] for the x and y dimensions, respectively.
Normalized texture coordinates cause the coordinates to be specified in the range
[0.0, 1.0) instead of [0, N), so the same 64×32 texture would be addressed by
normalized coordinates in the range [0, 1) in both the x and y dimensions.
Normalized texture coordinates are a natural fit to some applications‟ requirements,
if it is preferable for the texture coordinates to be independent of the texture size.
The addressing mode defines what happens when texture coordinates are out of
range. When using unnormalized texture coordinates, texture coordinates outside
the range [0, N) are clamped: Values below 0 are set to 0 and values greater or equal
to N are set to N-1. Clamping is also the default addressing mode when using
normalized texture coordinates: Values below 0.0 or above 1.0 are clamped to the
range [0.0, 1.0). For normalized coordinates, the “wrap” addressing mode also may
be specified. Wrap addressing is usually used when the texture contains a periodic
signal. It uses only the fractional part of the texture coordinate; for example, 1.25 is
treated the same as 0.25 and -1.25 is treated the same as 0.75.
Linear texture filtering may be done only for textures that are configured to return
floating-point data. It performs low-precision interpolation between neighboring
texels. When enabled, the texels surrounding a texture fetch location are read and
the return value of the texture fetch is interpolated based on where the texture
coordinates fell between the texels. Simple linear interpolation is performed for one-
Chapter 3. Programming Interface


30 CUDA Programming Guide Version 3.0

dimensional textures and bilinear interpolation is performed for two-dimensional
textures.
Appendix F gives more details on texture fetching.
3.2.4.3 Texture Binding
As explained in the reference manual, the runtime API has a low-level C-style
interface and a high-level C++-style interface. The texture type is defined in the
high-level API as a structure publicly derived from the textureReference type
defined in the low-level API as such:
struct textureReference {
int normalized;
enum cudaTextureFilterMode filterMode;
enum cudaTextureAddressMode addressMode[3];
struct cudaChannelFormatDesc channelDesc;
}
 normalized specifies whether texture coordinates are normalized or not; if it
is non-zero, all elements in the texture are addressed with texture coordinates in
the range [0,1] rather than in the range [0,width-1], [0,height-1], or
[0,depth-1] where width, height, and depth are the texture sizes;
 filterMode specifies the filtering mode, that is how the value returned when
fetching the texture is computed based on the input texture coordinates;
filterMode is equal to cudaFilterModePoint or
cudaFilterModeLinear; if it is cudaFilterModePoint, the returned
value is the texel whose texture coordinates are the closest to the input texture
coordinates; if it is cudaFilterModeLinear, the returned value is the linear
interpolation of the two (for a one-dimensional texture), four (for a
two-dimensional texture), or eight (for a three-dimensional texture) texels
whose texture coordinates are the closest to the input texture coordinates;
cudaFilterModeLinear is only valid for returned values of floating-point
type;
 addressMode specifies the addressing mode, that is how out-of-range texture
coordinates are handled; addressMode is an array of size three whose first,
second, and third elements specify the addressing mode for the first, second,
and third texture coordinates, respectively; the addressing mode is equal to
either cudaAddressModeClamp, in which case out-of-range texture
coordinates are clamped to the valid range, or cudaAddressModeWrap, in
which case out-of-range texture coordinates are wrapped to the valid range;
cudaAddressModeWrap is only supported for normalized texture
coordinates;
 channelDesc describes the format of the value that is returned when fetching
the texture; channelDesc is of the following type:
struct cudaChannelFormatDesc {
int x, y, z, w;
enum cudaChannelFormatKind f;
};
where x, y, z, and w are equal to the number of bits of each component of the
returned value and f is:
 cudaChannelFormatKindSigned if these components are of signed
integer type,
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 31

 cudaChannelFormatKindUnsigned if they are of unsigned integer
type,
 cudaChannelFormatKindFloat if they are of floating point type.
normalized, addressMode, and filterMode may be directly modified in host
code.
Before a kernel can use a texture reference to read from texture memory, the texture
reference must be bound to a texture using cudaBindTexture() or
cudaBindTextureToArray(). cudaUnbindTexture() is used to unbind a
texture reference.
The following code samples bind a texture reference to linear memory pointed to by
devPtr:
 Using the low-level API:
texture<float, 2, cudaReadModeElementType> texRef;
textureReference* texRefPtr;
cudaGetTextureReference(&texRefPtr, “texRef”);
cudaChannelFormatDesc channelDesc =
cudaCreateChannelDesc<float>();
cudaBindTexture2D(0, texRefPtr, devPtr, &channelDesc,
width, height, pitch);
 Using the high-level API:
texture<float, 2, cudaReadModeElementType> texRef;
cudaChannelFormatDesc channelDesc =
cudaCreateChannelDesc<float>();
cudaBindTexture2D(0, texRef, devPtr, &channelDesc,
width, height, pitch);
The following code samples bind a texture reference to a CUDA array cuArray:
 Using the low-level API:
texture<float, 2, cudaReadModeElementType> texRef;
textureReference* texRefPtr;
cudaGetTextureReference(&texRefPtr, “texRef”);
cudaChannelFormatDesc channelDesc;
cudaGetChannelDesc(&channelDesc, cuArray);
cudaBindTextureToArray(texRef, cuArray, &channelDesc);
 Using the high-level API:
texture<float, 2, cudaReadModeElementType> texRef;
cudaBindTextureToArray(texRef, cuArray);
The format specified when binding a texture to a texture reference must match the
parameters specified when declaring the texture reference; otherwise, the results of
texture fetches are undefined.
The following code sample applies some simple transformation kernel to a
// 2D float texture
texture<float, 2, cudaReadModeElementType> texRef;

// Simple transformation kernel
__global__ void transformKernel(float* output,
int width, int height, float theta)
{
// Calculate normalized texture coordinates
unsigned int x = blockIdx.x * blockDim.x + threadIdx.x;
Chapter 3. Programming Interface


32 CUDA Programming Guide Version 3.0

unsigned int y = blockIdx.y * blockDim.y + threadIdx.y;

float u = x / (float)width;
float v = y / (float)height;

// Transform coordinates
u -= 0.5f;
v -= 0.5f;
float tu = u * cosf(theta) – v * sinf(theta) + 0.5f;
float tv = v * cosf(theta) + u * sinf(theta) + 0.5f;

// Read from texture and write to global memory
output[y * width + x] = tex2D(tex, tu, tv);
}

// Host code
int main()
{
// Allocate CUDA array in device memory
cudaChannelFormatDesc channelDesc =
cudaCreateChannelDesc(32, 0, 0, 0,
cudaChannelFormatKindFloat);
cudaArray* cuArray;
cudaMallocArray(&cuArray, &channelDesc, width, height);

// Copy to device memory some data located at address h_data
// in host memory
cudaMemcpyToArray(cuArray, 0, 0, h_data, size,
cudaMemcpyHostToDevice);

// Set texture parameters
texRef.addressMode[0] = cudaAddressModeWrap;
texRef.addressMode[1] = cudaAddressModeWrap;
texRef.filterMode = cudaFilterModeLinear;
texRef.normalized = true;

// Bind the array to the texture
cudaBindTextureToArray(texRef, cuArray, channelDesc);

// Allocate result of transformation in device memory
float* output;
cudaMalloc((void**)&output, width * height * sizeof(float));

// Invoke kernel
dim3 dimBlock(16, 16);
dim3 dimGrid((width + dimBlock.x – 1) / dimBlock.x,
(height + dimBlock.y – 1) / dimBlock.y);
transformKernel<<<dimGrid, dimBlock>>>(output, width, height,
angle);

// Free device memory
cudaFreeArray(cuArray);
cudaFree(output);
}
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 33

3.2.5 Page-Locked Host Memory
The runtime also provides functions to allocate and free page-locked (also known as
pinned) host memory – as opposed to regular pageable host memory allocated by
malloc(): cudaHostAlloc() and cudaFreeHost().
Using page-locked host memory has several benefits:
 Copies between page-locked host memory and device memory can be
performed concurrently with kernel execution for some devices as mentioned in
Section 3.2.6;
 On some devices, page-locked host memory can be mapped into the address
space of the device, eliminating the need to copy it to or from device memory
as detailed in Section 3.2.5.3;
 On systems with a front-side bus, bandwidth between host memory and device
memory is higher if host memory is allocated as page-locked and even higher if
in addition it is allocated as write-combining as described in Section 3.2.5.2.
Page-locked host memory is a scarce resource however, so allocations in page-
locked memory will start failing long before allocations in pageable memory. In
addition, by reducing the amount of physical memory available to the operating
system for paging, allocating too much page-locked memory reduces overall system
performance.
The simple zero-copy SDK sample comes with a detailed document on the page-
locked memory APIs.
3.2.5.1 Portable Memory
A block of page-locked memory can be used by any host threads, but by default, the
benefits of using page-locked memory described above are only available for the
thread that allocates it. To make these advantages available to all threads, it needs to
be allocated by passing flag cudaHostAllocPortable to cudaHostAlloc().
3.2.5.2 Write-Combining Memory
By default page-locked host memory is allocated as cacheable. It can optionally be
allocated as write-combining instead by passing flag
cudaHostAllocWriteCombined to cudaHostAlloc(). Write-combining
memory frees up L1 and L2 cache resources, making more cache available to the
rest of the application. In addition, write-combining memory is not snooped during
transfers across the PCI Express bus, which can improve transfer performance by
up to 40%.
Reading from write-combining memory from the host is prohibitively slow, so
write-combining memory should in general be used for memory that the host only
writes to.
3.2.5.3 Mapped Memory
On some devices, a block of page-locked host memory can also be mapped into the
address space of the device by passing flag cudaHostAllocMapped to
cudaHostAlloc(). Such a block has therefore two addresses: one in host
memory and one in device memory. The host memory pointer is returned by
cudaHostAlloc() and the device memory pointer can be retrieved using
Chapter 3. Programming Interface


34 CUDA Programming Guide Version 3.0

cudaHostGetDevicePointer()and then used to access the block from within a
kernel.
Accessing host memory directly from within a kernel has several advantages:
 There is no need to allocate a block in device memory and copy data between
this block and the block in host memory; data transfers are implicitly performed
as needed by the kernel;
 There is no need to use streams (see Section 3.2.6.4) to overlap data transfers
with kernel execution; the kernel-originated data transfers automatically overlap
with kernel execution.
Since mapped page-locked memory is shared between host and device however, the
application must synchronize memory accesses using streams or events (see
Section 3.2.6) to avoid any potential read-after-write, write-after-read, or write-after-
write hazards.
A block of page-locked host memory can be allocated as both mapped and portable
(see Section 3.2.5.1), in which case each host thread that needs to map the block to
its device address space must call cudaHostGetDevicePointer() to retrieve a
device pointer, as device pointers will generally differ from one host thread to the
other.
To be able to retrieve the device pointer to any mapped page-locked memory within
a given host thread, page-locked memory mapping must be enabled by calling
cudaSetDeviceFlags() with the cudaDeviceMapHost flag before any other
CUDA calls is performed by the thread. Otherwise,
cudaHostGetDevicePointer() will return an error.
cudaHostGetDevicePointer() also returns an error if the device does not
support mapped page-locked host memory.
Applications may query whether a device supports mapped page-locked host
memory or not by calling cudaGetDeviceProperties() and checking the
canMapHostMemory property.
Note that atomic functions (Section 5.4.3B.10) operating on mapped page-locked
memory are not atomic from the point of view of the host or other devices.
3.2.6 Asynchronous Concurrent Execution
3.2.6.1 Concurrent Execution between Host and Device
In order to facilitate concurrent execution between host and device, some functions
are asynchronous: Control is returned to the host thread before the device has
completed the requested task. These are:
 Kernel launches;
 The functions that perform memory copies and are suffixed with Async;
 The functions that perform device ÷ device memory copies;
 The functions that set memory.
Programmers can globally disable asynchronous kernel launches for all CUDA
applications running on a system by setting the CUDA_LAUNCH_BLOCKING
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 35

environment variable to 1. This feature is provided for debugging purposes only and
should never be used as a way to make production software run reliably.
When an application is run via the CUDA debugger or the CUDA profiler, all
launches are synchronous.
3.2.6.2 Overlap of Data Transfer and Kernel Execution
Some devices of compute capability 1.1 and higher can perform copies between
page-locked host memory and device memory concurrently with kernel execution.
Applications may query this capability by calling cudaGetDeviceProperties()
and checking the deviceOverlap property. This capability is currently supported
only for memory copies that do not involve CUDA arrays or 2D arrays allocated
through cudaMallocPitch() (see Section 3.2.1).
3.2.6.3 Concurrent Kernel Execution
Some devices of compute capability 2.0 can execute multiple kernels concurrently.
Applications may query this capability by calling cudaGetDeviceProperties()
and checking the concurrentKernels property.
The maximum number of kernel launches that a device can execute concurrently is
four.
A kernel from one CUDA context cannot execute concurrently with a kernel from
another CUDA context.
Kernels that use many textures or a large amount of local memory are less likely to
execute concurrently with other kernels.
3.2.6.4 Concurrent Data Transfers
Devices of compute capability 2.0 can perform a copy from page-locked host
memory to device memory concurrently with a copy from device memory to page-
locked host memory.
3.2.6.5 Stream
Applications manage concurrency through streams. A stream is a sequence of
commands that execute in order. Different streams, on the other hand, may execute
their commands out of order with respect to one another or concurrently; this
behavior is not guaranteed and should therefore not be relied upon for correctness
(e.g. inter-kernel communication is undefined).
A stream is defined by creating a stream object and specifying it as the stream
parameter to a sequence of kernel launches and host ÷ device memory copies. The
following code sample creates two streams and allocates an array hostPtr of
float in page-locked memory.
cudaStream_t stream[2];
for (int i = 0; i < 2; ++i)
cudaStreamCreate(&stream[i]);
float* hostPtr;
cudaMallocHost((void**)&hostPtr, 2 * size);
Each of these streams is defined by the following code sample as a sequence of one
memory copy from host to device, one kernel launch, and one memory copy from
device to host:
for (int i = 0; i < 2; ++i)
Chapter 3. Programming Interface


36 CUDA Programming Guide Version 3.0

cudaMemcpyAsync(inputDevPtr + i * size, hostPtr + i * size,
size, cudaMemcpyHostToDevice, stream[i]);
for (int i = 0; i < 2; ++i)
MyKernel<<<100, 512, 0, stream[i]>>>
(outputDevPtr + i * size, inputDevPtr + i * size, size);
for (int i = 0; i < 2; ++i)
cudaMemcpyAsync(hostPtr + i * size, outputDevPtr + i * size,
size, cudaMemcpyDeviceToHost, stream[i]);
cudaThreadSynchronize();
Each stream copies its portion of input array hostPtr to array inputDevPtr in
device memory, processes inputDevPtr on the device by calling MyKernel(), and
copies the result outputDevPtr back to the same portion of hostPtr. Processing
hostPtr using two streams allows for the memory copies of one stream to overlap
with the kernel execution of the other stream. hostPtr must point to page-locked
host memory for any overlap to occur.
cudaThreadSynchronize() is called in the end to make sure all streams are
finished before proceeding further. It forces the runtime to wait until all preceding
device tasks in all streams have completed. cudaStreamSynchronize() forces the
runtime to wait until all preceding commands in a stream have completed. It can be
used to synchronize the host with a specific stream, allowing other streams to
continue executing on the device. cudaStreamQuery() provides applications
with a way to know if all preceding commands in a stream have completed. To
avoid unnecessary slowdowns, these functions are best used for timing purposes or
to isolate a launch or memory copy that is failing.
Streams are released by calling cudaStreamDestroy().
for (int i = 0; i < 2; ++i)
cudaStreamDestroy(stream[i]);
cudaStreamDestroy() waits for all preceding tasks in the given stream to
complete before destroying the stream and returning control to the host thread.
Two commands from different streams cannot run concurrently if either a page-
locked host memory allocation, a device memory allocation, a device memory set, a
device ÷ device memory copy, or any CUDA command to stream 0 (including
kernel launches and host ÷ device memory copies that do not specify any stream
parameter) is called in-between them by the host thread.
Any operation that requires a dependency check to see if a streamed kernel launch is
complete blocks all later kernel launches from any stream in the CUDA context
until the launch being checked is complete. Operations that require a dependency
check include any other commands within the same stream as the launch being
checked and any call to cudaStreamQuery() on that stream. Therefore,
applications should follow these guidelines to improve their potential for concurrent
kernel execution:
 All independent operations should be issued before dependent operations,
 Synchronization of any kind should be delayed as long as possible.
Switching between the L1/shared memory configurations described in
Section G.4.1 inserts a device-side synchronization barrier for all outstanding kernel
launches.
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 37

3.2.6.6 Event
The runtime also provides a way to closely monitor the device‟s progress, as well as
perform accurate timing, by letting the application asynchronously record events at
any point in the program and query when these events are actually recorded. An
event is recorded when all tasks – or optionally, all commands in a given stream –
preceding the event have completed. Events in stream zero are recorded after all
preceding tasks/commands from all streams are completed by the device.
The following code sample creates two events:
cudaEvent_t start, stop;
cudaEventCreate(&start);
cudaEventCreate(&stop);
These events can be used to time the code sample of the previous section the
following way:
cudaEventRecord(start, 0);
for (int i = 0; i < 2; ++i)
cudaMemcpyAsync(inputDev + i * size, inputHost + i * size,
size, cudaMemcpyHostToDevice, stream[i]);
for (int i = 0; i < 2; ++i)
MyKernel<<<100, 512, 0, stream[i]>>>
(outputDev + i * size, inputDev + i * size, size);
for (int i = 0; i < 2; ++i)
cudaMemcpyAsync(outputHost + i * size, outputDev + i * size,
size, cudaMemcpyDeviceToHost, stream[i]);
cudaEventRecord(stop, 0);
cudaEventSynchronize(stop);
float elapsedTime;
cudaEventElapsedTime(&elapsedTime, start, stop);
They are destroyed this way:
cudaEventDestroy(start);
cudaEventDestroy(stop);
3.2.6.7 Synchronous Calls
When a synchronous function is called, control is not returned to the host thread
before the device has completed the requested task. Whether the host thread will
then yield, block, or spin can be specified by calling cudaSetDeviceFlags()with
some specific flags (see reference manual for details) before any other CUDA calls is
performed by the host thread.
3.2.7 Graphics Interoperability
Some resources from OpenGL and Direct3D may be mapped into the address
space of CUDA, either to enable CUDA to read data written by OpenGL or
Direct3D, or to enable CUDA to write data for consumption by OpenGL or
Direct3D.
A resource must be registered to CUDA before it can be mapped using the
functions mentioned in Sections 3.2.7.1 and 3.2.7.2. These functions return a
pointer to a CUDA graphics resource of type struct cudaGraphicsResource.
Registering a resource is potentially high-overhead and therefore typically called only
Chapter 3. Programming Interface


38 CUDA Programming Guide Version 3.0

once per resource. A CUDA graphics resource is unregistered using
cudaGraphicsUnregisterResource().
Once a resource is registered to CUDA, it can be mapped and unmapped as many
times as necessary using cudaGraphicsMapResources() and
cudaGraphicsUnmapResources().
cudaGraphicsResourceSetMapFlags() can be called to specify usage hints
(write-only, read-only) that the CUDA driver can use to optimize resource
management.
A mapped resource can be read from or written to by kernels using the device
memory address returned by cudaGraphicsResourceGetMappedPointer()
for buffers and cudaGraphicsSubResourceGetMappedArray() for CUDA
arrays.
Accessing a resource through OpenGL or Direct3D while it is mapped to CUDA
produces undefined results.
Sections 3.2.7.1 and 3.2.7.2 give specifics for each graphics API and some code
samples.
3.2.7.1 OpenGL Interoperability
Interoperability with OpenGL requires that the CUDA device be specified by
cudaGLSetGLDevice() before any other runtime calls. Note that
cudaSetDevice()and cudaGLSetGLDevice() are mutually exclusive.
The OpenGL resources that may be mapped into the address space of CUDA are
OpenGL buffer, texture, and renderbuffer objects.
A buffer object is registered using cudaGraphicsGLRegisterBuffer(). In
CUDA, it appears as a device pointer and can therefore be read and written by
kernels or via cudaMemcpy() calls.
A texture or renderbuffer object is registered using
cudaGraphicsGLRegisterImage(). In CUDA, it appears as a CUDA array
and can therefore be bound to a texture reference and be read and written by
kernels or via cudaMemcpy2D() calls. cudaGraphicsGLRegisterImage()
supports all texture formats with a internal type of float (e.g. GL_RGBA_FLOAT32)
and unnormalized integer (e.g. GL_RGBA8UI). It does not currently support
normalized integer formats (e.g. GL_RGBA8). Please note that since GL_RGBA8UI is
an OpenGL 3.0 texture format, it can only be written by shaders, not the fixed
function pipeline.
The following code sample uses a kernel to dynamically modify a 2D
width x height grid of vertices stored in a vertex buffer object:
GLuint positionsVBO;
struct cudaGraphicsResource* positionsVBO_CUDA;

int main()
{
// Explicitly set device
cudaGLSetGLDevice(0);

// Initialize OpenGL and GLUT
...
glutDisplayFunc(display);
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 39


// Create buffer object and register it with CUDA
glGenBuffers(1, positionsVBO);
glBindBuffer(GL_ARRAY_BUFFER, &vbo);
unsigned int size = width * height * 4 * sizeof(float);
glBufferData(GL_ARRAY_BUFFER, size, 0, GL_DYNAMIC_DRAW);
glBindBuffer(GL_ARRAY_BUFFER, 0);
cudaGraphicsGLRegisterBuffer(&positionsVBO_CUDA,
positionsVBO,
cudaGraphicsMapFlagsWriteDiscard);

// Launch rendering loop
glutMainLoop();
}

void display()
{
// Map buffer object for writing from CUDA
float4* positions;
cudaGraphicsMapResources(1, &positionsVBO_CUDA, 0);
size_t num_bytes;
cudaGraphicsResourceGetMappedPointer((void**)&positions,
&num_bytes,
positionsVBO_CUDA));

// Execute kernel
dim3 dimBlock(16, 16, 1);
dim3 dimGrid(width / dimBlock.x, height / dimBlock.y, 1);
createVertices<<<dimGrid, dimBlock>>>(positions, time,
width, height);

// Unmap buffer object
cudaGraphicsUnmapResources(1, &positionsVBO_CUDA, 0);

// Render from buffer object
glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
glBindBuffer(GL_ARRAY_BUFFER, positionsVBO);
glVertexPointer(4, GL_FLOAT, 0, 0);
glEnableClientState(GL_VERTEX_ARRAY);
glDrawArrays(GL_POINTS, 0, width * height);
glDisableClientState(GL_VERTEX_ARRAY);

// Swap buffers
glutSwapBuffers();
glutPostRedisplay();
}

void deleteVBO()
{
cudaGraphicsUnregisterResource(positionsVBO_CUDA);
glDeleteBuffers(1, &positionsVBO);
}

__global__ void createVertices(float4* positions, float time,
unsigned int width, unsigned int height)
{
unsigned int x = blockIdx.x * blockDim.x + threadIdx.x;
Chapter 3. Programming Interface


40 CUDA Programming Guide Version 3.0

unsigned int y = blockIdx.y * blockDim.y + threadIdx.y;

// Calculate uv coordinates
float u = x / (float)width;
float v = y / (float)height;
u = u * 2.0f - 1.0f;
v = v * 2.0f - 1.0f;

// calculate simple sine wave pattern
float freq = 4.0f;
float w = sinf(u * freq + time)
* cosf(v * freq + time) * 0.5f;

// Write positions
positions[y * width + x] = make_float4(u, w, v, 1.0f);
}
On Windows and for Quadro GPUs, cudaWGLGetDevice() can be used to
retrieve the CUDA device associated to the handle returned by
wglEnumGpusNV(). Quadro GPUs offer higher performance OpenGL
interoperability than GeForce and Tesla GPUs in a multi-GPU configuration where
OpenGL rendering is performed on the Quadro GPU and CUDA computations are
performed on other GPUs in the system.
3.2.7.2 Direct3D Interoperability
Direct3D interoperability is supported for Direct3D 9, Direct3D 10, and
Direct3D 11.
A CUDA context may interoperate with only one Direct3D device at a time and the
CUDA context and Direct3D device must be created on the same GPU. Moreover,
the Direct3D device must be created with the
D3DCREATE_HARDWARE_VERTEXPROCESSING flag.
Interoperability with Direct3D requires that the Direct3D device be specified by
cudaD3D9SetDirect3DDevice(), cudaD3D10SetDirect3DDevice() and
cudaD3D11SetDirect3DDevice(), before any other runtime calls.
cudaD3D9GetDevice(), cudaD3D10GetDevice(), and
cudaD3D11GetDevice() can be used to retrieve the CUDA device associated to
some adapter.
The Direct3D resources that may be mapped into the address space of CUDA are
Direct3D buffers, textures, and surfaces. These resources are registered using
cudaGraphicsD3D9RegisterResource(),
cudaGraphicsD3D10RegisterResource(), and
cudaGraphicsD3D11RegisterResource().
The following code sample uses a kernel to dynamically modify a 2D
width x height grid of vertices stored in a vertex buffer object.
Direct3D 9 Version:
IDirect3D9* D3D;
IDirect3DDevice9* device;
struct CUSTOMVERTEX {
FLOAT x, y, z;
DWORD color;
};
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 41

IDirect3DVertexBuffer9* positionsVB;
struct cudaGraphicsResource* positionsVB_CUDA;

int main()
{
// Initialize Direct3D
D3D = Direct3DCreate9(D3D_SDK_VERSION);

// Get a CUDA-enabled adapter
unsigned int adapter = 0;
for (; adapter < g_pD3D->GetAdapterCount(); adapter++) {
D3DADAPTER_IDENTIFIER9 adapterId;
g_pD3D->GetAdapterIdentifier(adapter, 0, &adapterId);
int dev;
if (cudaD3D9GetDevice(&dev, adapterId.DeviceName)
== cudaSuccess)
break;
}

// Create device
...
D3D->CreateDevice(adapter, D3DDEVTYPE_HAL, hWnd,
D3DCREATE_HARDWARE_VERTEXPROCESSING,
&params, &device);

// Register device with CUDA
cudaD3D9SetDirect3DDevice(device);

// Create vertex buffer and register it with CUDA
unsigned int size = width * height * sizeof(CUSTOMVERTEX);
device->CreateVertexBuffer(size, 0, D3DFVF_CUSTOMVERTEX,
D3DPOOL_DEFAULT, &positionsVB, 0);
cudaGraphicsD3D9RegisterResource(&positionsVB_CUDA,
positionsVB,
cudaGraphicsRegisterFlagsNone);
cudaGraphicsResourceSetMapFlags(positionsVB_CUDA,
cudaGraphicsMapFlagsWriteDiscard);

// Launch rendering loop
while (...) {
...
Render();
...
}
}

void Render()
{
// Map vertex buffer for writing from CUDA
float4* positions;
cudaGraphicsMapResources(1, &positionsVB_CUDA, 0);
size_t num_bytes;
cudaGraphicsResourceGetMappedPointer((void**)&positions,
&num_bytes,
positionsVB_CUDA));

// Execute kernel
Chapter 3. Programming Interface


42 CUDA Programming Guide Version 3.0

dim3 dimBlock(16, 16, 1);
dim3 dimGrid(width / dimBlock.x, height / dimBlock.y, 1);
createVertices<<<dimGrid, dimBlock>>>(positions, time,
width, height);

// Unmap vertex buffer
cudaGraphicsUnmapResources(1, &positionsVB_CUDA, 0);

// Draw and present
...
}

void releaseVB()
{
cudaGraphicsUnregisterResource(positionsVB_CUDA);
positionsVB->Release();
}

__global__ void createVertices(float4* positions, float time,
unsigned int width, unsigned int height)
{
unsigned int x = blockIdx.x * blockDim.x + threadIdx.x;
unsigned int y = blockIdx.y * blockDim.y + threadIdx.y;

// Calculate uv coordinates
float u = x / (float)width;
float v = y / (float)height;
u = u * 2.0f - 1.0f;
v = v * 2.0f - 1.0f;

// Calculate simple sine wave pattern
float freq = 4.0f;
float w = sinf(u * freq + time)
* cosf(v * freq + time) * 0.5f;

// Write positions
positions[y * width + x] =
make_float4(u, w, v, __int_as_float(0xff00ff00));
}
Direct3D 10 Version:
ID3D10Device* device;
struct CUSTOMVERTEX {
FLOAT x, y, z;
DWORD color;
};
ID3D10Buffer* positionsVB;
struct cudaGraphicsResource* positionsVB_CUDA;

int main()
{
// Get a CUDA-enabled adapter
IDXGIFactory* factory;
CreateDXGIFactory(__uuidof(IDXGIFactory), (void**)&factory);
IDXGIAdapter* adapter = 0;
for (unsigned int i = 0; !adapter; ++i) {
if (FAILED(factory->EnumAdapters(i, &adapter))
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 43

break;
int dev;
if (cudaD3D10GetDevice(&dev, adapter) == cudaSuccess)
break;
adapter->Release();
}
factory->Release();

// Create swap chain and device
...
D3D10CreateDeviceAndSwapChain(adapter,
D3D10_DRIVER_TYPE_HARDWARE, 0,
D3D10_CREATE_DEVICE_DEBUG,
D3D10_SDK_VERSION,
&swapChainDesc, &swapChain,
&device);
adapter->Release();

// Register device with CUDA
cudaD3D10SetDirect3DDevice(device);

// Create vertex buffer and register it with CUDA
unsigned int size = width * height * sizeof(CUSTOMVERTEX);
D3D10_BUFFER_DESC bufferDesc;
bufferDesc.Usage = D3D10_USAGE_DEFAULT;
bufferDesc.ByteWidth = size;
bufferDesc.BindFlags = D3D10_BIND_VERTEX_BUFFER;
bufferDesc.CPUAccessFlags = 0;
bufferDesc.MiscFlags = 0;
device->CreateBuffer(&bufferDesc, 0, &positionsVB);
cudaGraphicsD3D10RegisterResource(&positionsVB_CUDA,
positionsVB,
cudaGraphicsRegisterFlagsNone);
cudaGraphicsResourceSetMapFlags(positionsVB_CUDA,
cudaGraphicsMapFlagsWriteDiscard);

// Launch rendering loop
while (...) {
...
Render();
...
}
}

void Render()
{
// Map vertex buffer for writing from CUDA
float4* positions;
cudaGraphicsMapResources(1, &positionsVB_CUDA, 0);
size_t num_bytes;
cudaGraphicsResourceGetMappedPointer((void**)&positions,
&num_bytes,
positionsVB_CUDA));

// Execute kernel
dim3 dimBlock(16, 16, 1);
dim3 dimGrid(width / dimBlock.x, height / dimBlock.y, 1);
Chapter 3. Programming Interface


44 CUDA Programming Guide Version 3.0

createVertices<<<dimGrid, dimBlock>>>(positions, time,
width, height);

// Unmap vertex buffer
cudaGraphicsUnmapResources(1, &positionsVB_CUDA, 0);

// Draw and present
...
}

void releaseVB()
{
cudaGraphicsUnregisterResource(positionsVB_CUDA);
positionsVB->Release();
}

__global__ void createVertices(float4* positions, float time,
unsigned int width, unsigned int height)
{
unsigned int x = blockIdx.x * blockDim.x + threadIdx.x;
unsigned int y = blockIdx.y * blockDim.y + threadIdx.y;

// Calculate uv coordinates
float u = x / (float)width;
float v = y / (float)height;
u = u * 2.0f - 1.0f;
v = v * 2.0f - 1.0f;

// Calculate simple sine wave pattern
float freq = 4.0f;
float w = sinf(u * freq + time)
* cosf(v * freq + time) * 0.5f;

// Write positions
positions[y * width + x] =
make_float4(u, w, v, __int_as_float(0xff00ff00));
}
Direct3D 11 Version:
ID3D11Device* device;
struct CUSTOMVERTEX {
FLOAT x, y, z;
DWORD color;
};
ID3D11Buffer* positionsVB;
struct cudaGraphicsResource* positionsVB_CUDA;

int main()
{
// Get a CUDA-enabled adapter
IDXGIFactory* factory;
CreateDXGIFactory(__uuidof(IDXGIFactory), (void**)&factory);
IDXGIAdapter* adapter = 0;
for (unsigned int i = 0; !adapter; ++i) {
if (FAILED(factory->EnumAdapters(i, &adapter))
break;
int dev;
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 45

if (cudaD3D11GetDevice(&dev, adapter) == cudaSuccess)
break;
adapter->Release();
}
factory->Release();

// Create swap chain and device
...
sFnPtr_D3D11CreateDeviceAndSwapChain(adapter,
D3D11_DRIVER_TYPE_HARDWARE,
0,
D3D11_CREATE_DEVICE_DEBUG,
featureLevels, 3,
D3D11_SDK_VERSION,
&swapChainDesc, &swapChain,
&device,
&featureLevel,
&deviceContext);
adapter->Release();

// Register device with CUDA
cudaD3D11SetDirect3DDevice(device);

// Create vertex buffer and register it with CUDA
unsigned int size = width * height * sizeof(CUSTOMVERTEX);
D3D11_BUFFER_DESC bufferDesc;
bufferDesc.Usage = D3D11_USAGE_DEFAULT;
bufferDesc.ByteWidth = size;
bufferDesc.BindFlags = D3D11_BIND_VERTEX_BUFFER;
bufferDesc.CPUAccessFlags = 0;
bufferDesc.MiscFlags = 0;
device->CreateBuffer(&bufferDesc, 0, &positionsVB);
cudaGraphicsD3D11RegisterResource(&positionsVB_CUDA,
positionsVB,
cudaGraphicsRegisterFlagsNone);
cudaGraphicsResourceSetMapFlags(positionsVB_CUDA,
cudaGraphicsMapFlagsWriteDiscard);

// Launch rendering loop
while (...) {
...
Render();
...
}
}

void Render()
{
// Map vertex buffer for writing from CUDA
float4* positions;
cudaGraphicsMapResources(1, &positionsVB_CUDA, 0);
size_t num_bytes;
cudaGraphicsResourceGetMappedPointer((void**)&positions,
&num_bytes,
positionsVB_CUDA));

// Execute kernel
Chapter 3. Programming Interface


46 CUDA Programming Guide Version 3.0

dim3 dimBlock(16, 16, 1);
dim3 dimGrid(width / dimBlock.x, height / dimBlock.y, 1);
createVertices<<<dimGrid, dimBlock>>>(positions, time,
width, height);

// Unmap vertex buffer
cudaGraphicsUnmapResources(1, &positionsVB_CUDA, 0);

// Draw and present
...
}

void releaseVB()
{
cudaGraphicsUnregisterResource(positionsVB_CUDA);
positionsVB->Release();
}

__global__ void createVertices(float4* positions, float time,
unsigned int width, unsigned int height)
{
unsigned int x = blockIdx.x * blockDim.x + threadIdx.x;
unsigned int y = blockIdx.y * blockDim.y + threadIdx.y;

// Calculate uv coordinates
float u = x / (float)width;
float v = y / (float)height;
u = u * 2.0f - 1.0f;
v = v * 2.0f - 1.0f;

// Calculate simple sine wave pattern
float freq = 4.0f;
float w = sinf(u * freq + time)
* cosf(v * freq + time) * 0.5f;

// Write positions
positions[y * width + x] =
make_float4(u, w, v, __int_as_float(0xff00ff00));
}
3.2.8 Error Handling
All runtime functions return an error code, but for an asynchronous function (see
Section 3.2.6), this error code cannot possibly report any of the asynchronous errors
that could occur on the device since the function returns before the device has
completed the task; the error code only reports errors that occur on the host prior
to executing the task, typically related to parameter validation; if an asynchronous
error occurs, it will be reported by some subsequent unrelated runtime function call.
The only way to check for asynchronous errors just after some asynchronous
function call is therefore to synchronize just after the call by calling
cudaThreadSynchronize() (or by using any other synchronization
mechanisms described in Section 3.2.6) and checking the error code returned by
cudaThreadSynchronize().
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 47

The runtime maintains an error variable for each host thread that is initialized to
cudaSuccess and is overwritten by the error code every time an error occurs (be
it a parameter validation error or an asynchronous error). cudaGetLastError()
returns this variable and resets it to cudaSuccess.
Kernel launches do not return any error code, so cudaGetLastError() must be
called just after the kernel launch to retrieve any pre-launch errors. To ensure that
any error returned by cudaGetLastError() does not originate from calls prior
to the kernel launch, one has to make sure that the runtime error variable is set to
cudaSuccess just before the kernel launch, for example, by calling
cudaGetLastError() just before the kernel launch. Kernel launches are
asynchronous, so to check for asynchronous errors, the application must
synchronize in-between the kernel launch and the call to cudaGetLastError().
Note that cudaErrorNotReady that may be returned by cudaStreamQuery()
and cudaEventQuery() is not considered an error and is therefore not reported
by cudaGetLastError().
3.2.9 Debugging using the Device Emulation Mode
CUDA-GDB can be used to debug devices of compute capability greater than 1.0
(see the CUDA-GDB user manual for supported platforms). The compiler and
runtime also supports an emulation mode for the purpose of debugging that can be
used even in the absence of any CUDA-enabled device. When compiling an
application in this mode (using the -deviceemu option), the device code is
compiled for and runs on the host, allowing the programmer to use the host‟s native
debugging support to debug the application as if it were a host application. The
preprocessor macro __DEVICE_EMULATION__ is defined in this mode. All code
for an application, including any libraries used, must be compiled consistently either
for device emulation or for device execution. Linking code compiled for device
emulation with code compiled for device execution causes the following runtime
error to be returned upon initialization: cudaErrorMixedDeviceExecution.
When running an application in device emulation mode, the programming model is
emulated by the runtime. For each thread in a thread block, the runtime creates a
thread on the host. The programmer needs to make sure that:
 The host is able to run up to the maximum number of threads per block, plus
one for the master thread.
 Enough memory is available to run all threads, knowing that each thread gets
256 KB of stack.
Many features provided through the device emulation mode make it a very effective
debugging tool:
 By using the host‟s native debugging support programmers can use all features
that the debugger supports, like setting breakpoints and inspecting data.
 Since device code is compiled to run on the host, the code can be augmented
with code that cannot run on the device, like input and output operations to
files or to the screen (printf(), etc.).
 Since all data resides on the host, any device- or host-specific data can be read
from either device or host code; similarly, any device or host function can be
called from either device or host code.
Chapter 3. Programming Interface


48 CUDA Programming Guide Version 3.0

 In case of incorrect usage of the synchronization intrinsic function, the runtime
detects dead lock situations.
Programmers must keep in mind that device emulation mode is emulating the
device, not simulating it. Therefore, device emulation mode is very useful in finding
algorithmic errors, but certain errors are hard to find:
 Race conditions are less likely to manifest themselves in device-emulation mode,
since the number of threads executing simultaneously is much smaller than on
an actual device.
 When dereferencing a pointer to global memory on the host or a pointer to
host memory on the device, device execution almost certainly fails in some
undefined way, whereas device emulation can produce correct results.
 Most of the time the same floating-point computation will not produce exactly
the same result when performed on the device as when performed on the host
in device emulation mode. This is expected since in general, all you need to get
different results for the same floating-point computation are slightly different
compiler options, let alone different compilers, different instruction sets, or
different architectures.
In particular, some host platforms store intermediate results of single-precision
floating-point calculations in extended precision registers, potentially resulting in
significant differences in accuracy when running in device emulation mode.
When this occurs, programmers can try any of the following methods, none of
which is guaranteed to work:
 Declare some floating-point variables as volatile to force single-precision
storage;
 Use the –ffloat-store compiler option of gcc,
 Use the /Op or /fp compiler options of the Visual C++ compiler,
 Use _FPU_GETCW() and _FPU_SETCW() on Linux or _controlfp()
on Windows to force single-precision floating-point computation for a
portion of the code by surrounding it with
unsigned int originalCW;
_FPU_GETCW(originalCW);
unsigned int cw = (originalCW & ~0x300) | 0x000;
_FPU_SETCW(cw);
or
unsigned int originalCW = _controlfp(0, 0);
_controlfp(_PC_24, _MCW_PC);
at the beginning, to store the current value of the control word and change
it to force the mantissa to be stored in 24 bits using, and with
_FPU_SETCW(originalCW);
or
_controlfp(originalCW, 0xfffff);
at the end, to restore the original control word.
Also, for single-precision floating-point numbers, compute devices of compute
capability 1.x do not support denormalized numbers (see Appendix G) as host
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 49

platforms usually do. This can lead to dramatically different results between
device emulation and device execution modes since some computation might
produce a finite result in one case and an infinite result in the other.
 The warp size is equal to 1 in device emulation mode (see Section 4.1 for the
definition of a warp). Therefore, the warp vote functions (described in
Section B.11) produce different results than in device execution mode.
3.3 Driver API
The driver API is a handle-based, imperative API: Most objects are referenced by
opaque handles that may be specified to functions to manipulate the objects.
The objects available in the driver API are summarized in Table 3-1.
Table 3-1. Objects Available in the CUDA Driver API
Object Handle Description
Device CUdevice CUDA-enabled device
Context CUcontext Roughly equivalent to a CPU process
Module CUmodule Roughly equivalent to a dynamic library
Function CUfunction Kernel
Heap memory CUdeviceptr Pointer to device memory
CUDA array CUarray Opaque container for one-dimensional or two-dimensional
data on the device, readable via texture references
Texture reference CUtexref Object that describes how to interpret texture memory data
The driver API is implemented in the nvcuda dynamic library and all its entry
points are prefixed with cu.
The driver API must be initialized with cuInit() before any function from the
driver API is called. A CUDA context must then be created that is attached to a
specific device and made current to the calling host thread as detailed in
Section 3.3.1.
Within a CUDA context, kernels are explicitly loaded as PTX or binary objects by
the host code as described in Section 3.3.2. Kernels written in C must therefore be
compiled separately into PTX or binary objects. Kernels are launched using API
entry points as described in Section 3.3.3.
Any application that wants to run on future device architectures must load PTX, not
binary code. This is because binary code is architecture-specific and therefore
incompatible with future architectures, whereas PTX code is compiled to binary
code at load time by the driver.
Here is the host code of the sample from Section 2.1 written using the driver API:
int main()
{
int N = ...;
size_t size = N * sizeof(float);

// Allocate input vectors h_A and h_B in host memory
float* h_A = (float*)malloc(size);
Chapter 3. Programming Interface


50 CUDA Programming Guide Version 3.0

float* h_B = (float*)malloc(size);

// Initialize input vectors
...

// Initialize
cuInit(0);

// Get number of devices supporting CUDA
int deviceCount = 0;
cuDeviceGetCount(&deviceCount);
if (deviceCount == 0) {
printf("There is no device supporting CUDA.\n");
exit (0);
}

// Get handle for device 0
CUdevice cuDevice;
cuDeviceGet(&cuDevice, 0);

// Create context
CUcontext cuContext;
cuCtxCreate(&cuContext, 0, cuDevice);

// Create module from binary file
CUmodule cuModule;
cuModuleLoad(&cuModule, “VecAdd.ptx”);

// Allocate vectors in device memory
CUdeviceptr d_A;
cuMemAlloc(&d_A, size);
CUdeviceptr d_B;
cuMemAlloc(&d_B, size);
CUdeviceptr d_C;
cuMemAlloc(&d_C, size);

// Copy vectors from host memory to device memory
cuMemcpyHtoD(d_A, h_A, size);
cuMemcpyHtoD(d_B, h_B, size);

// Get function handle from module
CUfunction vecAdd;
cuModuleGetFunction(&vecAdd, cuModule, "VecAdd");

// Invoke kernel
#define ALIGN_UP(offset, alignment) \
(offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1)
int offset = 0;
void* ptr;
ptr = (void*)(size_t)d_A;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(vecAdd, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ptr = (void*)(size_t)d_B;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(vecAdd, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 51

ptr = (void*)(size_t)d_C;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(vecAdd, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ALIGN_UP(offset, __alignof(N));
cuParamSeti(vecAdd, offset, N);
offset += sizeof(N);
cuParamSetSize(vecAdd, offset);
int threadsPerBlock = 256;
int blocksPerGrid =
(N + threadsPerBlock – 1) / threadsPerBlock;
cuFuncSetBlockShape(vecAdd, threadsPerBlock, 1, 1);
cuLaunchGrid(vecAdd, blocksPerGrid, 1);

...
}
Full code can be found in the vectorAddDrv SDK code sample.
3.3.1 Context
A CUDA context is analogous to a CPU process. All resources and actions
performed within the driver API are encapsulated inside a CUDA context, and the
system automatically cleans up these resources when the context is destroyed.
Besides objects such as modules and texture references, each context has its own
distinct 32-bit address space. As a result, CUdeviceptr values from different
contexts reference different memory locations.
A host thread may have only one device context current at a time. When a context is
created with cuCtxCreate(), it is made current to the calling host thread. CUDA
functions that operate in a context (most functions that do not involve device
enumeration or context management) will return
CUDA_ERROR_INVALID_CONTEXT if a valid context is not current to the thread.
Each host thread has a stack of current contexts. cuCtxCreate() pushes the new
context onto the top of the stack. cuCtxPopCurrent() may be called to detach
the context from the host thread. The context is then "floating" and may be pushed
as the current context for any host thread. cuCtxPopCurrent() also restores the
previous current context, if any.
A usage count is also maintained for each context. cuCtxCreate() creates a
context with a usage count of 1. cuCtxAttach() increments the usage count and
cuCtxDetach() decrements it. A context is destroyed when the usage count goes
to 0 when calling cuCtxDetach() or cuCtxDestroy().
Usage count facilitates interoperability between third party authored code operating
in the same context. For example, if three libraries are loaded to use the same
context, each library would call cuCtxAttach() to increment the usage count and
cuCtxDetach() to decrement the usage count when the library is done using the
context. For most libraries, it is expected that the application will have created a
context before loading or initializing the library; that way, the application can create
the context using its own heuristics, and the library simply operates on the context
handed to it. Libraries that wish to create their own contexts – unbeknownst to their
API clients who may or may not have created contexts of their own – would use
cuCtxPushCurrent() and cuCtxPopCurrent() as illustrated in Figure 3-3.
Chapter 3. Programming Interface


52 CUDA Programming Guide Version 3.0


Figure 3-3. Library Context Management
3.3.2 Module
Modules are dynamically loadable packages of device code and data, akin to DLLs in
Windows, that are output by nvcc (see Section 3.1). The names for all symbols,
including functions, global variables, and texture references, are maintained at
module scope so that modules written by independent third parties may interoperate
in the same CUDA context.
This code sample loads a module and retrieves a handle to some kernel:
CUmodule cuModule;
cuModuleLoad(&cuModule, “myModule.ptx”);
CUfunction myKernel;
cuModuleGetFunction(&myKernel, cuModule, “MyKernel”);
This code sample compiles and loads a new module from PTX code and parses
compilation errors:
#define ERROR_BUFFER_SIZE 100
CUmodule cuModule;
CUptxas_option options[3];
void* values[3];
char* PTXCode = “some PTX code”;
options[0] = CU_ASM_ERROR_LOG_BUFFER;
values[0] = (void*)malloc(ERROR_BUFFER_SIZE);
options[1] = CU_ASM_ERROR_LOG_BUFFER_SIZE_BYTES;
values[1] = (void*)ERROR_BUFFER_SIZE;
options[2] = CU_ASM_TARGET_FROM_CUCONTEXT;
values[2] = 0;
cuModuleLoadDataEx(&cuModule, PTXCode, 3, options, values);
for (int i = 0; i < values[1]; ++i) {
// Parse error string here
}
3.3.3 Kernel Execution
cuFuncSetBlockShape() sets the number of threads per block for a given
function, and how their threadIDs are assigned.
Library Initialization Call
cuCtxCreate()
Initialize
context
cuCtxPopCurrent()
Library Call
cuCtxPushCurrent()
Use
context
cuCtxPopCurrent()
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 53

cuFuncSetSharedSize() sets the size of shared memory for the function.
The cuParam*() family of functions is used to specify the parameters that will be
provided to the kernel the next time cuLaunchGrid() or cuLaunch() is
invoked to launch the kernel.
The second argument of each of the cuParam*() functions specifies the offset of
the parameter in the parameter stack. This offset must match the alignment
requirement for the parameter type in device code. Alignment requirements in
device code for the built-in vector types are listed in Table B-1. For all other basic
types, the alignment requirement in device code matches the alignment requirement
in host code and can therefore be obtained using __alignof(). The only
exception is when the host compiler aligns double and long long (and long on
a 64-bit system) on a one-word boundary instead of a two-word boundary (for
example, using gcc‟s compilation flag -mno-align-double) since in device code
these types are always aligned on a two-word boundary. Also, CUdeviceptr is an
integer, but represents a pointer, so its alignment requirement is
__alignof(void*). The following code sample uses a macro to adjust the offset
of each parameter to meet its alignment requirement.
#define ALIGN_UP(offset, alignment) \
(offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1)
int offset = 0;

int i;
ALIGN_UP(offset, __alignof(i));
cuParamSeti(cuFunction, offset, i);
offset += sizeof(i);

float4 f4;
ALIGN_UP(offset, 16); // float4’s alignment is 16
cuParamSetv(cuFunction, offset, &f4, sizeof(f4));
offset += sizeof(f4);

char c;
ALIGN_UP(offset, __alignof(c));
cuParamSeti(cuFunction, offset, c);
offset += sizeof(c);

float f;
ALIGN_UP(offset, __alignof(f));
cuParamSeti(cuFunction, offset, f);
offset += sizeof(f);

CUdeviceptr dptr;
// void* should be used to determine CUdeviceptr’s alignment
void* ptr = (void*)(size_t)dptr;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(cuFunction, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);

float2 f2;
ALIGN_UP(offset, 8); // float2’s alignment is 8
cuParamSetv(cuFunction, offset, &f2, sizeof(f2));
offset += sizeof(f2);

Chapter 3. Programming Interface


54 CUDA Programming Guide Version 3.0

cuParamSetSize(cuFunction, offset);

cuFuncSetBlockShape(cuFunction, blockWidth, blockHeight, 1);
cuLaunchGrid(cuFunction, gridWidth, gridHeight);
The alignment requirement of a structure is equal to the maximum of the alignment
requirements of its fields. The alignment requirement of a structure that contains
built-in vector types, CUdeviceptr, or non-aligned double and long long,
might therefore differ between device code and host code. Such a structure might
also be padded differently. The following structure, for example, is not padded at all
in host code, but it is padded in device code with 12 bytes after field f since the
alignment requirement for field f4 is 16.
typedef struct {
float f;
float4 f4;
} myStruct;
Any parameter of type myStruct must therefore be passed using separate calls to
cuParam*(), such as:
myStruct s;
int offset = 0;

cuParamSetv(cuFunction, offset, &s.f, sizeof(s.f));
offset += sizeof(s.f);

ALIGN_UP(offset, 16); // float4’s alignment is 16
cuParamSetv(cuFunction, offset, &s.f4, sizeof(s.f4));
offset += sizeof(s.f4);
3.3.4 Device Memory
Linear memory is allocated using cuMemAlloc() or cuMemAllocPitch() and
freed using cuMemFree().
Here is the host code of the sample from Section 3.2.1 written using the driver API:
// Host code
int main()
{
// Initialize
if (cuInit(0) != CUDA_SUCCESS)
exit (0);

// Get number of devices supporting CUDA
int deviceCount = 0;
cuDeviceGetCount(&deviceCount);
if (deviceCount == 0) {
printf("There is no device supporting CUDA.\n");
exit (0);
}

// Get handle for device 0
CUdevice cuDevice = 0;
cuDeviceGet(&cuDevice, 0);

// Create context
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 55

CUcontext cuContext;
cuCtxCreate(&cuContext, 0, cuDevice);

// Create module from binary file
CUmodule cuModule;
cuModuleLoad(&cuModule, “VecAdd.ptx”);

// Get function handle from module
CUfunction vecAdd;
cuModuleGetFunction(&vecAdd, cuModule, "VecAdd");

// Allocate vectors in device memory
size_t size = N * sizeof(float);
CUdeviceptr d_A;
cuMemAlloc(&d_A, size);
CUdeviceptr d_B;
cuMemAlloc(&d_B, size);
CUdeviceptr d_C;
cuMemAlloc(&d_C, size);

// Copy vectors from host memory to device memory
// h_A and h_B are input vectors stored in host memory
cuMemcpyHtoD(d_A, h_A, size);
cuMemcpyHtoD(d_B, h_B, size);

// Invoke kernel
#define ALIGN_UP(offset, alignment) \
(offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1)
int offset = 0;
void* ptr;
ptr = (void*)(size_t)d_A;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(vecAdd, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ptr = (void*)(size_t)d_B;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(vecAdd, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ptr = (void*)(size_t)d_C;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(vecAdd, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
cuParamSetSize(VecAdd, offset);
int threadsPerBlock = 256;
int blocksPerGrid =
(N + threadsPerBlock – 1) / threadsPerBlock;
cuFuncSetBlockShape(vecAdd, threadsPerBlock, 1, 1);
cuLaunchGrid(VecAdd, blocksPerGrid, 1);

// Copy result from device memory to host memory
// h_C contains the result in host memory
cuMemcpyDtoH(h_C, d_C, size);

// Free device memory
cuMemFree(d_A);
cuMemFree(d_B);
cuMemFree(d_C);
Chapter 3. Programming Interface


56 CUDA Programming Guide Version 3.0

}
Linear memory can also be allocated through cuMemAllocPitch(). This function
is recommended for allocations of 2D arrays as it makes sure that the allocation is
appropriately padded to meet the alignment requirements described in Section
5.3.2.1, therefore ensuring best performance when accessing the row addresses or
performing copies between 2D arrays and other regions of device memory (using
the cuMemcpy2D()). The returned pitch (or stride) must be used to access array
elements. The following code sample allocates a width×height 2D array of
floating-point values and shows how to loop over the array elements in device code:
// Host code (assuming cuModule has been loaded)
CUdeviceptr devPtr;
int pitch;
cuMemAllocPitch(&devPtr, &pitch,
width * sizeof(float), height, 4);
CUfunction myKernel;
cuModuleGetFunction(&myKernel, cuModule, “MyKernel”);
void* ptr = (void*)(size_t)devPtr;
cuParamSetv(myKernel, 0, &ptr, sizeof(ptr));
cuParamSetSize(myKernel, sizeof(ptr));
cuFuncSetBlockShape(myKernel, 512, 1, 1);
cuLaunchGrid(myKernel, 100, 1);

// Device code
__global__ void MyKernel(float* devPtr)
{
for (int r = 0; r < height; ++r) {
float* row = (float*)((char*)devPtr + r * pitch);
for (int c = 0; c < width; ++c) {
float element = row[c];
}
}
}
The following code sample allocates a width×height CUDA array of one 32-bit
floating-point component:
CUDA_ARRAY_DESCRIPTOR desc;
desc.Format = CU_AD_FORMAT_FLOAT;
desc.NumChannels = 1;
desc.Width = width;
desc.Height = height;
CUarray cuArray;
cuArrayCreate(&cuArray, &desc);
The reference manual lists all the various functions used to copy memory between
linear memory allocated with cuMemAlloc(), linear memory allocated with
cuMemAllocPitch(), and CUDA arrays.
The following code sample copies the 2D array to the CUDA array allocated in the
previous code samples:
CUDA_MEMCPY2D copyParam;
memset(&copyParam, 0, sizeof(copyParam));
copyParam.dstMemoryType = CU_MEMORYTYPE_ARRAY;
copyParam.dstArray = cuArray;
copyParam.srcMemoryType = CU_MEMORYTYPE_DEVICE;
copyParam.srcDevice = devPtr;
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 57

copyParam.srcPitch = pitch;
copyParam.WidthInBytes = width * sizeof(float);
copyParam.Height = height;
cuMemcpy2D(&copyParam);
The following code sample copies some host memory array to constant memory:
__constant__ float constData[256];
float data[256];
CUdeviceptr devPtr;
unsigned int bytes;
cuModuleGetGlobal(&devPtr, &bytes, cuModule, “constData”);
cuMemcpyHtoD(devPtr, data, bytes);
3.3.5 Shared Memory
The following code sample is the driver version of the host code of the sample from
Section 3.2.2. Note how the Matrix type must be declared differently in host code
since device pointers are represented as a handle of type CUdeviceptr.
In this sample, shared memory is statically allocated within the kernel as opposed to
allocated at runtime through cuFuncSetSharedSize().
// Matrices are stored in row-major order:
// M(row, col) = *(M.elements + row * M.stride + col)
typedef struct {
int width;
int height;
int stride;
#ifdef __CUDACC__
float* elements;
#else
CUdeviceptr elements;
#endif
} Matrix;

// Matrix multiplication - Host code
// Matrix dimensions are assumed to be multiples of BLOCK_SIZE
void MatMul(const Matrix A, const Matrix B, Matrix C)
{
// Load A and B to device memory
Matrix d_A;
d_A.width = d_A.stride = A.width; d_A.height = A.height;
size_t size = A.width * A.height * sizeof(float);
cuMemAlloc(&d_A.elements, size);
cuMemcpyHtoD(d_A.elements, A.elements, size);
Matrix d_B;
d_B.width = d_B.stride = B.width; d_B.height = B.height;
size = B.width * B.height * sizeof(float);
cuMemAlloc(&d_B.elements, size);
cuMemcpyHtoD(d_B.elements, B.elements, size);

// Allocate C in device memory
Matrix d_C;
d_C.width = d_C.stride = C.width; d_C.height = C.height;
size = C.width * C.height * sizeof(float);
cuMemAlloc(&d_C.elements, size);
Chapter 3. Programming Interface


58 CUDA Programming Guide Version 3.0


// Invoke kernel (assuming cuModule has been loaded)
CUfunction matMulKernel;
cuModuleGetFunction(&matMulKernel, cuModule, "MatMulKernel");
#define ALIGN_UP(offset, alignment) \
(offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1)
int offset = 0;
void* ptr;
ptr = (void*)(size_t)d_A;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(matMulKernel, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ptr = (void*)(size_t)d_B;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(matMulKernel, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ptr = (void*)(size_t)d_C;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(matMulKernel, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
cuParamSetSize(matMulKernel, offset);
cuFuncSetBlockShape(matMulKernel, BLOCK_SIZE, BLOCK_SIZE, 1);
cuLaunchGrid(matMulKernel,
B.width / dimBlock.x, A.height / dimBlock.y);

// Read C from device memory
cuMemcpyDtoH(C.elements, d_C.elements, size);

// Free device memory
cuMemFree(d_A.elements);
cuMemFree(d_B.elements);
cuMemFree(d_C.elements);
}
3.3.6 Multiple Devices
cuDeviceGetCount() and cuDeviceGet() provide a way to enumerate the
devices present in the system and other functions (described in the reference
manual) to retrieve their properties:
int deviceCount;
cuDeviceGetCount(&deviceCount);
int device;
for (int device = 0; device < deviceCount; ++device) {
CUdevice cuDevice;
cuDeviceGet(&cuDevice, device);
int major, minor;
cuDeviceComputeCapability(&major, &minor, cuDevice);
}
3.3.7 Texture Memory
Texure binding is done using cuTexRefSetAddress() for linear memory and
cuTexRefSetArray() for CUDA arrays.
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 59

If a module cuModule contains some texture reference texRef defined as
texture<float, 2, cudaReadModeElementType> texRef;
the following code sample retrieves texRef„s handle:
CUtexref cuTexRef;
cuModuleGetTexRef(&cuTexRef, cuModule, “texRef”);
The following code sample binds texRef to some linear memory pointed to by
devPtr:
CUDA_ARRAY_DESCRIPTOR desc;
cuTexRefSetAddress2D(cuTexRef, &desc, devPtr, pitch);
The following code samples bind texRef to a CUDA array cuArray:
cuTexRefSetArray(cuTexRef, cuArray, CU_TRSA_OVERRIDE_FORMAT);
The reference manual lists various functions used to set address mode, filter mode,
format, and other flags for some texture reference. The format specified when
binding a texture to a texture reference must match the parameters specified when
declaring the texture reference; otherwise, the results of texture fetches are
undefined.
The following code sample is the driver version of the host code of the sample from
Section 3.2.4.3.
// Host code
int main()
{
// Allocate CUDA array in device memory
CUarray cuArray;
CUDA_ARRAY_DESCRIPTOR desc;
desc.Format = CU_AD_FORMAT_FLOAT;
desc.NumChannels = 1;
desc.Width = width;
desc.Height = height;
cuArrayCreate(&cuArray, &desc);

// Copy to device memory some data located at address h_data
// in host memory
CUDA_MEMCPY2D copyParam;
memset(&copyParam, 0, sizeof(copyParam));
copyParam.dstMemoryType = CU_MEMORYTYPE_ARRAY;
copyParam.dstArray = cuArray;
copyParam.srcMemoryType = CU_MEMORYTYPE_HOST;
copyParam.srcHost = h_data;
copyParam.srcPitch = width * sizeof(float);
copyParam.WidthInBytes = copyParam.srcPitch;
copyParam.Height = height;
cuMemcpy2D(&copyParam);

// Set texture parameters
CUtexref texRef;
cuModuleGetTexRef(&texRef, cuModule, "texRef"));
cuTexRefSetAddressMode(texRef, 0, CU_TR_ADDRESS_MODE_WRAP);
cuTexRefSetAddressMode(texRef, 1, CU_TR_ADDRESS_MODE_WRAP);
cuTexRefSetFilterMode(texRef, CU_TR_FILTER_MODE_LINEAR);
cuTexRefSetFlags(texRef, CU_TRSF_NORMALIZED_COORDINATES);
cuTexRefSetFormat(texRef, CU_AD_FORMAT_FLOAT, 1);

Chapter 3. Programming Interface


60 CUDA Programming Guide Version 3.0

// Bind the array to the texture
cuTexRefSetArray(texRef, cuArray, CU_TRSA_OVERRIDE_FORMAT);

// Allocate result of transformation in device memory
CUdeviceptr output;
cuMemAlloc((void**)&output, width * height * sizeof(float));

// Invoke kernel (assuming cuModule has been loaded)
CUfunction transformKernel;
cuModuleGetFunction(&transformKernel,
cuModule, "transformKernel");
#define ALIGN_UP(offset, alignment) \
(offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1)
int offset = 0;
void* ptr = (void*)(size_t)output;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(transformKernel, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ALIGN_UP(offset, __alignof(width));
cuParamSeti(transformKernel, offset, width);
offset += sizeof(width);
ALIGN_UP(offset, __alignof(height));
cuParamSeti(transformKernel, offset, height);
offset += sizeof(height);
ALIGN_UP(offset, __alignof(angle));
cuParamSetf(transformKernel, offset, angle);
offset += sizeof(angle);
cuParamSetSize(transformKernel, offset));
cuParamSetTexRef(transformKernel,
CU_PARAM_TR_DEFAULT, texRef);
cuFuncSetBlockShape(transformKernel, 16, 16, 1);
cuLaunchGrid(transformKernel,
(width + dimBlock.x – 1) / dimBlock.x,
(height + dimBlock.y – 1) / dimBlock.y);

// Free device memory
cuArrayDestroy(cuArray);
cuMemFree(output);
}
3.3.8 Page-Locked Host Memory
Page-locked host memory can be allocated using cuMemHostAlloc() with
optional mutually non-exclusive flags:
 CU_MEMHOSTALLOC_PORTABLE to allocate memory that is portable across
CUDA contexts (see Section 3.2.5.1) 3.2.5.2;
 CU_MEMHOSTALLOC_WRITECOMBINED to allocate memory as write-
combining (see Section 3.2.5.2);
 CU_MEMHOSTALLOC_DEVICEMAP to allocate mapped page-locked memory
(see Section 3.2.5.3).
Page-locked host memory is freed using cuMemFreeHost().
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 61

Page-locked memory mapping is enabled for a CUDA context by creating the
context with the CU_CTX_MAP_HOST flag and device pointers to mapped page-
locked memory are retrieved using cuMemHostGetDevicePointer().
Applications may query whether a device supports mapped page-locked host
memory or not by checking the
CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY attribute using
cuDeviceGetAttribute().
3.3.9 Asynchronous Concurrent Execution
Applications may query if a device can perform copies between page-locked host
memory and device memory concurrently with kernel execution by checking the
CU_DEVICE_ATTRIBUTE_GPU_OVERLAP attribute using
cuDeviceGetAttribute().
Applications may query if a device supports multiple kernels running concurrently
by checking the CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS attribute using
cuDeviceGetAttribute().
3.3.9.1 Stream
The driver API provides functions similar to the runtime API to manage streams.
The following code sample is the driver version of the code sample from
Section 3.2.6.4.
CUstream stream[2];
for (int i = 0; i < 2; ++i)
cuStreamCreate(&stream[i], 0);
float* hostPtr;
cuMemAllocHost((void**)&hostPtr, 2 * size);

for (int i = 0; i < 2; ++i)
cuMemcpyHtoDAsync(inputDevPtr + i * size, hostPtr + i * size,
size, stream[i]);
for (int i = 0; i < 2; ++i) {
#define ALIGN_UP(offset, alignment) \
(offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1)
int offset = 0;
void* ptr;
ptr = (void*)(size_t)outputDevPtr;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(cuFunction, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ptr = (void*)(size_t)inputDevPtr;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(cuFunction, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ALIGN_UP(offset, __alignof(size));
cuParamSeti(cuFunction, offset, size);
offset += sizeof(int);
cuParamSetSize(cuFunction, offset);
cuFuncSetBlockShape(cuFunction, 512, 1, 1);
cuLaunchGridAsync(cuFunction, 100, 1, stream[i]);
}
Chapter 3. Programming Interface


62 CUDA Programming Guide Version 3.0

for (int i = 0; i < 2; ++i)
cuMemcpyDtoHAsync(hostPtr + i * size, outputDevPtr + i * size,
size, stream[i]);
cuCtxSynchronize();

for (int i = 0; i < 2; ++i)
cuStreamDestroy(&stream[i]);
3.3.9.2 Event Management
The driver API provides functions similar to the runtime API to manage events.
The following code sample is the driver version of the code sample from
Section 3.2.6.6.
CUevent start, stop;
cuEventCreate(&start);
cuEventCreate(&stop);

cuEventRecord(start, 0);
for (int i = 0; i < 2; ++i)
cuMemcpyHtoDAsync(inputDevPtr + i * size, hostPtr + i * size,
size, stream[i]);
for (int i = 0; i < 2; ++i) {
#define ALIGN_UP(offset, alignment) \
(offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1)
int offset = 0;
void* ptr;
ptr = (void*)(size_t)outputDevPtr;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(cuFunction, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ptr = (void*)(size_t)inputDevPtr;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(cuFunction, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ALIGN_UP(offset, __alignof(size));
cuParamSeti(cuFunction, offset, size);
offset += sizeof(size);
cuParamSetSize(cuFunction, offset);
cuFuncSetBlockShape(cuFunction, 512, 1, 1);
cuLaunchGridAsync(cuFunction, 100, 1, stream[i]);
}
for (int i = 0; i < 2; ++i)
cuMemcpyDtoHAsync(hostPtr + i * size, outputDevPtr + i * size,
size, stream[i]);
cuEventRecord(stop, 0);
cuEventSynchronize(stop);
float elapsedTime;
cuEventElapsedTime(&elapsedTime, start, stop);
They are destroyed this way:
cuEventDestroy(start);
cuEventDestroy(stop);
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 63

3.3.9.3 Synchronous Calls
Whether the host thread will yield, block, or spin on a synchronous function call can
be specified by calling cuCtxCreate() with some specific flags as described in the
reference manual.
3.3.10 Graphics Interoperability
The driver API provides functions similar to the runtime API to manage graphics
interoperability.
A resource must be registered to CUDA before it can be mapped using the
functions mentioned in Sections 3.3.10.1 and 3.3.10.2. These functions return a
CUDA graphics resource of type CUgraphicsResource. Registering a resource is
potentially high-overhead and therefore typically called only once per resource. A
CUDA graphics resource is unregistered using
cuGraphicsUnregisterResource().
Once a resource is registered to CUDA, it can be mapped and unmapped as many
times as necessary using cuGraphicsMapResources() and
cuGraphicsUnmapResources(). cuGraphicsResourceSetMapFlags()
can be called to specify usage hints (write-only, read-only) that the CUDA driver can
use to optimize resource management.
A mapped resource can be read from or written to by kernels using the device
memory address returned by cuGraphicsResourceGetMappedPointer() for
buffers and cuGraphicsSubResourceGetMappedArray() for CUDA arrays.
Accessing a resource through OpenGL or Direct3D while it is mapped to CUDA
produces undefined results.
Sections 3.3.10.1 and 3.3.10.2 give specifics for each graphics API and some code
samples.
3.3.10.1 OpenGL Interoperability
Interoperability with OpenGL requires that the CUDA context be specifically
created using cuGLCtxCreate() instead of cuCtxCreate().
The OpenGL resources that may be mapped into the address space of CUDA are
OpenGL buffer, texture, and renderbuffer objects. A buffer object is registered
using cuGraphicsGLRegisterBuffer(). A texture or renderbuffer object is
registered using cuGraphicsGLRegisterImage(). The same restrictions
described in Section 3.2.7.1 apply.
The following code sample is the driver version of the code sample from
Section 3.2.7.1.
CUfunction createVertices;
GLuint positionsVBO;
struct cudaGraphicsResource* positionsVBO_CUDA;

int main()
{
// Initialize driver API
...

Chapter 3. Programming Interface


64 CUDA Programming Guide Version 3.0

// Get handle for device 0
CUdevice cuDevice = 0;
cuDeviceGet(&cuDevice, 0);

// Create context
CUcontext cuContext;
cuGLCtxCreate(&cuContext, 0, cuDevice);

// Create module from binary file
CUmodule cuModule;
cuModuleLoad(&cuModule, “createVertices.ptx”);

// Get function handle from module
cuModuleGetFunction(&createVertices,
cuModule, "createVertices");

// Initialize OpenGL and GLUT
...
glutDisplayFunc(display);

// Create buffer object and register it with CUDA
glGenBuffers(1, positionsVBO);
glBindBuffer(GL_ARRAY_BUFFER, &vbo);
unsigned int size = width * height * 4 * sizeof(float);
glBufferData(GL_ARRAY_BUFFER, size, 0, GL_DYNAMIC_DRAW);
glBindBuffer(GL_ARRAY_BUFFER, 0);
cuGraphicsGLRegisterBuffer(&positionsVBO_CUDA,
positionsVBO,
cudaGraphicsMapFlagsWriteDiscard);

// Launch rendering loop
glutMainLoop();
}

void display()
{
// Map OpenGL buffer object for writing from CUDA
CUdeviceptr positions;
cuGraphicsMapResources(1, &positionsVBO_CUDA, 0);
size_t num_bytes;
cuGraphicsResourceGetMappedPointer((void**)&positions,
&num_bytes,
positionsVBO_CUDA));

// Execute kernel
#define ALIGN_UP(offset, alignment) \
(offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1)
int offset = 0;
void* ptr = (void*)(size_t)positions;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(createVertices, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ALIGN_UP(offset, __alignof(time));
cuParamSetf(createVertices, offset, time);
offset += sizeof(time);
ALIGN_UP(offset, __alignof(width));
cuParamSeti(createVertices, offset, width);
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 65

offset += sizeof(width);
ALIGN_UP(offset, __alignof(height));
cuParamSeti(createVertices, offset, height);
offset += sizeof(height);
cuParamSetSize(createVertices, offset);
int threadsPerBlock = 16;
cuFuncSetBlockShape(createVertices,
threadsPerBlock, threadsPerBlock, 1);
cuLaunchGrid(createVertices,
width / threadsPerBlock, height / threadsPerBlock);

// Unmap buffer object
cuGraphicsUnmapResources(1, &positionsVBO_CUDA, 0);

// Render from buffer object
glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
glBindBuffer(GL_ARRAY_BUFFER, positionsVBO);
glVertexPointer(4, GL_FLOAT, 0, 0);
glEnableClientState(GL_VERTEX_ARRAY);
glDrawArrays(GL_POINTS, 0, width * height);
glDisableClientState(GL_VERTEX_ARRAY);

// Swap buffers
glutSwapBuffers();
glutPostRedisplay();
}

void deleteVBO()
{
cuGraphicsUnregisterResource(positionsVBO_CUDA);
glDeleteBuffers(1, &positionsVBO);
}
On Windows and for Quadro GPUs, cuWGLGetDevice() can be used to retrieve
the CUDA device associated to the handle returned by wglEnumGpusNV().
3.3.10.2 Direct3D Interoperability
Interoperability with Direct3D requires that the Direct3D device be specified when
the CUDA context is created. This is done by creating the CUDA context using
cuD3D9CtxCreate() (resp. cuD3D10CtxCreate()) instead of
cuCtxCreate().
The Direct3D resources that may be mapped into the address space of CUDA are
Direct3D buffers, textures, and surfaces. These resources are registered using
cuGraphicsD3D9RegisterResource(),
cuGraphicsD3D10RegisterResource(), and
cuGraphicsD3D11RegisterResource().
The following code sample is the driver version of the host code of the sample from
Section 3.2.7.2.
Direct3D 9 Version:
IDirect3D9* D3D;
IDirect3DDevice9* device;
struct CUSTOMVERTEX {
FLOAT x, y, z;
DWORD color;
Chapter 3. Programming Interface


66 CUDA Programming Guide Version 3.0

};
IDirect3DVertexBuffer9* positionsVB;
struct cudaGraphicsResource* positionsVB_CUDA;

int main()
{
// Initialize Direct3D
D3D = Direct3DCreate9(D3D_SDK_VERSION);

// Get a CUDA-enabled adapter
unsigned int adapter = 0;
for (; adapter < g_pD3D->GetAdapterCount(); adapter++) {
D3DADAPTER_IDENTIFIER9 adapterId;
g_pD3D->GetAdapterIdentifier(adapter, 0, &adapterId);
int dev;
if (cuD3D9GetDevice(&dev, adapterId.DeviceName)
== cudaSuccess)
break;
}

// Create device
...
D3D->CreateDevice(adapter, D3DDEVTYPE_HAL, hWnd,
D3DCREATE_HARDWARE_VERTEXPROCESSING,
&params, &device);

// Initialize driver API
...

// Create context
CUdevice cuDevice;
CUcontext cuContext;
cuD3D9CtxCreate(&cuContext, &cuDevice, 0, &device);

// Create module from binary file
CUmodule cuModule;
cuModuleLoad(&cuModule, “createVertices.ptx”);

// Get function handle from module
cuModuleGetFunction(&createVertices,
cuModule, "createVertices");

// Create vertex buffer and register it with CUDA
unsigned int size = width * height * sizeof(CUSTOMVERTEX);
device->CreateVertexBuffer(size, 0, D3DFVF_CUSTOMVERTEX,
D3DPOOL_DEFAULT, &positionsVB, 0);
cuGraphicsD3D9RegisterResource(&positionsVB_CUDA,
positionsVB,
cudaGraphicsRegisterFlagsNone);
cuGraphicsResourceSetMapFlags(positionsVB_CUDA,
cudaGraphicsMapFlagsWriteDiscard);

// Launch rendering loop
while (...) {
...
Render();
...
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 67

}
}

void Render()
{
// Map vertex buffer for writing from CUDA
float4* positions;
cuGraphicsMapResources(1, &positionsVB_CUDA, 0);
size_t num_bytes;
cuGraphicsResourceGetMappedPointer((void**)&positions,
&num_bytes,
positionsVB_CUDA));

// Execute kernel
#define ALIGN_UP(offset, alignment) \
(offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1)
int offset = 0;
void* ptr = (void*)(size_t)positions;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(createVertices, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ALIGN_UP(offset, __alignof(time));
cuParamSetf(createVertices, offset, time);
offset += sizeof(time);
ALIGN_UP(offset, __alignof(width));
cuParamSeti(createVertices, offset, width);
offset += sizeof(width);
ALIGN_UP(offset, __alignof(height));
cuParamSeti(createVertices, offset, height);
offset += sizeof(height);
cuParamSetSize(createVertices, offset);
int threadsPerBlock = 16;
cuFuncSetBlockShape(createVertices,
threadsPerBlock, threadsPerBlock, 1);
cuLaunchGrid(createVertices,
width / threadsPerBlock, height / threadsPerBlock);

// Unmap vertex buffer
cuGraphicsUnmapResources(1, &positionsVB_CUDA, 0);

// Draw and present
...
}

void releaseVB()
{
cuGraphicsUnregisterResource(positionsVB_CUDA);
positionsVB->Release();
}
Direct3D 10 Version:
ID3D10Device* device;
struct CUSTOMVERTEX {
FLOAT x, y, z;
DWORD color;
};
ID3D10Buffer* positionsVB;
Chapter 3. Programming Interface


68 CUDA Programming Guide Version 3.0

struct cudaGraphicsResource* positionsVB_CUDA;

int main()
{
// Get a CUDA-enabled adapter
IDXGIFactory* factory;
CreateDXGIFactory(__uuidof(IDXGIFactory), (void**)&factory);
IDXGIAdapter* adapter = 0;
for (unsigned int i = 0; !adapter; ++i) {
if (FAILED(factory->EnumAdapters(i, &adapter))
break;
int dev;
if (cuD3D10GetDevice(&dev, adapter) == cudaSuccess)
break;
adapter->Release();
}
factory->Release();

// Create swap chain and device
...
D3D10CreateDeviceAndSwapChain(adapter,
D3D10_DRIVER_TYPE_HARDWARE, 0,
D3D10_CREATE_DEVICE_DEBUG,
D3D10_SDK_VERSION,
&swapChainDesc &swapChain,
&device);
adapter->Release();

// Initialize driver API
...

// Create context
CUdevice cuDevice;
CUcontext cuContext;
cuD3D10CtxCreate(&cuContext, &cuDevice, 0, &device);

// Create module from binary file
CUmodule cuModule;
cuModuleLoad(&cuModule, “createVertices.ptx”);

// Get function handle from module
cuModuleGetFunction(&createVertices,
cuModule, "createVertices");

// Create vertex buffer and register it with CUDA
unsigned int size = width * height * sizeof(CUSTOMVERTEX);
D3D10_BUFFER_DESC bufferDesc;
bufferDesc.Usage = D3D10_USAGE_DEFAULT;
bufferDesc.ByteWidth = size;
bufferDesc.BindFlags = D3D10_BIND_VERTEX_BUFFER;
bufferDesc.CPUAccessFlags = 0;
bufferDesc.MiscFlags = 0;
device->CreateBuffer(&bufferDesc, 0, &positionsVB);
cuGraphicsD3D10RegisterResource(&positionsVB_CUDA,
positionsVB,
cudaGraphicsRegisterFlagsNone);
cuGraphicsResourceSetMapFlags(positionsVB_CUDA,
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 69

cudaGraphicsMapFlagsWriteDiscard);

// Launch rendering loop
while (...) {
...
Render();
...
}
}

void Render()
{
// Map vertex buffer for writing from CUDA
float4* positions;
cuGraphicsMapResources(1, &positionsVB_CUDA, 0);
size_t num_bytes;
cuGraphicsResourceGetMappedPointer((void**)&positions,
&num_bytes,
positionsVB_CUDA));

// Execute kernel
#define ALIGN_UP(offset, alignment) \
(offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1)
int offset = 0;
void* ptr = (void*)(size_t)positions;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(createVertices, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ALIGN_UP(offset, __alignof(time));
cuParamSetf(createVertices, offset, time);
offset += sizeof(time);
ALIGN_UP(offset, __alignof(width));
cuParamSeti(createVertices, offset, width);
offset += sizeof(width);
ALIGN_UP(offset, __alignof(height));
cuParamSeti(createVertices, offset, height);
offset += sizeof(height);
cuParamSetSize(createVertices, offset);
int threadsPerBlock = 16;
cuFuncSetBlockShape(createVertices,
threadsPerBlock, threadsPerBlock, 1);
cuLaunchGrid(createVertices,
width / threadsPerBlock, height / threadsPerBlock);

// Unmap vertex buffer
cuGraphicsUnmapResources(1, &positionsVB_CUDA, 0);

// Draw and present
...
}

void releaseVB()
{
cuGraphicsUnregisterResource(positionsVB_CUDA);
positionsVB->Release();
}
Chapter 3. Programming Interface


70 CUDA Programming Guide Version 3.0

Direct3D 11 Version:
ID3D11Device* device;
struct CUSTOMVERTEX {
FLOAT x, y, z;
DWORD color;
};
ID3D11Buffer* positionsVB;
struct cudaGraphicsResource* positionsVB_CUDA;

int main()
{
// Get a CUDA-enabled adapter
IDXGIFactory* factory;
CreateDXGIFactory(__uuidof(IDXGIFactory), (void**)&factory);
IDXGIAdapter* adapter = 0;
for (unsigned int i = 0; !adapter; ++i) {
if (FAILED(factory->EnumAdapters(i, &adapter))
break;
int dev;
if (cuD3D11GetDevice(&dev, adapter) == cudaSuccess)
break;
adapter->Release();
}
factory->Release();

// Create swap chain and device
...
sFnPtr_D3D11CreateDeviceAndSwapChain(adapter,
D3D11_DRIVER_TYPE_HARDWARE,
0,
D3D11_CREATE_DEVICE_DEBUG,
featureLevels, 3,
D3D11_SDK_VERSION,
&swapChainDesc, &swapChain,
&device,
&featureLevel,
&deviceContext);
adapter->Release();

// Initialize driver API
...

// Create context
CUdevice cuDevice;
CUcontext cuContext;
cuD3D11CtxCreate(&cuContext, &cuDevice, 0, &device);

// Create module from binary file
CUmodule cuModule;
cuModuleLoad(&cuModule, “createVertices.ptx”);

// Get function handle from module
cuModuleGetFunction(&createVertices,
cuModule, "createVertices");

// Create vertex buffer and register it with CUDA
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 71

unsigned int size = width * height * sizeof(CUSTOMVERTEX);
D3D11_BUFFER_DESC bufferDesc;
bufferDesc.Usage = D3D11_USAGE_DEFAULT;
bufferDesc.ByteWidth = size;
bufferDesc.BindFlags = D3D10_BIND_VERTEX_BUFFER;
bufferDesc.CPUAccessFlags = 0;
bufferDesc.MiscFlags = 0;
device->CreateBuffer(&bufferDesc, 0, &positionsVB);
cuGraphicsD3D11RegisterResource(&positionsVB_CUDA,
positionsVB,
cudaGraphicsRegisterFlagsNone);
cuGraphicsResourceSetMapFlags(positionsVB_CUDA,
cudaGraphicsMapFlagsWriteDiscard);

// Launch rendering loop
while (...) {
...
Render();
...
}
}

void Render()
{
// Map vertex buffer for writing from CUDA
float4* positions;
cuGraphicsMapResources(1, &positionsVB_CUDA, 0);
size_t num_bytes;
cuGraphicsResourceGetMappedPointer((void**)&positions,
&num_bytes,
positionsVB_CUDA));

// Execute kernel
#define ALIGN_UP(offset, alignment) \
(offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1)
int offset = 0;
void* ptr = (void*)(size_t)positions;
ALIGN_UP(offset, __alignof(ptr));
cuParamSetv(createVertices, offset, &ptr, sizeof(ptr));
offset += sizeof(ptr);
ALIGN_UP(offset, __alignof(time));
cuParamSetf(createVertices, offset, time);
offset += sizeof(time);
ALIGN_UP(offset, __alignof(width));
cuParamSeti(createVertices, offset, width);
offset += sizeof(width);
ALIGN_UP(offset, __alignof(height));
cuParamSeti(createVertices, offset, height);
offset += sizeof(height);
cuParamSetSize(createVertices, offset);
int threadsPerBlock = 16;
cuFuncSetBlockShape(createVertices,
threadsPerBlock, threadsPerBlock, 1);
cuLaunchGrid(createVertices,
width / threadsPerBlock, height / threadsPerBlock);

// Unmap vertex buffer
Chapter 3. Programming Interface


72 CUDA Programming Guide Version 3.0

cuGraphicsUnmapResources(1, &positionsVB_CUDA, 0);

// Draw and present
...
}

void releaseVB()
{
cuGraphicsUnregisterResource(positionsVB_CUDA);
positionsVB->Release();
}
3.3.11 Error Handling
All driver functions return an error code, but for an asynchronous function (see
Section 3.2.6), this error code cannot possibly report any of the asynchronous errors
that could occur on the device since the function returns before the device has
completed the task; the error code only reports errors that occur on the host prior
to executing the task, typically related to parameter validation; if an asynchronous
error occurs, it will be reported by some subsequent unrelated runtime function call.
The only way to check for asynchronous errors just after some asynchronous
function call is therefore to synchronize just after the call by calling
cuCtxSynchronize() (or by using any other synchronization mechanisms
described in Section 3.3.9) and checking the error code returned by
cuCtxSynchronize().
3.4 Interoperability between Runtime and Driver
APIs
An application can mix runtime API code with driver API code within the limits
described in this section.
If a context is created and made current via the driver API, subsequent runtime calls
will pick up this context instead of creating a new one.
If the runtime is initialized (implicitly as mentioned in Section 3.2),
cuCtxAttach() can be used to retrieve the context created during initialization.
This context can be used by subsequent driver API calls.
Device memory can be allocated and freed using either API. CUdeviceptr can be
cast to regular pointers and vice-versa:
CUdeviceptr devPtr;
float* d_data;

// Allocation using driver API
cuMemAlloc(&devPtr, size);
d_data = (float*)devPtr;

// Allocation using runtime API
cudaMalloc((void**)&d_data, size);
devPtr = (CUdeviceptr)(size_t)d_data;
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 73

In particular, this means that applications written using the driver API can invoke
libraries written using the runtime API (such as CUFFT, CUBLAS, …).
All functions from the device and version management sections of the reference
manual can be used interchangeably.
The following features are not supported for applications that mix runtime API
code with driver API code:
 Device emulation,
 Context stack manipulation via cuCtxPopCurrent() and
cuCtxPushCurrent() (see Section 3.3.1).
3.5 Versioning and Compatibility
There are two version numbers that developers should care about when developing
a CUDA application: The compute capability that describes the general
specifications and features of the compute device (see Section 2.5) and the version
of the CUDA driver API that describes the features supported by the driver API
and runtime.
The version of the driver API is defined in the driver header file as
CUDA_VERSION. It allows developers to check whether their application requires a
newer driver than the one currently installed. This is important, because the driver
API is backward compatible, meaning that applications, plug-ins, and libraries
(including the C runtime) compiled against a particular version of the driver API will
continue to work on subsequent driver releases as illustrated in Figure 3-4. The
driver API is not forward compatible, which means that applications, plug-ins, and
libraries (including the C runtime) compiled against a particular version of the driver
API will not work on previous versions of the driver.
It is important to note that mixing and matching versions is not supported;
specifically:
 All applications, plug-ins, and libraries on a system must use the same version of
the CUDA driver API, since only one version of the CUDA driver can be
installed on a system.
 All plug-ins and libraries used by an application must use the same version of
the runtime.
 All plug-ins and libraries used by an application must use the same version of
any libraries that use the runtime (such as CUFFT, CUBLAS, …).

Chapter 3. Programming Interface


74 CUDA Programming Guide Version 3.0

1.0
Driver
Apps,
Libs &
Plug-ins
1.1
Driver
Apps,
Libs &
Plug-ins
2.0
Driver
Apps,
Libs &
Plug-ins
Compatible Incompatible
...
...

Figure 3-4. The Driver API is Backward, but Not Forward
Compatible
3.6 Compute Modes
On Tesla solutions running Linux, one can set any device in a system in one of the
three following modes using NVIDIA‟s System Management Interface (nvidia-smi),
which is a tool distributed as part of the Linux driver:
 Default compute mode: Multiple host threads can use the device (by calling
cudaSetDevice() on this device, when using the runtime API, or by making
current a context associated to the device, when using the driver API) at the
same time.
 Exclusive compute mode: Only one host thread can use the device at any given
time.
 Prohibited compute mode: No host thread can use the device.
This means, in particular, that a host thread using the runtime API without explicitly
calling cudaSetDevice() might be associated with a device other than device 0 if
device 0 turns out to be in prohibited compute mode or in exclusive compute mode
and used by another host thread. cudaSetValidDevices() can be used to set a
device from a prioritized list of devices.
Applications may query the compute mode of a device by calling
cudaGetDeviceProperties() and checking the computeMode property or
checking the CU_DEVICE_COMPUTE_MODE attribute using
cuDeviceGetAttribute().
3.7 Mode Switches
GPUs dedicate some DRAM memory to the so-called primary surface, which is used
to refresh the display device whose output is viewed by the user. When users initiate
Chapter 3. Programming Interface

CUDA Programming Guide Version 3.0 75

a mode switch of the display by changing the resolution or bit depth of the display
(using NVIDIA control panel or the Display control panel on Windows), the
amount of memory needed for the primary surface changes. For example, if the user
changes the display resolution from 1280x1024x32-bit to 1600x1200x32-bit, the
system must dedicate 7.68 MB to the primary surface rather than 5.24 MB. (Full-
screen graphics applications running with anti-aliasing enabled may require much
more display memory for the primary surface.) On Windows, other events that may
initiate display mode switches include launching a full-screen DirectX application,
hitting Alt+Tab to task switch away from a full-screen DirectX application, or
hitting Ctrl+Alt+Del to lock the computer.
If a mode switch increases the amount of memory needed for the primary surface,
the system may have to cannibalize memory allocations dedicated to CUDA
applications. Therefore, a mode switch results in any call to the CUDA runtime to
fail and return an invalid context error.






CUDA Programming Guide Version 3.0 77

Chapter 4.
Hardware Implementation
The CUDA architecture is built around a scalable array of multithreaded Streaming
Multiprocessors (SMs). When a CUDA program on the host CPU invokes a kernel
grid, the blocks of the grid are enumerated and distributed to multiprocessors with
available execution capacity. The threads of a thread block execute concurrently on
one multiprocessor, and multiple thread blocks can execute concurrently on one
multiprocessor. As thread blocks terminate, new blocks are launched on the vacated
multiprocessors.
A multiprocessor is designed to execute hundreds of threads concurrently. To
manage such a large amount of threads, it employs a unique architecture called
SIMT (Single-Instruction, Multiple-Thread) that is described in Section 4.1. To maximize
utilization of its functional units, it leverages thread-level parallelism by using
hardware multithreading as detailed in Section 4.2, more so than instruction-level
parallelism within a single thread (instructions are pipelined, but unlike CPU cores
they are executed in order and there is no branch prediction and no speculative
execution).
Sections 4.1 and 4.2 describe the architecture features of the streaming
multiprocessor that are common to all devices. Sections G.3.1 and G.4.1 provide the
specifics for devices of compute capabilities 1.x and 2.0, respectively.
4.1 SIMT Architecture
The multiprocessor creates, manages, schedules, and executes threads in groups of
32 parallel threads called warps. Individual threads composing a warp start together
at the same program address, but they have their own instruction address counter
and register state and are therefore free to branch and execute independently. The
term warp originates from weaving, the first parallel thread technology. A half-warp is
either the first or second half of a warp. A quarter-warp is either the first, second,
third, or fourth quarter of a warp.
When a multiprocessor is given one or more thread blocks to execute, it partitions
them into warps that get scheduled by a warp scheduler for execution. The way a block
is partitioned into warps is always the same; each warp contains threads of
consecutive, increasing thread IDs with the first warp containing thread 0.
Section 2.2 describes how thread IDs relate to thread indices in the block.
Chapter 4. Hardware Implementation


78 CUDA Programming Guide Version 3.0

A warp executes one common instruction at a time, so full efficiency is realized
when all 32 threads of a warp agree on their execution path. If threads of a warp
diverge via a data-dependent conditional branch, the warp serially executes each
branch path taken, disabling threads that are not on that path, and when all paths
complete, the threads converge back to the same execution path. Branch divergence
occurs only within a warp; different warps execute independently regardless of
whether they are executing common or disjoint code paths.
The SIMT architecture is akin to SIMD (Single Instruction, Multiple Data) vector
organizations in that a single instruction controls multiple processing elements. A
key difference is that SIMD vector organizations expose the SIMD width to the
software, whereas SIMT instructions specify the execution and branching behavior
of a single thread. In contrast with SIMD vector machines, SIMT enables
programmers to write thread-level parallel code for independent, scalar threads, as
well as data-parallel code for coordinated threads. For the purposes of correctness,
the programmer can essentially ignore the SIMT behavior; however, substantial
performance improvements can be realized by taking care that the code seldom
requires threads in a warp to diverge. In practice, this is analogous to the role of
cache lines in traditional code: Cache line size can be safely ignored when designing
for correctness but must be considered in the code structure when designing for
peak performance. Vector architectures, on the other hand, require the software to
coalesce loads into vectors and manage divergence manually.
If a non-atomic instruction executed by a warp writes to the same location in global
or shared memory for more than one of the threads of the warp, the number of
serialized writes that occur to that location varies depending on the compute
capability of the device (see Sections G.3.2, G.3.3, G.4.2, and G.4.3) and which
thread performs the final write is undefined.
If an atomic instruction (see Section B.10) executed by a warp reads, modifies, and
writes to the same location in global memory for more than one of the threads of
the warp, each read, modify, write to that location occurs and they are all serialized,
but the order in which they occur is undefined.
4.2 Hardware Multithreading
The execution context (program counters, registers, etc) for each warp processed by
a multiprocessor is maintained on-chip during the entire lifetime of the warp.
Switching from one execution context to another therefore has no cost, and at every
instruction issue time, the warp scheduler selects a warp that has threads ready to
execute (active threads) and issues the next instruction to those threads.
In particular, each multiprocessor has a set of 32-bit registers that are partitioned
among the warps, and a parallel data cache or shared memory that is partitioned among
the thread blocks.
The number of blocks and warps that can reside and be processed together on the
multiprocessor for a given kernel depends on the amount of registers and shared
memory used by the kernel and the amount of registers and shared memory
available on the multiprocessor. There are also a maximum number of resident
blocks and a maximum number of resident warps per multiprocessor. These limits
as well the amount of registers and shared memory available on the multiprocessor
Chapter 4: Hardware Implementation


CUDA Programming Guide Version 3.0 79

are a function of the compute capability of the device and are given in Appendix G.
If there are not enough registers or shared memory available per multiprocessor to
process at least one block, the kernel will fail to launch.
The total number of warps W
block
in a block is as follows:
) 1 , (
size
block
W
T
ceil W =
 T is the number of threads per block,
 W
size
is the warp size, which is equal to 32,
 ceil(x, y) is equal to x rounded up to the nearest multiple of y.
The total number of registers R
block
allocated for a block is as follows:
For devices of compute capability 1.x:
) , ) , ( (
T k size W block block
G R W G W ceil ceil R × × =

For devices of compute capability 2.0:
block T size k block
W G W R ceil R × × = ) , (
 G
W
is the warp allocation granularity, equal to 2 (compute capability 1.x only),
 R
k
is the number of registers used by the kernel,
 G
T
is the thread allocation granularity, equal to 256 for devices of compute
capability 1.0 and 1.1, and 512 for devices of compute capability 1.2 and 1.3,
and 64 for devices of compute capability 2.0.
The total amount of shared memory S
block
in bytes allocated for a block is as follows:
) , (
S k block
G S ceil S =
 S
k
is the amount of shared memory used by the kernel in bytes,
 G
S
is the shared memory allocation granularity, which is equal to 512 for devices
of compute capability 1.x and 128 for devices of compute capability 2.0.
4.3 Multiple Devices
In a system with multiple GPUs, all CUDA-enabled GPUs are accessible via the
CUDA driver and runtime as separate devices. There are however special
considerations as described below when the system is in SLI mode.
First, an allocation in one CUDA device on one GPU will consume memory on
other GPUs. Because of this, allocations may fail earlier than otherwise expected.
Second, when a Direct3D application runs in SLI Alternate Frame Rendering mode,
the Direct3D device(s) created by that application can be used for CUDA-Direct3D
interoperability (i.e., passed as a parameter to
cudaD3D[9|10]SetDirect3DDevice() when using the runtime API), but only
one CUDA device can be created at a time from one of these Direct3D devices.
This CUDA device only executes the CUDA work on one of the GPUs in the SLI
configuration. As a consequence, real interoperability only happens with the copy of
a Direct3D resource in that GPU (note: in AFR mode Direct3D resources that must
be in GPU memory are duplicated in the GPU memory of each GPU in the SLI
Chapter 4. Hardware Implementation


80 CUDA Programming Guide Version 3.0

configuration). In some cases this is not the desired behavior and an application may
need to forfeit use of the CUDA-Direct3D interoperability API and manually copy
the output of its CUDA work to Direct3D resources using the existing CUDA and
Direct3D API.






CUDA Programming Guide Version 3.0 81

Chapter 5.
Performance Guidelines
5.1 Overall Performance Optimization Strategies
Performance optimization revolves around three basic strategies:
 Maximize parallel execution to achieve maximum utilization;
 Optimize memory usage to achieve maximum memory throughput;
 Optimize instruction usage to achieve maximum instruction throughput.
Which strategies will yield the best performance gain for a particular portion of an
application depends on the performance limiters for that portion; optimizing
instruction usage of a kernel that is mostly limited by memory accesses will not yield
any significant performance gain, for example. Optimization efforts should
therefore be constantly directed by measuring and monitoring the performance
limiters, for example using the CUDA profiler. Also, comparing the floating-point
operation throughput or memory throughput – whichever makes more sense – of a
particular kernel to the corresponding peak theoretical throughput of the device
indicates how much room for improvement there is for the kernel.
5.2 Maximize Utilization
To maximize utilization the application should be structured in a way that it exposes
as much parallelism as possible and efficiently maps this parallelism to the various
components of the system to keep them busy most of the time.
5.2.1 Application Level
At a high level, the application should maximize parallel execution between the host,
the devices, and the bus connecting the host to the devices, by using asynchronous
functions calls and streams as described in Section 3.2.6. It should assign to each
processor the type of work it does best: serial workloads to the host; parallel
workloads to the devices.
For the parallel workloads, at points in the algorithm where parallelism is broken
because some threads need to synchronize in order to share data with each other,
there are two cases: Either these threads belong to the same block, in which case
Appendix E. Texture Fetching


82 CUDA Programming Guide Version 3.0

they should use __syncthreads() and share data through shared memory within
the same kernel invocation, or they belong to different blocks, in which case they
must share data through global memory using two separate kernel invocations, one
for writing to and one for reading from global memory. The second case is much
less optimal since it adds the overhead of extra kernel invocations and global
memory traffic. Its occurrence should therefore be minimized by mapping the
algorithm to the CUDA programming model in such a way that the computations
that require inter-thread communication are performed within a single thread block
as much as possible.
5.2.2 Device Level
At a lower level, the application should maximize parallel execution between the
multiprocessors of a device.
For devices of compute capability 1.x, only one kernel can execute on a device at
one time, so the kernel should be launched with at least as many thread blocks as
there are multiprocessors in the device.
For devices of compute capability 2.0, multiple kernels can execute concurrently on
a device, so maximum utilization can also be achieved by using streams to enable
enough kernels to execute concurrently as described in Section 3.2.6.
5.2.3 Multiprocessor Level
At an even lower level, the application should maximize parallel execution between
the various functional units within a multiprocessor.
As described in Section 4.2, a GPU multiprocessor relies on thread-level parallelism
to maximize utilization of its functional units. Utilization is therefore directly linked
to the number of resident warps. At every instruction issue time, a warp scheduler
selects a warp that is ready to execute and issues the next instruction to the active
threads of the warp. The number of clock cycles it takes for a warp to be ready to
execute its next instruction is called latency, and full utilization is achieved when the
warp scheduler always has some instruction to issue for some warp at every clock
cycle during that latency period, or in other words, when the latency of each warp is
completely “hidden” by other warps. How many instructions are required to hide
latency depends on the instruction throughput. For example, to hide a latency of L
clock cycles with basic single-precision floating-point arithmetic instructions
(scheduled on CUDA cores):
 L/4 (rounded up to nearest integer) instructions are required for devices of
compute capability 1.x since a multiprocessor issues one such instruction per
warp over 4 clock cycles, as mentioned in Section G.3.1,
 L/2 (rounded up to nearest integer) instructions are required for devices of
compute capability 2.0 since a multiprocessor issues the two instructions for a
pair of warps over 2 clock cycles, as mentioned in Section G.4.1.
The most common reason a warp is not ready to execute its next instruction is that
the instruction‟s input operands are not yet available.
Appendix E. Texture Fetching


CUDA Programming Guide Version 3.0 83

If all input operands are registers, latency is caused by register dependencies, i.e.,
some of the input operands are written by some previous instruction(s) whose
execution has not completed yet. In the case of a back-to-back register dependency
(i.e., some input operand is written by the previous instruction), the latency is equal
to the execution time of the previous instruction and the warp scheduler must
schedule instructions for different warps during that time. Execution time varies
depending on the instruction, but it is typically about 22 clock cycles, which
translates to 6 warps for devices of compute capability 1.x and 11 warps for devices
of compute capability 2.0.
If some input operand resides in off-chip memory, the latency is much higher: 400
to 800 clock cycles. The number of warps required to keep the warp scheduler busy
during such high latency periods depends on the kernel code; in general, more warps
are required if the ratio of the number of instructions with no off-chip memory
operands (i.e., arithmetic instructions most of the time) to the number of
instructions with off-chip memory operands is low (this ratio is commonly called
the arithmetic intensity of the program). If this ratio is 10, for example, then to hide
latencies of about 600 clock cycles, about 15 warps are required for devices of
compute capability 1.x and about 30 for devices of compute capability 2.0.
Another reason a warp is not ready to execute its next instruction is that it is waiting
at some memory fence (Section B.5) or synchronization point (Section B.6). A
synchronization point can force the multiprocessor to idle as more and more warps
wait for other warps in the same block to complete execution of instructions prior
to the synchronization point. Having multiple resident blocks per multiprocessor
can help reduce idling in this case, as warps from different blocks do not need to
wait for each other at synchronization points.
The number of blocks and warps residing on each multiprocessor for a given kernel
call depends on the execution configuration of the call (Section B.13), the memory
resources of the multiprocessor, and the resource requirements of the kernel as
described in Section 4.2. To assist programmers in choosing thread block size based
on register and shared memory requirements, the CUDA Software Development
Kit provides a spreadsheet, called the CUDA Occupancy Calculator, where
occupancy is defined as the ratio of the number of resident warps to the maximum
number of resident warps (given in Appendix G for various compute capabilities).
The number of registers used by a kernel can have a significant impact on the
number of resident warps. For example, for devices of compute capability 1.2, if a
kernel uses 16 registers and each block has 512 threads and requires very little
shared memory, then two blocks (i.e., 32 warps) can reside on the multiprocessor
since they require 2x512x16 registers, which exactly matches the number of registers
available on the multiprocessor. But as soon as the kernel uses one more register,
only one block (i.e., 16 warps) can be resident since two blocks would require
2x512x17 registers, which is more registers than are available on the multiprocessor.
Therefore, the compiler attempts to minimize register usage while keeping register
spilling (see Section 5.3.2.2) and the number of instructions to a minimum. Register
usage can be controlled using the -maxrregcount compiler option or launch
bounds as described in Section B.14.
The total amount of shared memory required for a block is equal to the sum of the
amount of statically allocated shared memory, the amount of dynamically allocated
shared memory, and for devices of compute capability 1.x, the amount of shared
memory used to pass the kernel‟s arguments (see Section B.1.4).
Appendix E. Texture Fetching


84 CUDA Programming Guide Version 3.0

Register, local, shared, and constant memory usages are reported by the compiler
when compiling with the --ptxas-options=-v option. Note that each double
variable (on devices that supports native double precision, i.e. devices of compute
capability 1.2 and higher) and each long long variable uses two registers.
However, devices of compute capability 1.2 and higher have at least twice as many
registers per multiprocessor as devices with lower compute capability.
The number of threads per block should be chosen as a multiple of the warp size to
avoid wasting computing resources with under-populated warps if possible.
As mentioned above, the effect of execution configuration on performance for a
given kernel call generally depends on the kernel code. Experimentation is therefore
recommended. Applications can also parameterize execution configurations based
on register file size and shared memory size, which depends on the compute
capability of the device, as well as on the number of multiprocessors and memory
bandwidth of the device, all of which can be queried using the runtime or driver
API (see reference manual).
5.3 Maximize Memory Throughput
The first step in maximizing overall memory throughput for the application is to
minimize data transfers with low bandwidth.
That means minimizing data transfers between the host and the device, as detailed
in Section 5.3.1, since these have much lower bandwidth than data transfers
between global memory and the device.
That also means minimizing data transfers between global memory and the device
by maximizing use of on-chip memory: shared memory and caches (i.e. L1/L2
caches available on devices of compute capability 2.0, texture cache and constant
cache available on all devices).
Shared memory is equivalent to a user-managed cache: The application explicitly
allocates and accesses it. As illustrated in Section 3.2.2, a typical programming
pattern is to stage data coming from device memory into shared memory; in other
words, to have each thread of a block:
 Load data from device memory to shared memory,
 Synchronize with all the other threads of the block so that each thread can
safely read shared memory locations that were populated by different threads,
 Process the data in shared memory,
 Synchronize again if necessary to make sure that shared memory has been
updated with the results,
 Write the results back to device memory.
For some applications (e.g. for which global memory accesses are data-dependent), a
traditional hardware-managed cache is more appropriate to exploit data locality. As
mentioned in Section G.4.1, for devices of compute capability 2.0, the same on-chip
memory is used for both L1 and shared memory, and how much of it is dedicated to
L1 versus shared memory is configurable for each kernel call.
The throughput of memory accesses by a kernel can vary by an order of magnitude
depending on access pattern for each type of memory. The next step in maximizing
Appendix E. Texture Fetching


CUDA Programming Guide Version 3.0 85

memory throughput is therefore to organize memory accesses as optimally as
possible based on the optimal memory access patterns described in Sections 5.3.2.1,
5.3.2.3, 5.3.2.4, and 5.3.2.5. This optimization is especially important for global
memory accesses as global memory bandwidth is low, so non-optimal global
memory accesses have a higher impact on performance.
5.3.1 Data Transfer between Host and Device
Applications should strive to minimize data transfer between the host and the
device. One way to accomplish this is to move more code from the host to the
device, even if that means running kernels with low parallelism computations.
Intermediate data structures may be created in device memory, operated on by the
device, and destroyed without ever being mapped by the host or copied to host
memory.
Also, because of the overhead associated with each transfer, batching many small
transfers into a single large transfer always performs better than making each
transfer separately.
On systems with a front-side bus, higher performance for data transfers between
host and device is achieved by using page-locked host memory as described in
Section 3.2.5.
In addition, when using mapped page-locked memory (Section 3.2.5.3), there is no
need to allocate any device memory and explicitly copy data between device and
host memory. Data transfers are implicitly performed each time the kernel accesses
the mapped memory. For maximum performance, these memory accesses must be
coalesced as with accesses to global memory (see Section 5.3.2.1). Assuming that
they are and that the mapped memory is read or written only once, using mapped
page-locked memory instead of explicit copies between device and host memory can
be a win for performance.
On integrated systems where device memory and host memory are physically the
same, any copy between host and device memory is superfluous and mapped page-
locked memory should be used instead. Applications may query whether a device is
integrated or not by calling cudaGetDeviceProperties() and checking the
integrated property or checking the CU_DEVICE_ATTRIBUTE_INTEGRATED
attribute using cuDeviceGetAttribute().
5.3.2 Device Memory Accesses
An instruction that accesses addressable memory (i.e., global, local, shared, constant,
or texture memory) might need to be re-issued multiple times depending on the
distribution of the memory addresses across the threads within the warp. How the
distribution affects the instruction throughput this way is specific to each type of
memory and described in the following sections. For example, for global memory,
as a general rule, the more scattered the addresses are, the more reduced the
throughput is.
Appendix E. Texture Fetching


86 CUDA Programming Guide Version 3.0

5.3.2.1 Global Memory
Global memory resides in device memory and device memory is accessed via 32-,
64-, or 128-byte memory transactions. These memory transactions must be naturally
aligned: Only the 32-, 64-, or 128-byte segments of device memory that are aligned
to their size (i.e. whose first address is a multiple of their size) can be read or written
by memory transactions.
When a warp executes an instruction that accesses global memory, it coalesces the
memory accesses of the threads within the warp into one or more of these memory
transactions depending on the size of the word accessed by each thread and the
distribution of the memory addresses across the threads. In general, the more
transactions are necessary, the more unused words are transferred in addition to the
words accessed by the threads, reducing the instruction throughput accordingly. For
example, if a 32-byte memory transaction is generated for each thread‟s 4-byte
access, throughput is divided by 8.
How many transactions are necessary and how throughput is ultimately affected
varies with the compute capability of the device. For devices of compute capability
1.0 and 1.1, the requirements on the distribution of the addresses across the threads
to get any coalescing at all are very strict. They are much more relaxed for devices of
higher compute capabilities. For devices of compute capability 2.0, the memory
transactions are cached, so data locality is exploited to reduce impact on throughput.
Sections G.3.2 and G.4.2 give more details on how global memory accesses are
handled for various compute capabilities.
To maximize global memory throughput, it is therefore important to maximize
coalescing by:
 Following the most optimal access patterns based on Sections G.3.2 and G.4.2,
 Using data types that meet the size and alignment requirement detailed in
Section 5.3.2.1.1,
 Padding data in some cases, for example, when accessing a two-dimensional
array as described in Section 5.3.2.1.2.
5.3.2.1.1 Size and Alignment Requirement
Global memory instructions support reading or writing words of size equal to 1, 2,
4, 8, or 16 bytes. Any access (via a variable or a pointer) to data residing in global
memory compiles to a single global memory instruction if and only if the size of the
data type is 1, 2, 4, 8, or 16 bytes and the data is naturally aligned (i.e. its address is a
multiple of that size).
If this size and alignment requirement is not fulfilled, the access compiles to
multiple instructions with interleaved access patterns that prevent these instructions
from fully coalescing. It is therefore recommended to use types that meet this
requirement for data that resides in global memory.
The alignment requirement is automatically fulfilled for the built-in types of
Section B.3.1 like float2 or float4.
For structures, the size and alignment requirements can be enforced by the compiler
using the alignment specifiers __align__(8) or __align__(16), such as
struct __align__(8) {
float x;
Appendix E. Texture Fetching


CUDA Programming Guide Version 3.0 87

float y;
};
or
struct __align__(16) {
float x;
float y;
float z;
};
Any address of a variable residing in global memory or returned by one of the
memory allocation routines from the driver or runtime API is always aligned to at
least 256 bytes.
Reading non-naturally aligned 8-byte or 16-byte words produces incorrect results
(off by a few words), so special care must be taken to maintain alignment of the
starting address of any value or array of values of these types. A typical case where
this might be easily overlooked is when using some custom global memory
allocation scheme, whereby the allocations of multiple arrays (with multiple calls to
cudaMalloc() or cuMemAlloc()) is replaced by the allocation of a single large
block of memory partitioned into multiple arrays, in which case the starting address
of each array is offset from the block‟s starting address.
5.3.2.1.2 Two-Dimensional Arrays
A common global memory access pattern is when each thread of index (tx,ty)
uses the following address to access one element of a 2D array of width width,
located at address BaseAddress of type type* (where type meets the
requirement described in Section 5.3.2.1.1):
BaseAddress + width * ty + tx
For these accesses to be fully coalesced, both the width of the thread block and the
width of the array must be a multiple of the warp size (or only half the warp size for
devices of compute capability 1.x).
In particular, this means that an array whose width is not a multiple of this size will
be accessed much more efficiently if it is actually allocated with a width rounded up
to the closest multiple of this size and its rows padded accordingly. The
cudaMallocPitch() and cuMemAllocPitch() functions and associated
memory copy functions described in the reference manual enable programmers to
write non-hardware-dependent code to allocate arrays that conform to these
constraints.
5.3.2.2 Local Memory
Local memory accesses only occur for some automatic variables as mentioned in
Section B.2.5. Automatic variables that the compiler is likely to place in local
memory are:
 Arrays for which it cannot determine that they are indexed with constant
quantities,
 Large structures or arrays that would consume too much register space,
 Any variable if the kernel uses more registers than available (this is also known
as register spilling).
Inspection of the PTX assembly code (obtained by compiling with the –ptx or
-keep option) will tell if a variable has been placed in local memory during the first
Appendix E. Texture Fetching


88 CUDA Programming Guide Version 3.0

compilation phases as it will be declared using the .local mnemonic and accessed
using the ld.local and st.local mnemonics. Even if it has not, subsequent
compilation phases might still decide otherwise though if they find it consumes too
much register space for the targeted architecture: Inspection of the cubin object
using cuobjdump will tell if this is the case. Also, the compiler reports total local
memory usage per kernel (lmem) when compiling with the --ptxas-options=-v
option. Note that some mathematical functions have implementation paths that
might access local memory.
The local memory space resides in device memory, so local memory accesses have
same high latency and low bandwidth as global memory accesses and are subject to
the same requirements for memory coalescing as described in Section 5.3.2.1. Local
memory is however organized such that consecutive 32-bit words are accessed by
consecutive thread IDs. Accesses are therefore fully coalesced as long as all threads
in a warp access the same relative address (e.g. same index in an array variable, same
member in a structure variable).
On devices of compute capability 2.0, local memory accesses are always cached in
L1 and L2 in the same way as global memory accesses (see Section G.4.2).
5.3.2.3 Shared Memory
Because it is on-chip, the shared memory space is much faster than the local and
global memory spaces. In fact, for all threads of a warp, accessing shared memory is
fast as long as there are no bank conflicts between the threads, as detailed below.
To achieve high bandwidth, shared memory is divided into equally-sized memory
modules, called banks, which can be accessed simultaneously. Any memory read or
write request made of n addresses that fall in n distinct memory banks can therefore
be serviced simultaneously, yielding an overall bandwidth that is n times as high as
the bandwidth of a single module.
However, if two addresses of a memory request fall in the same memory bank, there
is a bank conflict and the access has to be serialized. The hardware splits a memory
request with bank conflicts into as many separate conflict-free requests as necessary,
decreasing throughput by a factor equal to the number of separate memory requests.
If the number of separate memory requests is n, the initial memory request is said to
cause n-way bank conflicts.
To get maximum performance, it is therefore important to understand how memory
addresses map to memory banks in order to schedule the memory requests so as to
minimize bank conflicts. This is described in Sections G.3.3 and G.4.3 for devices of
compute capability 1.x and 2.0, respectively.
5.3.2.4 Constant Memory
The constant memory space resides in device memory and is cached in the constant
cache mentioned in Sections G.3.1 and G.4.1.
For devices of compute capability 1.x, a constant memory request for a warp is first
split into two requests, one for each half-warp, that are issued independently.
A request is then split into as many separate requests as there are different memory
addresses in the initial request, decreasing throughput by a factor equal to the
number of separate requests.
Appendix E. Texture Fetching


CUDA Programming Guide Version 3.0 89

The resulting requests are then serviced at the throughput of the constant cache in
case of a cache hit, or at the throughput of device memory otherwise.
5.3.2.5 Texture Memory
The texture memory space resides in device memory and is cached in texture cache,
so a texture fetch costs one memory read from device memory only on a cache
miss, otherwise it just costs one read from texture cache. The texture cache is
optimized for 2D spatial locality, so threads of the same warp that read texture
addresses that are close together in 2D will achieve best performance. Also, it is
designed for streaming fetches with a constant latency; a cache hit reduces DRAM
bandwidth demand but not fetch latency.
Reading device memory through texture fetching present some benefits that can
make it an advantageous alternative to reading device memory from global or
constant memory:
 If the memory reads do not follow the access patterns that global or constant
memory reads must respect to get good performance (see Sections 5.3.2.1 and
5.3.2.4), higher bandwidth can be achieved providing that there is locality in the
texture fetches (this is less likely for devices of compute capability 2.0 given that
global memory reads are cached on these devices);
 Addressing calculations are performed outside the kernel by dedicated units;
 Packed data may be broadcast to separate variables in a single operation;
 8-bit and 16-bit integer input data may be optionally converted to 32-bit
floating-point values in the range [0.0, 1.0] or [-1.0, 1.0] (see Section 3.2.4.1).
However, within the same kernel call, the texture cache is not kept coherent with
respect to global memory writes, so that any texture fetch to an address that has
been written to via a global write in the same kernel call returns undefined data. In
other words, a thread can safely read via texture some memory location only if this
memory location has been updated by a previous kernel call or memory copy, but
not if it has been previously updated by the same thread or another thread from the
same kernel call. This is only relevant when fetching from linear memory as a kernel
cannot write to CUDA arrays in any case.
5.4 Maximize Instruction Throughput
To maximize instruction throughput the application should:
 Minimize the use of arithmetic instructions with low throughput; this includes
trading precision for speed when it does not affect the end result, such as using
intrinsic instead of regular functions (intrinsic functions are listed in
Section C.2), single-precision instead of double-precision, or flushing
denormalized numbers to zero;
 Minimize divergent warps caused by control flow instructions as detailed in
Section 5.4.2;
 Reduce the number of instructions, for example, by optimizing out
synchronization points whenever possible as described in Section 5.4.3 or by
using restricted pointers as described in Section E.3.
In this section, throughputs are given in number of operations per clock cycle per
multiprocessor. For a warp size of 32, one instruction results in 32 operations.
Appendix E. Texture Fetching


90 CUDA Programming Guide Version 3.0

Therefore, if T is the number of operations per clock cycle, the instruction
throughput is one instruction every 32/T clock cycles.
All throughputs are for one multiprocessor. They must be multiplied by the number
of multiprocessors in the device to get throughput for the whole device.
5.4.1 Arithmetic Instructions
Table 5-1 gives the throughputs of the arithmetic instructions that are natively
supported in hardware for devices of various compute capabilities. For devices of
compute capability 2.0, two different warps execute half of the operations each
clock cycle (see Section G.4.1).
Table 5-1. Throughput of Native Arithmetic Instructions
(Operations per Clock Cycle per Multiprocessor)

Compute
Capability 1.x
Compute
Capability 2.0
32-bit floating-point
add, multiply, multiply-add
8 32
64-bit floating-point
add, multiply, multiply-add
1 16
32-bit integer
add, logical operation, shift, compare
8 32
24-bit integer multiply (__[u]mul24(x,y)) 8 Multiple instructions
32-bit integer
multiply, multiply-add, sum of absolute difference
Multiple instructions 32
32-bit floating-point
reciprocal, reciprocal square root,
base-2 logarithm (__log2f),
base-2 exponential (exp2f),
sine (__sinf), cosine (__cosf)
2 4
Type conversions 8 32

Other instructions and functions are implemented on top of the native instructions.
The implementation may be different for devices of compute capability 1.x and
devices of compute capability 2.0, and the number of native instructions after
compilation may fluctuate with every compiler version. For complicated functions,
there can be multiple code paths depending on input. cuobjdump can be used to
inspect a particular implementation in a cubin object.
The implementation of some functions are readily available on the CUDA header
files (math_functions.h, device_functions.h, …).
In general, code compiled with -ftz=true (denormalized numbers are flushed to
zero) tends to have higher performance than code compiled with -ftz=false.
Similarly, code compiled with -prec-div=false (less precise division) tends to
have higher performance code than code compiled with -prec-div=true, and
code compiled with -prec-sqrt=false (less precise square root) tends to have
Appendix E. Texture Fetching


CUDA Programming Guide Version 3.0 91

higher performance than code compiled with -prec-sqrt=true. The nvcc user
manual describes these compilation flags in more details.
Single-Precision Floating-Point Addition and Multiplication Intrinsics
__fadd_r[d,u], __fmul_r[d,u], and __fmaf_r[n,z,d,u] (see
Section C.2.1) compile to tens of instructions for devices of compute capability 1.x,
but map to a single native instruction for devices of compute capability 2.0.
Single-Precision Floating-Point Division
__fdividef(x, y) (see Section C.2.1) provides faster single-precision floating-
point division than the division operator.
Single-Precision Floating-Point Reciprocal Square Root
To preserve IEEE-754 semantics the compiler can optimize 1.0/sqrtf() into
rsqrtf() only when both reciprocal and square root are approximate, (i.e. with
-prec-div=false and -prec-sqrt=false). It is therefore recommended to
invoke rsqrtf() directly where desired.
Single-Precision Floating-Point Square Root
Single-precision floating-point square root is implemented as a reciprocal square
root followed by a reciprocal instead of a reciprocal square root followed by a
multiplication so that it gives correct results for 0 and infinity. Therefore, its
throughput is 1 operation per clock cycle for devices of compute capability 1.x and
2 operations per clock cycle for devices of compute capability 2.0.
Sine and Cosine
sinf(x), cosf(x), tanf(x), sincosf(x), and corresponding double-
precision instructions are much more expensive and even more so if the argument x
is large in magnitude.
More precisely, the argument reduction code (see math_functions.h for
implementation) comprises two code paths referred to as the fast path and the slow
path, respectively.
The fast path is used for arguments sufficiently small in magnitude and essentially
consists of a few multiply-add operations. The slow path is used for arguments large
in magnitude and consists of lengthy computations required to achieve correct
results over the entire argument range.
At present, the argument reduction code for the trigonometric functions selects the
fast path for arguments whose magnitude is less than 48039.0f for the single-
precision functions, and less than 2147483648.0 for the double-precision functions.
As the slow path requires more registers than the fast path, an attempt has been
made to reduce register pressure in the slow path by storing some intermediate
variables in local memory, which may affect performance because of local memory
high latency and bandwidth (see Section 5.3.2.2). At present, 28 bytes of local
memory are used by single-precision functions, and 44 bytes are used by double-
precision functions. However, the exact amount is subject to change.
Due to the lengthy computations and use of local memory in the slow path, the
throughput of these trigonometric functions is lower by one order of magnitude
when the slow path reduction is required as opposed to the fast path reduction.
Appendix E. Texture Fetching


92 CUDA Programming Guide Version 3.0

Integer Arithmetic
On devices of compute capability 1.x, 32-bit integer multiplication is implemented
using multiple instructions as it is not natively supported. 24-bit integer
multiplication is natively supported however via the __[u]mul24 intrinsic (see
Section C.2.3). Using __[u]mul24 instead of the 32-bit multiplication operator
whenever possible usually improves performance for instruction bound kernels. It
can have the opposite effect however in cases where the use of __[u]mul24
inhibits compiler optimizations.
On devices of compute capability 2.0, 32-bit integer multiplication is natively
supported, but 24-bit integer multiplication is not. __[u]mul24 is therefore
implemented using multiple instructions and should not be used.
Integer division and modulo operation are particularly costly and should be avoided
if possible or replaced with bitwise operations whenever possible: If n is a power of
2, (i/n) is equivalent to (i>>log2(n)) and (i%n) is equivalent to (i&(n-1));
the compiler will perform these conversions if n is literal.
__brev, __brevll, __popc, and __popcll (see Section C.2.3) compile to tens
of instructions for devices of compute capability 1.x, but __brev and __popc map
to a single instruction for devices of compute capability 2.0 and __brevll and
__popcll to just a few.
__clz, __clzll, __ffs, and __ffsll (see Section C.2.3) compile to fewer
instructions for devices of compute capability 2.0 than for devices of compute
capability 1.x.
Type Conversion
Sometimes, the compiler must insert conversion instructions, introducing additional
execution cycles. This is the case for:
 Functions operating on variables of type char or short whose operands
generally need to be converted to int,
 Double-precision floating-point constants (i.e. those constants defined without
any type suffix) used as input to single-precision floating-point computations (as
mandated by C/C++ standards).
This last case can be avoided by using single-precision floating-point constants,
defined with an f suffix such as 3.141592653589793f, 1.0f, 0.5f.
5.4.2 Control Flow Instructions
Any flow control instruction (if, switch, do, for, while) can significantly
impact the effective instruction throughput by causing threads of the same warp to
diverge (i.e. to follow different execution paths). If this happens, the different
executions paths have to be serialized, increasing the total number of instructions
executed for this warp. When all the different execution paths have completed, the
threads converge back to the same execution path.
To obtain best performance in cases where the control flow depends on the thread
ID, the controlling condition should be written so as to minimize the number of
divergent warps. This is possible because the distribution of the warps across the
block is deterministic as mentioned in Section 4.1. A trivial example is when the
Appendix E. Texture Fetching


CUDA Programming Guide Version 3.0 93

controlling condition only depends on (threadIdx / warpSize) where
warpSize is the warp size. In this case, no warp diverges since the controlling
condition is perfectly aligned with the warps.
Sometimes, the compiler may unroll loops or it may optimize out if or switch
statements by using branch predication instead, as detailed below. In these cases, no
warp can ever diverge. The programmer can also control loop unrolling using the
#pragma unroll directive (see Section E.2).
When using branch predication none of the instructions whose execution depends
on the controlling condition gets skipped. Instead, each of them is associated with a
per-thread condition code or predicate that is set to true or false based on the
controlling condition and although each of these instructions gets scheduled for
execution, only the instructions with a true predicate are actually executed.
Instructions with a false predicate do not write results, and also do not evaluate
addresses or read operands.
The compiler replaces a branch instruction with predicated instructions only if the
number of instructions controlled by the branch condition is less or equal to a
certain threshold: If the compiler determines that the condition is likely to produce
many divergent warps, this threshold is 7, otherwise it is 4.
5.4.3 Synchronization Instruction
Throughput for __syncthreads() is 8 operations per clock cycle for devices of
compute capability 1.x and 16 operations per clock cycle for devices of compute
capability 2.0.
Note that __syncthreads() can impact performance by forcing the
multiprocessor to idle as detailed in Section 5.2.3.
Because a warp executes one common instruction at a time, threads within a warp
are implicitly synchronized and this can sometimes be used to omit
__syncthreads() for better performance.
In the following code sample, for example, both calls to __syncthreads() are
required to get the expected result (i.e. result[i] = 2 * myArray[i] for
i > 0). Without synchronization, any of the two references to myArray[tid]
could return either 2 or the value initially stored in myArray, depending on whether
the memory read occurs before or after the memory write from
myArray[tid + 1] = 2.
// myArray is an array of integers located in global or shared
// memory
__global__ void MyKernel(int* result) {
int tid = threadIdx.x;
...
int ref1 = myArray[tid];
__syncthreads();
myArray[tid + 1] = 2;
__syncthreads();
int ref2 = myArray[tid];
result[tid] = ref1 * ref2;
...
}
Appendix E. Texture Fetching


94 CUDA Programming Guide Version 3.0

However, in the following slightly modified code sample, threads are guaranteed to
belong to the same warp, so that there is no need for any __syncthreads().
// myArray is an array of integers located in global or shared
// memory
__global__ void MyKernel(int* result) {
int tid = threadIdx.x;
...
if (tid < warpSize) {
int ref1 = myArray[tid];
myArray[tid + 1] = 2;
int ref2 = myArray[tid];
result[tid] = ref1 * ref2;
}
...
}
Simply removing the __syncthreads() is not enough however; myArray must
also be declared as volatile as described in Section B.2.4.







CUDA Programming Guide Version 3.0 95

Appendix A.
CUDA-Enabled GPUs
Table A-1 lists all CUDA-enabled devices with their compute capability, number of
multiprocessors, and number of CUDA cores.
These, as well as the clock frequency and the total amount of device memory, can
be queried using the runtime or driver API (see reference manual).
Table A-1. CUDA-Enabled Devices with Compute Capability,
Number of Multiprocessors, and Number of CUDA
Cores
Compute
Capability
Number of
Multiprocessors
Number of
CUDA Cores
GeForce GTX 295 1.3 2x30 2x240
GeForce GTX 285, GTX 280 1.3 30 240
GeForce GTX 260 1.3 24 192
GeForce 9800 GX2 1.1 2x16 2x128
GeForce GTS 250, GTS 150,
9800 GTX, 9800 GTX+,
8800 GTS 512, GTX 285M,
GTX 280M
1.1 16 128
GeForce 8800 Ultra, 8800 GTX 1.0 16 128
GeForce 9800 GT, 8800 GT,
GTX 260M, 9800M GTX
1.1 14 112
GeForce GT 240, GTS 360M,
GTS 350M
1.2 12 96
GeForce GT 130, 9600 GSO,
8800 GS, 8800M GTX, GTS 260M,
GTS 250M, 9800M GT
1.1 12 96
GeForce 8800 GTS 1.0 12 96
GeForce GT 335M 1.2 9 72
GeForce 9600 GT, 8800M GTS,
9800M GTS
1.1 8 64
GeForce GT 220, GT 330M,
GT 325M
1.2 6 48
GeForce 9700M GT, GT 240M,
GT 230M
1.1 6 48
Chapter 5. Performance Guidelines


96 CUDA Programming Guide Version 3.0

Compute
Capability
Number of
Multiprocessors
Number of
CUDA Cores
GeForce GT 120, 9500 GT,
8600 GTS, 8600 GT, 9700M GT,
9650M GS, 9600M GT, 9600M GS,
9500M GS, 8700M GT, 8600M GT,
8600M GS
1.1 4 32
GeForce 210, 310M, 305M 1.2 2 16
GeForce G100, 8500 GT, 8400 GS,
8400M GT, 9500M G, 9300M G,
8400M GS, 9400 mGPU,
9300 mGPU, 8300 mGPU,
8200 mGPU, 8100 mGPU,
G210M,G110M
1.1 2 16
GeForce 9300M GS, 9200M GS,
9100M G, 8400M G, G105M
1.1 1 8
Tesla S1070 1.3 4x30 4x240
Tesla C1060 1.3 30 240
Tesla S870 1.0 4x16 4x128
Tesla D870 1.0 2x16 2x128
Tesla C870 1.0 16 128
Quadro Plex 2200 D2 1.3 2x30 2x240
Quadro Plex 2100 D4 1.1 4x14 4x112
Quadro Plex 2100 Model S4 1.0 4x16 4x128
Quadro Plex 1000 Model IV 1.0 2x16 2x128
Quadro FX 5800 1.3 30 240
Quadro FX 4800 1.3 24 192
Quadro FX 4700 X2 1.1 2x14 2x112
Quadro FX 3700M, FX 3800M 1.1 16 128
Quadro FX 5600 1.0 16 128
Quadro FX 3700 1.1 14 112
Quadro FX 2800M 1.1 12 96
Quadro FX 4600 1.0 12 96
Quadro FX 1800M 1.2 9 72
Quadro FX 3600M 1.1 8 64
Quadro FX 880M, NVS 5100M 1.2 6 48
Quadro FX 2700M 1.1 6 48
Quadro FX 1700, FX 570,
NVS 320M, FX 1700M, FX 1600M,
FX 770M, FX 570M
1.1 4 32
Quadro FX 380 LP, FX 380M,
NVS 3100M, NVS 2100M
1.2 2 16
Quadro FX 370, NVS 290,
NVS 140M, NVS 135M, FX 360M
1.1 2 16
Quadro FX 370M, NVS 130M 1.1 1 8








CUDA Programming Guide Version 3.0 97

Appendix B.
C Language Extensions
B.1 Function Type Qualifiers
Function type qualifiers specify whether a function executes on the host or on the
device and whether it is callable from the host or from the device.
B.1.1 __device__
The __device__ qualifier declares a function that is:
 Executed on the device
 Callable from the device only.
B.1.2 __global__
The __global__ qualifier declares a function as being a kernel. Such a function is:
 Executed on the device,
 Callable from the host only.
B.1.3 __host__
The __host__ qualifier declares a function that is:
 Executed on the host,
 Callable from the host only.
It is equivalent to declare a function with only the __host__ qualifier or to declare
it without any of the __host__, __device__, or __global__ qualifier; in either
case the function is compiled for the host only.
However, the __host__ qualifier can also be used in combination with the
__device__ qualifier, in which case the function is compiled for both the host and
the device.
Appendix A. C Language Extensions


98 CUDA Programming Guide Version 3.0

B.1.4 Restrictions
__device__ and __global__ functions do not support recursion.
__device__ and __global__ functions cannot declare static variables inside
their body.
__device__ and __global__ functions cannot have a variable number of
arguments.
__device__ functions cannot have their address taken; function pointers to
__global__ functions, on the other hand, are supported.
The __global__ and __host__ qualifiers cannot be used together.
__global__ functions must have void return type.
Any call to a __global__ function must specify its execution configuration as
described in Section B.13.
A call to a __global__ function is asynchronous, meaning it returns before the
device has completed its execution.
__global__ function parameters are passed to the device:
 via shared memory and are limited to 256 bytes on devices of compute
capability 1.x,
 via constant memory and are limited to 4 KB on devices of compute capability
2.0.
B.2 Variable Type Qualifiers
Variable type qualifiers specify the memory location on the device of a variable.
B.2.1 __device__
The __device__ qualifier declares a variable that resides on the device.
At most one of the other type qualifiers defined in the next three sections may be
used together with __device__ to further specify which memory space the
variable belongs to. If none of them is present, the variable:
 Resides in global memory space,
 Has the lifetime of an application,
 Is accessible from all the threads within the grid and from the host through the
runtime library.
B.2.2 __constant__
The __constant__ qualifier, optionally used together with __device__,
declares a variable that:
 Resides in constant memory space,
Appendix A. C Language Extensions


CUDA Programming Guide Version 3.0 99

 Has the lifetime of an application,
 Is accessible from all the threads within the grid and from the host through the
runtime library.
B.2.3 __shared__
The __shared__ qualifier, optionally used together with __device__, declares a
variable that:
 Resides in the shared memory space of a thread block,
 Has the lifetime of the block,
 Is only accessible from all the threads within the block.
When declaring a variable in shared memory as an external array such as
extern __shared__ float shared[];
the size of the array is determined at launch time (see Section B.13). All variables
declared in this fashion, start at the same address in memory, so that the layout of
the variables in the array must be explicitly managed through offsets. For example, if
one wants the equivalent of
short array0[128];
float array1[64];
int array2[256];
in dynamically allocated shared memory, one could declare and initialize the arrays
the following way:
extern __shared__ char array[];
__device__ void func() // __device__ or __global__ function
{
short* array0 = (short*)array;
float* array1 = (float*)&array0[128];
int* array2 = (int*)&array1[64];
}
Note that pointers need to be aligned to the type they point to, so the following
code, for example, does not work since array1 is not aligned to 4 bytes.
extern __shared__ char array[];
__device__ void func() // __device__ or __global__ function
{
short* array0 = (short*)array;
float* array1 = (float*)&array0[127];
}
Alignment requirements for the built-in vector types are listed in Table B-1.
B.2.4 volatile
Only after the execution of a __threadfence_block(), __threadfence(),
or __syncthreads() (Sections B.5 and B.6) are prior writes to global or shared
memory guaranteed to be visible by other threads. As long as this requirement is
met, the compiler is free to optimize reads and writes to global or shared memory.
For example, in the code sample below, the first reference to myArray[tid]
Appendix A. C Language Extensions


100 CUDA Programming Guide Version 3.0

compiles into a global or shared memory read instruction, but the second reference
does not as the compiler simply reuses the result of the first read.
// myArray is an array of non-zero integers
// located in global or shared memory
__global__ void MyKernel(int* result) {
int tid = threadIdx.x;
int ref1 = myArray[tid] * 1;
myArray[tid + 1] = 2;
int ref2 = myArray[tid] * 1;
result[tid] = ref1 * ref2;
}
Therefore, ref2 cannot possibly be equal to 2 in thread tid as a result of thread
tid-1 overwriting myArray[tid] by 2.
This behavior can be changed using the volatile keyword: If a variable located in
global or shared memory is declared as volatile, the compiler assumes that its value
can be changed at any time by another thread and therefore any reference to this
variable compiles to an actual memory read instruction.
Note that even if myArray is declared as volatile in the code sample above, there is
no guarantee, in general, that ref2 will be equal to 2 in thread tid since thread
tid might read myArray[tid] into ref2 before thread tid-1 overwrites its
value by 2. Synchronization is required as mentioned in Section 5.4.3.
B.2.5 Restrictions
These qualifiers are not allowed on struct and union members, on formal
parameters and on local variables within a function that executes on the host.
__shared__ and __constant__ variables have implied static storage.
__device__, __shared__ and __constant__ variables cannot be defined as
external using the extern keyword. The only exception is for dynamically allocated
__shared__ variables as described in Section B.2.3.
__device__ and __constant__ variables are only allowed at file scope.
__constant__ variables cannot be assigned to from the device, only from the
host through host runtime functions (Sections 3.2.1 and 3.3.4).
__shared__ variables cannot have an initialization as part of their declaration.
An automatic variable declared in device code without any of these qualifiers
generally resides in a register. However in some cases the compiler might choose to
place it in local memory, which can have adverse performance consequences as
detailed in Section 5.3.2.2.
Pointers in code that is executed on the device are supported as long as the compiler
is able to resolve whether they point to either the shared memory space or the
global memory space, otherwise they are restricted to only point to memory
allocated or declared in the global memory space.
Dereferencing a pointer either to global or shared memory in code that is executed
on the host or to host memory in code that is executed on the device results in an
undefined behavior, most often in a segmentation fault and application termination.
Appendix A. C Language Extensions


CUDA Programming Guide Version 3.0 101

The address obtained by taking the address of a __device__, __shared__ or
__constant__ variable can only be used in device code. The address of a
__device__ or __constant__ variable obtained through
cudaGetSymbolAddress() as described in Section 3.3.4 can only be used in
host code.
B.3 Built-in Vector Types
B.3.1 char1, uchar1, char2, uchar2, char3, uchar3,
char4, uchar4, short1, ushort1, short2, ushort2,
short3, ushort3, short4, ushort4, int1, uint1, int2,
uint2, int3, uint3, int4, uint4, long1, ulong1,
long2, ulong2, long3, ulong3, long4, ulong4,
longlong1, longlong2, float1, float2, float3, float4,
double1, double2
These are vector types derived from the basic integer and floating-point types. They
are structures and the 1
st
, 2
nd
, 3
rd
, and 4
th
components are accessible through the
fields x, y, z, and w, respectively. They all come with a constructor function of the
form make_<type name>; for example,
int2 make_int2(int x, int y);
which creates a vector of type int2 with value (x, y).
In host code, the alignment requirement of a vector type is equal to the alignment
requirement of its base type. This is not always the case in device code as detailed in
Table B-1.
Table B-1. Alignment Requirements in Device Code
Type Alignment
char1, uchar1 1
char2, uchar2 2
char3, uchar3 1
char4, uchar4 4
short1, ushort1 2
short2, ushort2 4
short3, ushort3 2
short4, ushort4 8
int1, uint1 4
int2, uint2 8
int3, uint3 4
int4, uint4 16
long1, ulong1 Same as int1 or longlong1 depending on platform
Appendix A. C Language Extensions


102 CUDA Programming Guide Version 3.0

long2, ulong2 Same as int2 or longlong2 depending on platform
long3, ulong3 Same as int3 depending on platform
long4, ulong4 Same as int4 depending on platform
longlong1 8
longlong2 16
float1 4
float2 8
float3 4
float4 16
double1 8
double2 16
B.3.2 dim3
This type is an integer vector type based on uint3 that is used to specify
dimensions. When defining a variable of type dim3, any component left unspecified
is initialized to 1.
B.4 Built-in Variables
Built-in variables specify the grid and block dimensions and the block and thread
indices. They are only valid within functions that are executed on the device.
B.4.1 gridDim
This variable is of type dim3 (see Section B.3.2) and contains the dimensions of the
grid.
B.4.2 blockIdx
This variable is of type uint3 (see Section B.3.1) and contains the block index
within the grid.
B.4.3 blockDim
This variable is of type dim3 (see Section B.3.2) and contains the dimensions of the
block.
B.4.4 threadIdx
This variable is of type uint3 (see Section B.3.1) and contains the thread index
within the block.
Appendix A. C Language Extensions


CUDA Programming Guide Version 3.0 103

B.4.5 warpSize
This variable is of type int and contains the warp size in threads (see Section 4.1
for the definition of a warp).
B.4.6 Restrictions
 It is not allowed to take the address of any of the built-in variables.
 It is not allowed to assign values to any of the built-in variables.
B.5 Memory Fence Functions
void __threadfence_block();
waits until all global and shared memory accesses made by the calling thread prior to
__threadfence_block() are visible to all threads in the thread block.
void __threadfence();
waits until all global and shared memory accesses made by the calling thread prior to
__threadfence() are visible to:
 All threads in the thread block for shared memory accesses,
 All threads in the device for global memory accesses.
void __threadfence_system();
waits until all global and shared memory accesses made by the calling thread prior to
__threadfence_system() are visible to:
 All threads in the thread block for shared memory accesses,
 All threads in the device for global memory accesses,
 Host threads for page-locked host memory accesses (see Section 3.2.5.3).
__threadfence_system() is only supported by devices of compute capability
2.0.
In general, when a thread issues a series of writes to memory in a particular order,
other threads may see the effects of these memory writes in a different order.
__threadfence_block(), __threadfence(), and
__threadfence_system() can be used to enforce some ordering.
One use case is when threads consume some data produced by other threads as
illustrated by the following code sample of a kernel that computes the sum of an
array of N numbers in one call. Each block first sums a subset of the array and
stores the result in global memory. When all blocks are done, the last block done
reads each of these partial sums from global memory and sums them to obtain the
final result. In order to determine which block is finished last, each block atomically
increments a counter to signal that it is done with computing and storing its partial
sum (see Section B.10 about atomic functions). The last block is the one that
receives the counter value equal to gridDim.x-1. If no fence is placed between
storing the partial sum and incrementing the counter, the counter might increment
before the partial sum is stored and therefore, might reach gridDim.x-1 and let
Appendix A. C Language Extensions


104 CUDA Programming Guide Version 3.0

the last block start reading partial sums before they have been actually updated in
memory.
__device__ unsigned int count = 0;
__shared__ bool isLastBlockDone;
__global__ void sum(const float* array, unsigned int N,
float* result)
{
// Each block sums a subset of the input array
float partialSum = calculatePartialSum(array, N);

if (threadIdx.x == 0) {

// Thread 0 of each block stores the partial sum
// to global memory
result[blockIdx.x] = partialSum;

// Thread 0 makes sure its result is visible to
// all other threads
__threadfence();

// Thread 0 of each block signals that it is done
unsigned int value = atomicInc(&count, gridDim.x);

// Thread 0 of each block determines if its block is
// the last block to be done
isLastBlockDone = (value == (gridDim.x - 1));
}

// Synchronize to make sure that each thread reads
// the correct value of isLastBlockDone
__syncthreads();

if (isLastBlockDone) {

// The last block sums the partial sums
// stored in result[0 .. gridDim.x-1]
float totalSum = calculateTotalSum(result);

if (threadIdx.x == 0) {

// Thread 0 of last block stores total sum
// to global memory and resets count so that
// next kernel call works properly
result[0] = totalSum;
count = 0;
}
}
}
B.6 Synchronization Functions
void __syncthreads();
Appendix A. C Language Extensions


CUDA Programming Guide Version 3.0 105

waits until all threads in the thread block have reached this point and all global and
shared memory accesses made by these threads prior to __syncthreads() are
visible to all threads in the block.
__syncthreads() is used to coordinate communication between the threads of
the same block. When some threads within a block access the same addresses in
shared or global memory, there are potential read-after-write, write-after-read, or
write-after-write hazards for some of these memory accesses. These data hazards
can be avoided by synchronizing threads in-between these accesses.
__syncthreads() is allowed in conditional code but only if the conditional
evaluates identically across the entire thread block, otherwise the code execution is
likely to hang or produce unintended side effects.
Devices of compute capability 2.0 support three variations of __syncthreads()
described below.
int __syncthreads_count(int predicate);
is identical to __syncthreads() with the additional feature that it evaluates
predicate for all threads of the block and returns the number of threads for
which predicate evaluates to non-zero.
int __syncthreads_and(int predicate);
is identical to __syncthreads() with the additional feature that it evaluates
predicate for all threads of the block and returns non-zero if and only if
predicate evaluates to non-zero for all of them.
int __syncthreads_or(int predicate);
is identical to __syncthreads() with the additional feature that it evaluates
predicate for all threads of the block and returns non-zero if and only if
predicate evaluates to non-zero for any of them.
B.7 Mathematical Functions
Section C.1 contains a comprehensive list of the C/C++ standard library
mathematical functions that are currently supported in device code, along with their
respective error bounds. When executed in host code, a given function uses the C
runtime implementation if available.
For some of the functions of Section C.1, a less accurate, but faster version exists in
the device runtime component; it has the same name prefixed with __ (such as
__sinf(x)). These intrinsic functions are listed in Section C.2, along with their
respective error bounds.
The compiler has an option (-use_fast_math) that forces each function in Table
B-2 to compile to its intrinsic counterpart. In addition to reduce accuracy of the
affected functions, it may also cause some differences in special case handling. A
more robust approach is to selectively replace mathematical function calls by calls to
intrinsic functions only where it is merited by the performance gains and where
changed properties such as reduced accuracy and different special case handling can
be tolerated.
Table B-2. Functions Affected by –use_fast_math
Appendix A. C Language Extensions


106 CUDA Programming Guide Version 3.0

Operator/Function Device Function
x/y __fdividef(x,y)
sinf(x) __sinf(x)
cosf(x) __cosf(x)
tanf(x) __tanf(x)
sincosf(x,sptr,cptr) __sincosf(x,sptr,cptr)
logf(x) __logf(x)
log2f(x) __log2f(x)
log10f(x) __log10f(x)
expf(x) __expf(x)
exp10f(x) __exp10f(x)
powf(x,y) __powf(x,y)
B.8 Texture Functions
For texture functions, a combination of the texture reference‟s immutable (i.e.
compile-time) and mutable (i.e. runtime) attributes determine how the texture
coordinates are interpreted, what processing occurs during the texture fetch, and the
return value delivered by the texture fetch. Immutable attributes are described in
Section 3.2.4.1. Mutable attributes are described in Section 3.2.4.2. Texture fetching
is described in Appendix F.
Appendix G lists the maximum texture width, height, and depth depending on the
compute capability of the device.
B.8.1 tex1Dfetch()
template<class Type>
Type tex1Dfetch(
texture<Type, 1, cudaReadModeElementType> texRef,
int x);

float tex1Dfetch(
texture<unsigned char, 1, cudaReadModeNormalizedFloat> texRef,
int x);

float tex1Dfetch(
texture<signed char, 1, cudaReadModeNormalizedFloat> texRef,
int x);

float tex1Dfetch(
texture<unsigned short, 1, cudaReadModeNormalizedFloat> texRef,
int x);

float tex1Dfetch(
texture<signed short, 1, cudaReadModeNormalizedFloat> texRef,
int x);
fetch the region of linear memory bound to texture reference texRef using integer
texture coordinate x. No texture filtering and addressing modes are supported. For
Appendix A. C Language Extensions


CUDA Programming Guide Version 3.0 107

integer types, these functions may optionally promote the integer to single-precision
floating point.
Besides the functions shown above, 2-, and 4-tuples are supported; for example:
float4 tex1Dfetch(
texture<uchar4, 1, cudaReadModeNormalizedFloat> texRef,
int x);
fetches the region of linear memory bound to texture reference texRef using
texture coordinate x.
B.8.2 tex1D()
template<class Type, enum cudaTextureReadMode readMode>
Type tex1D(texture<Type, 1, readMode> texRef,
float x);
fetches the CUDA array bound to texture reference texRef using floating-point
texture coordinates x.
B.8.3 tex2D()
template<class Type, enum cudaTextureReadMode readMode>
Type tex2D(texture<Type, 2, readMode> texRef,
float x, float y);
fetches the CUDA array or the region of linear memory bound to texture reference
texRef using texture coordinates x and y.
B.8.4 tex3D()
template<class Type, enum cudaTextureReadMode readMode>
Type tex3D(texture<Type, 3, readMode> texRef,
float x, float y, float z);
fetches the CUDA array bound to texture reference texRef using texture
coordinates x, y, and z.
B.9 Time Function
clock_t clock();
when executed in device code, returns the value of a per-multiprocessor counter
that is incremented every clock cycle. Sampling this counter at the beginning and at
the end of a kernel, taking the difference of the two samples, and recording the
result per thread provides a measure for each thread of the number of clock cycles
taken by the device to completely execute the thread, but not of the number of
clock cycles the device actually spent executing thread instructions. The former
number is greater that the latter since threads are time sliced.
Appendix A. C Language Extensions


108 CUDA Programming Guide Version 3.0

B.10 Atomic Functions
An atomic function performs a read-modify-write atomic operation on one 32-bit or
64-bit word residing in global or shared memory. For example, atomicAdd()
reads a 32-bit word at some address in global or shared memory, adds a number to
it, and writes the result back to the same address. The operation is atomic in the
sense that it is guaranteed to be performed without interference from other threads.
In other words, no other thread can access this address until the operation is
complete.
Atomic operations only work with signed and unsigned integers with the exception
of atomicAdd() for devices of compute capability 2.0 and atomicExch() for all
devices, which also work for single-precision floating-point numbers.
Atomic functions can only be used in device functions and are only available for
devices of compute capability 1.1 and above.
Atomic functions operating on shared memory and atomic functions operating on
64-bit words are only available for devices of compute capability 1.2 and above.
Note that atomic functions operating on mapped page-locked memory
(Section 3.2.5.3) are not atomic from the point of view of the host or other devices.
B.10.1 Arithmetic Functions
B.10.1.1 atomicAdd()
int atomicAdd(int* address, int val);
unsigned int atomicAdd(unsigned int* address,
unsigned int val);
unsigned long long int atomicAdd(unsigned long long int* address,
unsigned long long int val);
float atomicAdd(float* address, float val);
reads the 32-bit or 64-bit word old located at the address address in global or
shared memory, computes (old + val), and stores the result back to memory at
the same address. These three operations are performed in one atomic transaction.
The function returns old.
64-bit words are only supported for global memory.
The floating-point version of atomicAdd() is only supported by devices of
compute capability 2.0.
B.10.1.2 atomicSub()
int atomicSub(int* address, int val);
unsigned int atomicSub(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global or shared
memory, computes (old - val), and stores the result back to memory at the
same address. These three operations are performed in one atomic transaction. The
function returns old.
Appendix A. C Language Extensions


CUDA Programming Guide Version 3.0 109

B.10.1.3 atomicExch()
int atomicExch(int* address, int val);
unsigned int atomicExch(unsigned int* address,
unsigned int val);
unsigned long long int atomicExch(unsigned long long int* address,
unsigned long long int val);
float atomicExch(float* address, float val);
reads the 32-bit or 64-bit word old located at the address address in global or
shared memory and stores val back to memory at the same address. These two
operations are performed in one atomic transaction. The function returns old.
64-bit words are only supported for global memory.
B.10.1.4 atomicMin()
int atomicMin(int* address, int val);
unsigned int atomicMin(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global or shared
memory, computes the minimum of old and val, and stores the result back to
memory at the same address. These three operations are performed in one atomic
transaction. The function returns old.
B.10.1.5 atomicMax()
int atomicMax(int* address, int val);
unsigned int atomicMax(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global or shared
memory, computes the maximum of old and val, and stores the result back to
memory at the same address. These three operations are performed in one atomic
transaction. The function returns old.
B.10.1.6 atomicInc()
unsigned int atomicInc(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global or shared
memory, computes ((old >= val) ? 0 : (old+1)), and stores the result
back to memory at the same address. These three operations are performed in one
atomic transaction. The function returns old.
B.10.1.7 atomicDec()
unsigned int atomicDec(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global or shared
memory, computes (((old == 0) | (old > val)) ? val : (old-1)),
and stores the result back to memory at the same address. These three operations
are performed in one atomic transaction. The function returns old.
B.10.1.8 atomicCAS()
int atomicCAS(int* address, int compare, int val);
unsigned int atomicCAS(unsigned int* address,
unsigned int compare,
unsigned int val);
Appendix A. C Language Extensions


110 CUDA Programming Guide Version 3.0

unsigned long long int atomicCAS(unsigned long long int* address,
unsigned long long int compare,
unsigned long long int val);
reads the 32-bit or 64-bit word old located at the address address in global or
shared memory, computes (old == compare ? val : old), and stores the
result back to memory at the same address. These three operations are performed in
one atomic transaction. The function returns old (Compare And Swap).
64-bit words are only supported for global memory.
B.10.2 Bitwise Functions
B.10.2.1 atomicAnd()

int atomicAnd(int* address, int val);
unsigned int atomicAnd(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global or shared
memory, computes (old & val), and stores the result back to memory at the
same address. These three operations are performed in one atomic transaction. The
function returns old.
B.10.2.2 atomicOr()

int atomicOr(int* address, int val);
unsigned int atomicOr(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global or shared
memory, computes (old | val), and stores the result back to memory at the
same address. These three operations are performed in one atomic transaction. The
function returns old.
B.10.2.3 atomicXor()

int atomicXor(int* address, int val);
unsigned int atomicXor(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global or shared
memory, computes (old ^ val), and stores the result back to memory at the
same address. These three operations are performed in one atomic transaction. The
function returns old.
B.11 Warp Vote Functions
Warp vote functions are only supported by devices of compute capability 1.2 and
higher (see Section 4.1 for the definition of a warp).
int __all(int predicate);
Appendix A. C Language Extensions


CUDA Programming Guide Version 3.0 111

evaluates predicate for all threads of the warp and returns non-zero if and only if
predicate evaluates to non-zero for all of them.
int __any(int predicate);
evaluates predicate for all threads of the warp and returns non-zero if and only if
predicate evaluates to non-zero for any of them.
unsigned int __ballot(int predicate);
evaluates predicate for all threads of the warp and returns an integer whose N
th

bit is set if and only if predicate evaluates to non-zero for the N
th
thread of the
warp. This function is only supported by devices of compute capability 2.0.
B.12 Profiler Counter Function
Each multiprocessor has a set of sixteen hardware counters that an application can
increment with a single instruction by calling the __prof_trigger() function.
void __prof_trigger(int counter);
increments by one per warp the per-multiprocessor hardware counter of index
counter. Counters 8 to 15 are reserved and should not be used by applications.
The value of counters 0, 1, …, 7 for the first multiprocessor can be obtained via the
CUDA profiler by listing prof_trigger_00, prof_trigger_01, …,
prof_trigger_07, etc. in the profiler.conf file (see the profiler manual for
more details). All counters are reset before each kernel call (note that when an
application is run via the CUDA debugger or the CUDA profiler, all launches are
synchronous).
B.13 Execution Configuration
Any call to a __global__ function must specify the execution configuration for that
call. The execution configuration defines the dimension of the grid and blocks that
will be used to execute the function on the device, as well as the associated stream
(see Section 3.3.9.1 for a description of streams).
When using the driver API, the execution configuration is specified through a series
of driver function calls as detailed in Section 3.3.3.
When using the runtime API (Section 3.2), the execution configuration is specified
by inserting an expression of the form <<< Dg, Db, Ns, S >>> between the
function name and the parenthesized argument list, where:
 Dg is of type dim3 (see Section B.3.2) and specifies the dimension and size of
the grid, such that Dg.x * Dg.y equals the number of blocks being launched;
Dg.z must be equal to 1;
 Db is of type dim3 (see Section B.3.2) and specifies the dimension and size of
each block, such that Db.x * Db.y * Db.z equals the number of threads
per block;
 Ns is of type size_t and specifies the number of bytes in shared memory that
is dynamically allocated per block for this call in addition to the statically
allocated memory; this dynamically allocated memory is used by any of the
Appendix A. C Language Extensions


112 CUDA Programming Guide Version 3.0

variables declared as an external array as mentioned in Section B.2.3; Ns is an
optional argument which defaults to 0;
 S is of type cudaStream_t and specifies the associated stream; S is an
optional argument which defaults to 0.
As an example, a function declared as
__global__ void Func(float* parameter);
must be called like this:
Func<<< Dg, Db, Ns >>>(parameter);
The arguments to the execution configuration are evaluated before the actual
function arguments and like the function arguments, are currently passed via shared
memory to the device.
The function call will fail if Dg or Db are greater than the maximum sizes allowed
for the device as specified in Appendix G, or if Ns is greater than the maximum
amount of shared memory available on the device, minus the amount of shared
memory required for static allocation, functions arguments (for devices of compute
capability 1.x), and execution configuration.
B.14 Launch Bounds
As discussed in detail in Section 5.2.3, the fewer registers a kernel uses, the more
threads and thread blocks are likely to reside on a multiprocessor, which can
improve performance.
Therefore, the compiler uses heuristics to minimize register usage while keeping
register spilling (see Section 5.3.2.2) and instruction count to a minimum. An
application can optionally aid these heuristics by providing additional information to
the compiler in the form of launch bounds that are specified using the
__launch_bounds__() qualifier in the definition of a __global__ function:
__global__ void
__launch_bounds__(maxThreadsPerBlock, minBlocksPerMultiprocessor)
MyKernel(...)
{
...
}
 maxThreadsPerBlock specifies the maximum number of threads per block
with which the application will ever launch MyKernel(); it compiles to the
.maxntid PTX directive;
 minBlocksPerMultiprocessor is optional and specifies the desired
minimum number of resident blocks per multiprocessor; it compiles to the
.minnctapersm PTX directive.
If launch bounds are specified, the compiler first derives from them the upper limit
L on the number of registers the kernel should use to ensure that
minBlocksPerMultiprocessor blocks (or a single block if
minBlocksPerMultiprocessor is not specified) of maxThreadsPerBlock
threads can reside on the multiprocessor (see Section 4.2 for the relationship
between the number of registers used by a kernel and the number of registers
Appendix A. C Language Extensions


CUDA Programming Guide Version 3.0 113

allocated per block). The compiler then optimizes register usage in the following
way:
 If the initial register usage is higher than L, the compiler reduces it further until
it becomes less or equal to L, usually at the expense of more local memory
usage and/or higher number of instructions;
 If the initial register usage is lower than L,
 If maxThreadsPerBlock is specified and
minBlocksPerMultiprocessor is not, the compiler uses
maxThreadsPerBlock to determine the register usage thresholds for the
transitions between n and n+1 resident blocks (i.e. when using one less
register makes room for an additional resident block as in the example of
Section 5.2.3) and then applies similar heuristics as when no launch bounds
are specified;
 If both minBlocksPerMultiprocessor and maxThreadsPerBlock
are specified, the compiler may increase register usage as high as L to
reduce the number of instructions and better hide single thread instruction
latency.
A kernel will fail to launch if it is executed with more threads per block than its
launch bound maxThreadsPerBlock.
Optimal launch bounds for a given kernel will usually differ across major
architecture revisions. The sample code below shows how this is typically handled in
device code using the __CUDA_ARCH__ macro introduced in Section 3.1.4.
#define THREADS_PER_BLOCK 256
#if __CUDA_ARCH__ >= 200
#define MY_KERNEL_MAX_THREADS (2 * THREADS_PER_BLOCK)
#define MY_KERNEL_MIN_BLOCKS 3
#else
#define MY_KERNEL_MAX_THREADS THREADS_PER_BLOCK
#define MY_KERNEL_MIN_BLOCKS 2
#endif

// Device code
__global__ void
__launch_bounds__(MY_KERNEL_MAX_THREADS, MY_KERNEL_MIN_BLOCKS)
MyKernel(...)
{
...
}
In the common case where MyKernel is invoked with the maximum number of
threads per block (specified as the first parameter of __launch_bounds__()), it
is tempting to use MY_KERNEL_MAX_THREADS as the number of threads per block
in the execution configuration:
// Host code
MyKernel<<<blocksPerGrid, MY_KERNEL_MAX_THREADS>>>(...);
This will not work however since __CUDA_ARCH__ is undefined in host code as
mentioned in Section 3.1.4, so MyKernel will launch with 256 threads per block
even when __CUDA_ARCH__ is greater or equal to 200. Instead the number of
threads per block should be determined:
Appendix A. C Language Extensions


114 CUDA Programming Guide Version 3.0

 Either at compile time using a macro that does not depend on
__CUDA_ARCH__, for example
// Host code
MyKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(...);
 Or at runtime based on the compute capability
// Host code
cudaGetDeviceProperties(&deviceProp, device);
int threadsPerBlock =
(deviceProp.major >= 2 ?
2 * THREADS_PER_BLOCK : THREADS_PER_BLOCK);
MyKernel<<<blocksPerGrid, threadsPerBlock>>>(...);
Register usage is reported by the --ptxas-options=-v compiler option. The
number of resident blocks can be derived from the occupancy reported by the
CUDA profiler (see Section 5.2.3 for a definition of occupancy).
Register usage can also be controlled for all __global__ functions in a file using
the -maxrregcount compiler option. The value of -maxrregcount is ignored
for functions with launch bounds.








CUDA Programming Guide Version 3.0 115

Appendix C.
Mathematical Functions
Functions from Section C.1 can be used in both host and device code whereas
functions from Section C.2 can only be used in device code.
Note that floating-point functions are overloaded, so that in general, there are three
prototypes for a given function <func-name>:
(1) double <func-name>(double), e.g. double log(double)
(2) float <func-name>(float), e.g. float log(float)
(3) float <func-name>f(float), e.g. float logf(float)
This means, in particular, that passing a float argument always results in a float
result (variants (2) and (3) above).
C.1 Standard Functions
This section lists all the mathematical standard library functions supported in device
code. It also specifies the error bounds of each function when executed on the
device. These error bounds also apply when the function is executed on the host in
the case where the host does not supply the function. They are generated from
extensive but not exhaustive tests, so they are not guaranteed bounds.
C.1.1 Single-Precision Floating-Point Functions
Addition and multiplication are IEEE-compliant, so have a maximum error of
0.5 ulp. However, on the device, the compiler often combines them into a single
multiply-add instruction (FMAD) and for devices of compute capability 1.x, FMAD
truncates the intermediate result of the multiplication as mentioned in Section G.2.
This combination can be avoided by using the __fadd_rn() and __fmul_rn()
intrinsic functions (see Section C.2).
The recommended way to round a single-precision floating-point operand to an
integer, with the result being a single-precision floating-point number is rintf(),
not roundf(). The reason is that roundf() maps to an 8-instruction sequence on
the device, whereas rintf() maps to a single instruction. truncf(), ceilf(),
and floorf() each map to a single instruction as well.
Appendix B. Mathematical Functions


116 CUDA Programming Guide Version 3.0

Table C-1. Mathematical Standard Library Functions with
Maximum ULP Error
The maximum error is stated as the absolute value of the difference
in ulps between a correctly rounded single-precision result and the
result returned by the CUDA library function.
Function Maximum ulp error
x+y 0 (IEEE-754 round-to-nearest-even)
(except for devices of compute capability 1.x when addition is
merged into an FMAD)
x*y 0 (IEEE-754 round-to-nearest-even)
(except for devices of compute capability 1.x when
multiplication is merged into an FMAD)
x/y 0 for compute capability ≥ 2 when compiled with
-prec-div=true
2 (full range), otherwise
1/x 0 for compute capability ≥ 2 when compiled with
-prec-div=true
1 (full range), otherwise
rsqrtf(x)
1/sqrtf(x)
2 (full range)
Applies to 1/sqrtf(x) only when it is converted to
rsqrtf(x) by the compiler.
sqrtf(x) 0 for compute capability ≥ 2 when compiled with
-prec-sqrt=true
3 (full range), otherwise
cbrtf(x) 1 (full range)
rcbrtf(x) 2 (full range)
hypotf(x,y) 3 (full range)
expf(x) 2 (full range)
exp2f(x) 2 (full range)
exp10f(x) 2 (full range)
expm1f(x) 1 (full range)
logf(x) 1 (full range)
log2f(x) 3 (full range)
log10f(x) 3 (full range)
log1pf(x) 2 (full range)
sinf(x) 2 (full range)
cosf(x) 2 (full range)
tanf(x) 4 (full range)
sincosf(x,sptr,cptr) 2 (full range)
sinpif(x) 2 (full range)
asinf(x) 4 (full range)
acosf(x) 3 (full range)
atanf(x) 2 (full range)
atan2f(y,x) 3 (full range)
sinhf(x) 3 (full range)
coshf(x) 2 (full range)
Appendix B. Mathematical Functions


CUDA Programming Guide Version 3.0 117

Function Maximum ulp error
tanhf(x) 2 (full range)
asinhf(x) 3 (full range)
acoshf(x) 4 (full range)
atanhf(x) 3 (full range)
powf(x,y) 8 (full range)
erff(x) 3 (full range)
erfcf(x) 6 (full range)
erfinvf(x) 5 (full range)
erfcinvf(x) 7 (full range)
lgammaf(x) 6 (outside interval -10.001 ... -2.264; larger inside)
tgammaf(x) 11 (full range)
fmaf(x,y,z) 0 (full range)
frexpf(x,exp) 0 (full range)
ldexpf(x,exp) 0 (full range)
scalbnf(x,n) 0 (full range)
scalblnf(x,l) 0 (full range)
logbf(x) 0 (full range)
ilogbf(x) 0 (full range)
fmodf(x,y) 0 (full range)
remainderf(x,y) 0 (full range)
remquof(x,y,iptr) 0 (full range)
modff(x,iptr) 0 (full range)
fdimf(x,y) 0 (full range)
truncf(x) 0 (full range)
roundf(x) 0 (full range)
rintf(x) 0 (full range)
nearbyintf(x) 0 (full range)
ceilf(x) 0 (full range)
floorf(x) 0 (full range)
lrintf(x) 0 (full range)
lroundf(x) 0 (full range)
llrintf(x) 0 (full range)
llroundf(x) 0 (full range)
signbit(x) N/A
isinf(x) N/A
isnan(x) N/A
isfinite(x) N/A
copysignf(x,y) N/A
fminf(x,y) N/A
fmaxf(x,y) N/A
fabsf(x) N/A
nanf(cptr) N/A
Appendix B. Mathematical Functions


118 CUDA Programming Guide Version 3.0

Function Maximum ulp error
nextafterf(x,y) N/A
C.1.2 Double-Precision Floating-Point Functions
The errors listed below only apply when compiling for devices with native double-
precision support. When compiling for devices without such support, such as
devices of compute capability 1.2 and lower, the double type gets demoted to
float by default and the double-precision math functions are mapped to their
single-precision equivalents.
The recommended way to round a double-precision floating-point operand to an
integer, with the result being a double-precision floating-point number is rint(),
not round(). The reason is that round() maps to an 8-instruction sequence on
the device, whereas rint() maps to a single instruction. trunc(), ceil(), and
floor() each map to a single instruction as well.
Table C-2. Mathematical Standard Library Functions with
Maximum ULP Error
The maximum error is stated as the absolute value of the difference
in ulps between a correctly rounded double-precision result and the
result returned by the CUDA library function.
Function Maximum ulp error
x+y 0 (IEEE-754 round-to-nearest-even)
x*y 0 (IEEE-754 round-to-nearest-even)
x/y 0 (IEEE-754 round-to-nearest-even)
1/x 0 (IEEE-754 round-to-nearest-even)
sqrt(x) 0 (IEEE-754 round-to-nearest-even)
rsqrt(x) 1 (full range)
cbrt(x) 1 (full range)
rcbrt(x) 1 (full range)
hypot(x,y) 2 (full range)
exp(x) 1 (full range)
exp2(x) 1 (full range)
exp10(x) 1 (full range)
expm1(x) 1 (full range)
log(x) 1 (full range)
log2(x) 1 (full range)
log10(x) 1 (full range)
log1p(x) 1 (full range)
sin(x) 2 (full range)
cos(x) 2 (full range)
tan(x) 2 (full range)
sincos(x,sptr,cptr) 2 (full range)
sinpi(x) 2 (full range)
Appendix B. Mathematical Functions


CUDA Programming Guide Version 3.0 119

Function Maximum ulp error
asin(x) 2 (full range)
acos(x) 2 (full range)
atan(x) 2 (full range)
atan2(y,x) 2 (full range)
sinh(x) 1 (full range)
cosh(x) 1 (full range)
tanh(x) 1 (full range)
asinh(x) 2 (full range)
acosh(x) 2 (full range)
atanh(x) 2 (full range)
pow(x,y) 2 (full range)
erf(x) 2 (full range)
erfc(x) 5 (full range)
erfinv(x) 8 (full range)
erfcinv(x) 8 (full range)
lgamma(x) 4 (outside interval -11.0001 ... -2.2637; larger inside)
tgamma(x) 8 (full range)
fma(x,y,z) 0 (IEEE-754 round-to-nearest-even)
frexp(x,exp) 0 (full range)
ldexp(x,exp) 0 (full range)
scalbn(x,n) 0 (full range)
scalbln(x,l) 0 (full range)
logb(x) 0 (full range)
ilogb(x) 0 (full range)
fmod(x,y) 0 (full range)
remainder(x,y) 0 (full range)
remquo(x,y,iptr) 0 (full range)
modf(x,iptr) 0 (full range)
fdim(x,y) 0 (full range)
trunc(x) 0 (full range)
round(x) 0 (full range)
rint(x) 0 (full range)
nearbyint(x) 0 (full range)
ceil(x) 0 (full range)
floor(x) 0 (full range)
lrint(x) 0 (full range)
lround(x) 0 (full range)
llrint(x) 0 (full range)
llround(x) 0 (full range)
signbit(x) N/A
isinf(x) N/A
isnan(x) N/A
Appendix B. Mathematical Functions


120 CUDA Programming Guide Version 3.0

Function Maximum ulp error
isfinite(x) N/A
copysign(x,y) N/A
fmin(x,y) N/A
fmax(x,y) N/A
fabs(x) N/A
nan(cptr) N/A
nextafter(x,y) N/A
C.1.3 Integer Functions
Integer min(x,y) and max(x,y) are supported and map to a single instruction on
the device.
C.2 Intrinsic Functions
This section lists the intrinsic functions that are only supported in device code.
Among these functions are the less accurate, but faster versions of some of the
functions of Section C.1; they have the same name prefixed with __ (such as
__sinf(x)).
Functions suffixed with _rn operate using the round-to-nearest-even rounding
mode.
Functions suffixed with _rz operate using the round-towards-zero rounding mode.
Functions suffixed with _ru operate using the round-up (to positive infinity)
rounding mode.
Functions suffixed with _rd operate using the round-down (to negative infinity)
rounding mode.
Unlike type conversion functions (such as __int2float_rn) that convert from
one type to another, type casting functions simply perform a type cast on the
argument, leaving the value unchanged. For example,
__int_as_float(0xC0000000) is equal to -2, __float_as_int(1.0f) is
equal to 0x3f800000.
C.2.1 Single-Precision Floating-Point Functions
__fadd_rn() and __fmul_rn() map to addition and multiplication operations
that the compiler never merges into FMADs. By contrast, additions and
multiplications generated from the '*' and '+' operators will frequently be combined
into FMADs.
Both the regular floating-point division and __fdividef(x,y) have the same
accuracy, but for 2
126
< y < 2
128
, __fdividef(x,y) delivers a result of zero,
whereas the regular division delivers the correct result to within the accuracy stated
in Table C-3. Also, for 2
126
< y < 2
128
, if x is infinity, __fdividef(x,y) delivers
Appendix B. Mathematical Functions


CUDA Programming Guide Version 3.0 121

a NaN (as a result of multiplying infinity by zero), while the regular division returns
infinity.
__saturate(x) returns 0 if x is less than 0, 1 if x is more than 1, and x
otherwise.
__float2ll_[rn,rz,ru,rd](x) (respectively
__float2ull_[rn,rz,ru,rd](x)) converts single-precision floating-point
parameter x to 64-bit signed (respectively unsigned) integer with specified IEEE-
754 rounding modes.
Table C-3. Single-Precision Floating-Point Intrinsic Functions
Supported by the CUDA Runtime Library with
Respective Error Bounds
Function Error bounds
__fadd_[rn,rz,ru,rd](x,y) IEEE-compliant.
__fmul_[rn,rz,ru,rd](x,y) IEEE-compliant.
__fmaf_[rn,rz,ru,rd](x,y,z) IEEE-compliant.
__frcp_[rn,rz,ru,rd](x) IEEE-compliant.
__fsqrt_[rn,rz,ru,rd](x) IEEE-compliant.
__fdiv_[rn,rz,ru,rd](x,y) IEEE-compliant.
__fdividef(x,y) For y in [2
-126
, 2
126
], the maximum ulp error is
2.
__expf(x) The maximum ulp error is
2 + floor(abs(1.16 * x)).
__exp10f(x) The maximum ulp error is
2 + floor(abs(2.95 * x)).
__logf(x) For x in [0.5, 2], the maximum absolute error
is 2
-21.41
, otherwise, the maximum ulp error is
3.
__log2f(x) For x in [0.5, 2], the maximum absolute error
is 2
-22
, otherwise, the maximum ulp error is 2.
__log10f(x) For x in [0.5, 2], the maximum absolute error
is 2
-24
, otherwise, the maximum ulp error is 3.
__sinf(x) For x in [-t, t], the maximum absolute error
is 2
-21.41
, and larger otherwise.
__cosf(x) For x in [-t, t], the maximum absolute error
is 2
-21.19
, and larger otherwise.
__sincosf(x,sptr,cptr) Same as sinf(x) and cosf(x).
__tanf(x) Derived from its implementation as
__sinf(x) * (1 / __cosf(x)).
__powf(x, y) Derived from its implementation as
exp2f(y * __log2f(x)).
__int_as_float(x) N/A
__float_as_int(x) N/A
__saturate(x) N/A
__float2int_[rn,rz,ru,rd](x) N/A
__float2uint_[rn,rz,ru,rd](x) N/A
Appendix B. Mathematical Functions


122 CUDA Programming Guide Version 3.0

__int2float_[rn,rz,ru,rd](x) N/A
__uint2float_[rn,rz,ru,rd](x) N/A
__float2ll_[rn,rz,ru,rd](x) N/A
__float2ull_[rn,rz,ru,rd](x) N/A
__ll2float_[rn,rz,ru,rd](x) N/A
__ull2float_[rn,rz,ru,rd](x) N/A
C.2.2 Double-Precision Floating-Point Functions
__dadd_rn() and __dmul_rn() map to addition and multiplication operations
that the compiler never merges into FMADs. By contrast, additions and
multiplications generated from the '*' and '+' operators will frequently be combined
into FMADs.
Table C-4. Double-Precision Floating-Point Intrinsic
Functions Supported by the CUDA Runtime
Library with Respective Error Bounds
Function Error bounds
__dadd_[rn,rz,ru,rd](x,y) IEEE-compliant.
__dmul_[rn,rz,ru,rd](x,y) IEEE-compliant.
__fma_[rn,rz,ru,rd](x,y,z) IEEE-compliant.
__double2float_[rn,rz](x) N/A
__double2int_[rn,rz,ru,rd](x) N/A
__double2uint_[rn,rz,ru,rd](x) N/A
__double2ll_[rn,rz,ru,rd](x) N/A
__double2ull_[rn,rz,ru,rd](x) N/A
__int2double_rn(x) N/A
__uint2double_rn(x) N/A
__ll2double_[rn,rz,ru,rd](x) N/A
__ull2double_[rn,rz,ru,rd](x) N/A
__double_as_longlong(x) N/A
__longlong_as_double(x) N/A
__double2hiint(x) N/A
__double2loint(x) N/A
__hiloint2double(x, ys) N/A
__ddiv_[rn,rz,ru,rd](x,y)(x,y) IEEE-compliant.
Requires compute capability ≥ 2.
__drcp_[rn,rz,ru,rd](x) IEEE-compliant.
Requires compute capability ≥ 2
__dsqrt_[rn,rz,ru,rd](x) IEEE-compliant.
Requires compute capability ≥ 2
__double2float_[ru,rd](x) N/A
Appendix B. Mathematical Functions


CUDA Programming Guide Version 3.0 123

C.2.3 Integer Functions
__[u]mul24(x,y) computes the product of the 24 least significant bits of the
integer parameters x and y and delivers the 32 least significant bits of the result. The
8 most significant bits of x or y are ignored.
__[u]mulhi(x,y) computes the product of the integer parameters x and y and
delivers the 32 most significant bits of the 64-bit result.
__[u]mul64hi(x,y) computes the product of the 64-bit integer parameters x
and y and delivers the 64 most significant bits of the 128-bit result.
__[u]sad(x,y,z) (Sum of Absolute Difference) returns the sum of integer
parameter z and the absolute value of the difference between integer parameters x
and y.
__clz(x) returns the number, between 0 and 32 inclusive, of consecutive zero bits
starting at the most significant bit (i.e. bit 31) of integer parameter x.
__clzll(x) returns the number, between 0 and 64 inclusive, of consecutive zero
bits starting at the most significant bit (i.e. bit 63) of 64-bit integer parameter x.
__ffs(x) returns the position of the first (least significant) bit set in integer
parameter x. The least significant bit is position 1. If x is 0, __ffs() returns 0.
Note that this is identical to the Linux function ffs.
__ffsll(x) returns the position of the first (least significant) bit set in 64-bit
integer parameter x. The least significant bit is position 1. If x is 0, __ffsll()
returns 0. Note that this is identical to the Linux function ffsll.
__popc(x) returns the number of bits that are set to 1 in the binary representation
of 32-bit integer parameter x.
__popcll(x) returns the number of bits that are set to 1 in the binary
representation of 64-bit integer parameter x.
__brev(x) reverses the bits of 32-bit unsigned integer parameter x, i.e. bit N of
the result corresponds to bit 31-N of x.
__brevll(x) reverses the bits of 64-bit unsigned long long parameter x, i.e. bit N
of the result corresponds to bit 63-N of x.







CUDA Programming Guide Version 3.0 125

Appendix D.
C++ Language Constructs
CUDA supports the following C++ language constructs for device code:
 Polymorphism
 Default Parameters
 Operator Overloading
 Namespaces
 Function Templates
 Classes for devices of compute capability 2.0
These C++ constructs are implemented as specified in “The C++ Programming
Langue” reference. It is valid to use any of these constructs in .cu CUDA files for
host, device, and kernel (__global__) functions. Any restrictions detailed in previous
parts of this programming guide, like the lack of support for recursion, still apply.
The following subsections provide examples of the various constructs.
D.1 Polymorphism
Generally, polymorphism is the ability to define that functions or operators behave
differently in different contexts. This is also referred to as function (and operator,
see below) overloading.
In practical terms, this means that it is permissible to define two different functions
within the same scope (namespace) as long as they have a distinguishable function
signature. That means that the two functions either consume a different number of
parameters or parameters of different types. When either of the multiple functions
gets invoked the compiler resolves to the function‟s implementation that matches
the function signature.
Because of implicit typecasting, a compiler may encounter multiple potential
matches for a function invocation and in that case the matching rules as described in
the C++ Language Standard apply. In practice this means that the compiler will pick
the closest match in case of multiple potential matches.
Example: The following is valid CUDA code:
__device__ void f(float x)
{
Appendix C. C++ Language Constructs


126 CUDA Programming Guide Version 3.0

// do something with x
}

__device__ void f(int i)
{
// do something with i
}

__device__ void f(double x, double y)
{
// do something with x and y
}
D.2 Default Parameters
With support for polymorphism as described in the previous subsection and the
function signature matching rules in place it becomes possible to provide support
for default values for function parameters.
Example:
__device__ void f(float x = 0.0f)
{
// do something with x
}
Kernel or other device functions can now invoke this version of f in one of two
ways:
f();
// or
float x = /* some value */;
f(x);

Default parameters can only be given for the last n parameters of a function.
D.3 Operator Overloading
Operator overloading allows programmers to define operators for new data-types.
Examples of overloadable operators in C++ are: +, -, *, /, +=, &, [], etc.
Example: The following is valid CUDA code, implementing the + operation
between two uchar4 vectors:
__device__ uchar4 operator+ (const uchar4 & a, const uchar4 & b)
{
uchar4 r;
r.x = a.x + b.x;
...
return r;
}
This new operator can now be used like this:
uchar4 a, b, c;
a = b = /* some initial value */;
Appendix C. C++ Language Constructs


CUDA Programming Guide Version 3.0 127

c = a + b;
D.4 Namespaces
Namespaces in C++ allow for the creation of a hierarchy of scopes of visibility. All
the symbols inside a namespace can be used within this namespaces without
additional syntax.
The use of namespaces can be used to solve the problem of name-clashes (two
different symbols using identical names), which commonly occurs when using
multiple function libraries from different sources.
Example: The following code defines two functions “f()” in two separate
namespaces (“nvidia” and “other”):
namespace nvidia {
__device__ void f(float x)
{ /* do something with x */ ;}
}

namespace other {
__device__ void f(float x)
{ /* do something with x */ ;}
}

The functions can now be used anywhere via fully qualified names:
nvidia::f(0.5f);
All the symbols in a namespace can be imported into another namespace (scope)
like this:
using namespace nvidia;
f(0.5f);
D.5 Function Templates
Function templates are a form of meta-programming that allows writing a generic
function in a data-type independent fashion. CUDA supports function templates to
the full extent of the C++ standard, including the following concepts:
 Implicit template parameter deduction.
 Explicit instantiation.
 Template specialization.
Example:
template <T>
__device__ bool f(T x)
{ return /* some clever code that turns x into a bool here */ }
This function will convert x of any data-type to a bool as long as the code in the
function‟s body can be compiled for the actually type (T) of the variable x.
f() can be invoked in two ways:
int x = 1;
Appendix C. C++ Language Constructs


128 CUDA Programming Guide Version 3.0

bool result = f(x);
This first type of invocation relies on the compiler‟s ability to implicitly deduce the
correct function type for T. In this case the compiler would deduce T to be int and
instantiate f<int>(x).
The second type of invoking the template function is via explicit instantiation like
this:
bool result = f<double>(0.5);

Function templates may be specialized:
template <T>
__device__ bool f(T x)
{ return false; }

template <>
__device__ bool
f<int>(T x)
{ return true; }

In this case the implementation for T representing the int type are specialized to
return true, all other types will be caught by the more general template and return
false.
The complete set of matching rules (for implicitly deducing template parameters)
and matching polymorphous functions apply as specified in the C++ standard.
D.6 Classes
Code compiled for devices with compute capability 2.0 and greater may make use of
C++ classes, as long as none of the member functions are virtual (this restriction
will be removed in some future release).
There are two common use cases for classes without virtual member functions:
 Small-data aggregations. E.g. data types like pixels (r, g, b, a), 2D and 3D points,
vectors, etc.
 Functor classes. The use of functors is necessitated by the fact that device-
function pointers are not supported and thus it is not possible to pass functions
as template parameters. A workaround for this restriction is the use of functor
classes (see code sample below).
D.6.1 Example 1 Pixel Data Type
The following is an example of a data type for RGBA pixels with 8 bit per channel
depth:
class PixelRGBA
{
public:
__device__
PixelRGBA(): r_(0), g_(0), b_(0), a_(0)
{ ; }
Appendix C. C++ Language Constructs


CUDA Programming Guide Version 3.0 129


__device__
PixelRGBA(unsigned char r, unsigned char g, unsigned char b,
unsigned char a = 255): r_(r), g_(g), b_(b), a_(a)
{ ; }

// other methods and operators left out for sake of brevity

private:
unsigned char r_, g_, b_, a_;

friend PixelRGBA operator+(const PixelRGBA &,
const PixelRGBA &);
};

__device__
PixelRGBA operator+(const PixelRGBA & p1, const PixelRGBA & p2)
{
return PixelRGBA(p1.r_ + p2.r_,
p1.g_ + p2.g_,
p1.b_ + p2.b_,
p1.a_ + p2.a_);
}

Other device code can now make use of this new data type as one would expect:
PixelRGBA p1, p2;

// [...] initialization of p1 and p2 here

PixelRGBA p3 = p1 + p2;
D.6.2 Example 2 Functor Class
The following example shows how functors may be used as function template
parameters to implement a set of vector arithmetic operations.
Here are two functors for float addition and subtraction:
class Add
{
public:
__device__
float
operator() (float a, float b)
const
{
return a + b;
}
};

class Sub
{
public:
__device__
float
Appendix C. C++ Language Constructs


130 CUDA Programming Guide Version 3.0

operator() (float a, float b)
const
{
return a - b;
}
};
The following templatized kernel makes use of the functors like the ones above in
order to implement operations on vectors of floats:
// Device code
template<class O>
__global__
void
VectorOperation(const float * A, const float * B,
float * C, unsigned int N, O op)
{
unsigned int iElement = blockDim.x * blockIdx.x + threadIdx.x;
if (iElement < N)
{
C[iElement] = op(A[iElement], B[iElement]);
}
}
The VectorOperation kernel may now be launched like this in order to get a
vector addition:
// Host code
VectorOperation<<<blocks, threads>>>(v1, v2, v3, N, Add());







CUDA Programming Guide Version 3.0 131

Appendix E.
NVCC Specifics
E.1 __noinline__
By default, a __device__ function is always inlined. The __noinline__
function qualifier however can be used as a hint for the compiler not to inline the
function if possible. The function body must still be in the same file where it is
called.
For devices of compute capability 1.x, the compiler will not honor the
__noinline__ qualifier for functions with pointer parameters and for functions
with large parameter lists.
E.2 #pragma unroll
By default, the compiler unrolls small loops with a known trip count. The #pragma
unroll directive however can be used to control unrolling of any given loop. It
must be placed immediately before the loop and only applies to that loop. It is
optionally followed by a number that specifies how many times the loop must be
unrolled.
For example, in this code sample:
#pragma unroll 5
for (int i = 0; i < n; ++i)
the loop will be unrolled 5 times. The compiler will also insert code to ensure
correctness (in the example above, to ensure that there will only be n iterations if n
is less than 5, for example). It is up to the programmer to make sure that the
specified unroll number gives the best performance.
#pragma unroll 1 will prevent the compiler from ever unrolling a loop.
If no number is specified after #pragma unroll, the loop is completely unrolled
if its trip count is constant, otherwise it is not unrolled at all.
Appendix D. NVCC Specifics


132 CUDA Programming Guide Version 3.0

E.3 __restrict__
nvcc supports restricted pointers via the __restrict__ keyword.
Restricted pointers were introduced in C99 to alleviate the aliasing problem that
exists in C-type languages, and which inhibits all kind of optimization from code re-
ordering to common sub-expression elimination.
Here is an example subject to the aliasing issue, where use of restricted pointer can
help the compiler to reduce the number of instructions:
void foo(const float* a,
const float* b,
float* c)
{
c[0] = a[0] * b[0];
c[1] = a[0] * b[0];
c[2] = a[0] * b[0] * a[1];
c[3] = a[0] * a[1];
c[4] = a[0] * b[0];
c[5] = b[0];
...
}
In C-type languages, the pointers a, b, and c may be aliased, so any write through c
could modify elements of a or b. This means that to guarantee functional
correctness, the compiler cannot load a[0] and b[0] into registers, multiply them,
and store the result to both c[0] and c[1], because the results would differ from
the abstract execution model if, say, a[0] is really the same location as c[0]. So
the compiler cannot take advantage of the common sub-expression. Likewise,
the compiler cannot just reorder the computation of c[4] into the proximity of the
computation of c[0] and c[1] because the preceding write to c[3] could change
the inputs to the computation of c[4].
By making a, b, and c restricted pointers, the programmer asserts to the compiler
that the pointers are in fact not aliased, which in this case means writes through c
would never overwrite elements of a or b. This changes the function prototype as
follows:
void foo(const float* __restrict__ a,
const float* __restrict__ b,
float* __restrict__ c);
Note that all pointer arguments need to be made restricted for the compiler
optimizer to derive any benefit. With the __restrict keywords added, the
compiler can now reorder and do common sub-expression elimination at will, while
retaining functionality identical with the abstract execution model:
void foo(const float* __restrict__ a,
const float* __restrict__ b,
float* __restrict__ c)
{
float t0 = a[0];
float t1 = b[0];
float t2 = t0 * t2;
float t3 = a[1];
c[0] = t2;
Appendix D. NVCC Specifics


CUDA Programming Guide Version 3.0 133

c[1] = t2;
c[4] = t2;
c[2] = t2 * t3;
c[3] = t0 * t3;
c[5] = t1;
...
}
The effects here are a reduced number of memory accesses and reduced number of
computations. This is balanced by an increase in register pressure due to "cached"
loads and common sub-expressions.
Since register pressure is a critical issue in many CUDA codes, use of restricted
pointers can have negative performance impact on CUDA code, due to reduced
occupancy.








CUDA Programming Guide Version 3.0 135

Appendix F.
Texture Fetching
This appendix gives the formula used to compute the value returned by the texture
functions of Section B.8 depending on the various attributes of the texture reference
(see Section 3.2.4).
The texture bound to the texture reference is represented as an array T of N texels
for a one-dimensional texture, M N × texels for a two-dimensional texture, or
L M N × × texels for a three-dimensional texture. It is fetched using texture
coordinates x , y , and z .
A texture coordinate must fall within T ‟s valid addressing range before it can be
used to address T . The addressing mode specifies how an out-of-range texture
coordinate x is remapped to the valid range. If x is non-normalized, only the
clamp addressing mode is supported and x is replaced by 0 if 0 < x and 1 ÷ N if
x N s . If x is normalized:
 In clamp addressing mode, x is replaced by 0 if 0 < x and
N
1
1÷ if x s 1 ,
 In wrap addressing mode, x is replaced by ) (x frac , where
) ( ) ( x floor x x frac ÷ = and ) (x floor is the largest integer not greater than x .
In the remaining of the appendix, x , , and z are the non-normalized texture
coordinates remapped to ‟s valid addressing range. , , and z are derived
from the normalized texture coordinates xˆ , yˆ
, and zˆ as such: x N x ˆ = , y M y ˆ = , and
z L z ˆ = .
y
T x y
Appendix F. CUDA-Enabled GPUs


136 CUDA Programming Guide Version 3.0

F.1 Nearest-Point Sampling
In this filtering mode, the value returned by the texture fetch is
 ] [ ) ( i T x tex = for a one-dimensional texture,
 ] , [ ) , ( j i T y x tex = for a two-dimensional texture,
 ] , , [ ) , , ( k j i T z y x tex = for a three-dimensional texture,
where ) (x floor i = , ) (y floor j = , and ) (z floor k = .
Figure D-1 illustrates nearest-point sampling for a one-dimensional texture with
4 = N .
For integer textures, the value returned by the texture fetch can be optionally
remapped to [0.0, 1.0] (see Section 3.2.4.1).


Figure F-1. Nearest-Point Sampling of a One-Dimensional
Texture of Four Texels
F.2 Linear Filtering
In this filtering mode, which is only available for floating-point textures, the value
returned by the texture fetch is
 ] 1 [ ] [ ) 1 ( ) ( + + ÷ = i T i T x tex o o for a one-dimensional texture,
0 4 1 2 3
T[0]
T[1]
T[2]
T[3]
x
0 1 0.25 0.5 0.75
Non-Normalized
Normalized
tex(x)
Appendix D. C++ Language Constructs


CUDA Programming Guide Version 3.0 137

 ] 1 , 1 [ ] 1 , [ ) 1 ( ] , 1 [ ) 1 ( ] , [ ) 1 )( 1 ( ) , ( + + + + ÷ + + ÷ + ÷ ÷ = j i T j i T j i T j i T y x tex o| | o | o | o
for a two-dimensional texture,
 = ) , , ( z y x tex

] 1 , 1 , 1 [ ] 1 , 1 , [ ) 1 (
] 1 , , 1 [ ) 1 ( ] 1 , , [ ) 1 )( 1 (
] , 1 , 1 [ ) 1 ( ] , 1 , [ ) 1 ( ) 1 (
] , , 1 [ ) 1 )( 1 ( ] , , [ ) 1 )( 1 )( 1 (
+ + + + + + ÷
+ + + ÷ + + ÷ ÷
+ + + ÷ + + ÷ ÷
+ + ÷ ÷ + ÷ ÷ ÷
k j i T k j i T
k j i T k j i T
k j i T k j i T
k j i T k j i T
o|¸ |¸ o
¸ | o ¸ | o
¸ o| ¸ | o
¸ | o ¸ | o

for a three-dimensional texture,
where:
 ) (
B
x floor i = , ) (
B
x frac = o , 5 . 0 ÷ = x x
B
,
 ) (
B
y floor j = , ) (
B
y frac = | , 5 . 0 ÷ = y y
B
,
 ) (
B
z floor k = , ) (
B
z frac = ¸ , 5 . 0 ÷ = z z
B
.
o , | , and ¸ are stored in 9-bit fixed point format with 8 bits of fractional value
(so 1.0 is exactly represented).
Figure F-2 illustrates nearest-point sampling for a one-dimensional texture with
4 = N .

Figure F-2. Linear Filtering of a One-Dimensional Texture of
Four Texels in Clamp Addressing Mode
0 4 1 2 3
T[0]
T[1]
T[2]
T[3]
tex(x)
x
0 1 0.25 0.5 0.75
Non-Normalized
Normalized
Appendix F. CUDA-Enabled GPUs


138 CUDA Programming Guide Version 3.0

F.3 Table Lookup
A table lookup ) (x TL where x spans the interval ] , 0 [ R can be implemented as
) 5 . 0
1
( ) ( +
÷
= x
R
N
tex x TL in order to ensure that ] 0 [ ) 0 ( T TL = and ] 1 [ ) ( ÷ = N T R TL .
Figure F-3 illustrates the use of texture filtering to implement a table lookup with
4 = R or 1 = R from a one-dimensional texture with .


Figure F-3. One-Dimensional Table Lookup Using Linear
Filtering

4 = N
0 4 4/3 8/3
T[0]
T[1]
T[2]
T[3]
TL(x)
x
0 1 1/3 2/3






CUDA Programming Guide Version 3.0 139

Appendix G.
Compute Capabilities
The general specifications and features of a compute device depend on its compute
capability (see Section 2.5).
Section G.1 gives the features and technical specifications associated to each
compute capability.
Section G.2 reviews the compliance with the IEEE floating-point standard.
Section G.3 and 0 give more details on the architecture of devices of compute
capability 1.x and 2.0, respectively.
G.1 Features and Technical Specifications
Compute Capability
Feature Support
(Unlisted features are supported
for all compute capabilities)
1.0 1.1 1.2 1.3 2.0
Integer atomic functions operating on
32-bit words in global memory
(Section B.10)
No yes
Integer atomic functions operating on
64-bit words in global memory
(Section B.10)
No Yes Integer atomic functions operating on
32-bit words in shared memory
(Section B.10)
Warp vote functions (Section B.11)
Double-precision floating-point
numbers
No Yes
Floating-point atomic addition
operating on 32-bit words in global
and shared memory (Section B.10)
No Yes
__ballot() (Section B.11)
__threadfence_system() (Section B.5)
__syncthreads_count(),
__syncthreads_and(),
__syncthreads_or() (Section B.6)
Appendix G. Compute Capabilities


140 CUDA Programming Guide Version 3.0



Compute Capability
Technical Specifications 1.0 1.1 1.2 1.3 2.0
Maximum x- or y-dimension of a grid
of thread blocks
65535
Maximum number of threads per
block
512 1024
Maximum x- or y-dimension of a
block
512 1024
Maximum z-dimension of a block 64
Warp size 32
Maximum number of resident blocks
per multiprocessor
8
Maximum number of resident warps
per multiprocessor
24 32 48
Maximum number of resident threads
per multiprocessor
768 1024 1536
Number of 32-bit registers per
multiprocessor
8 K 16 K 32 K
Maximum amount of shared memory
per multiprocessor
16 KB 48 KB
Number of shared memory banks 16 32
Amount of local memory per thread 16 KB 512 KB
Constant memory size 64 KB
Cache working set per multiprocessor
for constant memory
8 KB
Cache working set per multiprocessor
for texture memory
Device dependent, between 6 KB and 8 KB
Maximum width for a 1D texture
reference bound to a CUDA array
8192 32768
Maximum width for a 1D texture
reference bound to linear memory
2
27

Maximum width and height for a 2D
texture reference bound to linear
memory or a CUDA array
65536 x 32768
65536 x
65536
Maximum width, height, and depth
for a 3D texture reference bound to
linear memory or a CUDA array
2048 x 2048 x 2048
4096 x
4096 x
4096
Maximum number of instructions per
kernel
2 million

G.2 Floating-Point Standard
All compute devices follow the IEEE 754-2008 standard for binary floating-point
arithmetic with the following deviations:
Appendix G. Compute Capabilities


CUDA Programming Guide Version 3.0 141

 There is no dynamically configurable rounding mode; however, most of the
operations support multiple IEEE rounding modes, exposed via device
intrinsics;
 There is no mechanism for detecting that a floating-point exception has
occurred and all operations behave as if the IEEE-754 exceptions are always
masked, and deliver the masked response as defined by IEEE-754 if there is an
exceptional event; for the same reason, while SNaN encodings are supported,
they are not signaling and are handled as quiet;
 The result of a single-precision floating-point operation involving one or more
input NaNs is the quiet NaN of bit pattern 0x7fffffff;
 Double-precision floating-point absolute value and negation are not compliant
with IEEE-754 with respect to NaNs; these are passed through unchanged;
 For single-precision floating-point numbers on devices of compute
capability 1.x:
 Denormalized numbers are not supported; floating-point arithmetic and
comparison instructions convert denormalized operands to zero prior to
the floating-point operation;
 Underflowed results are flushed to zero;
 Some instructions are not IEEE-compliant:
 Addition and multiplication are often combined into a single multiply-
add instruction (FMAD), which truncates (i.e. without rounding) the
intermediate mantissa of the multiplication;
 Division is implemented via the reciprocal in a non-standard-compliant
way;
 Square root is implemented via the reciprocal square root in a non-
standard-compliant way;
 For addition and multiplication, only round-to-nearest-even and
round-towards-zero are supported via static rounding modes; directed
rounding towards +/- infinity is not supported;
To mitigate the impact of these restrictions, IEEE-compliant software (and
therefore slower) implementations are provided through the following
intrinsics (c.f. Section C.2.1):
 __fmaf_r{n,z,u,d}(float, float, float): single-precision
fused multiply-add with IEEE rounding modes,
 __frcp_r[n,z,u,d](float): single-precision reciprocal with
IEEE rounding modes,
 __fdiv_r[n,z,u,d](float, float): single-precision division
with IEEE rounding modes,
 __fsqrt_r[n,z,u,d](float): single-precision square root with
IEEE rounding modes,
 __fadd_r[u,d](float, float): single-precision addition with
IEEE directed rounding,
 __fmul_r[u,d](float, float): single-precision multiplication
with IEEE directed rounding;
 For double-precision floating-point numbers on devices of compute
capability 1.x:
Appendix G. Compute Capabilities


142 CUDA Programming Guide Version 3.0

 Round-to-nearest-even is the only supported IEEE rounding mode for
reciprocal, division, and square root.
When compiling for devices without native double-precision floating-point support,
i.e. devices of compute capability 1.2 and lower, each double variable is converted
to single-precision floating-point format (but retains its size of 64 bits) and double-
precision floating-point arithmetic gets demoted to single-precision floating-point
arithmetic.
For devices of compute capability 2.0 and higher, code must be compiled with
-ftz=false, -prec-div=true, and -prec-sqrt=true to ensure IEEE
compliance (this is the default setting; see the nvcc user manual for description of
these compilation flags); code compiled with -ftz=true, -prec-div=false,
and -prec-sqrt=false comes closest to the code generated for devices of
compute capability 1.x.
Addition and multiplication are often combined into a single multiply-add
instruction:
 FMAD for single precision on devices of compute capability 1.x,
 FFMA for single precision on devices of compute capability 2.0.
As mentioned above, FMAD truncates the mantissa prior to use it in the addition.
FFMA, on the other hand, is an IEEE-754(2008) compliant fused multiply-add
instruction, so the full-width product is being used in the addition and a single
rounding occurs during generation of the final result. While FFMA in general has
superior numerical properties compared to FMAD, the switch from FMAD to
FFMA can cause slight changes in numeric results and can in rare circumstances
lead to slighty larger error in final results.
In accordance to the IEEE-754R standard, if one of the input parameters to
fminf(), fmin(), fmaxf(), or fmax() is NaN, but not the other, the result is
the non-NaN parameter.
The conversion of a floating-point value to an integer value in the case where the
floating-point value falls outside the range of the integer format is left undefined by
IEEE-754. For compute devices, the behavior is to clamp to the end of the
supported range. This is unlike the x86 architecture behavior.
G.3 Compute Capability 1.x
G.3.1 Architecture
For devices of compute capability 1.x, a multiprocessor consists of:
 8 CUDA cores for integer and single-precision floating-point arithmetic
operations,
 1 double-precision floating-point unit for double-precision floating-point
arithmetic operations,
 2 special function units for single-precision floating-point transcendental
functions (these units can also handle single-precision floating-point
multiplications),
Appendix G. Compute Capabilities


CUDA Programming Guide Version 3.0 143

 1 warp scheduler.
To execute an instruction for all threads of a warp, the warp scheduler must
therefore issue the instruction over:
 4 clock cycles for an integer or single-precision floating-point arithmetic
instruction,
 32 clock cycles for a double-precision floating-point arithmetic instruction,
 16 clock cycles for a single-precision floating-point transcendental instruction.
A multiprocessor also has a read-only constant cache that is shared by all functional
units and speeds up reads from the constant memory space, which resides in device
memory.
Multiprocessors are grouped into Texture Processor Clusters (TPCs). The number of
multiprocessors per TPC is:
 2 for devices of compute capabilities 1.0 and 1.1,
 3 for devices of compute capabilities 1.2 and 1.3.
Each TPC has a read-only texture cache that is shared by all multiprocessors and
speeds up reads from the texture memory space, which resides in device memory.
Each multiprocessor accesses the texture cache via a texture unit that implements
the various addressing modes and data filtering mentioned in Section 3.2.4.
The local and global memory spaces reside in device memory and are not cached.
G.3.2 Global Memory
A global memory request for a warp is split into two memory requests, one for each
half-warp, that are issued independently. Sections G.3.2.1 and G.3.2.2 describe how
the memory accesses of threads within a half-warp are coalesced into one or more
memory transactions depending on the compute capability of the device. Figure G-1
shows some examples of global memory accesses and corresponding memory
transactions based on compute capability.
The resulting memory transactions are serviced at the throughput of device
memory.
G.3.2.1 Devices of Compute Capability 1.0 and 1.1
To coalesce, the memory request for a half-warp must satisfy the following
conditions:
 The size of the words accessed by the threads must be 4, 8, or 16 bytes;
 If this size is:
 4, all 16 words must lie in the same 64-byte segment,
 8, all 16 words must lie in the same 128-byte segment,
 16, the first 8 words must lie in the same 128-byte segment and the last 8
words in the following 128-byte segment;
 Threads must access the words in sequence: The k
th
thread in the half-warp
must access the k
th
word.
If the half-warp meets these requirements, a 64-byte memory transaction, a 128-byte
memory transaction, or two 128-byte memory transactions are issued if the size of
Appendix G. Compute Capabilities


144 CUDA Programming Guide Version 3.0

the words accessed by the threads is 4, 8, or 16, respectively. Coalescing is achieved
even if the warp is divergent, i.e. there are some inactive threads that do not actually
access memory.
If the half-warp does not meet these requirements, 16 separate 32-byte memory
transactions are issued.
G.3.2.2 Devices of Compute Capability 1.2 and 1.3
Threads can access any words in any order, including the same words, and a single
memory transaction for each segment addressed by the half-warp is issued. This is
in contrast with devices of compute capabilities 1.0 and 1.1 where threads need to
access words in sequence and coalescing only happens if the half-warp addresses a
single segment.
More precisely, the following protocol is used to determine the memory transactions
necessary to service all threads in a half-warp:
 Find the memory segment that contains the address requested by the lowest
numbered active thread. The segment size depends on the size of the words
accessed by the threads:
 32 bytes for 1-byte words,
 64 bytes for 2-byte words,
 128 bytes for 4-, 8- and 16-byte words.
 Find all other active threads whose requested address lies in the same segment.
 Reduce the transaction size, if possible:
 If the transaction size is 128 bytes and only the lower or upper half is used,
reduce the transaction size to 64 bytes;
 If the transaction size is 64 bytes (originally or after reduction from 128
bytes) and only the lower or upper half is used, reduce the transaction size
to 32 bytes.
 Carry out the transaction and mark the serviced threads as inactive.
 Repeat until all threads in the half-warp are serviced.
G.3.3 Shared Memory
Shared memory has 16 banks that are organized such that successive 32-bit words
are assigned to successive banks, i.e. interleaved. Each bank has a bandwidth of 32
bits per two clock cycles.
A shared memory request for a warp is split into two memory requests, one for each
half-warp, that are issued independently. As a consequence, there can be no bank
conflict between a thread belonging to the first half of a warp and a thread
belonging to the second half of the same warp.
If a non-atomic instruction executed by a warp writes to the same location in shared
memory for more than one of the threads of the warp, only one thread per half-
warp performs a write and which thread performs the final write is undefined.
Appendix G. Compute Capabilities


CUDA Programming Guide Version 3.0 145

G.3.3.1 32-Bit Strided Access
A common access pattern is for each thread to access a 32-bit word from an array
indexed by the thread ID tid and with some stride s:
__shared__ float shared[32];
float data = shared[BaseIndex + s * tid];
In this case, threads tid and tid+n access the same bank whenever s*n is a
multiple of the number of banks (i.e. 16) or, equivalently, whenever n is a multiple
of 16/d where d is the greatest common divisor of 16 and s. As a consequence,
there will be no bank conflict only if half the warp size (i.e. 16) is less than or equal
to 16/d., that is only if d is equal to 1, i.e. s is odd.
Figure G-2 shows some examples of strided access for devices of compute
capability 2.0. The same examples apply for devices of compute capability 1.x, but
with 16 banks instead of 32.
G.3.3.2 32-Bit Broadcast Access
Shared memory features a broadcast mechanism whereby a 32-bit word can be read
and broadcast to several threads simultaneously when servicing one memory read
request. This reduces the number of bank conflicts when several threads read from
an address within the same 32-bit word. More precisely, a memory read request
made of several addresses is serviced in several steps over time by servicing one
conflict-free subset of these addresses per step until all addresses have been
serviced; at each step, the subset is built from the remaining addresses that have yet
to be serviced using the following procedure:
 Select one of the words pointed to by the remaining addresses as the broadcast
word;
 Include in the subset:
 All addresses that are within the broadcast word,
 One address for each bank (other than the broadcasting bank) pointed to
by the remaining addresses.
Which word is selected as the broadcast word and which address is picked up for
each bank at each cycle are unspecified.
A common conflict-free case is when all threads of a half-warp read from an address
within the same 32-bit word.
Figure G-3 shows some examples of memory read accesses that involve the
broadcast mechanism. The same examples apply for devices of compute capability
1.x, but with 16 banks instead of 32.
G.3.3.3 8-Bit and 16-Bit Access
8-bit and 16-bit accesses typically generate bank conflicts. For example, there are
bank conflicts if an array of char is accessed the following way:
__shared__ char shared[32];
char data = shared[BaseIndex + tid];
because shared[0], shared[1], shared[2], and shared[3], for example,
belong to the same bank. There are no bank conflicts however, if the same array is
accessed the following way:
char data = shared[BaseIndex + 4 * tid];
Appendix G. Compute Capabilities


146 CUDA Programming Guide Version 3.0

G.3.3.4 Larger Than 32-Bit Access
Accesses that are larger than 32-bit per thread are split into 32-bit accesses that
typically generate bank conflicts.
For example, there are 2-way bank conflicts for arrays of doubles accessed as
follows:
__shared__ double shared[32];
double data = shared[BaseIndex + tid];
as the memory request is compiled into two separate 32-bit requests with a stride of
two. One way to avoid bank conflicts in this case is two split the double operands
like in the following sample code:
__shared__ int shared_lo[32];
__shared__ int shared_hi[32];

double dataIn;
shared_lo[BaseIndex + tid] = __double2loint(dataIn);
shared_hi[BaseIndex + tid] = __double2hiint(dataIn);

double dataOut =
__hiloint2double(shared_hi[BaseIndex + tid],
shared_lo[BaseIndex + tid]);
This might not always improve performance however and does perform worse on
devices of compute capabilities 2.0.
The same applies to structure assignments. The following code, for example:
__shared__ struct type shared[32];
struct type data = shared[BaseIndex + tid];
results in:
 Three separate reads without bank conflicts if type is defined as
struct type {
float x, y, z;
};
since each member is accessed with an odd stride of three 32-bit words;
 Two separate reads with bank conflicts if type is defined as
struct type {
float x, y;
};
since each member is accessed with an even stride of two 32-bit words.
G.4 Compute Capability 2.0
G.4.1 Architecture
For devices of compute capability 2.0, a multiprocessor consists of:
 32 CUDA cores for integer and floating-point arithmetic operations,
 4 special function units for single-precision floating-point transcendental
functions,
Appendix G. Compute Capabilities


CUDA Programming Guide Version 3.0 147

 2 warp schedulers.
At every instruction issue time, the first scheduler issues an instruction for some
warp with an odd ID and the second scheduler issues an instruction for some warp
with an even ID. The only exception is when a scheduler issues a double-precision
floating-point instruction: In this case, the other scheduler cannot issue any
instruction.
A warp scheduler can issue an instruction to only half of the CUDA cores. To
execute an instruction for all threads of a warp, a warp scheduler must therefore
issue the instruction over:
 2 clock cycles for an integer or floating-point arithmetic instruction,
 2 clock cycles for a double-precision floating-point arithmetic instruction,
 8 clock cycles for a single-precision floating-point transcendental instruction.
A multiprocessor also has a read-only uniform cache that is shared by all functional
units and speeds up reads from the constant memory space, which resides in device
memory.
There is an L1 cache for each multiprocessor and an L2 cache shared by all
multiprocessors, both of which are used to cache accesses to local or global
memory, including temporary register spills. The cache behavior (e.g. whether reads
are cached in both L1 and L2 or in L2 only) can be partially configured on a per-
access basis using modifiers to the load or store instruction.
The same on-chip memory is used for both L1 and shared memory: It can be
configured as 48 KB of shared memory with 16 KB of L1 cache (default setting) or
as 16 KB of shared memory with 48 KB of L1 cache using
cudaFuncSetCacheConfig() or cuFuncSetCacheConfig():
// Device code
__global__ void MyKernel()
{
...
}

// Host code

// Runtime API
// cudaFuncCachePreferShared: shared memory is 48 KB
// cudaFuncCachePreferL1: shared memory is 16 KB
// cudaFuncCachePreferNone: no preference
cudaFuncSetCacheConfig(MyKernel, cudaFuncCachePreferShared)

// Driver API
// CU_FUNC_CACHE_PREFER_SHARED: shared memory is 48 KB
// CU_FUNC_CACHE_PREFER_L1: shared memory is 16 KB
// CU_FUNC_CACHE_PREFER_NONE: no preference
CUfunction myKernel;
cuFuncSetCacheConfig(myKernel, CU_FUNC_CACHE_PREFER_SHARED)

Multiprocessors are grouped into Graphics Processor Clusters (GPCs). A GPC includes
four multiprocessors.
Each multiprocessor has a read-only texture cache to speed up reads from the
texture memory space, which resides in device memory. It accesses the texture cache
Appendix G. Compute Capabilities


148 CUDA Programming Guide Version 3.0

via a texture unit that implements the various addressing modes and data filtering
mentioned in Section 3.2.4.
G.4.2 Global Memory
Global memory accesses are cached. Using the –dlcm compilation flag, they can be
configured at compile time to be cached in both L1 and L2 (-Xptxas -dlcm=ca)
or in L2 only (-Xptxas -dlcm=cg).
A cache line in L1 or L2 is 128 bytes and maps to a 128-byte aligned segment in
device memory.
If the size of the words accessed by each thread is more than 4 bytes, a memory
request by a warp is first split into separate 128-byte memory requests that are
issued independently:
 Two memory requests, one for each half-warp, if the size is 8 bytes,
 Four memory requests, one for each quarter-warp, if the size is 16 bytes.
Each memory request is then broken down into cache line requests that are issued
independently. A cache line request is serviced at the throughput of L1 or L2 cache
in case of a cache hit, or at the throughput of device memory, otherwise.
If a non-atomic instruction executed by a warp writes to the same location in global
memory for more than one of the threads of the warp, only one thread performs a
write and which thread does it is undefined.

Appendix G. Compute Capabilities


CUDA Programming Guide Version 3.0 149


Figure G-1. Examples of Global Memory Accesses by a Warp,
4-Byte Word per Thread, and Associated Memory
Transactions Based on Compute Capability
128 160 192 256 224 96 288 Addresses:
0 31

Threads:
Compute capability:
Memory transactions:
1.0 and 1.1
8 x 32B at 128
8 x 32B at 160
8 x 32B at 192
8 x 32B at 224
1.2 and 1.3
1 x 64B at 128
1 x 64B at 192


2.0
1 x 128B at 128
Aligned and non-sequential
Uncached Cached
128 160 192 256 224 96 288 Addresses:
0 31

Threads:
Compute capability:
Memory transactions:
1.0 and 1.1
1 x 64B at 128
1 x 64B at 192
1.2 and 1.3
1 x 64B at 128
1 x 64B at 192


2.0
1 x 128B at 128
Aligned and sequential
Uncached Cached
128 160 192 256 224 96 288 Addresses:
0 31

Threads:
Compute capability:
Memory transactions:
1.0 and 1.1
8 x 32B at 128
8 x 32B at 160
8 x 32B at 192
8 x 32B at 224
1.2 and 1.3
1 x 128B at 128
1 x 64B at 192
1 x 32B at 256


2.0
1 x 128B at 128
1 x 128B at 256
Misaligned and sequential
Uncached Cached
Appendix G. Compute Capabilities


150 CUDA Programming Guide Version 3.0

G.4.3 Shared Memory
Shared memory has 32 banks that are organized such that successive 32-bit words
are assigned to successive banks, i.e. interleaved. Each bank has a bandwidth of 32
bits per two clock cycles. Therefore, unlike for devices of lower compute capability,
there may be bank conflicts between a thread belonging to the first half of a warp
and a thread belonging to the second half of the same warp.
A bank conflict only occurs if two or more threads access any bytes within different
32-bit words belonging to the same bank. If two or more threads access any bytes
within the same 32-bit word, there is no bank conflict between these threads: For
read accesses, the word is broadcast to the requesting threads (unlike for devices of
compute capability 1.x, multiple words can be broadcast in a single transaction); for
write accesses, each byte is written by only one of the threads (which thread
performs the write is undefined).
This means, in particular, that unlike for devices of compute capability 1.x, there are
no bank conflicts if an array of char is accessed as follows, for example:
__shared__ char shared[32];
char data = shared[BaseIndex + tid];
G.4.3.1 32-Bit Strided Access
A common access pattern is for each thread to access a 32-bit word from an array
indexed by the thread ID tid and with some stride s:
__shared__ float shared[32];
float data = shared[BaseIndex + s * tid];
In this case, threads tid and tid+n access the same bank whenever s*n is a
multiple of the number of banks (i.e. 32) or, equivalently, whenever n is a multiple
of 32/d where d is the greatest common divisor of 32 and s. As a consequence,
there will be no bank conflict only if the warp size (i.e. 32) is less than or equal to
32/d., that is only if d is equal to 1, i.e. s is odd.
Figure G-2 shows some examples of strided access.
G.4.3.2 Larger Than 32-Bit Access
64-bit and 128-bit accesses are specifically handled to minimize bank conflicts as
described below.
Other accesses larger than 32-bit are split into 32-bit, 64-bit, or 128-bit accesses.
The following code, for example:
struct type {
float x, y, z;
};

__shared__ struct type shared[32];
struct type data = shared[BaseIndex + tid];
results in three separate 32-bit reads without bank conflicts since each member is
accessed with a stride of three 32-bit words.
64-Bit Accesses
For 64-bit accesses, a bank conflict only occurs if two or more threads in either of
the half-warps access different addresses belonging to the same bank.
Appendix G. Compute Capabilities


CUDA Programming Guide Version 3.0 151

Unlike for devices of compute capability 1.x, there are no bank conflicts for arrays
of doubles accessed as follows, for example:
__shared__ double shared[32];
double data = shared[BaseIndex + tid];
128-Bit Accesses
The majority of 128-bit accesses will cause 2-way bank conflicts, even if no two
threads in a quarter-warp access different addresses belonging to the same bank.
Therefore, to determine the ways of bank conflicts, one must add 1 to the
maximum number of threads in a quarter-warp that access different addresses
belonging to the same bank.
G.4.4 Constant Memory
In addition to the constant memory space supported by devices of all compute
capabilities (where __constant__ variables reside), devices of compute capability
2.0 support the LDU (LoaD Uniform) instruction that the compiler use to load any
variable that is:
 pointing to global memory,
 read-only in the kernel (programmer can enforce this using the const
keyword),
 not dependent on thread ID.


Appendix G. Compute Capabilities


152 CUDA Programming Guide Version 3.0


Left: Linear addressing with a stride of one 32-bit word (no bank conflict).
Middle: Linear addressing with a stride of two 32-bit words (2-way bank conflicts).
Right: Linear addressing with a stride of three 32-bit words (no bank conflict).
Figure G-2 Examples of Strided Shared Memory Accesses for
Devices of Compute Capability 2.0
31

30

29

28

27

26

25

24

23

22

21

20

19

18

17

16

15

14

13

12

11

10

9

8

7

6

5

4

3

2

1

0

Banks: Threads:
0
1
2
4
3
5
6
7
8
9
10
12
11
13
14
15
16
17
18
20
19
21
22
23
24
25
26
28
27
29
30
31 31

30

29

28

27

26

25

24

23

22

21

20

19

18

17

16

15

14

13

12

11

10

9

8

7

6

5

4

3

2

1

0

Banks: Threads:
0
1
2
4
3
5
6
7
8
9
10
12
11
13
14
15
16
17
18
20
19
21
22
23
24
25
26
28
27
29
30
31 31

30

29

28

27

26

25

24

23

22

21

20

19

18

17

16

15

14

13

12

11

10

9

8

7

6

5

4

3

2

1

0

Banks: Threads:
0
1
2
4
3
5
6
7
8
9
10
12
11
13
14
15
16
17
18
20
19
21
22
23
24
25
26
28
27
29
30
31
Appendix G. Compute Capabilities


CUDA Programming Guide Version 3.0 153


Left: Conflict-free access via random permutation.
Middle: Conflict-free access since threads 3, 4, 6, 7, and 9 access the same word within bank 5.
Right: Conflict-free broadcast access (all threads access the same word).
Figure G-3 Examples of Irregular and Colliding Shared
Memory Accesses for Devices of Compute
Capability 2.0
31

30

29

28

27

26

25

24

23

22

21

20

19

18

17

16

15

14

13

12

11

10

9

8

7

6

5

4

3

2

1

0

Banks: Threads:
0
1
2
4
3
5
6
7
8
9
10
12
11
13
14
15
16
17
18
20
19
21
22
23
24
25
26
28
27
29
30
31 31

30

29

28

27

26

25

24

23

22

21

20

19

18

17

16

15

14

13

12

11

10

9

8

7

6

5

4

3

2

1

0

Banks: Threads:
0
1
2
4
3
5
6
7
8
9
10
12
11
13
14
15
16
17
18
20
19
21
22
23
24
25
26
28
27
29
30
31 31

30

29

28

27

26

25

24

23

22

21

20

19

18

17

16

15

14

13

12

11

10

9

8

7

6

5

4

3

2

1

0

Banks: Threads:
0
1
2
4
3
5
6
7
8
9
10
12
11
13
14
15
16
17
18
20
19
21
22
23
24
25
26
28
27
29
30
31

NVIDIA Corporation
2701 San Tomas Expressway
Santa Clara, CA 95050
www.nvidia.com
Notice
ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND
OTHER DOCUMENTS (TOGETHER AND SEPARATELY, “MATERIALS”) ARE BEING PROVIDED “AS IS.” NVIDIA
MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE
MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT,
MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE.
Information furnished is believed to be accurate and reliable. However, NVIDIA Corporation assumes no
responsibility for the consequences of use of such information or for any infringement of patents or other
rights of third parties that may result from its use. No license is granted by implication or otherwise under any
patent or patent rights of NVIDIA Corporation. Specifications mentioned in this publication are subject to
change without notice. This publication supersedes and replaces all information previously supplied. NVIDIA
Corporation products are not authorized for use as critical components in life support devices or systems
without express written approval of NVIDIA Corporation.
Trademarks
NVIDIA, the NVIDIA logo, GeForce, Tesla, and Quadro are trademarks or registered trademarks of NVIDIA
Corporation. Other company and product names may be trademarks of the respective companies with which
they are associated.
OpenCL is trademark of Apple Inc. used under license to the Khronos Group Inc.
Copyright
© 2006-2010 NVIDIA Corporation. All rights reserved.
This work incorporates portions of on an earlier work: Scalable Parallel Programming with CUDA, in ACM
Queue, VOL 6, No. 2 (March/April 2008), © ACM, 2008. http://mags.acm.org/queue/20080304/?u1=texterity"

ii

CUDA Programming Guide Version 3.0

Table of Contents

Chapter 1. Introduction ..................................................................................... 1 1.1 1.2 1.3 1.4 2.1 2.2 2.3 2.4 2.5 3.1 From Graphics Processing to General-Purpose Parallel Computing ................... 1 CUDA™: a General-Purpose Parallel Computing Architecture ........................... 3 A Scalable Programming Model ..................................................................... 4 Document’s Structure ................................................................................... 5 Kernels ........................................................................................................ 7 Thread Hierarchy.......................................................................................... 8 Memory Hierarchy ...................................................................................... 10 Heterogeneous Programming ...................................................................... 11 Compute Capability .................................................................................... 14 Compilation with NVCC ............................................................................... 15 Compilation Workflow .......................................................................... 16 Binary Compatibility ............................................................................. 16 PTX Compatibility ................................................................................ 16 Application Compatibility ...................................................................... 17 C/C++ Compatibility ............................................................................ 18 Device Memory .................................................................................... 18 Shared Memory ................................................................................... 21 Multiple Devices................................................................................... 27 Texture Memory .................................................................................. 28 Texture Reference Declaration ....................................................... 29 Runtime Texture Reference Attributes ............................................ 29 Texture Binding ............................................................................ 30 Portable Memory ........................................................................... 33

Chapter 2. Programming Model ......................................................................... 7

Chapter 3. Programming Interface ................................................................. 15 3.1.1 3.1.2 3.1.3 3.1.4 3.1.5 3.2 3.2.1 3.2.2 3.2.3 3.2.4

CUDA C ..................................................................................................... 18

3.2.4.1 3.2.4.2 3.2.4.3 3.2.5 3.2.5.1

Page-Locked Host Memory ................................................................... 33

CUDA Programming Guide Version 3.0

iii

............. 57 Multiple Devices..3....................................9.................2................ 34 Overlap of Data Transfer and Kernel Execution ..............7 3...........9 3......................3..................1 3.....3.....6 3.......................5..............................................................10............1 3...............................................7.................3......8 3..................................................6 3.........3...... 51 Module ................5...2................ 58 Page-Locked Host Memory ................................6.............0 ..................2...........................2.......... 46 Debugging using the Device Emulation Mode ...................... 63 Direct3D Interoperability ............................................................6..........................................2.9 Write-Combining Memory .....3 3.......................................... 47 Context .............2 3.............................. 60 Asynchronous Concurrent Execution ...............................9....6..7................................. 72 Interoperability between Runtime and Driver APIs ................................................................3.....2......5 3.. 49 3.............................2 3.4 3..9............................3................................................1 3.....................................................2.............. 35 Event ...3... 65 Driver API ................. 72 iv CUDA Programming Guide Version 3............3.................6 3..........3 3..................... 35 Concurrent Data Transfers .................................6..........................................2......................................................2......................10........................ 61 Stream ..... 35 Stream ...................... 63 OpenGL Interoperability ................................................................................................................ 63 Error Handling ................2 3....... 35 Concurrent Kernel Execution .....3.................. 34 Graphics Interoperability ....................................................6................................................................................................ 52 Device Memory ...........................3.3..8 3..................................2..................7 3....................3.................................. 37 Error Handling .3............2 3...............3 3.....2 3...............5 3...... 61 Event Management .. 58 Texture Memory ........3......................................... 54 Shared Memory .......10 3.............................................2...... 38 Direct3D Interoperability .................................................2 3...7 3............................ 37 OpenGL Interoperability .........................2...4 Graphics Interoperability .................... 52 Kernel Execution ....................6............... 40 Asynchronous Concurrent Execution .................................. 33 Mapped Memory ........................ 62 Synchronous Calls .....6......................................11 3......1 3.3 3........................1 3.................................4 3........3.................................. 33 Concurrent Execution between Host and Device....3 3........2................2............................... 37 Synchronous Calls ...2................3....

..... 81 Maximize Utilization .........................2...........4 5................................................. 97 __global__ ........................................ 93 Appendix A...................................................1 Function Type Qualifiers ........... 81 Device Level ............................2........3 5...... 98 CUDA Programming Guide Version 3............... C Language Extensions .......... 95 Appendix B...........................2....................................2 4........................4....... 78 Multiple Devices ....................................4............................ 85 Global Memory.... 89 Arithmetic Instructions .....3.............4 5...3.......................................2........................................ 77 Chapter 5......................................................................................1......................2 B................. 97 Restrictions .. 81 Application Level.................................................................................................. Hardware Implementation .2 5..........................................................................4..........3................. 73 Compute Modes ................................................................3.............................3 5.......................................................2 Versioning and Compatibility ......1 B........................ 98 B................3 Maximize Instruction Throughput .....5 3...3................1 5.... 85 Device Memory Accesses ................. 79 Overall Performance Optimization Strategies .............4 B.................................2 5..............................................................................................................3................................. 81 5..............1 4..................................................2 Variable Type Qualifiers .................... 82 Data Transfer between Host and Device ......................... 97 __host__ ........................................ 77 Hardware Multithreading ........................1 5................ 87 Shared Memory ................................................................................................ 92 Synchronization Instruction ............................1 5.................. 88 Texture Memory ............3......................................................................3 5...2 Maximize Memory Throughput .................3........... 74 SIMT Architecture... 97 B....................................................... Performance Guidelines .......2 5..................................... 89 Chapter 4................ 97 __device__ ...................................................................................................... 82 Multiprocessor Level ......1........0 v ......................1.......3 B.........................................................................7 4. 90 Control Flow Instructions ...............................................1 5...................................................5 5............................... 74 Mode Switches ................................................................2.... 86 Local Memory .................... 84 5......................................... 88 Constant Memory .....................................2.......6 3.........1........3 5................2.................................................................1 5...............................................2............ CUDA-Enabled GPUs ....................................

......... float2.................2 B.4................... longlong2.................................1..........................5 B................2 B................. long4...1..............3 __device__ ............... long3.. 98 __constant__ ........................ 108 Arithmetic Functions ............................4 B................. 99 Restrictions ........... 102 blockDim .........3 B............ uchar3................................ uint2........ uchar1... short3...........3.............2...............4 B................. 104 Mathematical Functions .......... ulong2..........................................................................8..................... 102 blockIdx ............................................................................. int1............................................................ ulong4. 109 atomicMax() ....................................... double2 ................. 100 Built-in Vector Types ................... uint3. 107 tex3D() .............................4 B..................................1.............10 Time Function .......................................................... ulong3.....2................9 B.... double1.................... 107 Atomic Functions .............4............2................ short2........ uint1....3 B............ char2........ 105 Texture Functions......................10.....1 B..................................................... 103 Restrictions .......................... uchar2....................................... short1. 108 atomicAdd().......................8...................................................................... ushort2....8........................................ 101 B.....1........ float1................2......5 B......... 107 tex2D() .................2 B..................8 dim3 .....3 B.....................2 B....................................................................................................... 101 B................................................. 109 atomicMin() .......... uchar4.....4.................................10............................................5 B.B......4 B..... long1.............................................................................................0 .....................10....................... 106 tex1Dfetch() ............................10..............5 B......................4........ 106 tex1D() ..........................................3 B................................1 B.. 102 warpSize ......................4 B....6 B...........................................2 B........................ float4................................................................................1 B... 108 atomicExch() ..................1 vi CUDA Programming Guide Version 3....... ulong1................... 102 threadIdx .........................................10.... float3........... char3..........................10.................1 B............................................4............................................................ 102 Memory Fence Functions .... long2....................... 103 Built-in Variables ................................ uint4................... 109 B.................................................. 98 __shared__ ......2.............. 107 B.... 108 atomicSub() ........ int4............. 99 volatile .1............... ushort3.........1 char1...................... ushort4..8............. short4... 103 Synchronization Functions ............6 B....... ushort1........ 102 gridDim .. longlong1......................................... int2..............4............................3.7 B........... int3.............................. char4......

.....................................2 C...........................................................................................................................................5 D...................................................................6..................................................7 B....................2.......................1....14 C.............................................. 123 Appendix C........ 109 atomicDec() .. 125 Polymorphism .....................................................................................................10..................................... 132 CUDA Programming Guide Version 3.........1 C....2...........................1 D................. 120 Appendix D........ 127 Function Templates ............................... 126 Operator Overloading ..........................10.1.... 110 Warp Vote Functions ................ 110 Bitwise Functions ....................................... 120 Double-Precision Floating-Point Functions ....... 122 Integer Functions ...3 C.......... 131 __noinline__ .................2 C...................................6......................12 B.....................................2 C..2...............................................................................................................................2..........2 B..............0 vii ...................10...................................................................................... 110 atomicOr() ........................1 atomicInc() ................................1...................................... 126 Namespaces......... 111 Execution Configuration ............................. NVCC Specifics ......................................................1 E....... 115 C..................................................................................................................1.....................................................................................6 Intrinsic Functions ................................................ 109 atomicCAS() ................................................ 131 __restrict__ ....................... 127 Classes ....................6 B.......1 C...................4 D.......................................B...................10.................... 125 Default Parameters ....................2 E.. 118 Integer Functions ......... 112 Standard Functions ...2 D................... 110 Profiler Counter Function ...................................................................3 Appendix E...............................3 D......... 115 Single-Precision Floating-Point Functions ..3 B.. 128 Example 1 Pixel Data Type ................13 B.................10.10....8 B.............................................................................................................................. 109 atomicAnd()........2 B.................................... C++ Language Constructs ............... 128 Example 2 Functor Class ...................2 E.................1..............3 D........................ 120 Single-Precision Floating-Point Functions ..1................. 131 #pragma unroll .......................................... 115 Double-Precision Floating-Point Functions ....2............................................................11 B..........................................2.....1 B....................................... 110 atomicXor() ..............1 D............................................................... 111 Launch Bounds........... 129 D.. Mathematical Functions ...................................10......................

...3........................................2 and 1.................................................................................................................. Texture Fetching ........ 143 Devices of Compute Capability 1........................3........................................................ 146 Appendix G......................0 ....... 148 Shared Memory ...................3................... 136 Linear Filtering ................... 150 Larger Than 32-Bit Access ...........................................3. Compute Capabilities .....Appendix F..2 F........................2 G.............3................4 G......................... 135 F...............................................3 Shared Memory ...............................3 G.....1 G.............4...............3 Nearest-Point Sampling...3 G.............3.......... 142 Global Memory .............4... 136 Table Lookup .......................................................3.............................................. 140 Compute Capability 1.......2..4............... 144 Compute Capability 2................... 146 Global Memory ....................................................................3..............1 G.. 139 G..... 145 Larger Than 32-Bit Access .....................................................................3.................................................................0 and 1.........................................3 ....................... 143 Devices of Compute Capability 1............4.................1 ................................................................. 145 8-Bit and 16-Bit Access ................ 145 32-Bit Broadcast Access....2 G....3.............1 G...................................... 150 32-Bit Strided Access .....3.................2 G.................... 150 G..................3........... 144 32-Bit Strided Access ...................... 151 viii CUDA Programming Guide Version 3......... 138 Features and Technical Specifications .........................3............................ 139 Floating-Point Standard ..................0 .................................. 142 Architecture.............1 G.........................................3...............2 G.....4.............. 146 Architecture....2 G............x .....................3...........................2.3 G.............1 F...............................2 G...4 G................................4 Constant Memory .........................................1 G................1 G...........................................4.........

...... 13 Matrix Multiplication without Shared Memory .. CUDA is Designed to Support Various Languages or Application Programming Interfaces ................................................ Figure 3-2.............................. 52 The Driver API is Backward.....0.............. Figure 2-2. 23 Matrix Multiplication with Shared Memory............. 137 Figure E-3.................................. 9 Memory Hierarchy ....................... 138 Figure G-1......... Figure 2-3....................................0 ............................ Automatic Scalability ............... 11 Heterogeneous Programming ...... but Not Forward Compatible .................. 152 Figure G-3 Examples of Irregular and Broadcast Shared Memory Accesses for Devices of Compute Capability 2.......................................................................... 27 Library Context Management ......................... 5 Grid of Thread Blocks ................0 ix ........ 4 Figure 1-4..................... Figure 2-1................ 74 Nearest-Point Sampling of a One-Dimensional Texture of Four Texels ............... Figure 3-1..... and Associated Memory Transactions Based on Compute Capability ..................... 3 Figure 1-3............................................ One-Dimensional Table Lookup Using Linear Filtering ....... Figure 3-3.............................................................................................................................. Examples of Global Memory Accesses by a Warp. 149 Figure G-2 Examples of Strided Shared Memory Accesses for Devices of Compute Capability 2.List of Figures Figure 1-1.............................. 4-Byte Word per Thread....................... 136 Figure E-2........................................................ Floating-Point Operations per Second and Memory Bandwidth for the CPU and GPU 2 Figure 1-2... Figure 3-4........................................... 153 CUDA Programming Guide Version 3................................. Linear Filtering of a One-Dimensional Texture of Four Texels in Clamp Addressing Mode ... The GPU Devotes More Transistors to Data Processing .................... Figure E-1..............................................

.

high-definition 3D graphics.Chapter 1. manycore processor with tremendous computational horsepower and very high memory bandwidth. multithreaded. the programmable Graphic Processor Unit or GPU has evolved into a highly parallel. CUDA Programming Guide Version 3.1 From Graphics Processing to General-Purpose Parallel Computing Driven by the insatiable market demand for realtime.0 1 . as illustrated by Figure 1-1. Introduction 1.

highly parallel computation – exactly what graphics rendering is about – and therefore designed such that more transistors are devoted to data processing rather than data caching and flow control.0 . Floating-Point Operations per Second and Memory Bandwidth for the CPU and GPU The reason behind the discrepancy in floating-point capability between the CPU and the GPU is that the GPU is specialized for compute-intensive. Introduction GT200 G80 G80 Ultra G92 G71 NV40 G70 3. 2 CUDA Programming Guide Version 3.0 GHz Core2 Duo Jun 2005 Mar 2006 Nov May 2007 3.2 GHz Harpertown NV35 NV30 Jan Jun 2003 Apr 2004 Jun 2008 GT200 = GeForce GTX 280 G92 = GeForce 9800 GTX G80 = GeForce 8800 GTX G71 = GeForce 7900 GTX G70 = GeForce 7800 GTX NV40 = GeForce 6800 Ultra NV35 = GeForce FX 5950 Ultra NV30 = GeForce FX 5800 G80 Ultra G80 G71 NV40 Harpertown Woodcrest Prescott EE Northwood NV30 Figure 1-1. as schematically illustrated by Figure 1-2.Chapter 1.

Many applications that process large data sets can use a data-parallel programming model to speed up the computations. other languages or application programming interfaces are supported. As illustrated by Figure 1-3. OpenCL. a general purpose parallel computing architecture – with a new parallel programming model and instruction set architecture – that leverages the parallel compute engine in NVIDIA GPUs to solve many complex computational problems in a more efficient way than on a CPU. stereo vision. from general signal processing or physics simulation to computational finance or computational biology. many algorithms outside the field of image rendering and processing are accelerated by data-parallel processing. there is a lower requirement for sophisticated flow control. image scaling. Data-parallel processing maps data elements to parallel processing threads. Because the same program is executed for each data element. Similarly. NVIDIA introduced CUDA™. Introduction Control ALU ALU ALU ALU Cache DRAM DRAM CPU GPU Figure 1-2. 1. and because it is executed on many data elements and has high arithmetic intensity. In 3D rendering.0 3 . video encoding and decoding.2 CUDA™: a General-Purpose Parallel Computing Architecture In November 2006.Chapter 1. the memory access latency can be hidden with calculations instead of big data caches. The GPU Devotes More Transistors to Data Processing More specifically. CUDA comes with a software environment that allows developers to use C as a high-level programming language. large sets of pixels and vertices are mapped to parallel threads. and pattern recognition can map image blocks and pixels to parallel processing threads. such as CUDA FORTRAN. image and media processing applications such as post-processing of rendered images. and DirectCompute. the GPU is especially well-suited to address problems that can be expressed as data-parallel computations – the same program is executed on many data elements in parallel – with high arithmetic intensity – the ratio of arithmetic operations to memory operations. CUDA Programming Guide Version 3. In fact.

Furthermore.Chapter 1. 4 CUDA Programming Guide Version 3. At its core are three key abstractions – a hierarchy of thread groups. and barrier synchronization – that are simply exposed to the programmer as a minimal set of language extensions. Introduction Figure 1-3. in any order. their parallelism continues to scale with Moore‟s law. This decomposition preserves language expressivity by allowing threads to cooperate when solving each sub-problem. The CUDA parallel programming model is designed to overcome this challenge while maintaining a low learning curve for programmers familiar with standard programming languages such as C. and each sub-problem into finer pieces that can be solved cooperatively in parallel by all threads within the block. CUDA is Designed to Support Various Languages or Application Programming Interfaces 1. The challenge is to develop application software that transparently scales its parallelism to leverage the increasing number of processor cores. shared memories. They guide the programmer to partition the problem into coarse sub-problems that can be solved independently in parallel by blocks of threads. each block of threads can be scheduled on any of the available processor cores. and only the runtime system needs to know the physical processor count. nested within coarse-grained data parallelism and task parallelism. and at the same time enables automatic scalability.0 .3 A Scalable Programming Model The advent of multicore CPUs and manycore GPUs means that mainstream processor chips are now parallel systems. These abstractions provide fine-grained data parallelism and thread parallelism. Indeed. concurrently or sequentially. so that a compiled CUDA program can execute on any number of processor cores as illustrated by Figure 1-4. much as 3D graphics applications transparently scale their parallelism to manycore GPUs with widely varying numbers of cores.

Chapter 2 outlines the CUDA programming model. Chapter 3 describes the programming interface. mainstream GeForce GPUs (see Appendix A for a list of all CUDA-enabled GPUs). Chapter 5 gives some guidance on how to achieve maximum performance. Chapter 4 describes the hardware implementation. 5 CUDA Programming Guide Version 3.4 Document’s Structure This document is organized into the following chapters:       Chapter 1 is a general introduction to CUDA.0 . Figure 1-4. Appendix A lists all CUDA-enabled devices. Automatic Scalability 1. Multithreaded CUDA Program Block 0 Block 4 Block 1 Block 5 Block 5 Block 2 Block 6 Block 6 Block 3 Block 7 GPU with 2 Cores Core 0 Core 1 GPU with 4 Cores Core 0 Core 1 Core 2 Core 3 Block 0 Block 1 Block 0 Block 1 Block 2 Block 3 Block 2 Block 3 Block 4 Block 5 Block 6 Block 7 Block 4 Block 5 Block 6 Block 7 A multithreaded program is partitioned into blocks of threads that execute independently from each other.Chapter 1. so that a GPU with more cores will automatically execute the program in less time than a GPU with fewer cores. Introduction This scalable programming model allows the CUDA architecture to span a wide market range by simply scaling the number of processors and memory partitions: from the high-performance enthusiast GeForce GPUs and professional Quadro and Tesla computing products to a variety of inexpensive.

0 . Appendix C lists the mathematical functions supported in CUDA. as well as more architectural details. Introduction       Appendix B is a detailed description of all extensions to the C language.Chapter 1. 6 CUDA Programming Guide Version 3. Appendix E lists the specific keywords and directives supported by nvcc. Appendix G gives the technical specifications of various devices. Appendix D lists the C++ constructs supported in device code. Appendix F gives more details on texture fetching.

2. Each thread that executes the kernel is given a unique thread ID that is accessible within the kernel through the built-in threadIdx variable. // Kernel invocation with N threads VecAdd<<<1. Full code for the vector addition example used in this chapter and the next can be found in the vectorAdd SDK code sample. are executed N times in parallel by N different CUDA threads. float* C) { int i = threadIdx. as opposed to only once like regular C functions. N>>>(A. when called.13). that. B. float* B. the following sample code adds two vectors A and B of size N and stores the result into vector C: // Kernel definition __global__ void VecAdd(float* A. Programming Model This chapter introduces the main concepts behind the CUDA programming model by outlining how they are exposed in C. An extensive description of CUDA C is given in Section 3. } int main() { .0 7 .2. C[i] = A[i] + B[i].x. A kernel is defined using the __global__ declaration specifier and the number of CUDA threads that execute that kernel for a given kernel call is specified using a new <<<…>>> execution configuration syntax (see Appendix B. As an illustration. } Here. CUDA Programming Guide Version 3.1 Kernels CUDA C extends C by allowing the programmer to define C functions.Chapter 2. called kernels... C). each of the N threads that execute VecAdd() performs one pair-wise addition.

int j = threadIdx. for a two-dimensional block of size (Dx. Dy. } int main() { . Dy). B. C). Programming Model 2. or three-dimensional thread index. y) is (x + y Dx). Blocks are organized into a one-dimensional or two-dimensional grid of thread blocks as illustrated by Figure 2-1. the thread ID of a thread of index (x. dim3 threadsPerBlock(N. This provides a natural way to invoke computation across the elements in a domain such as a vector. C[i][j] = A[i][j] + B[i][j]. or volume. they are the same. However. two-dimensional. Dz).Chapter 2. // Kernel invocation with one block of N * N * 1 threads int numBlocks = 1. the following code adds two matrices A and B of size NxN and stores the result into matrix C: // Kernel definition __global__ void MatAdd(float A[N][N]. float B[N][N]. threadsPerBlock>>>(A. threadIdx is a 3-component vector.y.. z) is (x + y Dx + z Dx Dy). N). a thread block may contain up to 512 threads. so that the total number of threads is equal to the number of threads per block times the number of blocks. float C[N][N]) { int i = threadIdx. a kernel can be executed by multiple equally-shaped thread blocks. matrix. MatAdd<<<numBlocks. two-dimensional. or three-dimensional thread block. so that threads can be identified using a one-dimensional. The index of a thread and its thread ID relate to each other in a straightforward way: For a one-dimensional block. 8 CUDA Programming Guide Version 3. y.. The number of thread blocks in a grid is usually dictated by the size of the data being processed or the number of processors in the system.x. } There is a limit to the number of threads per block. since all threads of a block are expected to reside on the same processor core and must share the limited memory resources of that core. As an example. which it can greatly exceed. On current GPUs. forming a one-dimensional.2 Thread Hierarchy For convenience.0 . the thread ID of a thread of index (x. for a threedimensional block of size (Dx.

0) Thread (0. The dimension of the thread block is accessible within the kernel through the built-in blockDim variable.0 9 . 2) Figure 2-1. Two-dimensional blocks or grids can be specified as in the example above. 2) Thread (3. Grid of Thread Blocks The number of threads per block and the number of blocks per grid specified in the <<<…>>> syntax can be of type int or dim3. CUDA Programming Guide Version 3. 0) Block (0. int j = blockIdx.y. 1) Block (1. 0) Block (2. float C[N][N]) { int i = blockIdx. 1) Block (1. 1) Thread (0.x * blockDim. if (i < N && j < N) C[i][j] = A[i][j] + B[i][j]. float B[N][N]. the code becomes as follows.x + threadIdx.y + threadIdx. 2) Thread (1. 1) Thread (0. 0) Thread (3. 0) Block (1. Extending the previous MatAdd() example to handle multiple blocks. Each block within the grid can be identified by a one-dimensional or twodimensional index accessible within the kernel through the built-in blockIdx variable. 1) Thread (1. 1) Thread (3.y * blockDim. 0) Thread (1. 2) Thread (2.Chapter 2: Programming Model Grid Block (0.x. // Kernel definition __global__ void MatAdd(float A[N][N]. 1) Thread (2. 0) Thread (2. 1) Block (2.

y). and texture memory spaces are optimized for different memory usages (see Sections 5.3.2.5). Texture memory also offers different addressing modes. and texture memory spaces are persistent across kernel launches by the same application. All threads have access to the same global memory. For efficient cooperation..0 . Thread blocks are required to execute independently: It must be possible to execute them in any order. } A thread block size of 16x16 (256 threads).2 gives an example of using shared memory.2.2. constant. dim3 numBlocks(N / threadsPerBlock. The global. Programming Model } int main() { . this example assumes that the number of threads per grid in each dimension is evenly divisible by the number of threads per block in that dimension. for some specific data formats (see Section 3. N / threadsPerBlock. 2. The grid is created with enough blocks to have one thread per matrix element as before. 5. For simplicity.x. This independence requirement allows thread blocks to be scheduled in any order across any number of cores as illustrated by Figure 1-4. enabling programmers to write code that scales with the number of cores. one can specify synchronization points in the kernel by calling the __syncthreads() intrinsic function. Each thread block has shared memory visible to all threads of the block and with the same lifetime as the block. as well as data filtering. Threads within a block can cooperate by sharing data through some shared memory and by synchronizing their execution to coordinate memory accesses.2.. although that need not be the case.3. // Kernel invocation dim3 threadsPerBlock(16. B. There are also two additional read-only memory spaces accessible by all threads: the constant and texture memory spaces.3. constant. threadsPerBlock>>>(A. Each thread has private local memory. C). 16).3 Memory Hierarchy CUDA threads may access data from multiple memory spaces during their execution as illustrated by Figure 2-2.1. and 5.2. MatAdd<<<numBlocks.Chapter 2. Section 3. The global. is a common choice. __syncthreads() acts as a barrier at which all threads in the block must wait before any is allowed to proceed.4. the shared memory is expected to be a low-latency memory near each processor core (much like an L1 cache) and __syncthreads() is expected to be lightweight. 10 CUDA Programming Guide Version 3. although arbitrary in this case.4). in parallel or in series. More precisely.

the CUDA programming model assumes that the CUDA threads execute on a physically separate device that operates as a coprocessor to the host running the C program. CUDA Programming Guide Version 3.4 Heterogeneous Programming As illustrated by Figure 2-3. 0) Block (1. 1) Block (0. 1) Grid 1 Block (0. 0) Block (1. 2) Block (1. for example. 0) Block (2. 0) Global memory Block (0.Chapter 2: Programming Model Thread Per-thread local memory Thread Block Per-block shared memory Grid 0 Block (0. 1) Block (1. Memory Hierarchy 2. This is the case.0 11 . when the kernels execute on a GPU and the rest of the C program executes on a CPU. 2) Figure 2-2. 0) Block (0. 1) Block (2. 1) Block (1.

a program manages the global. and texture memory spaces visible to kernels through calls to the CUDA runtime (described in Chapter 3). Therefore. referred to as host memory and device memory. constant.Chapter 2. Programming Model The CUDA programming model also assumes that both the host and the device maintain their own separate memory spaces in DRAM.0 . respectively. 12 CUDA Programming Guide Version 3. This includes device memory allocation and deallocation as well as data transfer between host and device memory.

0 13 . 0) Block (2. 0) Block (1. 1) Block (1. 0) Block (0.Chapter 2: Programming Model C Program Sequential Execution Serial code Host Parallel kernel Kernel0<<<>>>() Device Grid 0 Block (0. 1) Block (2. 1) Block (1. 0) Block (0. 1) Serial code Host Parallel kernel Kernel1<<<>>>() Device Grid 1 Block (0. 2) Serial code executes on the host while parallel code executes on the device. Figure 2-3. Heterogeneous Programming CUDA Programming Guide Version 3. 1) Block (0. 2) Block (1. 0) Block (1.

Prior devices are all of compute capability 1.Chapter 2.0 . 14 CUDA Programming Guide Version 3. The minor revision number corresponds to an incremental improvement to the core architecture. Appendix G gives the technical specifications of each compute capability. Devices with the same major revision number are of the same core architecture.x (Their major revision number is 1). The major revision number of devices based on the Fermi architecture is 2. Appendix A lists of all CUDA-enabled devices along with their compute capability.5 Compute Capability The compute capability of a device is defined by a major revision number and a minor revision number. Programming Model 2. possibly including new features.

page-locked host memory.3 assumes knowledge of these concepts and describes how they are exposed by the driver API.1 Compilation with NVCC Kernels can be written using the CUDA instruction set architecture. CUDA C also supports device emulation. The runtime API is built on top of the CUDA driver API. and module management are all implicit and resulting code is more concise. which facilitates debugging (see Section 3. etc. CUDA C exposes the CUDA programming model as a minimal set of extensions to the C language. 3. CUDA C comes with a runtime API and both the runtime API and the driver API provide functions to allocate and deallocate device memory. Section 3. It is however usually more CUDA Programming Guide Version 3. which is described in the PTX reference manual. An application typically uses either one or the other. CUDA arrays. Section 3. manage systems with multiple devices.1. device enumeration. Any source file that contains some of these extensions must be compiled with nvcc as outlined in Section 3.0 15 . context.2. asynchronous execution. The CUDA driver API is a lower-level C API that provides functions to load kernels as modules of CUDA binary or assembly code. Initialization. shared memory. Programming Interface Two interfaces are currently supported to write CUDA programs: CUDA C and the CUDA driver API. the CUDA driver API requires more code. These extensions allow programmers to define a kernel as a C function and use some new syntax to specify the grid and block dimension each time the function is called.8). In contrast. texture memory.4.Chapter 3. but offers a better level of control and is language-independent since it handles binary or assembly code.2 continues the description of CUDA C started in Chapter 2. transfer data between host memory and device memory. Binary and assembly codes are usually obtained by compiling kernels written in C. to inspect their parameters. It also introduces concepts that are common to both CUDA C and the driver API: linear memory. interoperability with graphics APIs. called PTX. is harder to program and debug. and to launch them. but it can use both under the limitations described in Section 3.

Binary compatibility is guaranteed from one minor revision to the next one.3 PTX Compatibility Some PTX instructions are only supported on devices of higher compute capabilities.2 Binary Compatibility Binary code is architecture-specific.3. code that executes on the host) and device code (i. as detailed in Section 3. In other words.13) into the necessary CUDA C runtime function calls to load and launch each compiled kernel. A complete description can be found in the nvcc user manual. 3.  Or link to the generated host code.y is only guaranteed to execute on devices of compute capability X. but allow applications to benefit from latest compiler improvements. code that executes on the device).1. kernels must be compiled into binary code by nvcc to execute on the device. a cubin object generated for compute capability X.1 (and described in more details in Section B. Applications can then: Either load and execute the PTX code or cubin object on the device using the CUDA driver API (see Section 3.4. 3. Programming Interface effective to use a high-level programming language such as C.Chapter 3. but not from one minor revision to the previous one or across major revisions.z where z≥y. This section gives an overview of nvcc workflow and command options.1. Just-in-time compilation increases application load time. It is also the only way for applications to run on devices that did not exist at the time the application was compiled.e. For example. This is called just-in-time compilation.3) and ignore the generated host code (if any). compiling with –code=sm_13 produces binary code for devices of compute capability 1.1. In both cases. Any PTX code loaded by an application at runtime is compiled further to binary code by the device driver. nvcc‟s basic workflow consists in separating device code from host code and compiling the device code into an assembly form (PTX code) and/or binary form (cubin object). A cubin object is generated using the compiler option –code that specifies the targeted architecture: For example.0 . atomic instructions on global memory are only supported 16 CUDA Programming Guide Version 3.1. nvcc is a compiler driver that simplifies the process of compiling C or PTX code: It provides simple and familiar command line options and executes them by invoking the collection of tools that implement the different compilation stages.  3.1 Compilation Workflow Source files compiled with nvcc can include a mix of host code (i.e. the generated host code includes the PTX code and/or cubin object as a global initialized data array and a translation of the <<<…>>> syntax introduced in Section 2. The generated host code is output either as C code that is left to be compiled using another tool or as object code directly by letting nvcc invoke the host compiler during the last compilation stage.

1 and above. will be: 1.3. nvcc x. Which PTX and binary code gets embedded in a CUDA C application is controlled by the –arch and –code compiler options or the –gencode compiler option as detailed in the nvcc user manual.code=sm_10 –gencode arch=compute_11. For example. “arch=sm_13” is a shorthand for “arch=compute_13 code=compute_13.1. 1. Host code is generated to automatically select at runtime the most appropriate code to load and execute.1 (second -gencode option). for example.2. which are only supported in devices of compute capability 1.0.1 PTX code for devices with compute capabilities 2.4 Application Compatibility To execute code on devices of specific compute capability. PTX code produced for some specific compute capability can always be compiled to binary code of greater or equal compute capability.0 17 . must be compiled with “-arch=sm_13” (or higher compute capability).1. __CUDA_ARCH__ is equal to 110. It is only defined for device code.1 binary code for devices with compute capability 1. 1.cu –gencode arch=compute_10. an application must load binary or PTX code that is compatible with this compute capability as described in Sections 3. code that contains double-precision arithmetic.sm_13” (which is the same as “gencode arch=compute_13. Programming Interface on devices of compute capability 1.1. CUDA Programming Guide Version 3. and gencode compiler options. The __CUDA_ARCH__ macro can be used to differentiate various code paths based on compute capability.sm_11\’ embeds binary code compatible with compute capability 1.1.Chapter 3. otherwise double-precision arithmetic will get demoted to single-precision arithmetic.3 and above.  binary code obtained by compiling 1. in the above example. to be able to execute code on future architectures with higher compute capability – for which no binary code can be generated yet –. For example.  Applications using the driver API must compile code to separate files and explicitly load and execute the most appropriate file at runtime.  1. –code.2 and 3. for example.0 (first –gencode option) and PTX and binary code compatible with compute capability 1. an application must load PTX code that will be compiled just-intime for these devices. So. which.0 binary code for devices with compute capability 1. The –arch compiler option specifies the compute capability that is assumed when compiling C to PTX code.0 or higher.1 and higher. When compiling with “arch=compute_11” for example. x. 3.code=\’compute_11.code=\’compute_13. The nvcc user manual lists various shorthands for the –arch. In particular.3.sm_13\’”). double-precision instructions are only supported on devices of compute capability 1.cu can have an optimized code path that uses atomic operations.

On system with multiple devices. Therefore only runtime functions calls made by the host thread (memory copies. Device memory can be allocated either as linear memory or as CUDA arrays. any resource (memory. Once the runtime has been initialized in a host thread.5 C/C++ Compatibility The front end of the compiler processes CUDA source files according to C++ syntax rules. kernel launches.1) is created under the hood as part of initialization and made current to the host thread. One needs to keep this in mind when timing runtime function calls and when interpreting the error code from the first call into the runtime. void pointers (e.4.3. and copy device memory. Kernels can only operate out of device memory. This section continues with an introduction to the runtime. deallocate. As a consequence of the use of C++ syntax rules. 3. The runtime is implemented in the cudart dynamic library and all its entry points are prefixed with cuda. A complete description of all extensions can be found in Appendix B and a complete description of the runtime in the CUDA reference manual. only a subset of C++ is fully supported for the device code as described in detail in Appendix D.3.0 . etc. However. 3.1 Device Memory As mentioned in Section 2. and it cannot be made current to any other host thread. Full C++ is supported for the host code. kernels are executed on device 0 by default as detailed in Section 3. The core language extensions have been introduced in Chapter 2.2. Programming Interface 3.) allocated via some runtime function call in the host thread is only valid within the context of the host thread. as well as transfer data between host memory and device memory.Chapter 3.. the CUDA programming model assumes a system composed of a host and a device.2. …) can operate on these resources.2 CUDA C CUDA C provides a simple path for users familiar with the C programming language to easily write programs for execution by the device. It consists of a minimal set of extensions to the C language and a runtime library. stream. There is no explicit initialization function for the runtime. returned by malloc()) cannot be assigned to non-void pointers without a typecast. nvcc also support specific keywords and directives detailed in Appendix E. so the runtime provides functions to allocate. event. it initializes the first time a runtime function is called (more specifically any function other than functions from the device and version management sections of the reference manual). 18 CUDA Programming Guide Version 3. each with their own separate memory.g. This is because a CUDA context (see Section 3.1.

cudaMalloc((void**)&d_A.Chapter 3. size). cudaMemcpyDeviceToHost).. size. d_C. float* h_B = (float*)malloc(size). // Copy vectors from host memory to device memory cudaMemcpy(d_A.x and 40-bit address space of devices of compute capability 2. Linear memory is typically allocated using cudaMalloc() and freed using cudaFree() and data transfer between host memory and device memory are typically done using cudaMemcpy(). cudaMemcpyHostToDevice). d_B. int N) { int i = blockDim.4. if (i < N) C[i] = A[i] + B[i]. cudaMalloc((void**)&d_B.x. size_t size = N * sizeof(float). // Allocate vectors in device memory float* d_A. } // Host code int main() { int N = . cudaMalloc((void**)&d_C. d_C.2. // Copy result from device memory to host memory // h_C contains the result in host memory cudaMemcpy(h_C.. float* B.x * blockIdx. // Allocate input vectors h_A and h_B in host memory float* h_A = (float*)malloc(size). float* d_B. int blocksPerGrid = (N + threadsPerBlock – 1) / threadsPerBlock. in a binary tree.. cudaMemcpyHostToDevice). // Free device memory CUDA Programming Guide Version 3. cudaMemcpy(d_B. N).. threadsPerBlock>>>(d_A. size. h_B. Programming Interface CUDA arrays are opaque memory layouts optimized for texture fetching.. float* d_C. so separately allocated entities can reference one another via pointers.0 19 . h_A. In the vector addition code sample of Section 2. for example.x + threadIdx.1. // Initialize input vectors . size). Linear memory exists on the device in a 32-bit address space for devices of compute capability 1. // Invoke kernel int threadsPerBlock = 256. size). float* C.0. the vectors need to be copied from host memory to device memory: // Device code __global__ void VecAdd(float* A. size. VecAdd<<<blocksPerGrid. They are described in Section 3.

size_t pitch = devPitchedPtr. 512>>>(devPitchedPtr. cudaMallocPitch((void**)&devPtr. 20 CUDA Programming Guide Version 3. cudaMalloc3D(&devPitchedPtr.height. Programming Interface cudaFree(d_A). c < width. ++z) { char* slice = devPtr + z * slicePitch. cudaExtent extent) { char* devPtr = devPitchedPtr. for (int c = 0. therefore ensuring best performance when accessing the row addresses or performing copies between 2D arrays and other regions of device memory (using the cudaMemcpy2D() and cudaMemcpy3D() functions). MyKernel<<<100. ++x) { float element = row[x]. width * sizeof(float). // Free host memory .ptr. ++r) { float* row = (float*)((char*)devPtr + r * pitch). cudaFree(d_C). 512>>>(devPtr. extent). The returned pitch (or stride) must be used to access array elements. size_t slicePitch = pitch * extent. &pitch.height. for (int y = 0. cudaExtent extent = make_cudaExtent(64. 64). // Device code __global__ void MyKernel(float* devPtr. y < extent. } } } The following code sample allocates a width×height×depth 3D array of floating-point values and shows how to loop over the array elements in device code: // Host code cudaPitchedPtr devPitchedPtr. The following code sample allocates a width×height 2D array of floating-point values and shows how to loop over the array elements in device code: // Host code float* devPtr. pitch).. 64.width. x < extent. int pitch) { for (int r = 0. z < extent. } Linear memory can also be allocated through cudaMallocPitch() and cudaMalloc3D().3. // Device code __global__ void MyKernel(cudaPitchedPtr devPitchedPtr. extent).1. ++c) { float element = row[c]. r < height.Chapter 3. height). cudaFree(d_B). These functions are recommended for allocations of 2D or 3D arrays as it makes sure that the allocation is appropriately padded to meet the alignment requirements described in Section 5. for (int x = 0.pitch.0 .2. int pitch. for (int z = 0.depth.. MyKernel<<<100. ++y) { float* row = (float*)(slice + y * pitch).

} Matrix. const Matrix B.3. sizeof(data)). 3. Any opportunity to replace global memory accesses by shared memory accesses should therefore be exploited as illustrated by the following matrix multiplication example. Matrix C) { // Load A and B to device memory Matrix d_A.2. and memory allocated for variables declared in global or constant memory space. int height.height times. cudaGetSymbolAddress() is used to retrieve the address pointing to the memory allocated for a variable declared in global memory space. Matrix).0 21 . The following code sample is a straightforward implementation of matrix multiplication that does not take advantage of shared memory.3. float data[256].2 and detailed in Section 5. A is therefore read B.width + col) typedef struct { int width.2.2 Shared Memory As detailed in Section B.elements + row * M.Chapter 3. the following code sample copies some host memory array to constant memory: __constant__ float constData[256]. Each thread reads one row of A and one column of B and computes the corresponding element of C as illustrated in Figure 3-1. linear memory allocated with cudaMallocPitch() or cudaMalloc3D(). data. // Thread block size #define BLOCK_SIZE 16 // Forward declaration of the matrix multiplication kernel __global__ void MatMulKernel(const Matrix. As an example. float* elements. CUDA arrays.Host code // Matrix dimensions are assumed to be multiples of BLOCK_SIZE void MatMul(const Matrix A. Programming Interface } } } } The reference manual lists all the various functions used to copy memory between linear memory allocated with cudaMalloc().2 shared memory is allocated using the __shared__ qualifier. Shared memory is expected to be much faster than global memory as mentioned in Section 2. cudaMemcpyToSymbol(constData. const Matrix. The size of the allocated memory is obtained through cudaGetSymbolSize(). // Matrix multiplication . CUDA Programming Guide Version 3. col) = *(M.width times from global memory and B is read A. // Matrices are stored in row-major order: // M(row.

width = A. size. cudaMemcpyHostToDevice). e < A. dim3 dimGrid(B.elements). // Free device memory cudaFree(d_A.height. // Read C from device memory cudaMemcpy(C.height = A.height * sizeof(float). size).height. d_B.elements. dimBlock>>>(d_A.elements.height.width = C.width = B. Matrix B.width. } 22 CUDA Programming Guide Version 3.width + col] = Cvalue. B.height = C. cudaMemcpyDeviceToHost). BLOCK_SIZE).width. Matrix d_B.x + threadIdx. cudaMalloc((void**)&d_B.height * sizeof(float).y + threadIdx.height = B.height / dimBlock.height * sizeof(float).y).width * C.elements[row * A.elements.width * B. } // Matrix multiplication kernel called by MatMul() __global__ void MatMulKernel(Matrix A. size = C. A. MatMulKernel<<<dimGrid.x. ++e) Cvalue += A. cudaMalloc((void**)&d_A.elements. Matrix C) { // Each thread computes one element of C // by accumulating results into Cvalue float Cvalue = 0.elements). size_t size = A. d_C. size.elements[e * B.width.elements).Chapter 3. d_B. cudaFree(d_C. cudaMemcpy(d_B. d_A.width / dimBlock.width. Cd.width * A.elements. for (int e = 0.width + e] * B. cudaMemcpy(d_A. size).x.0 .elements. size = B. cudaFree(d_B.elements[row * C. int row = blockIdx.elements. A. cudaMemcpyHostToDevice). Programming Interface d_A. d_C). C. d_C. int col = blockIdx. // Allocate C in device memory Matrix d_C.y * blockDim. size. d_B.elements. size).x * blockDim.y. // Invoke kernel dim3 dimBlock(BLOCK_SIZE. cudaMalloc((void**)&d_C.elements.width + col].

A. these two rectangular matrices are divided into as many square matrices of dimension block_size as necessary and Csub is computed as the sum of the products of these square matrices. Programming Interface 0 col B 0 A C row A.height 23 B. Each of these products is performed by first loading the two corresponding square matrices from global memory to shared memory with one thread loading one element of each matrix. Matrix Multiplication without Shared Memory The following code sample is an implementation of matrix multiplication that does take advantage of shared memory. CUDA Programming Guide Version 3.width A. In order to fit into the device‟s resources.Chapter 3.width) that has the same column indices as Csub.height-1 B. and the submatrix of B of dimension (block_size. In this implementation.width Figure 3-1.width. Csub is equal to the product of two rectangular matrices: the sub-matrix of A of dimension (A. block_size) that has the same line indices as Csub.height B. As illustrated in Figure 3-2. and then by having each thread compute one element of the product.0 A. Each thread accumulates the result of each of these products into a register and once done writes the result to global memory. each thread block is responsible for computing one square sub-matrix Csub of C and each thread within the block is responsible for computing one element of Csub.width-1 .

1) are used to get and set elements and build any submatrix from a matrix.width = BLOCK_SIZE. Asub. int col) { return A.elements[row * A. Matrix). // Matrix multiplication .0 . Matrix C) { // Load A and B to device memory Matrix d_A. int height. int row. int row. int col) { Matrix Asub.width / block_size) times from global memory and B is read (A. Asub.stride = A. __device__ functions (see Section B.stride + col]. int stride. const Matrix. // Get a matrix element __device__ float GetElement(const Matrix A. 24 CUDA Programming Guide Version 3. } // Set a matrix element __device__ void SetElement(Matrix A. } // Get the BLOCK_SIZExBLOCK_SIZE sub-matrix Asub of A that is // located col sub-matrices to the right and row sub-matrices down // from the upper-left corner of A __device__ Matrix GetSubMatrix(Matrix A.1. int col. float value) { A. Asub. } Matrix. Asub. return Asub.stride + col] = value.elements[row * A.stride + col) typedef struct { int width.elements = &A. col) = *(M.elements + row * M.height / block_size) times. int row. Programming Interface By blocking the computation this way. float* elements.elements[A.stride. } // Thread block size #define BLOCK_SIZE 16 // Forward declaration of the matrix multiplication kernel __global__ void MatMulKernel(const Matrix.Chapter 3.Host code // Matrix dimensions are assumed to be multiples of BLOCK_SIZE void MatMul(const Matrix A. const Matrix B. The Matrix type from the previous code sample is augmented with a stride field. so that sub-matrices can be efficiently represented with the same type. we take advantage of fast shared memory and save a lot of global memory bandwidth since A is only read (B.height = BLOCK_SIZE.stride * BLOCK_SIZE * row + BLOCK_SIZE * col]. // Matrices are stored in row-major order: // M(row.

// Read C from device memory cudaMemcpy(C. cudaMalloc((void**)&d_C. Matrix C) { // Block row and column int blockRow = blockIdx.height = C. Programming Interface d_A.elements.width * C.height. size = C. m < (A.stride = B. cudaFree(d_C.elements. dimBlock>>>(d_A. MatMulKernel<<<dimGrid. cudaMemcpy(d_B. size. int col = threadIdx.height = B. d_B. size). d_C. d_A.elements. size.Chapter 3. cudaMalloc((void**)&d_A.y.height. size.elements.elements.stride = A.height * sizeof(float).height = A. blockRow.elements. A.x. // Invoke kernel dim3 dimBlock(BLOCK_SIZE. size_t size = A.width * B. Matrix d_B. cudaMemcpyHostToDevice).0 25 . cudaFree(d_B.width / dimBlock.elements. d_C. // Free device memory cudaFree(d_A. Matrix B. size).y. BLOCK_SIZE).width * A. // Each thread block computes one sub-matrix Csub of C Matrix Csub = GetSubMatrix(C. cudaMalloc((void**)&d_B.elements).width = d_C.x.width = d_A. B.width. A. // Thread row and column within Csub int row = threadIdx.width. cudaMemcpyDeviceToHost).stride = C.height * sizeof(float).x.elements). blockCol).elements.height.width.height / dimBlock. int blockCol = blockIdx. cudaMemcpyHostToDevice). d_B.y).height * sizeof(float).elements). cudaMemcpy(d_A. // Loop over all the sub-matrices of A and B that are // required to compute Csub // Multiply each pair of sub-matrices together // and accumulate the results for (int m = 0. d_B. size = B.elements.width = d_B. } // Matrix multiplication kernel called by MatMul() __global__ void MatMulKernel(Matrix A.width / BLOCK_SIZE). d_C). d_C. dim3 dimGrid(B. size). // Allocate C in device memory Matrix d_C. ++m) { CUDA Programming Guide Version 3. // Each thread computes one element of Csub // by accumulating results into Cvalue float Cvalue = 0.

} // Write Csub to device memory // Each thread writes one element SetElement(Csub. blockCol). // Multiply Asub and Bsub together for (int e = 0. // Load Asub and Bsub from device memory to shared memory // Each thread loads one element of each sub-matrix As[row][col] = GetElement(Asub. e < BLOCK_SIZE. Cvalue). // Get sub-matrix Bsub of B Matrix Bsub = GetSubMatrix(B. // Shared memory used to store Asub and Bsub respectively __shared__ float As[BLOCK_SIZE][BLOCK_SIZE]. col). // Synchronize to make sure the sub-matrices are loaded // before starting the computation __syncthreads(). blockRow. Bs[row][col] = GetElement(Bsub. col).0 . m). col. // Synchronize to make sure that the preceding // computation is done before loading two new // sub-matrices of A and B in the next iteration __syncthreads().Chapter 3. row. m. row. } 26 CUDA Programming Guide Version 3. row. __shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE]. ++e) Cvalue += As[row][e] * Bs[e][col]. Programming Interface // Get sub-matrix Asub of A Matrix Asub = GetSubMatrix(A.

These devices can be enumerated. Programming Interface blockCol B BLOCK_SIZE 0 0 A C col Csub BLOCK_SIZE-1 BLOCK_SIZE blockRow BLOCK_SIZE row BLOCK_SIZE-1 BLOCK_SIZE BLOCK_SIZE BLOCK_SIZE B. but by design. Matrix Multiplication with Shared Memory 3. Several host threads can execute device code on the same device. multiple host threads are required to execute device code on multiple devices. cudaGetDeviceCount(&deviceCount). their properties can be queried. int deviceCount. int device. The following code sample enumerates all devices in the system and retrieves their properties.2.width Figure 3-2. for (device = 0. It also determines the number of CUDA-enabled devices. device < deviceCount. if (dev == 0) { CUDA Programming Guide Version 3. As a consequence. cudaGetDeviceProperties(&deviceProp.height 27 B.0 A.width A. any CUDA resources created through the runtime in one host thread cannot be used by the runtime from another host thread. Also.Chapter 3. a host thread can execute device code on only one device at any given time. ++device) { cudaDeviceProp deviceProp.height . and one of them can be selected for kernel executions.3 Multiple Devices A host system can have multiple devices. device).

2 or 4 components that may be signed or unsigned 8-.2. each of which has 1. 16. A texture reference has several attributes.\n").major == 9999 && deviceProp. called a texture. Any subsequent API call reinitializes the runtime.minor == 9999) printf("There is no device supporting CUDA. short for “texture elements. two-dimensional. Elements of the array are called texels.8. 16-bit floats (currently only supported through the driver API).4. After a device has been selected. or three-dimensional and composed of elements. Programming Interface if (deviceProp. Several distinct texture references might be bound to the same texture or to textures that overlap in memory. The first parameter of a texture fetch specifies an object called a texture reference. deviceCount). as well as how the input coordinates are interpreted and what processing should be done. } } By default.6 for exceptions). it must be bound through runtime functions to some region of memory. or a threedimensional array using three texture coordinates. 3. CUDA arrays are opaque memory layouts optimized for texture fetching. Any other device can be selected by calling cudaSetDevice() first.2. They are one-dimensional.3. described in Section B.4 Texture Memory CUDA supports a subset of the texturing hardware that the GPU uses for graphics to access texture memory. As detailed in Section 3. a two-dimensional array using two texture coordinates.” Other attributes define the input and output data types of the texture fetch. One of them is its dimensionality that specifies whether the texture is addressed as a one-dimensional array using one texture coordinate. either implicitly or explicitly. else printf("There are %d devices supporting CUDA\n". 28 CUDA Programming Guide Version 3. the device associated to the host thread is implicitly selected as device 0 as soon as a non-device management runtime function is called (see Section 3. Texture memory is read from kernels using device functions called texture fetches. A texture can be any region of linear memory or a CUDA array. cudaThreadExit() cleans up all runtime-related resources associated with the calling host thread. before it can be used by a kernel. any subsequent explicit call to cudaSetDevice() will fail up until cudaThreadExit() is called. A texture reference defines which part of texture memory is fetched.0 .2.or 32-bit integers. or 32-bit floats. CUDA arrays are only readable by kernels through texture fetching and may only be bound to texture references with the same number of packed components. Reading data from texture memory instead of global memory can have several performance benefits as described in Section 5.Chapter 3.5.3. else if (deviceCount == 1) printf("There is 1 device supporting CUDA\n").

2-. textures are referenced using floating-point coordinates in the range [0.  Dim specifies the dimensionality of the texture reference and is equal to 1.0 are clamped to the range [0. the “wrap” addressing mode also may be specified.0) instead of [0. For normalized coordinates. a texture that is 6432 in size will be referenced with coordinates in the range [0. 1.0. for example. where:  Type specifies the type of data that is returned when fetching the texture. N). 3. Programming Interface 3.0] for signed integer type. When using unnormalized texture coordinates. 1. By default. if it is cudaReadModeNormalizedFloat and Type is a 16-bit or 8-bit integer type. 31] for the x and y dimensions. ReadMode is an optional argument which defaults to cudaReadModeElementType. Dim is an optional argument which defaults to 1.3.0. Dim. 2.0 29 . texture coordinates outside the range [0. A texture reference is declared at file scope as a variable of type texture: texture<Type. Simple linear interpolation is performed for one- CUDA Programming Guide Version 3. 1) in both the x and y dimensions. they are specified when declaring the texture reference.25 is treated the same as 0. Normalized texture coordinates are a natural fit to some applications‟ requirements.2 Runtime Texture Reference Attributes The other attributes of a texture reference are mutable and can be changed at runtime through the host runtime. 1. The addressing mode defines what happens when texture coordinates are out of range. 1. 63] and [0.0). the texels surrounding a texture fetch location are read and the return value of the texture fetch is interpolated based on where the texture coordinates fell between the texels. 1. They specify whether texture coordinates are normalized or not.Chapter 3. Clamping is also the default addressing mode when using normalized texture coordinates: Values below 0. and 4-component vector types defined in Section B. if it is preferable for the texture coordinates to be independent of the texture size.0. respectively. an unsigned 8-bit texture element with the value 0xff reads as 1. if it is cudaReadModeElementType. Normalized texture coordinates cause the coordinates to be specified in the range [0.4.4. the value is actually returned as floating-point type and the full range of the integer type is mapped to [0.0 or above 1.  ReadMode is equal to cudaReadModeNormalizedFloat or cudaReadModeElementType. or 3. so the same 6432 texture would be addressed by normalized coordinates in the range [0.25 and -1. N) are clamped: Values below 0 are set to 0 and values greater or equal to N are set to N-1. For example. Linear texture filtering may be done only for textures that are configured to return floating-point data. no conversion is performed. It performs low-precision interpolation between neighboring texels. the addressing mode.0. When enabled.0] for unsigned integer type and [-1. as detailed below.75. N) where N is the size of the texture in the dimension corresponding to the coordinate. Wrap addressing is usually used when the texture contains a periodic signal. ReadMode> texRef. and texture filtering.25 is treated the same as 0.2. for example.2.1 Texture Reference Declaration Some of the attributes of a texture reference are immutable and must be known at compile time.1. Type is restricted to the basic integer and single-precision floating-point types and any of the 1-. It uses only the fractional part of the texture coordinate.

if it is cudaFilterModeLinear.  filterMode specifies the filtering mode. that is how out-of-range texture coordinates are handled. cudaAddressModeWrap is only supported for normalized texture coordinates. cudaFilterModeLinear is only valid for returned values of floating-point type. all elements in the texture are addressed with texture coordinates in the range [0. }. Appendix F gives more details on texture fetching. and depth are the texture sizes. height. 3. z. if it is cudaFilterModePoint.depth-1] where width. channelDesc is of the following type: struct cudaChannelFormatDesc { int x. or eight (for a three-dimensional texture) texels whose texture coordinates are the closest to the input texture coordinates.2.  addressMode specifies the addressing mode. Programming Interface dimensional textures and bilinear interpolation is performed for two-dimensional textures. respectively. or [0. four (for a two-dimensional texture).4. normalized specifies whether texture coordinates are normalized or not. and third texture coordinates.0 . the addressing mode is equal to either cudaAddressModeClamp. the runtime API has a low-level C-style interface and a high-level C++-style interface. channelDesc. addressMode is an array of size three whose first. and third elements specify the addressing mode for the first. that is how the value returned when fetching the texture is computed based on the input texture coordinates. the returned value is the texel whose texture coordinates are the closest to the input texture coordinates.1] rather than in the range [0. The texture type is defined in the high-level API as a structure publicly derived from the textureReference type defined in the low-level API as such: struct textureReference { int enum cudaTextureFilterMode enum cudaTextureAddressMode struct cudaChannelFormatDesc }  normalized. or cudaAddressModeWrap. second. if it is non-zero. in which case out-of-range texture coordinates are clamped to the valid range. where x.height-1]. y. 30 CUDA Programming Guide Version 3. second. y. filterMode is equal to cudaFilterModePoint or cudaFilterModeLinear. enum cudaChannelFormatKind f. z. addressMode[3].3 Texture Binding As explained in the reference manual.Chapter 3. the returned value is the linear interpolation of the two (for a one-dimensional texture). w. in which case out-of-range texture coordinates are wrapped to the valid range. filterMode. [0.width-1]. and w are equal to the number of bits of each component of the returned value and f is:  cudaChannelFormatKindSigned if these components are of signed integer type.  channelDesc describes the format of the value that is returned when fetching the texture.

cudaReadModeElementType> texRef. Before a kernel can use a texture reference to read from texture memory.  Using the high-level API: texture<float. otherwise. cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc<float>(). cudaChannelFormatDesc channelDesc. 2. “texRef”). cudaBindTextureToArray(texRef. &channelDesc. devPtr. float theta) { // Calculate normalized texture coordinates unsigned int x = blockIdx. 2. cuArray). CUDA Programming Guide Version 3. cudaReadModeElementType> texRef. devPtr. textureReference* texRefPtr. The following code samples bind a texture reference to a CUDA array cuArray:  Using the low-level API: texture<float. pitch). texRefPtr. normalized. cuArray. &channelDesc. 2. “texRef”). height.x + threadIdx.x. cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc<float>(). cuArray). cudaBindTexture2D(0. cudaGetTextureReference(&texRefPtr. cudaUnbindTexture() is used to unbind a texture reference. cudaBindTexture2D(0.  cudaChannelFormatKindFloat if they are of floating point type. 2. The following code sample applies some simple transformation kernel to a // 2D float texture texture<float. width. 2.  Using the high-level API: texture<float. The following code samples bind a texture reference to linear memory pointed to by devPtr:  Using the low-level API: texture<float. addressMode. Programming Interface  cudaChannelFormatKindUnsigned if they are of unsigned integer type. the texture reference must be bound to a texture using cudaBindTexture() or cudaBindTextureToArray(). cudaReadModeElementType> texRef. The format specified when binding a texture to a texture reference must match the parameters specified when declaring the texture reference. textureReference* texRefPtr.Chapter 3. cudaReadModeElementType> texRef. the results of texture fetches are undefined. pitch). height. int width. &channelDesc). and filterMode may be directly modified in host code. texRef. width.x * blockDim.0 31 . cudaGetTextureReference(&texRefPtr. // Simple transformation kernel __global__ void transformKernel(float* output. cudaGetChannelDesc(&channelDesc. cudaReadModeElementType> texRef. int height. cudaBindTextureToArray(texRef.

cudaMemcpyHostToDevice). 0.y. texRef.filterMode = cudaFilterModeLinear. 16). tv).normalized = true.x – 1) / dimBlock. // Free device memory cudaFreeArray(cuArray). channelDesc).addressMode[0] = cudaAddressModeWrap. size. // Transform coordinates u -= 0. // Set texture parameters texRef. width.5f.0 . h_data. 0. height. texRef. // Invoke kernel dim3 dimBlock(16. width * height * sizeof(float)). transformKernel<<<dimGrid. float v = y / (float)height. // Allocate result of transformation in device memory float* output.5f. // Bind the array to the texture cudaBindTextureToArray(texRef.y * blockDim. width. &channelDesc. cudaArray* cuArray.5f. float tv = v * cosf(theta) + u * sinf(theta) + 0.Chapter 3. dim3 dimGrid((width + dimBlock. cudaMallocArray(&cuArray. tu.y + threadIdx. v -= 0.x. cuArray. 0.y). Programming Interface unsigned int y = blockIdx. } 32 CUDA Programming Guide Version 3. float u = x / (float)width. cudaChannelFormatKindFloat).5f. angle).addressMode[1] = cudaAddressModeWrap. cudaFree(output). 0. 0. // Read from texture and write to global memory output[y * width + x] = tex2D(tex. cudaMalloc((void**)&output. height). texRef. // Copy to device memory some data located at address h_data // in host memory cudaMemcpyToArray(cuArray.y – 1) / dimBlock. float tu = u * cosf(theta) – v * sinf(theta) + 0. (height + dimBlock. } // Host code int main() { // Allocate CUDA array in device memory cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc(32. dimBlock>>>(output.

0 33 . making more cache available to the rest of the application.2 Write-Combining Memory By default page-locked host memory is allocated as cacheable. Such a block has therefore two addresses: one in host memory and one in device memory.  On systems with a front-side bus.6. by reducing the amount of physical memory available to the operating system for paging.3 Mapped Memory On some devices. eliminating the need to copy it to or from device memory as detailed in Section 3.5. The host memory pointer is returned by cudaHostAlloc() and the device memory pointer can be retrieved using CUDA Programming Guide Version 3.  On some devices.2. it needs to be allocated by passing flag cudaHostAllocPortable to cudaHostAlloc().2. allocating too much page-locked memory reduces overall system performance. 3.  The simple zero-copy SDK sample comes with a detailed document on the pagelocked memory APIs.5. In addition. which can improve transfer performance by up to 40%. Programming Interface 3.5.5 Page-Locked Host Memory The runtime also provides functions to allocate and free page-locked (also known as pinned) host memory – as opposed to regular pageable host memory allocated by malloc(): cudaHostAlloc() and cudaFreeHost().1 Portable Memory A block of page-locked memory can be used by any host threads. the benefits of using page-locked memory described above are only available for the thread that allocates it. a block of page-locked host memory can also be mapped into the address space of the device by passing flag cudaHostAllocMapped to cudaHostAlloc(). write-combining memory is not snooped during transfers across the PCI Express bus. page-locked host memory can be mapped into the address space of the device. Using page-locked host memory has several benefits: Copies between page-locked host memory and device memory can be performed concurrently with kernel execution for some devices as mentioned in Section 3.2. 3. bandwidth between host memory and device memory is higher if host memory is allocated as page-locked and even higher if in addition it is allocated as write-combining as described in Section 3. Page-locked host memory is a scarce resource however.2.Chapter 3.2.5.2.3. but by default.2. so allocations in pagelocked memory will start failing long before allocations in pageable memory. Reading from write-combining memory from the host is prohibitively slow. To make these advantages available to all threads.2. so write-combining memory should in general be used for memory that the host only writes to. It can optionally be allocated as write-combining instead by passing flag cudaHostAllocWriteCombined to cudaHostAlloc().5. 3. In addition. Write-combining memory frees up L1 and L2 cache resources.

page-locked memory mapping must be enabled by calling cudaSetDeviceFlags() with the cudaDeviceMapHost flag before any other CUDA calls is performed by the thread.6 3. Note that atomic functions (Section 5.  There is no need to use streams (see Section 3. Programmers can globally disable asynchronous kernel launches for all CUDA applications running on a system by setting the CUDA_LAUNCH_BLOCKING  34 CUDA Programming Guide Version 3.5. write-after-read. in which case each host thread that needs to map the block to its device address space must call cudaHostGetDevicePointer() to retrieve a device pointer.1 Asynchronous Concurrent Execution Concurrent Execution between Host and Device In order to facilitate concurrent execution between host and device.  The functions that set memory. as device pointers will generally differ from one host thread to the other.2. or write-afterwrite hazards. 3. cudaHostGetDevicePointer() also returns an error if the device does not support mapped page-locked host memory. Since mapped page-locked memory is shared between host and device however.6.  A block of page-locked host memory can be allocated as both mapped and portable (see Section 3.0 . the application must synchronize memory accesses using streams or events (see Section 3.1). These are: Kernel launches.3B.2. Programming Interface cudaHostGetDevicePointer()and then used to access the block from within a kernel.6.2. the kernel-originated data transfers automatically overlap with kernel execution. data transfers are implicitly performed as needed by the kernel. To be able to retrieve the device pointer to any mapped page-locked memory within a given host thread.10) operating on mapped page-locked memory are not atomic from the point of view of the host or other devices.Chapter 3.2. cudaHostGetDevicePointer() will return an error.6) to avoid any potential read-after-write. Applications may query whether a device supports mapped page-locked host memory or not by calling cudaGetDeviceProperties() and checking the canMapHostMemory property.4.  The functions that perform device  device memory copies. Accessing host memory directly from within a kernel has several advantages: There is no need to allocate a block in device memory and copy data between this block and the block in host memory. some functions are asynchronous: Control is returned to the host thread before the device has completed the requested task.2.  The functions that perform memory copies and are suffixed with Async.4) to overlap data transfers with kernel execution. Otherwise.

cudaStream_t stream[2].6. for (int i = 0. ++i) CUDA Programming Guide Version 3.3 Concurrent Kernel Execution Some devices of compute capability 2.0 can execute multiple kernels concurrently.6. cudaMallocHost((void**)&hostPtr.1 and higher can perform copies between page-locked host memory and device memory concurrently with kernel execution.5 Stream Applications manage concurrency through streams. 3. Applications may query this capability by calling cudaGetDeviceProperties() and checking the deviceOverlap property.0 can perform a copy from page-locked host memory to device memory concurrently with a copy from device memory to pagelocked host memory. 3. 3.g. ++i) cudaStreamCreate(&stream[i]). A stream is a sequence of commands that execute in order.1). The following code sample creates two streams and allocates an array hostPtr of float in page-locked memory. 3.4 Concurrent Data Transfers Devices of compute capability 2. Programming Interface environment variable to 1. may execute their commands out of order with respect to one another or concurrently. 2 * size). on the other hand. Each of these streams is defined by the following code sample as a sequence of one memory copy from host to device. Different streams. and one memory copy from device to host: for (int i = 0. This feature is provided for debugging purposes only and should never be used as a way to make production software run reliably. A stream is defined by creating a stream object and specifying it as the stream parameter to a sequence of kernel launches and host  device memory copies. all launches are synchronous. float* hostPtr.6. i < 2.2 Overlap of Data Transfer and Kernel Execution Some devices of compute capability 1. This capability is currently supported only for memory copies that do not involve CUDA arrays or 2D arrays allocated through cudaMallocPitch() (see Section 3.2.Chapter 3. i < 2. The maximum number of kernel launches that a device can execute concurrently is four.2. one kernel launch.2.6.0 35 . When an application is run via the CUDA debugger or the CUDA profiler. Applications may query this capability by calling cudaGetDeviceProperties() and checking the concurrentKernels property.2. Kernels that use many textures or a large amount of local memory are less likely to execute concurrently with other kernels.2. inter-kernel communication is undefined). this behavior is not guaranteed and should therefore not be relied upon for correctness (e. A kernel from one CUDA context cannot execute concurrently with a kernel from another CUDA context.

i < 2. Synchronization of any kind should be delayed as long as possible. size).0 . i < 2. cudaStreamSynchronize() forces the runtime to wait until all preceding commands in a stream have completed. allowing other streams to continue executing on the device.4. cudaMemcpyDeviceToHost. It can be used to synchronize the host with a specific stream.1 inserts a device-side synchronization barrier for all outstanding kernel launches. Operations that require a dependency check include any other commands within the same stream as the launch being checked and any call to cudaStreamQuery() on that stream. or any CUDA command to stream 0 (including kernel launches and host  device memory copies that do not specify any stream parameter) is called in-between them by the host thread. hostPtr + i * size. Two commands from different streams cannot run concurrently if either a pagelocked host memory allocation. is called in the end to make sure all streams are finished before proceeding further. and copies the result outputDevPtr back to the same portion of hostPtr. stream[i]). these functions are best used for timing purposes or to isolate a launch or memory copy that is failing. To avoid unnecessary slowdowns. for (int i = 0. cudaThreadSynchronize(). stream[i]>>> (outputDevPtr + i * size. inputDevPtr + i * size. Therefore. ++i) cudaStreamDestroy(stream[i]). Each stream copies its portion of input array hostPtr to array inputDevPtr in device memory. for (int i = 0. 0. a device memory allocation. It forces the runtime to wait until all preceding device tasks in all streams have completed. cudaStreamDestroy() waits for all preceding tasks in the given stream to complete before destroying the stream and returning control to the host thread. size. cudaMemcpyHostToDevice. a device memory set.   36 CUDA Programming Guide Version 3. Programming Interface cudaMemcpyAsync(inputDevPtr + i * size. Processing hostPtr using two streams allows for the memory copies of one stream to overlap with the kernel execution of the other stream.Chapter 3. Any operation that requires a dependency check to see if a streamed kernel launch is complete blocks all later kernel launches from any stream in the CUDA context until the launch being checked is complete. i < 2. hostPtr must point to page-locked host memory for any overlap to occur. Switching between the L1/shared memory configurations described in Section G. ++i) MyKernel<<<100. applications should follow these guidelines to improve their potential for concurrent kernel execution: All independent operations should be issued before dependent operations. stream[i]). size. for (int i = 0. cudaStreamQuery() provides applications with a way to know if all preceding commands in a stream have completed. outputDevPtr + i * size. 512. ++i) cudaMemcpyAsync(hostPtr + i * size. processes inputDevPtr on the device by calling MyKernel(). a device  device memory copy. cudaThreadSynchronize() Streams are released by calling cudaStreamDestroy().

An event is recorded when all tasks – or optionally. inputHost + i * size.2. cudaMemcpyDeviceToHost. block. ++i) cudaMemcpyAsync(inputDev + i * size. 0).2.7.7 Synchronous Calls When a synchronous function is called. by letting the application asynchronously record events at any point in the program and query when these events are actually recorded. 3.6. size. stream[i]). These events can be used to time the code sample of the previous section the following way: cudaEventRecord(start. size. cudaMemcpyHostToDevice. as well as perform accurate timing. either to enable CUDA to read data written by OpenGL or Direct3D.0 37 . 0. Whether the host thread will then yield.2. A resource must be registered to CUDA before it can be mapped using the functions mentioned in Sections 3. cudaEventSynchronize(stop). Registering a resource is potentially high-overhead and therefore typically called only CUDA Programming Guide Version 3. cudaEventCreate(&start).2. 0).7. inputDev + i * size. cudaEventCreate(&stop). stop). Events in stream zero are recorded after all preceding tasks/commands from all streams are completed by the device.Chapter 3. size). or to enable CUDA to write data for consumption by OpenGL or Direct3D. or spin can be specified by calling cudaSetDeviceFlags()with some specific flags (see reference manual for details) before any other CUDA calls is performed by the host thread. all commands in a given stream – preceding the event have completed.6. i < 2. for (int i = 0. 3. cudaEventElapsedTime(&elapsedTime. ++i) cudaMemcpyAsync(outputHost + i * size. ++i) MyKernel<<<100. for (int i = 0. control is not returned to the host thread before the device has completed the requested task. outputDev + i * size. cudaEventDestroy(stop). The following code sample creates two events: cudaEvent_t start. They are destroyed this way: cudaEventDestroy(start). 512.7 Graphics Interoperability Some resources from OpenGL and Direct3D may be mapped into the address space of CUDA. cudaEventRecord(stop. stop. float elapsedTime. stream[i]).2. Programming Interface 3.1 and 3. start.2. stream[i]>>> (outputDev + i * size. i < 2. for (int i = 0. These functions return a pointer to a CUDA graphics resource of type struct cudaGraphicsResource. i < 2.6 Event The runtime also provides a way to closely monitor the device‟s progress.

A mapped resource can be read from or written to by kernels using the device memory address returned by cudaGraphicsResourceGetMappedPointer() for buffers and cudaGraphicsSubResourceGetMappedArray() for CUDA arrays. read-only) that the CUDA driver can use to optimize resource management. Accessing a resource through OpenGL or Direct3D while it is mapped to CUDA produces undefined results. 3. // Initialize OpenGL and GLUT .. A CUDA graphics resource is unregistered using cudaGraphicsUnregisterResource().7. cudaGraphicsResourceSetMapFlags() can be called to specify usage hints (write-only.. Note that cudaSetDevice()and cudaGLSetGLDevice() are mutually exclusive.2. 38 CUDA Programming Guide Version 3.0 texture format. it appears as a device pointer and can therefore be read and written by kernels or via cudaMemcpy() calls. A buffer object is registered using cudaGraphicsGLRegisterBuffer(). The following code sample uses a kernel to dynamically modify a 2D width x height grid of vertices stored in a vertex buffer object: GLuint positionsVBO. it can be mapped and unmapped as many times as necessary using cudaGraphicsMapResources() and cudaGraphicsUnmapResources().1 OpenGL Interoperability Interoperability with OpenGL requires that the CUDA device be specified by cudaGLSetGLDevice() before any other runtime calls.Chapter 3. Please note that since GL_RGBA8UI is an OpenGL 3. It does not currently support normalized integer formats (e. Sections 3. glutDisplayFunc(display). int main() { // Explicitly set device cudaGLSetGLDevice(0). The OpenGL resources that may be mapped into the address space of CUDA are OpenGL buffer. Once a resource is registered to CUDA. and renderbuffer objects. Programming Interface once per resource.7. In CUDA.g. it appears as a CUDA array and can therefore be bound to a texture reference and be read and written by kernels or via cudaMemcpy2D() calls. A texture or renderbuffer object is registered using cudaGraphicsGLRegisterImage(). In CUDA. not the fixed function pipeline. cudaGraphicsGLRegisterImage() supports all texture formats with a internal type of float (e. GL_RGBA8UI).2.0 . struct cudaGraphicsResource* positionsVBO_CUDA. it can only be written by shaders.2. texture.g.7.1 and 3.g. GL_RGBA8).2 give specifics for each graphics API and some code samples. GL_RGBA_FLOAT32) and unnormalized integer (e.

0). &positionsVBO_CUDA. // Render from buffer object glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT).x. &vbo). &num_bytes. time. 1). glDeleteBuffers(1. Programming Interface // Create buffer object and register it with CUDA glGenBuffers(1. cudaGraphicsGLRegisterBuffer(&positionsVBO_CUDA. // Unmap buffer object cudaGraphicsUnmapResources(1.x + threadIdx. &positionsVBO). width * height). dimBlock>>>(positions. positionsVBO. glDrawArrays(GL_POINTS. createVertices<<<dimGrid. GL_DYNAMIC_DRAW). 0).Chapter 3. height / dimBlock. 0). cudaGraphicsMapResources(1.x.0 39 . } void deleteVBO() { cudaGraphicsUnregisterResource(positionsVBO_CUDA). cudaGraphicsMapFlagsWriteDiscard). } __global__ void createVertices(float4* positions. 1). // Execute kernel dim3 dimBlock(16. glVertexPointer(4. cudaGraphicsResourceGetMappedPointer((void**)&positions. } void display() { // Map buffer object for writing from CUDA float4* positions. positionsVBO). float time. glBufferData(GL_ARRAY_BUFFER. 0. glBindBuffer(GL_ARRAY_BUFFER. GL_FLOAT.x * blockDim. &positionsVBO_CUDA. unsigned int size = width * height * 4 * sizeof(float). 0. unsigned int width. width. // Swap buffers glutSwapBuffers(). CUDA Programming Guide Version 3. dim3 dimGrid(width / dimBlock. glBindBuffer(GL_ARRAY_BUFFER. positionsVBO). // Launch rendering loop glutMainLoop(). glEnableClientState(GL_VERTEX_ARRAY). height). glBindBuffer(GL_ARRAY_BUFFER. unsigned int height) { unsigned int x = blockIdx. positionsVBO_CUDA)).y. size_t num_bytes. 0). size. glutPostRedisplay(). 16. 0. glDisableClientState(GL_VERTEX_ARRAY).

and Direct3D 11.1. y. } On Windows and for Quadro GPUs. Programming Interface unsigned int y = blockIdx. 3. v = v * 2.2. cudaD3D9GetDevice().0f.0f . cudaWGLGetDevice() can be used to retrieve the CUDA device associated to the handle returned by wglEnumGpusNV().0f .Chapter 3. w. cudaD3D10GetDevice(). float v = y / (float)height. u = u * 2. DWORD color. cudaGraphicsD3D10RegisterResource(). // Write positions positions[y * width + x] = make_float4(u. Direct3D 10. and cudaD3D11GetDevice() can be used to retrieve the CUDA device associated to some adapter.y + threadIdx. The Direct3D resources that may be mapped into the address space of CUDA are Direct3D buffers. struct CUSTOMVERTEX { FLOAT x.y * blockDim. z.0f. before any other runtime calls. 40 CUDA Programming Guide Version 3. // Calculate uv coordinates float u = x / (float)width. IDirect3DDevice9* device. v.0 . cudaD3D10SetDirect3DDevice() and cudaD3D11SetDirect3DDevice().0f). textures. Interoperability with Direct3D requires that the Direct3D device be specified by cudaD3D9SetDirect3DDevice(). Quadro GPUs offer higher performance OpenGL interoperability than GeForce and Tesla GPUs in a multi-GPU configuration where OpenGL rendering is performed on the Quadro GPU and CUDA computations are performed on other GPUs in the system. and cudaGraphicsD3D11RegisterResource(). A CUDA context may interoperate with only one Direct3D device at a time and the CUDA context and Direct3D device must be created on the same GPU. Direct3D 9 Version: IDirect3D9* D3D.1.0f. and surfaces. The following code sample uses a kernel to dynamically modify a 2D width x height grid of vertices stored in a vertex buffer object.5f. the Direct3D device must be created with the D3DCREATE_HARDWARE_VERTEXPROCESSING flag. 1. Moreover. These resources are registered using cudaGraphicsD3D9RegisterResource(). }.7. float w = sinf(u * freq + time) * cosf(v * freq + time) * 0. // calculate simple sine wave pattern float freq = 4.y.2 Direct3D Interoperability Direct3D interoperability is supported for Direct3D 9.

0).. D3DPOOL_DEFAULT.. cudaGraphicsResourceSetMapFlags(positionsVB_CUDA. . cudaGraphicsMapResources(1. // Get a CUDA-enabled adapter unsigned int adapter = 0.. &adapterId).DeviceName) == cudaSuccess) break. 0. positionsVB. D3D->CreateDevice(adapter. hWnd. positionsVB_CUDA)). device->CreateVertexBuffer(size.. 0). cudaGraphicsResourceGetMappedPointer((void**)&positions. // Register device with CUDA cudaD3D9SetDirect3DDevice(device). &device).. adapterId. adapter < g_pD3D->GetAdapterCount(). &params. if (cudaD3D9GetDevice(&dev. int dev. &positionsVB_CUDA. Programming Interface IDirect3DVertexBuffer9* positionsVB. cudaGraphicsRegisterFlagsNone).. } } void Render() { // Map vertex buffer for writing from CUDA float4* positions. g_pD3D->GetAdapterIdentifier(adapter. // Execute kernel CUDA Programming Guide Version 3. } // Create device . D3DCREATE_HARDWARE_VERTEXPROCESSING.. adapter++) { D3DADAPTER_IDENTIFIER9 adapterId. D3DDEVTYPE_HAL. D3DFVF_CUSTOMVERTEX. cudaGraphicsMapFlagsWriteDiscard). cudaGraphicsD3D9RegisterResource(&positionsVB_CUDA.0 41 . size_t num_bytes. &positionsVB. // Create vertex buffer and register it with CUDA unsigned int size = width * height * sizeof(CUSTOMVERTEX). &num_bytes.) { . struct cudaGraphicsResource* positionsVB_CUDA. for (. // Launch rendering loop while (.Chapter 3. 0.. int main() { // Initialize Direct3D D3D = Direct3DCreate9(D3D_SDK_VERSION). Render().

dimBlock>>>(positions. } Direct3D 10 Version: ID3D10Device* device.0 . __int_as_float(0xff00ff00)). 1). unsigned int height) { unsigned int x = blockIdx. w.. // Calculate uv coordinates float u = x / (float)width. float time.0f.y.0f .x + threadIdx. (void**)&factory). width. } __global__ void createVertices(float4* positions. dim3 dimGrid(width / dimBlock. height). ++i) { if (FAILED(factory->EnumAdapters(i. u = u * 2.y * blockDim. unsigned int y = blockIdx.1.0f . v. positionsVB->Release(). int main() { // Get a CUDA-enabled adapter IDXGIFactory* factory. // Unmap vertex buffer cudaGraphicsUnmapResources(1.. 1).0f. IDXGIAdapter* adapter = 0.x. struct cudaGraphicsResource* positionsVB_CUDA. // Calculate simple sine wave pattern float freq = 4. struct CUSTOMVERTEX { FLOAT x. unsigned int width.y. y. ID3D10Buffer* positionsVB. for (unsigned int i = 0. // Draw and present .0f. &adapter)) 42 CUDA Programming Guide Version 3. !adapter. } void releaseVB() { cudaGraphicsUnregisterResource(positionsVB_CUDA).1. time. }. createVertices<<<dimGrid. DWORD color. float w = sinf(u * freq + time) * cosf(v * freq + time) * 0. float v = y / (float)height.x. v = v * 2. 0).5f.Chapter 3. Programming Interface dim3 dimBlock(16. CreateDXGIFactory(__uuidof(IDXGIFactory). &positionsVB_CUDA.y + threadIdx. 16. z.x * blockDim. // Write positions positions[y * width + x] = make_float4(u. height / dimBlock.

if (cudaD3D10GetDevice(&dev.MiscFlags = 0. &num_bytes. // Register device with CUDA cudaD3D10SetDirect3DDevice(device)... 1). cudaGraphicsMapFlagsWriteDiscard). cudaGraphicsResourceSetMapFlags(positionsVB_CUDA.BindFlags = D3D10_BIND_VERTEX_BUFFER. &positionsVB). height / dimBlock. positionsVB_CUDA)). 1). 16.y. cudaGraphicsMapResources(1. D3D10CreateDeviceAndSwapChain(adapter. D3D10_BUFFER_DESC bufferDesc. &swapChain. bufferDesc. cudaGraphicsRegisterFlagsNone). cudaGraphicsD3D10RegisterResource(&positionsVB_CUDA. dim3 dimGrid(width / dimBlock. bufferDesc. device->CreateBuffer(&bufferDesc..Usage = D3D10_USAGE_DEFAULT. // Launch rendering loop while (. CUDA Programming Guide Version 3.. 0. // Create vertex buffer and register it with CUDA unsigned int size = width * height * sizeof(CUSTOMVERTEX). D3D10_SDK_VERSION.CPUAccessFlags = 0. int dev. } factory->Release(). D3D10_CREATE_DEVICE_DEBUG.. // Execute kernel dim3 dimBlock(16.ByteWidth = size. bufferDesc.. adapter->Release(). &positionsVB_CUDA. } } void Render() { // Map vertex buffer for writing from CUDA float4* positions.x.Chapter 3. size_t num_bytes. &device). adapter) == cudaSuccess) break.0 43 . .. D3D10_DRIVER_TYPE_HARDWARE. adapter->Release(). cudaGraphicsResourceGetMappedPointer((void**)&positions. Render(). // Create swap chain and device . Programming Interface break. bufferDesc..) { . 0. positionsVB. 0). bufferDesc. &swapChainDesc.

0f . ID3D11Buffer* positionsVB. unsigned int height) { unsigned int x = blockIdx. ++i) { if (FAILED(factory->EnumAdapters(i.1. 0).0 .1.y. &positionsVB_CUDA. height). v = v * 2.x * blockDim.5f. v. u = u * 2.0f . !adapter.0f. // Calculate simple sine wave pattern float freq = 4.y * blockDim. IDXGIAdapter* adapter = 0. width. CreateDXGIFactory(__uuidof(IDXGIFactory). w. struct cudaGraphicsResource* positionsVB_CUDA. int main() { // Get a CUDA-enabled adapter IDXGIFactory* factory. z..y + threadIdx. y. // Write positions positions[y * width + x] = make_float4(u.Chapter 3. positionsVB->Release(). float v = y / (float)height. unsigned int width. float time. struct CUSTOMVERTEX { FLOAT x. dimBlock>>>(positions.x. } Direct3D 11 Version: ID3D11Device* device. __int_as_float(0xff00ff00)). 44 CUDA Programming Guide Version 3. // Calculate uv coordinates float u = x / (float)width. unsigned int y = blockIdx. Programming Interface createVertices<<<dimGrid. }. (void**)&factory). float w = sinf(u * freq + time) * cosf(v * freq + time) * 0. time. DWORD color.. } __global__ void createVertices(float4* positions. &adapter)) break.0f.x + threadIdx. } void releaseVB() { cudaGraphicsUnregisterResource(positionsVB_CUDA). // Draw and present . // Unmap vertex buffer cudaGraphicsUnmapResources(1. for (unsigned int i = 0.0f. int dev.

&positionsVB_CUDA. &deviceContext).. cudaGraphicsResourceGetMappedPointer((void**)&positions. cudaGraphicsRegisterFlagsNone)... // Create vertex buffer and register it with CUDA unsigned int size = width * height * sizeof(CUSTOMVERTEX).0 45 . bufferDesc. D3D11_DRIVER_TYPE_HARDWARE. bufferDesc. D3D11_SDK_VERSION..CPUAccessFlags = 0. D3D11_CREATE_DEVICE_DEBUG. cudaGraphicsMapFlagsWriteDiscard). size_t num_bytes. 0.MiscFlags = 0. &swapChainDesc. Render(). // Register device with CUDA cudaD3D11SetDirect3DDevice(device). D3D11_BUFFER_DESC bufferDesc.ByteWidth = size. . 0. // Create swap chain and device . cudaGraphicsMapResources(1. adapter->Release(). adapter->Release(). &device. // Launch rendering loop while (. Programming Interface if (cudaD3D11GetDevice(&dev..BindFlags = D3D11_BIND_VERTEX_BUFFER. 0). sFnPtr_D3D11CreateDeviceAndSwapChain(adapter. bufferDesc. adapter) == cudaSuccess) break. } } void Render() { // Map vertex buffer for writing from CUDA float4* positions.Chapter 3. 3. &swapChain.Usage = D3D11_USAGE_DEFAULT.) { . &positionsVB). featureLevels. device->CreateBuffer(&bufferDesc. &num_bytes.. cudaGraphicsD3D11RegisterResource(&positionsVB_CUDA. // Execute kernel CUDA Programming Guide Version 3. bufferDesc.. } factory->Release(). bufferDesc. &featureLevel.. positionsVB_CUDA)). positionsVB. cudaGraphicsResourceSetMapFlags(positionsVB_CUDA.

float v = y / (float)height. &positionsVB_CUDA. float time. 1).y + threadIdx. Programming Interface dim3 dimBlock(16. // Draw and present . width. dim3 dimGrid(width / dimBlock.8 Error Handling All runtime functions return an error code.x * blockDim. the error code only reports errors that occur on the host prior to executing the task.6).0f..x. 0). it will be reported by some subsequent unrelated runtime function call.5f. __int_as_float(0xff00ff00)).0f . u = u * 2.. float w = sinf(u * freq + time) * cosf(v * freq + time) * 0.2.y * blockDim. positionsVB->Release(). } void releaseVB() { cudaGraphicsUnregisterResource(positionsVB_CUDA).1.6) and checking the error code returned by cudaThreadSynchronize().0f.x + threadIdx. dimBlock>>>(positions.Chapter 3.1.2. // Calculate simple sine wave pattern float freq = 4. unsigned int width. time. // Write positions positions[y * width + x] = make_float4(u. if an asynchronous error occurs. createVertices<<<dimGrid.y. v. 46 CUDA Programming Guide Version 3. unsigned int height) { unsigned int x = blockIdx. The only way to check for asynchronous errors just after some asynchronous function call is therefore to synchronize just after the call by calling cudaThreadSynchronize() (or by using any other synchronization mechanisms described in Section 3. but for an asynchronous function (see Section 3.x. } 3. 16.0 .0f . // Calculate uv coordinates float u = x / (float)width. v = v * 2. 1). height / dimBlock. } __global__ void createVertices(float4* positions. w.2.y. this error code cannot possibly report any of the asynchronous errors that could occur on the device since the function returns before the device has completed the task.0f. // Unmap vertex buffer cudaGraphicsUnmapResources(1. typically related to parameter validation. height). unsigned int y = blockIdx.

Kernel launches do not return any error code. knowing that each thread gets 256 KB of stack.Chapter 3. 3. For each thread in a thread block. All code for an application. the application must synchronize in-between the kernel launch and the call to cudaGetLastError().0 47 . The programmer needs to make sure that: The host is able to run up to the maximum number of threads per block. must be compiled consistently either for device emulation or for device execution. plus one for the master thread. Kernel launches are asynchronous. like setting breakpoints and inspecting data.  Since device code is compiled to run on the host. any device or host function can be called from either device or host code. similarly.9 Debugging using the Device Emulation Mode CUDA-GDB can be used to debug devices of compute capability greater than 1. Programming Interface The runtime maintains an error variable for each host thread that is initialized to cudaSuccess and is overwritten by the error code every time an error occurs (be it a parameter validation error or an asynchronous error). the device code is compiled for and runs on the host.). the runtime creates a thread on the host.0 (see the CUDA-GDB user manual for supported platforms). including any libraries used. one has to make sure that the runtime error variable is set to cudaSuccess just before the kernel launch.or host-specific data can be read from either device or host code. like input and output operations to files or to the screen (printf(). When running an application in device emulation mode. so to check for asynchronous errors. Many features provided through the device emulation mode make it a very effective debugging tool:  By using the host‟s native debugging support programmers can use all features that the debugger supports. Note that cudaErrorNotReady that may be returned by cudaStreamQuery() and cudaEventQuery() is not considered an error and is therefore not reported by cudaGetLastError(). Linking code compiled for device emulation with code compiled for device execution causes the following runtime error to be returned upon initialization: cudaErrorMixedDeviceExecution. by calling cudaGetLastError() just before the kernel launch. the programming model is emulated by the runtime. the code can be augmented with code that cannot run on the device. When compiling an application in this mode (using the -deviceemu option). allowing the programmer to use the host‟s native debugging support to debug the application as if it were a host application. The compiler and runtime also supports an emulation mode for the purpose of debugging that can be used even in the absence of any CUDA-enabled device. so cudaGetLastError() must be called just after the kernel launch to retrieve any pre-launch errors. etc. To ensure that any error returned by cudaGetLastError() does not originate from calls prior to the kernel launch. cudaGetLastError() returns this variable and resets it to cudaSuccess.  Since all data resides on the host. for example. any device.  Enough memory is available to run all threads.2. The preprocessor macro __DEVICE_EMULATION__ is defined in this mode.  CUDA Programming Guide Version 3.

none of which is guaranteed to work:  Declare some floating-point variables as volatile to force single-precision storage. _FPU_GETCW(originalCW). to store the current value of the control word and change it to force the mantissa to be stored in 24 bits using. Programming Interface In case of incorrect usage of the synchronization intrinsic function. or unsigned int originalCW = _controlfp(0. _controlfp(_PC_24. or _controlfp(originalCW.  Use _FPU_GETCW() and _FPU_SETCW() on Linux or _controlfp() on Windows to force single-precision floating-point computation for a portion of the code by surrounding it with  unsigned int originalCW.  Most of the time the same floating-point computation will not produce exactly the same result when performed on the device as when performed on the host in device emulation mode. all you need to get different results for the same floating-point computation are slightly different compiler options.  Use the /Op or /fp compiler options of the Visual C++ compiler.x do not support denormalized numbers (see Appendix G) as host 48 CUDA Programming Guide Version 3. since the number of threads executing simultaneously is much smaller than on an actual device. let alone different compilers. potentially resulting in significant differences in accuracy when running in device emulation mode. at the beginning. some host platforms store intermediate results of single-precision floating-point calculations in extended precision registers. Therefore. device emulation mode is very useful in finding algorithmic errors. programmers can try any of the following methods. whereas device emulation can produce correct results. the runtime detects dead lock situations. 0). unsigned int cw = (originalCW & ~0x300) | 0x000. When this occurs. _MCW_PC). not simulating it.  Use the –ffloat-store compiler option of gcc.  When dereferencing a pointer to global memory on the host or a pointer to host memory on the device. and with _FPU_SETCW(originalCW). different instruction sets. Also. 0xfffff). at the end. or different architectures. _FPU_SETCW(cw). to restore the original control word.Chapter 3. device execution almost certainly fails in some undefined way. Programmers must keep in mind that device emulation mode is emulating the device. for single-precision floating-point numbers. but certain errors are hard to find:  Race conditions are less likely to manifest themselves in device-emulation mode.0 . This is expected since in general. compute devices of compute capability 1. In particular.

imperative API: Most objects are referenced by opaque handles that may be specified to functions to manipulate the objects.3 Driver API The driver API is a handle-based. Kernels are launched using API entry points as described in Section 3.3. Table 3-1. Therefore.1. This is because binary code is architecture-specific and therefore incompatible with future architectures. This can lead to dramatically different results between device emulation and device execution modes since some computation might produce a finite result in one case and an infinite result in the other. Within a CUDA context. Any application that wants to run on future device architectures must load PTX. not binary code.11) produce different results than in device execution mode. the warp vote functions (described in Section B.  The warp size is equal to 1 in device emulation mode (see Section 4. size_t size = N * sizeof(float). Kernels written in C must therefore be compiled separately into PTX or binary objects.0 49 . A CUDA context must then be created that is attached to a specific device and made current to the calling host thread as detailed in Section 3..3. Objects Available in the CUDA Driver API Object Device Context Module Function Heap memory CUDA array Texture reference Handle CUdevice CUcontext CUmodule CUfunction CUdeviceptr CUarray CUtexref Description CUDA-enabled device Roughly equivalent to a CPU process Roughly equivalent to a dynamic library Kernel Pointer to device memory Opaque container for one-dimensional or two-dimensional data on the device. The driver API must be initialized with cuInit() before any function from the driver API is called. The objects available in the driver API are summarized in Table 3-1.. kernels are explicitly loaded as PTX or binary objects by the host code as described in Section 3. 3. Here is the host code of the sample from Section 2.3.1 written using the driver API: int main() { int N = . CUDA Programming Guide Version 3.1 for the definition of a warp). // Allocate input vectors h_A and h_B in host memory float* h_A = (float*)malloc(size).2. Programming Interface platforms usually do.3.. readable via texture references Object that describes how to interpret texture memory data The driver API is implemented in the nvcuda dynamic library and all its entry points are prefixed with cu. whereas PTX code is compiled to binary code at load time by the driver.Chapter 3.

\n"). cuParamSetv(vecAdd. cuModuleGetFunction(&vecAdd. 50 CUDA Programming Guide Version 3. "VecAdd"). cuMemAlloc(&d_C. Programming Interface float* h_B = (float*)malloc(size). offset. ptr = (void*)(size_t)d_B. size). // Create context CUcontext cuContext. &ptr. &ptr. // Create module from binary file CUmodule cuModule. ptr = (void*)(size_t)d_A. // Initialize cuInit(0). h_A. cuParamSetv(vecAdd. cuModuleLoad(&cuModule. 0. // Invoke kernel #define ALIGN_UP(offset. offset += sizeof(ptr). cuMemAlloc(&d_B. ALIGN_UP(offset.. ALIGN_UP(offset.Chapter 3. h_B. __alignof(ptr)). size). } // Get handle for device 0 CUdevice cuDevice. if (deviceCount == 0) { printf("There is no device supporting CUDA. exit (0). void* ptr. cuModule. size). cuDeviceGetCount(&deviceCount). cuCtxCreate(&cuContext. // Get number of devices supporting CUDA int deviceCount = 0. CUdeviceptr d_B. sizeof(ptr)). // Allocate vectors in device memory CUdeviceptr d_A. // Copy vectors from host memory to device memory cuMemcpyHtoD(d_A. offset. __alignof(ptr)).. cuMemcpyHtoD(d_B. sizeof(ptr)).0 . alignment) \ (offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1) int offset = 0. cuDevice). // Get function handle from module CUfunction vecAdd. CUdeviceptr d_C. 0). offset += sizeof(ptr). size). cuMemAlloc(&d_A. size). cuDeviceGet(&cuDevice. “VecAdd.ptx”). // Initialize input vectors .

Chapter 3. Programming Interface

ptr = (void*)(size_t)d_C; ALIGN_UP(offset, __alignof(ptr)); cuParamSetv(vecAdd, offset, &ptr, sizeof(ptr)); offset += sizeof(ptr); ALIGN_UP(offset, __alignof(N)); cuParamSeti(vecAdd, offset, N); offset += sizeof(N); cuParamSetSize(vecAdd, offset); int threadsPerBlock = 256; int blocksPerGrid = (N + threadsPerBlock – 1) / threadsPerBlock; cuFuncSetBlockShape(vecAdd, threadsPerBlock, 1, 1); cuLaunchGrid(vecAdd, blocksPerGrid, 1); ... }

Full code can be found in the vectorAddDrv SDK code sample.

3.3.1

Context
A CUDA context is analogous to a CPU process. All resources and actions performed within the driver API are encapsulated inside a CUDA context, and the system automatically cleans up these resources when the context is destroyed. Besides objects such as modules and texture references, each context has its own distinct 32-bit address space. As a result, CUdeviceptr values from different contexts reference different memory locations. A host thread may have only one device context current at a time. When a context is created with cuCtxCreate(), it is made current to the calling host thread. CUDA functions that operate in a context (most functions that do not involve device enumeration or context management) will return CUDA_ERROR_INVALID_CONTEXT if a valid context is not current to the thread. Each host thread has a stack of current contexts. cuCtxCreate() pushes the new context onto the top of the stack. cuCtxPopCurrent() may be called to detach the context from the host thread. The context is then "floating" and may be pushed as the current context for any host thread. cuCtxPopCurrent() also restores the previous current context, if any. A usage count is also maintained for each context. cuCtxCreate() creates a context with a usage count of 1. cuCtxAttach() increments the usage count and cuCtxDetach() decrements it. A context is destroyed when the usage count goes to 0 when calling cuCtxDetach() or cuCtxDestroy(). Usage count facilitates interoperability between third party authored code operating in the same context. For example, if three libraries are loaded to use the same context, each library would call cuCtxAttach() to increment the usage count and cuCtxDetach() to decrement the usage count when the library is done using the context. For most libraries, it is expected that the application will have created a context before loading or initializing the library; that way, the application can create the context using its own heuristics, and the library simply operates on the context handed to it. Libraries that wish to create their own contexts – unbeknownst to their API clients who may or may not have created contexts of their own – would use cuCtxPushCurrent() and cuCtxPopCurrent() as illustrated in Figure 3-3.

CUDA Programming Guide Version 3.0

51

Chapter 3. Programming Interface

Library Initialization Call
cuCtxCreate() Initialize context cuCtxPopCurrent()

Library Call
cuCtxPushCurrent() Use context cuCtxPopCurrent()

Figure 3-3. Library Context Management

3.3.2

Module
Modules are dynamically loadable packages of device code and data, akin to DLLs in Windows, that are output by nvcc (see Section 3.1). The names for all symbols, including functions, global variables, and texture references, are maintained at module scope so that modules written by independent third parties may interoperate in the same CUDA context. This code sample loads a module and retrieves a handle to some kernel:
CUmodule cuModule; cuModuleLoad(&cuModule, “myModule.ptx”); CUfunction myKernel; cuModuleGetFunction(&myKernel, cuModule, “MyKernel”);

This code sample compiles and loads a new module from PTX code and parses compilation errors:
#define ERROR_BUFFER_SIZE 100 CUmodule cuModule; CUptxas_option options[3]; void* values[3]; char* PTXCode = “some PTX code”; options[0] = CU_ASM_ERROR_LOG_BUFFER; values[0] = (void*)malloc(ERROR_BUFFER_SIZE); options[1] = CU_ASM_ERROR_LOG_BUFFER_SIZE_BYTES; values[1] = (void*)ERROR_BUFFER_SIZE; options[2] = CU_ASM_TARGET_FROM_CUCONTEXT; values[2] = 0; cuModuleLoadDataEx(&cuModule, PTXCode, 3, options, values); for (int i = 0; i < values[1]; ++i) { // Parse error string here }

3.3.3

Kernel Execution
cuFuncSetBlockShape() sets the number of threads per block for a given

function, and how their threadIDs are assigned.
CUDA Programming Guide Version 3.0

52

Chapter 3. Programming Interface

cuFuncSetSharedSize() sets the size of shared memory for the function.

The cuParam*() family of functions is used to specify the parameters that will be provided to the kernel the next time cuLaunchGrid() or cuLaunch() is invoked to launch the kernel. The second argument of each of the cuParam*() functions specifies the offset of the parameter in the parameter stack. This offset must match the alignment requirement for the parameter type in device code. Alignment requirements in device code for the built-in vector types are listed in Table B-1. For all other basic types, the alignment requirement in device code matches the alignment requirement in host code and can therefore be obtained using __alignof(). The only exception is when the host compiler aligns double and long long (and long on a 64-bit system) on a one-word boundary instead of a two-word boundary (for example, using gcc‟s compilation flag -mno-align-double) since in device code these types are always aligned on a two-word boundary. Also, CUdeviceptr is an integer, but represents a pointer, so its alignment requirement is __alignof(void*). The following code sample uses a macro to adjust the offset of each parameter to meet its alignment requirement.
#define ALIGN_UP(offset, alignment) \ (offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1) int offset = 0; int i; ALIGN_UP(offset, __alignof(i)); cuParamSeti(cuFunction, offset, i); offset += sizeof(i); float4 f4; ALIGN_UP(offset, 16); // float4’s alignment is 16 cuParamSetv(cuFunction, offset, &f4, sizeof(f4)); offset += sizeof(f4); char c; ALIGN_UP(offset, __alignof(c)); cuParamSeti(cuFunction, offset, c); offset += sizeof(c); float f; ALIGN_UP(offset, __alignof(f)); cuParamSeti(cuFunction, offset, f); offset += sizeof(f); CUdeviceptr dptr; // void* should be used to determine CUdeviceptr’s alignment void* ptr = (void*)(size_t)dptr; ALIGN_UP(offset, __alignof(ptr)); cuParamSetv(cuFunction, offset, &ptr, sizeof(ptr)); offset += sizeof(ptr); float2 f2; ALIGN_UP(offset, 8); // float2’s alignment is 8 cuParamSetv(cuFunction, offset, &f2, sizeof(f2)); offset += sizeof(f2);

CUDA Programming Guide Version 3.0

53

f). offset). &s. } myStruct. The following structure. Such a structure might also be padded differently. int offset = 0. float4 f4. cuDeviceGetCount(&deviceCount). // Create context 54 CUDA Programming Guide Version 3. might therefore differ between device code and host code.f4).Chapter 3. // Get number of devices supporting CUDA int deviceCount = 0. exit (0). sizeof(s. or non-aligned double and long long. offset.0 . ALIGN_UP(offset. The alignment requirement of a structure is equal to the maximum of the alignment requirements of its fields. 0).f4)). typedef struct { float f. sizeof(s. cuParamSetv(cuFunction. cuFuncSetBlockShape(cuFunction. is not padded at all in host code. 3.f. 16).f)).\n"). gridWidth. if (deviceCount == 0) { printf("There is no device supporting CUDA. offset. cuDeviceGet(&cuDevice. for example. The alignment requirement of a structure that contains built-in vector types. such as: myStruct s. offset += sizeof(s.1 written using the driver API: // Host code int main() { // Initialize if (cuInit(0) != CUDA_SUCCESS) exit (0).f4. offset += sizeof(s.4 Device Memory Linear memory is allocated using cuMemAlloc() or cuMemAllocPitch() and freed using cuMemFree().3. blockWidth. // float4’s alignment is 16 cuParamSetv(cuFunction. gridHeight). } // Get handle for device 0 CUdevice cuDevice = 0. 1). Programming Interface cuParamSetSize(cuFunction. cuLaunchGrid(cuFunction.2. Here is the host code of the sample from Section 3. Any parameter of type myStruct must therefore be passed using separate calls to cuParam*(). &s. blockHeight. CUdeviceptr. but it is padded in device code with 12 bytes after field f since the alignment requirement for field f4 is 16.

int threadsPerBlock = 256. size). &ptr. CUdeviceptr d_C. // Get function handle from module CUfunction vecAdd. offset += sizeof(ptr). __alignof(ptr)).0 55 . CUDA Programming Guide Version 3. cuParamSetSize(VecAdd. size). cuLaunchGrid(VecAdd. void* ptr. cuMemAlloc(&d_A. 0. size). cuParamSetv(vecAdd. size). offset += sizeof(ptr). alignment) \ (offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1) int offset = 0. offset. cuDevice). cuMemFree(d_C).ptx”). // Allocate vectors in device memory size_t size = N * sizeof(float). d_C. cuMemcpyHtoD(d_B. cuMemFree(d_B). CUdeviceptr d_B. cuCtxCreate(&cuContext. “VecAdd. ptr = (void*)(size_t)d_A. cuModuleLoad(&cuModule. // Copy vectors from host memory to device memory // h_A and h_B are input vectors stored in host memory cuMemcpyHtoD(d_A. // Free device memory cuMemFree(d_A). cuParamSetv(vecAdd. sizeof(ptr)).Chapter 3. __alignof(ptr)). sizeof(ptr)). cuParamSetv(vecAdd. h_A. cuFuncSetBlockShape(vecAdd. threadsPerBlock. &ptr. offset. ptr = (void*)(size_t)d_C. // Copy result from device memory to host memory // h_C contains the result in host memory cuMemcpyDtoH(h_C. cuModule. offset). offset += sizeof(ptr). size). Programming Interface CUcontext cuContext. "VecAdd"). cuModuleGetFunction(&vecAdd. size). blocksPerGrid. ALIGN_UP(offset. ALIGN_UP(offset. cuMemAlloc(&d_B. CUdeviceptr d_A. // Create module from binary file CUmodule cuModule. // Invoke kernel #define ALIGN_UP(offset. 1). ALIGN_UP(offset. sizeof(ptr)). &ptr. int blocksPerGrid = (N + threadsPerBlock – 1) / threadsPerBlock. 1. offset. ptr = (void*)(size_t)d_B. cuMemAlloc(&d_C. 1). h_B. __alignof(ptr)).

sizeof(ptr)). desc. 1. 1).Height = height.dstMemoryType = CU_MEMORYTYPE_ARRAY. This function is recommended for allocations of 2D arrays as it makes sure that the allocation is appropriately padded to meet the alignment requirements described in Section 5. CUarray cuArray.2. void* ptr = (void*)(size_t)devPtr. 4). 0. 0.0 . copyParam. ++c) { float element = row[c]. desc. The following code sample copies the 2D array to the CUDA array allocated in the previous code samples: CUDA_MEMCPY2D copyParam. Programming Interface } Linear memory can also be allocated through cuMemAllocPitch(). 100. cuFuncSetBlockShape(myKernel. and CUDA arrays. int pitch. width * sizeof(float). &desc). cuParamSetv(myKernel. The following code sample allocates a width×height 2D array of floating-point values and shows how to loop over the array elements in device code: // Host code (assuming cuModule has been loaded) CUdeviceptr devPtr.dstArray = cuArray. copyParam.Format = CU_AD_FORMAT_FLOAT. 56 CUDA Programming Guide Version 3. memset(&copyParam. cuModule. for (int c = 0. desc. cuLaunchGrid(myKernel. &ptr. 1). copyParam. sizeof(ptr)). CUfunction myKernel. The reference manual lists all the various functions used to copy memory between linear memory allocated with cuMemAlloc(). &pitch. copyParam. sizeof(copyParam)). therefore ensuring best performance when accessing the row addresses or performing copies between 2D arrays and other regions of device memory (using the cuMemcpy2D()).3. cuParamSetSize(myKernel. 512. linear memory allocated with cuMemAllocPitch().1. r < height. // Device code __global__ void MyKernel(float* devPtr) { for (int r = 0. ++r) { float* row = (float*)((char*)devPtr + r * pitch).NumChannels = 1. “MyKernel”).srcMemoryType = CU_MEMORYTYPE_DEVICE. cuModuleGetFunction(&myKernel.srcDevice = devPtr. cuMemAllocPitch(&devPtr. c < width.Width = width. } } } The following code sample allocates a width×height CUDA array of one 32-bit floating-point component: CUDA_ARRAY_DESCRIPTOR desc. desc. cuArrayCreate(&cuArray. The returned pitch (or stride) must be used to access array elements. height.Chapter 3.

width * A. cuMemcpyHtoD(devPtr. cuMemcpyHtoD(d_A.width * C. size = C. &bytes.height * sizeof(float).Chapter 3.elements.width. d_C. d_A.Height = height.3. shared memory is statically allocated within the kernel as opposed to allocated at runtime through cuFuncSetSharedSize(). Matrix d_B. float data[256].5 Shared Memory The following code sample is the driver version of the host code of the sample from Section 3.height = B. unsigned int bytes.height. // Matrices are stored in row-major order: // M(row. CUdeviceptr devPtr.elements. d_C. cuMemAlloc(&d_B. // Allocate C in device memory Matrix d_C. cuMemcpy2D(&copyParam). cuModuleGetGlobal(&devPtr.elements. #ifdef __CUDACC__ float* elements. B. cuModule. d_A. #else CUdeviceptr elements.elements. int height. size).WidthInBytes = width * sizeof(float). size = B. cuMemAlloc(&d_C. Matrix C) { // Load A and B to device memory Matrix d_A. 3.height = A.width * B.elements.stride = B.srcPitch = pitch. // Matrix multiplication .width = d_B.height * sizeof(float). The following code sample copies some host memory array to constant memory: __constant__ float constData[256]. A.height. copyParam.2.stride = A.width = d_C. size).Host code // Matrix dimensions are assumed to be multiples of BLOCK_SIZE void MatMul(const Matrix A.width. size). data.elements.height = C. cuMemAlloc(&d_A. int stride. const Matrix B. d_B.width = d_A.width.stride = C. CUDA Programming Guide Version 3. size). Programming Interface copyParam. d_B.height. size_t size = A.0 57 . bytes). #endif } Matrix.elements + row * M.stride + col) typedef struct { int width. In this sample. “constData”).2. col) = *(M. cuMemcpyHtoD(d_B. copyParam.elements. Note how the Matrix type must be declared differently in host code since device pointers are represented as a handle of type CUdeviceptr.height * sizeof(float). size).

cuParamSetv(matMulKernel. ++device) { CUdevice cuDevice. cuMemFree(d_B. ALIGN_UP(offset. sizeof(ptr)). sizeof(ptr)). ptr = (void*)(size_t)d_B. int device. offset.0 58 . // Read C from device memory cuMemcpyDtoH(C. } 3. 1). cuLaunchGrid(matMulKernel. &ptr. __alignof(ptr)).6 Multiple Devices cuDeviceGetCount() and cuDeviceGet() provide a way to enumerate the devices present in the system and other functions (described in the reference manual) to retrieve their properties: int deviceCount.3. cuDeviceGet(&cuDevice.elements). A. int major. BLOCK_SIZE. &minor. void* ptr. "MatMulKernel"). alignment) \ (offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1) int offset = 0. Programming Interface // Invoke kernel (assuming cuModule has been loaded) CUfunction matMulKernel. offset. ALIGN_UP(offset. __alignof(ptr)). offset). sizeof(ptr)). ptr = (void*)(size_t)d_A.y). cuParamSetSize(matMulKernel.height / dimBlock. for (int device = 0. ALIGN_UP(offset. cuMemFree(d_C. offset += sizeof(ptr). BLOCK_SIZE. cuDeviceComputeCapability(&major.elements. cuParamSetv(matMulKernel. cuFuncSetBlockShape(matMulKernel. offset += sizeof(ptr).Chapter 3. cuModule. ptr = (void*)(size_t)d_C. } 3. B.elements). device < deviceCount.x.7 Texture Memory Texure binding is done using cuTexRefSetAddress() for linear memory and cuTexRefSetArray() for CUDA arrays.3. offset. cuModuleGetFunction(&matMulKernel. &ptr. device). cuDevice). offset += sizeof(ptr). &ptr.elements.width / dimBlock. size). cuDeviceGetCount(&deviceCount). d_C. CUDA Programming Guide Version 3. #define ALIGN_UP(offset. __alignof(ptr)).elements). // Free device memory cuMemFree(d_A. minor. cuParamSetv(matMulKernel.

2. "texRef")). 0. desc.Height = height.srcPitch = width * sizeof(float). copyParam. “texRef”). 1. // Copy to device memory some data located at address h_data // in host memory CUDA_MEMCPY2D copyParam. cuModule. otherwise.3.WidthInBytes = copyParam. filter mode. memset(&copyParam. &desc. desc.srcMemoryType = CU_MEMORYTYPE_HOST. the results of texture fetches are undefined. CU_TR_FILTER_MODE_LINEAR). The reference manual lists various functions used to set address mode.srcHost = h_data. devPtr. desc. cuTexRefSetFlags(texRef. // Set texture parameters CUtexref texRef. desc. CUDA_ARRAY_DESCRIPTOR desc.NumChannels = 1. The following code sample is the driver version of the host code of the sample from Section 3. the following code sample retrieves texRef„s handle: CUtexref cuTexRef. The following code sample binds texRef to some linear memory pointed to by devPtr: CUDA_ARRAY_DESCRIPTOR desc.Width = width. The following code samples bind texRef to a CUDA array cuArray: cuTexRefSetArray(cuTexRef. cuModule. cuMemcpy2D(&copyParam).srcPitch. and other flags for some texture reference. cuTexRefSetFormat(texRef. cuArray.0 59 . cuTexRefSetAddress2D(cuTexRef. cuModuleGetTexRef(&cuTexRef. copyParam.Chapter 3. cuModuleGetTexRef(&texRef. CUDA Programming Guide Version 3. format.dstArray = cuArray. cuTexRefSetFilterMode(texRef. copyParam. cuArrayCreate(&cuArray. CU_TRSA_OVERRIDE_FORMAT). The format specified when binding a texture to a texture reference must match the parameters specified when declaring the texture reference. sizeof(copyParam)). CU_TR_ADDRESS_MODE_WRAP). copyParam. 1).Height = height. CU_TRSF_NORMALIZED_COORDINATES). // Host code int main() { // Allocate CUDA array in device memory CUarray cuArray. copyParam. 2. cudaReadModeElementType> texRef. &desc). Programming Interface If a module cuModule contains some texture reference texRef defined as texture<float. CU_TR_ADDRESS_MODE_WRAP). 0.dstMemoryType = CU_MEMORYTYPE_ARRAY. copyParam. pitch). cuTexRefSetAddressMode(texRef. cuTexRefSetAddressMode(texRef.Format = CU_AD_FORMAT_FLOAT.4. copyParam. CU_AD_FORMAT_FLOAT.

Programming Interface // Bind the array to the texture cuTexRefSetArray(texRef. offset. (height + dimBlock. width). cuMemFree(output). ALIGN_UP(offset. offset += sizeof(angle). &ptr. cuParamSetf(transformKernel.5. "transformKernel"). cuLaunchGrid(transformKernel. CU_PARAM_TR_DEFAULT. alignment) \ (offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1) int offset = 0.1) 3. #define ALIGN_UP(offset. offset.8 Page-Locked Host Memory Page-locked host memory can be allocated using cuMemHostAlloc() with optional mutually non-exclusive flags:  CU_MEMHOSTALLOC_PORTABLE to allocate memory that is portable across CUDA contexts (see Section 3. cuParamSeti(transformKernel.5.0 . cuModule.2. CU_TRSA_OVERRIDE_FORMAT). cuModuleGetFunction(&transformKernel. __alignof(height)). void* ptr = (void*)(size_t)output.2. texRef). offset += sizeof(height).2).x – 1) / dimBlock. 16.y). cuParamSetSize(transformKernel. height).Chapter 3. __alignof(angle)). (width + dimBlock. // Free device memory cuArrayDestroy(cuArray).3. cuFuncSetBlockShape(transformKernel.5. ALIGN_UP(offset. width * height * sizeof(float)). cuArray.y – 1) / dimBlock.2. } 3. offset. sizeof(ptr)).3). cuMemAlloc((void**)&output.2. 1).  CU_MEMHOSTALLOC_WRITECOMBINED to allocate memory as writecombining (see Section 3. 16. 60 CUDA Programming Guide Version 3.  CU_MEMHOSTALLOC_DEVICEMAP to allocate mapped page-locked memory (see Section 3. ALIGN_UP(offset.2. offset)). offset. __alignof(width)).5.x. offset += sizeof(width). offset += sizeof(ptr). // Invoke kernel (assuming cuModule has been loaded) CUfunction transformKernel. angle). ALIGN_UP(offset. Page-locked host memory is freed using cuMemFreeHost(). cuParamSetTexRef(transformKernel. cuParamSetv(transformKernel. // Allocate result of transformation in device memory CUdeviceptr output. cuParamSeti(transformKernel. __alignof(ptr)).

stream[i]). ++i) cuMemcpyHtoDAsync(inputDevPtr + i * size. Programming Interface Page-locked memory mapping is enabled for a CUDA context by creating the context with the CU_CTX_MAP_HOST flag and device pointers to mapped pagelocked memory are retrieved using cuMemHostGetDevicePointer(). 1. size). 3. ALIGN_UP(offset.9 Asynchronous Concurrent Execution Applications may query if a device can perform copies between page-locked host memory and device memory concurrently with kernel execution by checking the CU_DEVICE_ATTRIBUTE_GPU_OVERLAP attribute using cuDeviceGetAttribute(). ptr = (void*)(size_t)inputDevPtr. 512. offset).9. cuLaunchGridAsync(cuFunction. alignment) \ (offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1) int offset = 0.6. cuParamSetSize(cuFunction. i < 2. offset. void* ptr. __alignof(ptr)).1 Stream The driver API provides functions similar to the runtime API to manage streams. __alignof(ptr)). offset += sizeof(ptr). 1. ALIGN_UP(offset. cuFuncSetBlockShape(cuFunction. Applications may query if a device supports multiple kernels running concurrently by checking the CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS attribute using cuDeviceGetAttribute(). 3. offset += sizeof(int). ++i) { #define ALIGN_UP(offset. ALIGN_UP(offset. } CUDA Programming Guide Version 3. 2 * size). ++i) cuStreamCreate(&stream[i].3. The following code sample is the driver version of the code sample from Section 3. size. cuParamSetv(cuFunction. cuParamSetv(cuFunction. ptr = (void*)(size_t)outputDevPtr.3. i < 2.Chapter 3.0 61 . for (int i = 0. 1). &ptr. 0). cuParamSeti(cuFunction. stream[i]). CUstream stream[2]. offset. offset. i < 2. for (int i = 0. cuMemAllocHost((void**)&hostPtr. Applications may query whether a device supports mapped page-locked host memory or not by checking the CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY attribute using cuDeviceGetAttribute(). sizeof(ptr)). __alignof(size)).4. float* hostPtr. sizeof(ptr)). &ptr.2. for (int i = 0. hostPtr + i * size. offset += sizeof(ptr). 100.

ptr = (void*)(size_t)inputDevPtr. They are destroyed this way: cuEventDestroy(start). The following code sample is the driver version of the code sample from Section 3. offset. size.Chapter 3.9. ALIGN_UP(offset. ALIGN_UP(offset.2. sizeof(ptr)). offset += sizeof(size). cuParamSetv(cuFunction. for (int i = 0. ++i) { #define ALIGN_UP(offset. offset. ++i) cuMemcpyDtoHAsync(hostPtr + i * size. offset += sizeof(ptr). cuEventRecord(stop. __alignof(ptr)). ALIGN_UP(offset. &ptr. outputDevPtr + i * size. i < 2. cuEventCreate(&stop). ++i) cuMemcpyHtoDAsync(inputDevPtr + i * size. for (int i = 0. i < 2. cuEventRecord(start. cuLaunchGridAsync(cuFunction. start. &ptr. stream[i]). 1. cuParamSeti(cuFunction.0 . i < 2. 0). cuEventCreate(&start). cuParamSetSize(cuFunction. __alignof(size)). Programming Interface for (int i = 0. stream[i]). sizeof(ptr)). 62 CUDA Programming Guide Version 3. size). stream[i]). CUevent start. 512. 3. stop). 100. outputDevPtr + i * size. cuFuncSetBlockShape(cuFunction.6. offset += sizeof(ptr). cuEventElapsedTime(&elapsedTime.6. size. offset). ++i) cuMemcpyDtoHAsync(hostPtr + i * size. stream[i]). i < 2. offset. stop. 1. for (int i = 0. cuEventSynchronize(stop). 1). __alignof(ptr)). float elapsedTime. cuParamSetv(cuFunction. 0). } for (int i = 0. hostPtr + i * size. alignment) \ (offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1) int offset = 0. i < 2. void* ptr. cuEventDestroy(stop). ++i) cuStreamDestroy(&stream[i]).2 Event Management The driver API provides functions similar to the runtime API to manage events. size. cuCtxSynchronize().3. ptr = (void*)(size_t)outputDevPtr.

Once a resource is registered to CUDA. read-only) that the CUDA driver can use to optimize resource management.3.3.3.1 OpenGL Interoperability Interoperability with OpenGL requires that the CUDA context be specifically created using cuGLCtxCreate() instead of cuCtxCreate().10. struct cudaGraphicsResource* positionsVBO_CUDA.7. CUDA Programming Guide Version 3.3.10..10 Graphics Interoperability The driver API provides functions similar to the runtime API to manage graphics interoperability.9.1 apply.3. 3. Sections 3.2.0 63 . Registering a resource is potentially high-overhead and therefore typically called only once per resource.10.Chapter 3. block.10.1 and 3. A resource must be registered to CUDA before it can be mapped using the functions mentioned in Sections 3. cuGraphicsResourceSetMapFlags() can be called to specify usage hints (write-only. The following code sample is the driver version of the code sample from Section 3. The OpenGL resources that may be mapped into the address space of CUDA are OpenGL buffer. or spin on a synchronous function call can be specified by calling cuCtxCreate() with some specific flags as described in the reference manual. The same restrictions described in Section 3. Programming Interface 3.2. A mapped resource can be read from or written to by kernels using the device memory address returned by cuGraphicsResourceGetMappedPointer() for buffers and cuGraphicsSubResourceGetMappedArray() for CUDA arrays.7.3. CUfunction createVertices. These functions return a CUDA graphics resource of type CUgraphicsResource.3 Synchronous Calls Whether the host thread will yield. and renderbuffer objects..3. A texture or renderbuffer object is registered using cuGraphicsGLRegisterImage(). A CUDA graphics resource is unregistered using cuGraphicsUnregisterResource().2.1 and 3. it can be mapped and unmapped as many times as necessary using cuGraphicsMapResources() and cuGraphicsUnmapResources().2 give specifics for each graphics API and some code samples. GLuint positionsVBO. 3. texture. A buffer object is registered using cuGraphicsGLRegisterBuffer().1.10. int main() { // Initialize driver API . Accessing a resource through OpenGL or Direct3D while it is mapped to CUDA produces undefined results.

time). 64 CUDA Programming Guide Version 3. 0). "createVertices"). offset. alignment) \ (offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1) int offset = 0.Chapter 3. // Create buffer object and register it with CUDA glGenBuffers(1. cuGraphicsMapResources(1. // Create context CUcontext cuContext. offset += sizeof(ptr). offset += sizeof(time).. // Execute kernel #define ALIGN_UP(offset. cuDeviceGet(&cuDevice. void* ptr = (void*)(size_t)positions. cuParamSetv(createVertices. offset. positionsVBO_CUDA)). // Initialize OpenGL and GLUT . 0..0 . cuGraphicsGLRegisterBuffer(&positionsVBO_CUDA. cudaGraphicsMapFlagsWriteDiscard). &positionsVBO_CUDA. unsigned int size = width * height * 4 * sizeof(float). size_t num_bytes. &num_bytes.ptx”). offset. cuModuleLoad(&cuModule. &ptr. // Launch rendering loop glutMainLoop(). glBindBuffer(GL_ARRAY_BUFFER. size. “createVertices. 0). 0. &vbo). GL_DYNAMIC_DRAW). cuParamSetf(createVertices. positionsVBO. __alignof(time)). glBufferData(GL_ARRAY_BUFFER. glBindBuffer(GL_ARRAY_BUFFER. 0). __alignof(width)). // Create module from binary file CUmodule cuModule. cuParamSeti(createVertices. width). cuGraphicsResourceGetMappedPointer((void**)&positions. // Get function handle from module cuModuleGetFunction(&createVertices. ALIGN_UP(offset. } void display() { // Map OpenGL buffer object for writing from CUDA CUdeviceptr positions. sizeof(ptr)). cuDevice). __alignof(ptr)). ALIGN_UP(offset. Programming Interface // Get handle for device 0 CUdevice cuDevice = 0. cuModule. positionsVBO). cuGLCtxCreate(&cuContext. ALIGN_UP(offset. glutDisplayFunc(display).

cuGraphicsD3D10RegisterResource(). glVertexPointer(4. positionsVBO). The following code sample is the driver version of the host code of the sample from Section 3. } void deleteVBO() { cuGraphicsUnregisterResource(positionsVBO_CUDA).0 65 . int threadsPerBlock = 16. offset). GL_FLOAT. // Unmap buffer object cuGraphicsUnmapResources(1. 0). offset. The Direct3D resources that may be mapped into the address space of CUDA are Direct3D buffers. &positionsVBO). CUDA Programming Guide Version 3. 3. and surfaces. glutPostRedisplay(). z. Direct3D 9 Version: IDirect3D9* D3D. These resources are registered using cuGraphicsD3D9RegisterResource(). glEnableClientState(GL_VERTEX_ARRAY). offset += sizeof(height). cuLaunchGrid(createVertices.10. cuWGLGetDevice() can be used to retrieve the CUDA device associated to the handle returned by wglEnumGpusNV(). // Swap buffers glutSwapBuffers(). IDirect3DDevice9* device. and cuGraphicsD3D11RegisterResource(). threadsPerBlock. cuD3D10CtxCreate()) instead of cuCtxCreate(). glDrawArrays(GL_POINTS. textures. DWORD color. Programming Interface offset += sizeof(width). y.2. struct CUSTOMVERTEX { FLOAT x. 0. 0. // Render from buffer object glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT). This is done by creating the CUDA context using cuD3D9CtxCreate() (resp. 0). glBindBuffer(GL_ARRAY_BUFFER. cuParamSetSize(createVertices.3.2 Direct3D Interoperability Interoperability with Direct3D requires that the Direct3D device be specified when the CUDA context is created. 1). threadsPerBlock. width * height). cuFuncSetBlockShape(createVertices. glDisableClientState(GL_VERTEX_ARRAY). glDeleteBuffers(1.2. width / threadsPerBlock.Chapter 3.7. &positionsVBO_CUDA. } On Windows and for Quadro GPUs. height / threadsPerBlock). __alignof(height)). height). cuParamSeti(createVertices. ALIGN_UP(offset.

&positionsVB.DeviceName) == cudaSuccess) break. Render(). int main() { // Initialize Direct3D D3D = Direct3DCreate9(D3D_SDK_VERSION). cuD3D9CtxCreate(&cuContext.. adapterId. &device). // Create vertex buffer and register it with CUDA unsigned int size = width * height * sizeof(CUSTOMVERTEX). &device). . adapter++) { D3DADAPTER_IDENTIFIER9 adapterId. cudaGraphicsMapFlagsWriteDiscard). 0.. 66 CUDA Programming Guide Version 3. cudaGraphicsRegisterFlagsNone). &params. cuModuleLoad(&cuModule. D3DCREATE_HARDWARE_VERTEXPROCESSING. 0. &cuDevice. // Initialize driver API . adapter < g_pD3D->GetAdapterCount(). // Create context CUdevice cuDevice. device->CreateVertexBuffer(size. cuGraphicsD3D9RegisterResource(&positionsVB_CUDA.Chapter 3. positionsVB. "createVertices"). CUcontext cuContext.ptx”). Programming Interface }.... cuModule. // Get function handle from module cuModuleGetFunction(&createVertices. for (.. if (cuD3D9GetDevice(&dev. D3D->CreateDevice(adapter.. // Get a CUDA-enabled adapter unsigned int adapter = 0. int dev. 0. D3DDEVTYPE_HAL. } // Create device .. “createVertices. // Create module from binary file CUmodule cuModule.0 ..) { . 0). // Launch rendering loop while (.. D3DPOOL_DEFAULT. &adapterId). D3DFVF_CUSTOMVERTEX. g_pD3D->GetAdapterIdentifier(adapter. struct cudaGraphicsResource* positionsVB_CUDA. IDirect3DVertexBuffer9* positionsVB. hWnd. cuGraphicsResourceSetMapFlags(positionsVB_CUDA.

cuGraphicsResourceGetMappedPointer((void**)&positions. offset += sizeof(height). CUDA Programming Guide Version 3. positionsVB_CUDA)). 0). offset). cuParamSetf(createVertices. width / threadsPerBlock. // Execute kernel #define ALIGN_UP(offset. threadsPerBlock. cuFuncSetBlockShape(createVertices. cuParamSetSize(createVertices. }.. // Unmap vertex buffer cuGraphicsUnmapResources(1. offset += sizeof(ptr). &positionsVB_CUDA. __alignof(time)). cuParamSetv(createVertices. width). offset. positionsVB->Release(). ALIGN_UP(offset. } Direct3D 10 Version: ID3D10Device* device. cuGraphicsMapResources(1.. &positionsVB_CUDA. alignment) \ (offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1) int offset = 0. ALIGN_UP(offset. y. threadsPerBlock. __alignof(height)).0 67 . offset += sizeof(time). offset. ALIGN_UP(offset. &num_bytes. time). 1). offset. cuParamSeti(createVertices. cuParamSeti(createVertices. int threadsPerBlock = 16. // Draw and present . 0). &ptr. __alignof(width)). z. cuLaunchGrid(createVertices.Chapter 3. height / threadsPerBlock). Programming Interface } } void Render() { // Map vertex buffer for writing from CUDA float4* positions. size_t num_bytes. sizeof(ptr)). } void releaseVB() { cuGraphicsUnregisterResource(positionsVB_CUDA). offset. __alignof(ptr)). struct CUSTOMVERTEX { FLOAT x. void* ptr = (void*)(size_t)positions. DWORD color. offset += sizeof(width). ALIGN_UP(offset. ID3D10Buffer* positionsVB. height).

. adapter->Release(). cudaGraphicsRegisterFlagsNone).MiscFlags = 0. adapter) == cudaSuccess) break. “createVertices. device->CreateBuffer(&bufferDesc. bufferDesc.Chapter 3.CPUAccessFlags = 0. int main() { // Get a CUDA-enabled adapter IDXGIFactory* factory. &cuDevice. D3D10_DRIVER_TYPE_HARDWARE. 0.BindFlags = D3D10_BIND_VERTEX_BUFFER. 0. // Create module from binary file CUmodule cuModule. if (cuD3D10GetDevice(&dev. &device). IDXGIAdapter* adapter = 0. CreateDXGIFactory(__uuidof(IDXGIFactory). CUcontext cuContext. "createVertices"). D3D10_BUFFER_DESC bufferDesc. cuModuleLoad(&cuModule. &device). positionsVB. cuModule. 68 CUDA Programming Guide Version 3. } factory->Release(). D3D10_CREATE_DEVICE_DEBUG. int dev. bufferDesc. // Create vertex buffer and register it with CUDA unsigned int size = width * height * sizeof(CUSTOMVERTEX).ptx”).Usage = D3D10_USAGE_DEFAULT. bufferDesc. &positionsVB).. cuGraphicsResourceSetMapFlags(positionsVB_CUDA. for (unsigned int i = 0. cuD3D10CtxCreate(&cuContext. bufferDesc. // Get function handle from module cuModuleGetFunction(&createVertices. &swapChainDesc &swapChain. &adapter)) break. D3D10CreateDeviceAndSwapChain(adapter. Programming Interface struct cudaGraphicsResource* positionsVB_CUDA. (void**)&factory). // Create swap chain and device . // Create context CUdevice cuDevice. adapter->Release(). bufferDesc. D3D10_SDK_VERSION. 0.0 .. cuGraphicsD3D10RegisterResource(&positionsVB_CUDA.. // Initialize driver API . ++i) { if (FAILED(factory->EnumAdapters(i.ByteWidth = size. !adapter.

.. size_t num_bytes. offset += sizeof(height). cuParamSetf(createVertices. // Unmap vertex buffer cuGraphicsUnmapResources(1. 0).. threadsPerBlock. positionsVB->Release(). &positionsVB_CUDA. positionsVB_CUDA)). &ptr. } } void Render() { // Map vertex buffer for writing from CUDA float4* positions.. cuParamSeti(createVertices. __alignof(time)). // Launch rendering loop while (. void* ptr = (void*)(size_t)positions.. 1). int threadsPerBlock = 16. __alignof(width)). cuGraphicsResourceGetMappedPointer((void**)&positions. ALIGN_UP(offset. &positionsVB_CUDA. width). threadsPerBlock. alignment) \ (offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1) int offset = 0. .. // Draw and present . // Execute kernel #define ALIGN_UP(offset. 0). offset. sizeof(ptr)). height / threadsPerBlock).) { .0 69 . offset. } void releaseVB() { cuGraphicsUnregisterResource(positionsVB_CUDA). Render(). ALIGN_UP(offset. offset += sizeof(width). ALIGN_UP(offset. __alignof(height)). cuGraphicsMapResources(1. offset). width / threadsPerBlock. cuFuncSetBlockShape(createVertices. offset += sizeof(time). cuParamSeti(createVertices. offset. cuParamSetv(createVertices. Programming Interface cudaGraphicsMapFlagsWriteDiscard).Chapter 3. offset += sizeof(ptr)... __alignof(ptr)). time). height). ALIGN_UP(offset. &num_bytes. cuLaunchGrid(createVertices. offset. cuParamSetSize(createVertices. } CUDA Programming Guide Version 3.

&featureLevel. &deviceContext).ptx”). adapter) == cudaSuccess) break. CUcontext cuContext.. cuModule. 0. struct CUSTOMVERTEX { FLOAT x. if (cuD3D11GetDevice(&dev. "createVertices").0 . z. sFnPtr_D3D11CreateDeviceAndSwapChain(adapter. &adapter)) break. y. cuD3D11CtxCreate(&cuContext.Chapter 3. (void**)&factory). }. DWORD color. D3D11_SDK_VERSION. &cuDevice. D3D11_DRIVER_TYPE_HARDWARE. int dev. // Create swap chain and device . D3D11_CREATE_DEVICE_DEBUG. IDXGIAdapter* adapter = 0. CreateDXGIFactory(__uuidof(IDXGIFactory). // Create vertex buffer and register it with CUDA 70 CUDA Programming Guide Version 3. ++i) { if (FAILED(factory->EnumAdapters(i. &swapChain. &swapChainDesc. for (unsigned int i = 0. int main() { // Get a CUDA-enabled adapter IDXGIFactory* factory. ID3D11Buffer* positionsVB. } factory->Release(). 0. // Create context CUdevice cuDevice. featureLevels. // Create module from binary file CUmodule cuModule. Programming Interface Direct3D 11 Version: ID3D11Device* device.. &device). 3. adapter->Release(). !adapter. // Initialize driver API .. // Get function handle from module cuModuleGetFunction(&createVertices. “createVertices. adapter->Release().. struct cudaGraphicsResource* positionsVB_CUDA. &device. cuModuleLoad(&cuModule.

bufferDesc. cuParamSetv(createVertices. threadsPerBlock. offset += sizeof(ptr).. offset. ALIGN_UP(offset. __alignof(time)). offset). &positionsVB).. 0. ALIGN_UP(offset. cuFuncSetBlockShape(createVertices. positionsVB_CUDA)). // Unmap vertex buffer CUDA Programming Guide Version 3. __alignof(ptr)). } } void Render() { // Map vertex buffer for writing from CUDA float4* positions.ByteWidth = size. offset += sizeof(height). cuLaunchGrid(createVertices. int threadsPerBlock = 16.Chapter 3. &ptr. bufferDesc. &num_bytes. width / threadsPerBlock. height / threadsPerBlock). offset. cuParamSetf(createVertices. // Launch rendering loop while (. height). 1). ALIGN_UP(offset. positionsVB. alignment) \ (offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1) int offset = 0. cuParamSeti(createVertices. sizeof(ptr)).CPUAccessFlags = 0. cuGraphicsResourceGetMappedPointer((void**)&positions. &positionsVB_CUDA.. bufferDesc. device->CreateBuffer(&bufferDesc.Usage = D3D11_USAGE_DEFAULT. __alignof(width)). cudaGraphicsRegisterFlagsNone).) { . threadsPerBlock. bufferDesc. void* ptr = (void*)(size_t)positions. 0).MiscFlags = 0.. cuGraphicsResourceSetMapFlags(positionsVB_CUDA. __alignof(height)). cuParamSeti(createVertices.. cuGraphicsD3D11RegisterResource(&positionsVB_CUDA. cuGraphicsMapResources(1. . offset += sizeof(time).BindFlags = D3D10_BIND_VERTEX_BUFFER. ALIGN_UP(offset. Render(). offset. Programming Interface unsigned int size = width * height * sizeof(CUSTOMVERTEX).0 71 . offset. cuParamSetSize(createVertices. size_t num_bytes. bufferDesc. // Execute kernel #define ALIGN_UP(offset. offset += sizeof(width). time). width).. cudaGraphicsMapFlagsWriteDiscard). D3D11_BUFFER_DESC bufferDesc.

cuCtxAttach() can be used to retrieve the context created during initialization. devPtr = (CUdeviceptr)(size_t)d_data. positionsVB->Release().6). Device memory can be allocated and freed using either API. } void releaseVB() { cuGraphicsUnregisterResource(positionsVB_CUDA). // Allocation using driver API cuMemAlloc(&devPtr.. size). // Allocation using runtime API cudaMalloc((void**)&d_data. float* d_data. but for an asynchronous function (see Section 3. 72 CUDA Programming Guide Version 3.3. If a context is created and made current via the driver API. subsequent runtime calls will pick up this context instead of creating a new one.2. &positionsVB_CUDA.. d_data = (float*)devPtr. CUdeviceptr can be cast to regular pointers and vice-versa: CUdeviceptr devPtr. } 3.2). typically related to parameter validation. If the runtime is initialized (implicitly as mentioned in Section 3. this error code cannot possibly report any of the asynchronous errors that could occur on the device since the function returns before the device has completed the task. The only way to check for asynchronous errors just after some asynchronous function call is therefore to synchronize just after the call by calling cuCtxSynchronize() (or by using any other synchronization mechanisms described in Section 3. the error code only reports errors that occur on the host prior to executing the task.3.11 Error Handling All driver functions return an error code. size).0 . if an asynchronous error occurs. it will be reported by some subsequent unrelated runtime function call. 3. This context can be used by subsequent driver API calls.4 Interoperability between Runtime and Driver APIs An application can mix runtime API code with driver API code within the limits described in this section.Chapter 3.9) and checking the error code returned by cuCtxSynchronize(). // Draw and present . 0). Programming Interface cuGraphicsUnmapResources(1.

CUBLAS.  3. plug-ins.  CUDA Programming Guide Version 3. plug-ins. and libraries (including the C runtime) compiled against a particular version of the driver API will not work on previous versions of the driver.5 Versioning and Compatibility There are two version numbers that developers should care about when developing a CUDA application: The compute capability that describes the general specifications and features of the compute device (see Section 2. …).Chapter 3. This is important. …). this means that applications written using the driver API can invoke libraries written using the runtime API (such as CUFFT.0 73 . because the driver API is backward compatible. The driver API is not forward compatible. Programming Interface In particular. The following features are not supported for applications that mix runtime API code with driver API code: Device emulation. plug-ins.  All plug-ins and libraries used by an application must use the same version of the runtime.1). meaning that applications.5) and the version of the CUDA driver API that describes the features supported by the driver API and runtime.3. It is important to note that mixing and matching versions is not supported. and libraries on a system must use the same version of the CUDA driver API. and libraries (including the C runtime) compiled against a particular version of the driver API will continue to work on subsequent driver releases as illustrated in Figure 3-4. The version of the driver API is defined in the driver header file as CUDA_VERSION. It allows developers to check whether their application requires a newer driver than the one currently installed. All functions from the device and version management sections of the reference manual can be used interchangeably.  Context stack manipulation via cuCtxPopCurrent() and cuCtxPushCurrent() (see Section 3. specifically: All applications. CUBLAS. since only one version of the CUDA driver can be installed on a system. which means that applications.  All plug-ins and libraries used by an application must use the same version of any libraries that use the runtime (such as CUFFT.

. which is a tool distributed as part of the Linux driver: Default compute mode: Multiple host threads can use the device (by calling cudaSetDevice() on this device. cudaSetValidDevices() can be used to set a device from a prioritized list of devices. Libs & Plug-ins Apps. Figure 3-4.7 Mode Switches GPUs dedicate some DRAM memory to the so-called primary surface. one can set any device in a system in one of the three following modes using NVIDIA‟s System Management Interface (nvidia-smi). which is used to refresh the display device whose output is viewed by the user.  Prohibited compute mode: No host thread can use the device. or by making current a context associated to the device.Chapter 3.0 .1 Driver Incompatible 2. that a host thread using the runtime API without explicitly calling cudaSetDevice() might be associated with a device other than device 0 if device 0 turns out to be in prohibited compute mode or in exclusive compute mode and used by another host thread. Libs & Plug-ins Apps. in particular... When users initiate 74 CUDA Programming Guide Version 3. 1. 3. This means.  Exclusive compute mode: Only one host thread can use the device at any given time. but Not Forward Compatible 3.6 Compute Modes On Tesla solutions running Linux. Programming Interface Apps.0 Driver . when using the runtime API. Libs & Plug-ins . when using the driver API) at the same time.0 Driver Compatible 1. The Driver API is Backward..  Applications may query the compute mode of a device by calling cudaGetDeviceProperties() and checking the computeMode property or checking the CU_DEVICE_COMPUTE_MODE attribute using cuDeviceGetAttribute().

(Fullscreen graphics applications running with anti-aliasing enabled may require much more display memory for the primary surface. the system must dedicate 7. hitting Alt+Tab to task switch away from a full-screen DirectX application. Therefore. If a mode switch increases the amount of memory needed for the primary surface. CUDA Programming Guide Version 3.) On Windows.0 75 .Chapter 3. the amount of memory needed for the primary surface changes.24 MB. a mode switch results in any call to the CUDA runtime to fail and return an invalid context error. if the user changes the display resolution from 1280x1024x32-bit to 1600x1200x32-bit.68 MB to the primary surface rather than 5. the system may have to cannibalize memory allocations dedicated to CUDA applications. Programming Interface a mode switch of the display by changing the resolution or bit depth of the display (using NVIDIA control panel or the Display control panel on Windows). For example. or hitting Ctrl+Alt+Del to lock the computer. other events that may initiate display mode switches include launching a full-screen DirectX application.

.

each warp contains threads of consecutive. it leverages thread-level parallelism by using hardware multithreading as detailed in Section 4.0 77 . A half-warp is either the first or second half of a warp. new blocks are launched on the vacated multiprocessors. increasing thread IDs with the first warp containing thread 0. When a multiprocessor is given one or more thread blocks to execute. The threads of a thread block execute concurrently on one multiprocessor. schedules.1 SIMT Architecture The multiprocessor creates. CUDA Programming Guide Version 3. Multiple-Thread) that is described in Section 4. or fourth quarter of a warp.1 and 4.1 and G. To manage such a large amount of threads. A multiprocessor is designed to execute hundreds of threads concurrently.0. 4. but unlike CPU cores they are executed in order and there is no branch prediction and no speculative execution).1 provide the specifics for devices of compute capabilities 1. but they have their own instruction address counter and register state and are therefore free to branch and execute independently.3. Section 2.1. As thread blocks terminate. respectively. Sections 4. Sections G.2 describe the architecture features of the streaming multiprocessor that are common to all devices.2. Individual threads composing a warp start together at the same program address. To maximize utilization of its functional units. When a CUDA program on the host CPU invokes a kernel grid.2 describes how thread IDs relate to thread indices in the block. the first parallel thread technology.x and 2.4. third. it partitions them into warps that get scheduled by a warp scheduler for execution. and multiple thread blocks can execute concurrently on one multiprocessor. second.Chapter 4. and executes threads in groups of 32 parallel threads called warps. The way a block is partitioned into warps is always the same. A quarter-warp is either the first. manages. it employs a unique architecture called SIMT (Single-Instruction. Hardware Implementation The CUDA architecture is built around a scalable array of multithreaded Streaming Multiprocessors (SMs). the blocks of the grid are enumerated and distributed to multiprocessors with available execution capacity. more so than instruction-level parallelism within a single thread (instructions are pipelined. The term warp originates from weaving.

the threads converge back to the same execution path. and when all paths complete.4. modify.2. The SIMT architecture is akin to SIMD (Single Instruction. so full efficiency is realized when all 32 threads of a warp agree on their execution path.10) executed by a warp reads. G.2 Hardware Multithreading The execution context (program counters. G. each read. A key difference is that SIMD vector organizations expose the SIMD width to the software.3. Multiple Data) vector organizations in that a single instruction controls multiple processing elements. different warps execute independently regardless of whether they are executing common or disjoint code paths. Branch divergence occurs only within a warp. The number of blocks and warps that can reside and be processed together on the multiprocessor for a given kernel depends on the amount of registers and shared memory used by the kernel and the amount of registers and shared memory available on the multiprocessor. In contrast with SIMD vector machines. Switching from one execution context to another therefore has no cost. Hardware Implementation A warp executes one common instruction at a time. write to that location occurs and they are all serialized. For the purposes of correctness. each multiprocessor has a set of 32-bit registers that are partitioned among the warps. whereas SIMT instructions specify the execution and branching behavior of a single thread. and a parallel data cache or shared memory that is partitioned among the thread blocks. and at every instruction issue time.4. the warp serially executes each branch path taken. In particular. the warp scheduler selects a warp that has threads ready to execute (active threads) and issues the next instruction to those threads. on the other hand. the number of serialized writes that occur to that location varies depending on the compute capability of the device (see Sections G. There are also a maximum number of resident blocks and a maximum number of resident warps per multiprocessor.2. disabling threads that are not on that path. If a non-atomic instruction executed by a warp writes to the same location in global or shared memory for more than one of the threads of the warp. this is analogous to the role of cache lines in traditional code: Cache line size can be safely ignored when designing for correctness but must be considered in the code structure when designing for peak performance. modifies.0 . 4. SIMT enables programmers to write thread-level parallel code for independent.3) and which thread performs the final write is undefined. Vector architectures.3. however.3. and G.Chapter 4. scalar threads. require the software to coalesce loads into vectors and manage divergence manually. If an atomic instruction (see Section B. as well as data-parallel code for coordinated threads. registers. the programmer can essentially ignore the SIMT behavior. and writes to the same location in global memory for more than one of the threads of the warp. If threads of a warp diverge via a data-dependent conditional branch. etc) for each warp processed by a multiprocessor is maintained on-chip during the entire lifetime of the warp. but the order in which they occur is undefined. substantial performance improvements can be realized by taking care that the code seldom requires threads in a warp to diverge. In practice. These limits as well the amount of registers and shared memory available on the multiprocessor 78 CUDA Programming Guide Version 3.

G S )   Sk is the amount of shared memory used by the kernel in bytes.0 and 1.3 Multiple Devices In a system with multiple GPUs. but only one CUDA device can be created at a time from one of these Direct3D devices. real interoperability only happens with the copy of a Direct3D resource in that GPU (note: in AFR mode Direct3D resources that must be in GPU memory are duplicated in the GPU memory of each GPU in the SLI CUDA Programming Guide Version 3. which is equal to 32. Second. the Direct3D device(s) created by that application can be used for CUDA-Direct3D interoperability (i.x only). which is equal to 512 for devices of compute capability 1. an allocation in one CUDA device on one GPU will consume memory on other GPUs. There are however special considerations as described below when the system is in SLI mode.1) Wsize T is the number of threads per block. As a consequence. all CUDA-enabled GPUs are accessible via the CUDA driver and runtime as separate devices.0: Rblock  ceil( Rk  Wsize .2 and 1.3.x and 128 for devices of compute capability 2. The total number of registers Rblock allocated for a block is as follows: For devices of compute capability 1. passed as a parameter to cudaD3D[9|10]SetDirect3DDevice() when using the runtime API). allocations may fail earlier than otherwise expected. This CUDA device only executes the CUDA work on one of the GPUs in the SLI configuration.  ceil(x. GT )  Wblock GW is the warp allocation granularity. GS is the shared memory allocation granularity.  GT is the thread allocation granularity. y) is equal to x rounded up to the nearest multiple of y.1..e.x: Rblock  ceil(ceil(Wblock . and 64 for devices of compute capability 2.0.  Wsize is the warp size. GT ) For devices of compute capability 2. The total number of warps Wblock in a block is as follows: Wblock  ceil (  T . equal to 256 for devices of compute capability 1. equal to 2 (compute capability 1. First. GW )  Wsize  Rk . If there are not enough registers or shared memory available per multiprocessor to process at least one block. Because of this. the kernel will fail to launch. 4. when a Direct3D application runs in SLI Alternate Frame Rendering mode. The total amount of shared memory Sblock in bytes allocated for a block is as follows:  S block  ceil ( S k . and 512 for devices of compute capability 1.0 79 .Chapter 4: Hardware Implementation are a function of the compute capability of the device and are given in Appendix G.  Rk is the number of registers used by the kernel.0.

80 CUDA Programming Guide Version 3.Chapter 4.0 . In some cases this is not the desired behavior and an application may need to forfeit use of the CUDA-Direct3D interoperability API and manually copy the output of its CUDA work to Direct3D resources using the existing CUDA and Direct3D API. Hardware Implementation configuration).

and the bus connecting the host to the devices.  5. 5.1 Overall Performance Optimization Strategies Performance optimization revolves around three basic strategies: Maximize parallel execution to achieve maximum utilization. for example.6.1 Application Level At a high level. for example using the CUDA profiler. comparing the floating-point operation throughput or memory throughput – whichever makes more sense – of a particular kernel to the corresponding peak theoretical throughput of the device indicates how much room for improvement there is for the kernel. in which case CUDA Programming Guide Version 3.  Optimize memory usage to achieve maximum memory throughput. For the parallel workloads. the devices. Also.2.2. Which strategies will yield the best performance gain for a particular portion of an application depends on the performance limiters for that portion.0 81 .  Optimize instruction usage to achieve maximum instruction throughput. optimizing instruction usage of a kernel that is mostly limited by memory accesses will not yield any significant performance gain. It should assign to each processor the type of work it does best: serial workloads to the host. at points in the algorithm where parallelism is broken because some threads need to synchronize in order to share data with each other. there are two cases: Either these threads belong to the same block. by using asynchronous functions calls and streams as described in Section 3. Optimization efforts should therefore be constantly directed by measuring and monitoring the performance limiters.Chapter 5.2 Maximize Utilization To maximize utilization the application should be structured in a way that it exposes as much parallelism as possible and efficiently maps this parallelism to the various components of the system to keep them busy most of the time. the application should maximize parallel execution between the host. parallel workloads to the devices. Performance Guidelines 5.

0 .x.x since a multiprocessor issues one such instruction per warp over 4 clock cycles.2. As described in Section 4. in which case they must share data through global memory using two separate kernel invocations. The number of clock cycles it takes for a warp to be ready to execute its next instruction is called latency.Appendix E. one for writing to and one for reading from global memory. when the latency of each warp is completely “hidden” by other warps. so maximum utilization can also be achieved by using streams to enable enough kernels to execute concurrently as described in Section 3. a GPU multiprocessor relies on thread-level parallelism to maximize utilization of its functional units.3 Multiprocessor Level At an even lower level.3.6. as mentioned in Section G.2. The second case is much less optimal since it adds the overhead of extra kernel invocations and global memory traffic.2. the application should maximize parallel execution between the various functional units within a multiprocessor. or in other words.1. The most common reason a warp is not ready to execute its next instruction is that the instruction‟s input operands are not yet available.2 Device Level At a lower level. and full utilization is achieved when the warp scheduler always has some instruction to issue for some warp at every clock cycle during that latency period. Its occurrence should therefore be minimized by mapping the algorithm to the CUDA programming model in such a way that the computations that require inter-thread communication are performed within a single thread block as much as possible. For example. so the kernel should be launched with at least as many thread blocks as there are multiprocessors in the device.  82 CUDA Programming Guide Version 3.1. Texture Fetching they should use __syncthreads() and share data through shared memory within the same kernel invocation. How many instructions are required to hide latency depends on the instruction throughput.0 since a multiprocessor issues the two instructions for a pair of warps over 2 clock cycles. multiple kernels can execute concurrently on a device. For devices of compute capability 1. At every instruction issue time. Utilization is therefore directly linked to the number of resident warps. or they belong to different blocks.4.2. only one kernel can execute on a device at one time. as mentioned in Section G. a warp scheduler selects a warp that is ready to execute and issues the next instruction to the active threads of the warp. to hide a latency of L clock cycles with basic single-precision floating-point arithmetic instructions (scheduled on CUDA cores): L/4 (rounded up to nearest integer) instructions are required for devices of compute capability 1.0. 5.  L/2 (rounded up to nearest integer) instructions are required for devices of compute capability 2. For devices of compute capability 2. the application should maximize parallel execution between the multiprocessors of a device. 5.

0. 32 warps) can reside on the multiprocessor since they require 2x512x16 registers. in general. the amount of shared memory used to pass the kernel‟s arguments (see Section B. The number of blocks and warps residing on each multiprocessor for a given kernel call depends on the execution configuration of the call (Section B. the latency is equal to the execution time of the previous instruction and the warp scheduler must schedule instructions for different warps during that time.e. Register usage can be controlled using the -maxrregcount compiler option or launch bounds as described in Section B.14.13).5) or synchronization point (Section B.2. CUDA Programming Guide Version 3. arithmetic instructions most of the time) to the number of instructions with off-chip memory operands is low (this ratio is commonly called the arithmetic intensity of the program).x.x and 11 warps for devices of compute capability 2.. Having multiple resident blocks per multiprocessor can help reduce idling in this case. then to hide latencies of about 600 clock cycles.4). latency is caused by register dependencies. If some input operand resides in off-chip memory. if a kernel uses 16 registers and each block has 512 threads and requires very little shared memory. But as soon as the kernel uses one more register.e. the memory resources of the multiprocessor.2. as warps from different blocks do not need to wait for each other at synchronization points. for devices of compute capability 1. The number of registers used by a kernel can have a significant impact on the number of resident warps.2) and the number of instructions to a minimum. i. and for devices of compute capability 1. some input operand is written by the previous instruction). where occupancy is defined as the ratio of the number of resident warps to the maximum number of resident warps (given in Appendix G for various compute capabilities).. A synchronization point can force the multiprocessor to idle as more and more warps wait for other warps in the same block to complete execution of instructions prior to the synchronization point. the latency is much higher: 400 to 800 clock cycles. for example.. For example. about 15 warps are required for devices of compute capability 1. Another reason a warp is not ready to execute its next instruction is that it is waiting at some memory fence (Section B. To assist programmers in choosing thread block size based on register and shared memory requirements. Therefore. the CUDA Software Development Kit provides a spreadsheet. The total amount of shared memory required for a block is equal to the sum of the amount of statically allocated shared memory. Execution time varies depending on the instruction.e. and the resource requirements of the kernel as described in Section 4. called the CUDA Occupancy Calculator. some of the input operands are written by some previous instruction(s) whose execution has not completed yet. 16 warps) can be resident since two blocks would require 2x512x17 registers. In the case of a back-to-back register dependency (i.. Texture Fetching If all input operands are registers.e.0 83 . If this ratio is 10. the compiler attempts to minimize register usage while keeping register spilling (see Section 5. The number of warps required to keep the warp scheduler busy during such high latency periods depends on the kernel code.2.x and about 30 for devices of compute capability 2.. the amount of dynamically allocated shared memory.e. which exactly matches the number of registers available on the multiprocessor.Appendix E. which is more registers than are available on the multiprocessor. only one block (i. but it is typically about 22 clock cycles.6). more warps are required if the ratio of the number of instructions with no off-chip memory operands (i. which translates to 6 warps for devices of compute capability 1. then two blocks (i.1.3.0.

2.g. That also means minimizing data transfers between global memory and the device by maximizing use of on-chip memory: shared memory and caches (i. The number of threads per block should be chosen as a multiple of the warp size to avoid wasting computing resources with under-populated warps if possible. shared.0. The next step in maximizing 84 CUDA Programming Guide Version 3.2. for devices of compute capability 2.3 Maximize Memory Throughput The first step in maximizing overall memory throughput for the application is to minimize data transfers with low bandwidth.  Process the data in shared memory. the effect of execution configuration on performance for a given kernel call generally depends on the kernel code. in other words. and how much of it is dedicated to L1 versus shared memory is configurable for each kernel call. i. Experimentation is therefore recommended. As mentioned in Section G.2 and higher) and each long long variable uses two registers. since these have much lower bandwidth than data transfers between global memory and the device. which depends on the compute capability of the device. a traditional hardware-managed cache is more appropriate to exploit data locality.e. devices of compute capability 1. As mentioned above. Shared memory is equivalent to a user-managed cache: The application explicitly allocates and accesses it.  Synchronize again if necessary to make sure that shared memory has been updated with the results.4. However. Note that each double variable (on devices that supports native double precision. the same on-chip memory is used for both L1 and shared memory. devices of compute capability 1. texture cache and constant cache available on all devices).0 . a typical programming pattern is to stage data coming from device memory into shared memory.Appendix E. Texture Fetching Register. 5.e.1. as detailed in Section 5.2 and higher have at least twice as many registers per multiprocessor as devices with lower compute capability. for which global memory accesses are data-dependent). as well as on the number of multiprocessors and memory bandwidth of the device. to have each thread of a block: Load data from device memory to shared memory.0. local.1. and constant memory usages are reported by the compiler when compiling with the --ptxas-options=-v option. As illustrated in Section 3.  Write the results back to device memory. L1/L2 caches available on devices of compute capability 2. all of which can be queried using the runtime or driver API (see reference manual). For some applications (e.3. Applications can also parameterize execution configurations based on register file size and shared memory size. That means minimizing data transfers between the host and the device. Synchronize with all the other threads of the block so that each thread can safely read shared memory locations that were populated by different threads.   The throughput of memory accesses by a kernel can vary by an order of magnitude depending on access pattern for each type of memory.

Applications may query whether a device is integrated or not by calling cudaGetDeviceProperties() and checking the integrated property or checking the CU_DEVICE_ATTRIBUTE_INTEGRATED attribute using cuDeviceGetAttribute().0 85 .5.3.Appendix E. CUDA Programming Guide Version 3. constant. so non-optimal global memory accesses have a higher impact on performance.1). 5. In addition. Texture Fetching memory throughput is therefore to organize memory accesses as optimally as possible based on the optimal memory access patterns described in Sections 5. any copy between host and device memory is superfluous and mapped pagelocked memory should be used instead.e. For maximum performance.4.2 Device Memory Accesses An instruction that accesses addressable memory (i. local. On integrated systems where device memory and host memory are physically the same. using mapped page-locked memory instead of explicit copies between device and host memory can be a win for performance.5. Data transfers are implicitly performed each time the kernel accesses the mapped memory.3)..2.2. the more reduced the throughput is. 5. On systems with a front-side bus.2. 5.3. or texture memory) might need to be re-issued multiple times depending on the distribution of the memory addresses across the threads within the warp. Also. operated on by the device.5. the more scattered the addresses are. shared. Intermediate data structures may be created in device memory. there is no need to allocate any device memory and explicitly copy data between device and host memory. and destroyed without ever being mapped by the host or copied to host memory. because of the overhead associated with each transfer. For example. batching many small transfers into a single large transfer always performs better than making each transfer separately.3. How the distribution affects the instruction throughput this way is specific to each type of memory and described in the following sections. Assuming that they are and that the mapped memory is read or written only once.2.3. and 5. higher performance for data transfers between host and device is achieved by using page-locked host memory as described in Section 3. when using mapped page-locked memory (Section 3.3.3. This optimization is especially important for global memory accesses as global memory bandwidth is low.1. these memory accesses must be coalesced as with accesses to global memory (see Section 5.2.2. 5.3. for global memory.1 Data Transfer between Host and Device Applications should strive to minimize data transfer between the host and the device. as a general rule. even if that means running kernels with low parallelism computations. global. One way to accomplish this is to move more code from the host to the device.2.3.

They are much more relaxed for devices of higher compute capabilities. For devices of compute capability 2.Appendix E.0. or 16 bytes. To maximize global memory throughput.2 and G. the requirements on the distribution of the addresses across the threads to get any coalescing at all are very strict.0 .2.3. the memory transactions are cached. The alignment requirement is automatically fulfilled for the built-in types of Section B. 4. it coalesces the memory accesses of the threads within the warp into one or more of these memory transactions depending on the size of the word accessed by each thread and the distribution of the memory addresses across the threads.1.1.1. 2.2. or 128-byte memory transactions. 8.3. It is therefore recommended to use types that meet this requirement for data that resides in global memory. 86 CUDA Programming Guide Version 3.   5. For devices of compute capability 1.1. If this size and alignment requirement is not fulfilled.2 and G. reducing the instruction throughput accordingly.1. 64-. when accessing a two-dimensional array as described in Section 5.0 and 1. whose first address is a multiple of their size) can be read or written by memory transactions. or 16 bytes and the data is naturally aligned (i.3.3.e.e. These memory transactions must be naturally aligned: Only the 32-. Sections G.3. the access compiles to multiple instructions with interleaved access patterns that prevent these instructions from fully coalescing. 2. How many transactions are necessary and how throughput is ultimately affected varies with the compute capability of the device. In general. For structures. Any access (via a variable or a pointer) to data residing in global memory compiles to a single global memory instruction if and only if the size of the data type is 1.  Padding data in some cases.1 Global Memory Global memory resides in device memory and device memory is accessed via 32-.4. if a 32-byte memory transaction is generated for each thread‟s 4-byte access.2 give more details on how global memory accesses are handled for various compute capabilities. 64-.4.2. For example. so data locality is exploited to reduce impact on throughput.2. the more unused words are transferred in addition to the words accessed by the threads. the size and alignment requirements can be enforced by the compiler using the alignment specifiers __align__(8) or __align__(16).3. throughput is divided by 8. When a warp executes an instruction that accesses global memory.1 like float2 or float4. Using data types that meet the size and alignment requirement detailed in Section 5.2. its address is a multiple of that size). it is therefore important to maximize coalescing by: Following the most optimal access patterns based on Sections G.2. for example. Texture Fetching 5. such as struct __align__(8) { float x.1 Size and Alignment Requirement Global memory instructions support reading or writing words of size equal to 1.3. 4. 8. or 128-byte segments of device memory that are aligned to their size (i. the more transactions are necessary.

Texture Fetching float y. so special care must be taken to maintain alignment of the starting address of any value or array of values of these types.5. }. this means that an array whose width is not a multiple of this size will be accessed much more efficiently if it is actually allocated with a width rounded up to the closest multiple of this size and its rows padded accordingly.Appendix E. Reading non-naturally aligned 8-byte or 16-byte words produces incorrect results (off by a few words). Automatic variables that the compiler is likely to place in local memory are: Arrays for which it cannot determine that they are indexed with constant quantities.  Any variable if the kernel uses more registers than available (this is also known as register spilling). Inspection of the PTX assembly code (obtained by compiling with the –ptx or -keep option) will tell if a variable has been placed in local memory during the first  CUDA Programming Guide Version 3. 5.0 87 . located at address BaseAddress of type type* (where type meets the requirement described in Section 5.1. 5.2 Local Memory Local memory accesses only occur for some automatic variables as mentioned in Section B. Any address of a variable residing in global memory or returned by one of the memory allocation routines from the driver or runtime API is always aligned to at least 256 bytes. float z.2. float y. The cudaMallocPitch() and cuMemAllocPitch() functions and associated memory copy functions described in the reference manual enable programmers to write non-hardware-dependent code to allocate arrays that conform to these constraints. both the width of the thread block and the width of the array must be a multiple of the warp size (or only half the warp size for devices of compute capability 1.2.3.ty) uses the following address to access one element of a 2D array of width width.1): BaseAddress + width * ty + tx For these accesses to be fully coalesced.2.3.1. or struct __align__(16) { float x.  Large structures or arrays that would consume too much register space. In particular. }.2 Two-Dimensional Arrays A common global memory access pattern is when each thread of index (tx. in which case the starting address of each array is offset from the block‟s starting address.3. whereby the allocations of multiple arrays (with multiple calls to cudaMalloc() or cuMemAlloc()) is replaced by the allocation of a single large block of memory partitioned into multiple arrays. A typical case where this might be easily overlooked is when using some custom global memory allocation scheme.x).2.

the shared memory space is much faster than the local and global memory spaces.0. On devices of compute capability 2. which can be accessed simultaneously.3. the compiler reports total local memory usage per kernel (lmem) when compiling with the --ptxas-options=-v option. 5. A request is then split into as many separate requests as there are different memory addresses in the initial request.1.4. If the number of separate memory requests is n.4.2.4.3. Local memory is however organized such that consecutive 32-bit words are accessed by consecutive thread IDs. decreasing throughput by a factor equal to the number of separate memory requests.3.Appendix E. decreasing throughput by a factor equal to the number of separate requests.3 for devices of compute capability 1. accessing shared memory is fast as long as there are no bank conflicts between the threads. it is therefore important to understand how memory addresses map to memory banks in order to schedule the memory requests so as to minimize bank conflicts. Note that some mathematical functions have implementation paths that might access local memory. shared memory is divided into equally-sized memory modules.4 Constant Memory The constant memory space resides in device memory and is cached in the constant cache mentioned in Sections G. subsequent compilation phases might still decide otherwise though if they find it consumes too much register space for the targeted architecture: Inspection of the cubin object using cuobjdump will tell if this is the case. yielding an overall bandwidth that is n times as high as the bandwidth of a single module. same index in an array variable.local mnemonics. same member in a structure variable). as detailed below. Any memory read or write request made of n addresses that fall in n distinct memory banks can therefore be serviced simultaneously. Texture Fetching compilation phases as it will be declared using the . Accesses are therefore fully coalesced as long as all threads in a warp access the same relative address (e.3 Shared Memory Because it is on-chip. Also. called banks. 88 CUDA Programming Guide Version 3.2. there is a bank conflict and the access has to be serialized.g. However. The local memory space resides in device memory. To get maximum performance.3.2). The hardware splits a memory request with bank conflicts into as many separate conflict-free requests as necessary. a constant memory request for a warp is first split into two requests. 5.x.3 and G. Even if it has not. so local memory accesses have same high latency and low bandwidth as global memory accesses and are subject to the same requirements for memory coalescing as described in Section 5.2.x and 2. for all threads of a warp. respectively.3.1. For devices of compute capability 1.0 . In fact. local memory accesses are always cached in L1 and L2 in the same way as global memory accesses (see Section G. if two addresses of a memory request fall in the same memory bank.local mnemonic and accessed using the ld.local and st. This is described in Sections G.0. that are issued independently. the initial memory request is said to cause n-way bank conflicts.1 and G. To achieve high bandwidth. one for each half-warp.

Appendix E.

Texture Fetching

The resulting requests are then serviced at the throughput of the constant cache in case of a cache hit, or at the throughput of device memory otherwise.

5.3.2.5

Texture Memory
The texture memory space resides in device memory and is cached in texture cache, so a texture fetch costs one memory read from device memory only on a cache miss, otherwise it just costs one read from texture cache. The texture cache is optimized for 2D spatial locality, so threads of the same warp that read texture addresses that are close together in 2D will achieve best performance. Also, it is designed for streaming fetches with a constant latency; a cache hit reduces DRAM bandwidth demand but not fetch latency. Reading device memory through texture fetching present some benefits that can make it an advantageous alternative to reading device memory from global or constant memory: If the memory reads do not follow the access patterns that global or constant memory reads must respect to get good performance (see Sections 5.3.2.1 and 5.3.2.4), higher bandwidth can be achieved providing that there is locality in the texture fetches (this is less likely for devices of compute capability 2.0 given that global memory reads are cached on these devices);  Addressing calculations are performed outside the kernel by dedicated units;  Packed data may be broadcast to separate variables in a single operation;  8-bit and 16-bit integer input data may be optionally converted to 32-bit floating-point values in the range [0.0, 1.0] or [-1.0, 1.0] (see Section 3.2.4.1). However, within the same kernel call, the texture cache is not kept coherent with respect to global memory writes, so that any texture fetch to an address that has been written to via a global write in the same kernel call returns undefined data. In other words, a thread can safely read via texture some memory location only if this memory location has been updated by a previous kernel call or memory copy, but not if it has been previously updated by the same thread or another thread from the same kernel call. This is only relevant when fetching from linear memory as a kernel cannot write to CUDA arrays in any case.

5.4

Maximize Instruction Throughput
To maximize instruction throughput the application should: Minimize the use of arithmetic instructions with low throughput; this includes trading precision for speed when it does not affect the end result, such as using intrinsic instead of regular functions (intrinsic functions are listed in Section C.2), single-precision instead of double-precision, or flushing denormalized numbers to zero;  Minimize divergent warps caused by control flow instructions as detailed in Section 5.4.2;  Reduce the number of instructions, for example, by optimizing out synchronization points whenever possible as described in Section 5.4.3 or by using restricted pointers as described in Section E.3. In this section, throughputs are given in number of operations per clock cycle per multiprocessor. For a warp size of 32, one instruction results in 32 operations.

CUDA Programming Guide Version 3.0

89

Appendix E.

Texture Fetching Therefore, if T is the number of operations per clock cycle, the instruction throughput is one instruction every 32/T clock cycles. All throughputs are for one multiprocessor. They must be multiplied by the number of multiprocessors in the device to get throughput for the whole device.

5.4.1

Arithmetic Instructions
Table 5-1 gives the throughputs of the arithmetic instructions that are natively supported in hardware for devices of various compute capabilities. For devices of compute capability 2.0, two different warps execute half of the operations each clock cycle (see Section G.4.1).

Table 5-1.

Throughput of Native Arithmetic Instructions (Operations per Clock Cycle per Multiprocessor)
Compute Capability 1.x Compute Capability 2.0
32 16 32 Multiple instructions 32

32-bit floating-point add, multiply, multiply-add 64-bit floating-point add, multiply, multiply-add 32-bit integer add, logical operation, shift, compare 24-bit integer multiply (__[u]mul24(x,y)) 32-bit integer multiply, multiply-add, sum of absolute difference 32-bit floating-point reciprocal, reciprocal square root, base-2 logarithm (__log2f), base-2 exponential (exp2f), sine (__sinf), cosine (__cosf) Type conversions

8 1 8 8 Multiple instructions

2

4

8

32

Other instructions and functions are implemented on top of the native instructions. The implementation may be different for devices of compute capability 1.x and devices of compute capability 2.0, and the number of native instructions after compilation may fluctuate with every compiler version. For complicated functions, there can be multiple code paths depending on input. cuobjdump can be used to inspect a particular implementation in a cubin object. The implementation of some functions are readily available on the CUDA header files (math_functions.h, device_functions.h, …). In general, code compiled with -ftz=true (denormalized numbers are flushed to zero) tends to have higher performance than code compiled with -ftz=false. Similarly, code compiled with -prec-div=false (less precise division) tends to have higher performance code than code compiled with -prec-div=true, and code compiled with -prec-sqrt=false (less precise square root) tends to have

90

CUDA Programming Guide Version 3.0

Appendix E.

Texture Fetching

higher performance than code compiled with -prec-sqrt=true. The nvcc user manual describes these compilation flags in more details.

Single-Precision Floating-Point Addition and Multiplication Intrinsics
__fadd_r[d,u], __fmul_r[d,u], and __fmaf_r[n,z,d,u] (see

Section C.2.1) compile to tens of instructions for devices of compute capability 1.x, but map to a single native instruction for devices of compute capability 2.0.

Single-Precision Floating-Point Division
__fdividef(x, y) (see Section C.2.1) provides faster single-precision floating-

point division than the division operator.

Single-Precision Floating-Point Reciprocal Square Root
To preserve IEEE-754 semantics the compiler can optimize 1.0/sqrtf() into rsqrtf() only when both reciprocal and square root are approximate, (i.e. with -prec-div=false and -prec-sqrt=false). It is therefore recommended to invoke rsqrtf() directly where desired.

Single-Precision Floating-Point Square Root
Single-precision floating-point square root is implemented as a reciprocal square root followed by a reciprocal instead of a reciprocal square root followed by a multiplication so that it gives correct results for 0 and infinity. Therefore, its throughput is 1 operation per clock cycle for devices of compute capability 1.x and 2 operations per clock cycle for devices of compute capability 2.0.

Sine and Cosine
sinf(x), cosf(x), tanf(x), sincosf(x), and corresponding double-

precision instructions are much more expensive and even more so if the argument x is large in magnitude. More precisely, the argument reduction code (see math_functions.h for implementation) comprises two code paths referred to as the fast path and the slow path, respectively. The fast path is used for arguments sufficiently small in magnitude and essentially consists of a few multiply-add operations. The slow path is used for arguments large in magnitude and consists of lengthy computations required to achieve correct results over the entire argument range. At present, the argument reduction code for the trigonometric functions selects the fast path for arguments whose magnitude is less than 48039.0f for the singleprecision functions, and less than 2147483648.0 for the double-precision functions. As the slow path requires more registers than the fast path, an attempt has been made to reduce register pressure in the slow path by storing some intermediate variables in local memory, which may affect performance because of local memory high latency and bandwidth (see Section 5.3.2.2). At present, 28 bytes of local memory are used by single-precision functions, and 44 bytes are used by doubleprecision functions. However, the exact amount is subject to change. Due to the lengthy computations and use of local memory in the slow path, the throughput of these trigonometric functions is lower by one order of magnitude when the slow path reduction is required as opposed to the fast path reduction.

CUDA Programming Guide Version 3.0

91

32-bit integer multiplication is natively supported.3) compile to fewer instructions for devices of compute capability 2. On devices of compute capability 2. the compiler must insert conversion instructions. 0. defined with an f suffix such as 3.2. Using __[u]mul24 instead of the 32-bit multiplication operator whenever possible usually improves performance for instruction bound kernels.2.0 92 .3) compile to tens of instructions for devices of compute capability 1.0 than for devices of compute capability 1.  Double-precision floating-point constants (i. Integer division and modulo operation are particularly costly and should be avoided if possible or replaced with bitwise operations whenever possible: If n is a power of 2.0f. 32-bit integer multiplication is implemented using multiple instructions as it is not natively supported.e. switch. those constants defined without any type suffix) used as input to single-precision floating-point computations (as mandated by C/C++ standards). 24-bit integer multiplication is natively supported however via the __[u]mul24 intrinsic (see Section C. __ffs.4. It can have the opposite effect however in cases where the use of __[u]mul24 inhibits compiler optimizations.2. __brev. but 24-bit integer multiplication is not.0.0 and __brevll and __popcll to just a few. for.3).x. while) can significantly impact the effective instruction throughput by causing threads of the same warp to diverge (i. to follow different execution paths). and __popcll (see Section C. the controlling condition should be written so as to minimize the number of divergent warps. __[u]mul24 is therefore implemented using multiple instructions and should not be used. increasing the total number of instructions executed for this warp.Appendix E.x. __clz.1. __popc. introducing additional execution cycles. This is the case for: Functions operating on variables of type char or short whose operands generally need to be converted to int.x. To obtain best performance in cases where the control flow depends on the thread ID. If this happens.2 Control Flow Instructions Any flow control instruction (if. and __ffsll (see Section C. A trivial example is when the CUDA Programming Guide Version 3. (i/n) is equivalent to (i>>log2(n)) and (i%n) is equivalent to (i&(n-1)).5f. the compiler will perform these conversions if n is literal. the threads converge back to the same execution path.e. __clzll. When all the different execution paths have completed. the different executions paths have to be serialized.  5. 1.141592653589793f. do. __brevll. Texture Fetching Integer Arithmetic On devices of compute capability 1. but __brev and __popc map to a single instruction for devices of compute capability 2. This last case can be avoided by using single-precision floating-point constants. This is possible because the distribution of the warps across the block is deterministic as mentioned in Section 4. Type Conversion Sometimes.

. both calls to __syncthreads() are required to get the expected result (i. Because a warp executes one common instruction at a time. In the following code sample. no warp diverges since the controlling condition is perfectly aligned with the warps. each of them is associated with a per-thread condition code or predicate that is set to true or false based on the controlling condition and although each of these instructions gets scheduled for execution. In this case. __syncthreads().Appendix E. Texture Fetching controlling condition only depends on (threadIdx / warpSize) where warpSize is the warp size.. and also do not evaluate addresses or read operands. The compiler replaces a branch instruction with predicated instructions only if the number of instructions controlled by the branch condition is less or equal to a certain threshold: If the compiler determines that the condition is likely to produce many divergent warps. result[i] = 2 * myArray[i] for i > 0). depending on whether the memory read occurs before or after the memory write from myArray[tid + 1] = 2. only the instructions with a true predicate are actually executed. In these cases. Sometimes. int ref1 = myArray[tid]. no warp can ever diverge. . int ref2 = myArray[tid]. // myArray is an array of integers located in global or shared // memory __global__ void MyKernel(int* result) { int tid = threadIdx.3 Synchronization Instruction Throughput for __syncthreads() is 8 operations per clock cycle for devices of compute capability 1. the compiler may unroll loops or it may optimize out if or switch statements by using branch predication instead. . Instructions with a false predicate do not write results. myArray[tid + 1] = 2..e.. 5.0 93 . any of the two references to myArray[tid] could return either 2 or the value initially stored in myArray.0.4. otherwise it is 4. threads within a warp are implicitly synchronized and this can sometimes be used to omit __syncthreads() for better performance. Instead. as detailed below. Without synchronization. When using branch predication none of the instructions whose execution depends on the controlling condition gets skipped. __syncthreads().2.2). result[tid] = ref1 * ref2.x and 16 operations per clock cycle for devices of compute capability 2. The programmer can also control loop unrolling using the #pragma unroll directive (see Section E. this threshold is 7.x. } CUDA Programming Guide Version 3. Note that __syncthreads() can impact performance by forcing the multiprocessor to idle as detailed in Section 5.3. for example.

.x. Texture Fetching However. result[tid] = ref1 * ref2. 94 CUDA Programming Guide Version 3. if (tid < warpSize) { int ref1 = myArray[tid]. int ref2 = myArray[tid].. // myArray is an array of integers located in global or shared // memory __global__ void MyKernel(int* result) { int tid = threadIdx.0 .2. so that there is no need for any __syncthreads(). myArray[tid + 1] = 2.. . in the following slightly modified code sample. threads are guaranteed to belong to the same warp. myArray must also be declared as volatile as described in Section B.Appendix E. } . } Simply removing the __syncthreads() is not enough however..4.

9800M GTS GeForce GT 220.1 1. 8800 GTX GeForce 9800 GT. GTX 260M. GTS 350M GeForce GT 130. and Number of CUDA Cores Compute Capability GeForce GTX 295 GeForce GTX 285. GTX 285M.3 1. GTS 360M. can be queried using the runtime or driver API (see reference manual). CUDA-Enabled Devices with Compute Capability.2 1. 8800 GT.Appendix A. GT 240M.3 1.2 1.1 12 9 8 6 6 96 72 64 48 48 CUDA Programming Guide Version 3. GTS 260M. 8800 GTS 512. GT 330M. GTS 250M. 9800M GTX GeForce GT 240.1 1. Table A-1.1 1. 9600 GSO.2 1.1 Number of Multiprocessors 2x30 30 24 2x16 16 Number of CUDA Cores 2x240 240 192 2x128 128 1. Number of Multiprocessors. GT 230M 1. 9800 GTX. These.0 1.3 1. GTX 280 GeForce GTX 260 GeForce 9800 GX2 GeForce GTS 250. 8800M GTS. 9800 GTX+. GTS 150. and number of CUDA cores. GTX 280M GeForce 8800 Ultra. as well as the clock frequency and the total amount of device memory. 9800M GT GeForce 8800 GTS GeForce GT 335M GeForce 9600 GT. 8800 GS. CUDA-Enabled GPUs Table A-1 lists all CUDA-enabled devices with their compute capability. 8800M GTX.0 1. number of multiprocessors.1 16 14 12 12 128 112 96 96 1.0 95 . GT 325M GeForce 9700M GT.

1 1. FX 1700M.3 1.1 1.G110M GeForce 9300M GS. 9100M G. NVS 135M. G210M.1 1.3 1. 305M GeForce G100.1 2 2 1 16 16 8 96 CUDA Programming Guide Version 3. 9200M GS. 9400 mGPU.2 1. FX 570M Quadro FX 380 LP. 8600 GTS.0 1.0 1.3 1. 9600M GT. 9500 GT. 8300 mGPU.0 1. 9500M G. 8400M GT.1 1.2 1. NVS 3100M. 8600M GS GeForce 210. NVS 2100M Quadro FX 370. FX 360M Quadro FX 370M.0 1.1 1 4x30 30 4x16 2x16 16 2x30 4x14 4x16 2x16 30 24 2x14 16 16 14 12 12 9 8 6 6 4 8 4x240 240 4x128 2x128 128 2x240 4x112 4x128 2x128 240 192 2x112 128 128 112 96 96 72 64 48 48 32 1. 310M.1 Number of Multiprocessors 4 Number of CUDA Cores 32 1. 9500M GS.3 1.3 1. 8500 GT.0 .Chapter 5. FX 570. 8400M G. 8700M GT. FX 380M. FX 3800M Quadro FX 5600 Quadro FX 3700 Quadro FX 2800M Quadro FX 4600 Quadro FX 1800M Quadro FX 3600M Quadro FX 880M. 8400 GS. 9300M G. 8400M GS.1 1.1 1.2 1. NVS 5100M Quadro FX 2700M Quadro FX 1700. NVS 140M. Performance Guidelines Compute Capability GeForce GT 120. 8600 GT.0 1. 9600M GS. FX 770M.1 1. 9650M GS. NVS 130M 1. 8600M GT.0 1.1 1. 9700M GT. G105M Tesla S1070 Tesla C1060 Tesla S870 Tesla D870 Tesla C870 Quadro Plex 2200 D2 Quadro Plex 2100 D4 Quadro Plex 2100 Model S4 Quadro Plex 1000 Model IV Quadro FX 5800 Quadro FX 4800 Quadro FX 4700 X2 Quadro FX 3700M. 8100 mGPU.2 1. NVS 320M. NVS 290.1 1. 9300 mGPU. 8200 mGPU.0 1. FX 1600M.1 2 2 16 16 1.

3 __host__ The __host__ qualifier declares a function that is: Executed on the host.1 Function Type Qualifiers Function type qualifiers specify whether a function executes on the host or on the device and whether it is callable from the host or from the device. in which case the function is compiled for both the host and the device. C Language Extensions B. in either case the function is compiled for the host only.  Callable from the host only. or __global__ qualifier. __device__. It is equivalent to declare a function with only the __host__ qualifier or to declare it without any of the __host__.0 97 .  B.1.  B. B.1.1 __device__ The __device__ qualifier declares a function that is: Executed on the device  Callable from the device only.  Callable from the host only.  However.Appendix B. the __host__ qualifier can also be used in combination with the __device__ qualifier.1. Such a function is: Executed on the device.2 __global__ The __global__ qualifier declares a function as being a kernel. CUDA Programming Guide Version 3.

Appendix A.

C Language Extensions

B.1.4

Restrictions
__device__ and __global__ functions do not support recursion. __device__ and __global__ functions cannot declare static variables inside

their body.
__device__ and __global__ functions cannot have a variable number of

arguments.
__device__ functions cannot have their address taken; function pointers to __global__ functions, on the other hand, are supported.

The __global__ and __host__ qualifiers cannot be used together.
__global__ functions must have void return type.

Any call to a __global__ function must specify its execution configuration as described in Section B.13. A call to a __global__ function is asynchronous, meaning it returns before the device has completed its execution.
__global__ function parameters are passed to the device:

via shared memory and are limited to 256 bytes on devices of compute capability 1.x,  via constant memory and are limited to 4 KB on devices of compute capability 2.0.

B.2
B.2.1

Variable Type Qualifiers
Variable type qualifiers specify the memory location on the device of a variable.

__device__
The __device__ qualifier declares a variable that resides on the device. At most one of the other type qualifiers defined in the next three sections may be used together with __device__ to further specify which memory space the variable belongs to. If none of them is present, the variable: Resides in global memory space,  Has the lifetime of an application,  Is accessible from all the threads within the grid and from the host through the runtime library.

B.2.2

__constant__
The __constant__ qualifier, optionally used together with __device__, declares a variable that:

Resides in constant memory space,
CUDA Programming Guide Version 3.0

98

Appendix A.

C Language Extensions

Has the lifetime of an application,  Is accessible from all the threads within the grid and from the host through the runtime library.

B.2.3

__shared__
The __shared__ qualifier, optionally used together with __device__, declares a variable that: Resides in the shared memory space of a thread block,  Has the lifetime of the block,  Is only accessible from all the threads within the block. When declaring a variable in shared memory as an external array such as
 extern __shared__ float shared[];

the size of the array is determined at launch time (see Section B.13). All variables declared in this fashion, start at the same address in memory, so that the layout of the variables in the array must be explicitly managed through offsets. For example, if one wants the equivalent of
short array0[128]; float array1[64]; int array2[256];

in dynamically allocated shared memory, one could declare and initialize the arrays the following way:
extern __shared__ char array[]; __device__ void func() // __device__ or __global__ function { short* array0 = (short*)array; float* array1 = (float*)&array0[128]; int* array2 = (int*)&array1[64]; }

Note that pointers need to be aligned to the type they point to, so the following code, for example, does not work since array1 is not aligned to 4 bytes.
extern __shared__ char array[]; __device__ void func() // __device__ or __global__ function { short* array0 = (short*)array; float* array1 = (float*)&array0[127]; }

Alignment requirements for the built-in vector types are listed in Table B-1.

B.2.4

volatile
Only after the execution of a __threadfence_block(), __threadfence(), or __syncthreads() (Sections B.5 and B.6) are prior writes to global or shared memory guaranteed to be visible by other threads. As long as this requirement is met, the compiler is free to optimize reads and writes to global or shared memory. For example, in the code sample below, the first reference to myArray[tid]

CUDA Programming Guide Version 3.0

99

Appendix A.

C Language Extensions compiles into a global or shared memory read instruction, but the second reference does not as the compiler simply reuses the result of the first read.
// myArray is an array of non-zero integers // located in global or shared memory __global__ void MyKernel(int* result) { int tid = threadIdx.x; int ref1 = myArray[tid] * 1; myArray[tid + 1] = 2; int ref2 = myArray[tid] * 1; result[tid] = ref1 * ref2; }

Therefore, ref2 cannot possibly be equal to 2 in thread tid as a result of thread tid-1 overwriting myArray[tid] by 2. This behavior can be changed using the volatile keyword: If a variable located in global or shared memory is declared as volatile, the compiler assumes that its value can be changed at any time by another thread and therefore any reference to this variable compiles to an actual memory read instruction. Note that even if myArray is declared as volatile in the code sample above, there is no guarantee, in general, that ref2 will be equal to 2 in thread tid since thread tid might read myArray[tid] into ref2 before thread tid-1 overwrites its value by 2. Synchronization is required as mentioned in Section 5.4.3.

B.2.5

Restrictions
These qualifiers are not allowed on struct and union members, on formal parameters and on local variables within a function that executes on the host.
__shared__ and __constant__ variables have implied static storage. __device__, __shared__ and __constant__ variables cannot be defined as external using the extern keyword. The only exception is for dynamically allocated __shared__ variables as described in Section B.2.3. __device__ and __constant__ variables are only allowed at file scope. __constant__ variables cannot be assigned to from the device, only from the

host through host runtime functions (Sections 3.2.1 and 3.3.4).
__shared__ variables cannot have an initialization as part of their declaration.

An automatic variable declared in device code without any of these qualifiers generally resides in a register. However in some cases the compiler might choose to place it in local memory, which can have adverse performance consequences as detailed in Section 5.3.2.2. Pointers in code that is executed on the device are supported as long as the compiler is able to resolve whether they point to either the shared memory space or the global memory space, otherwise they are restricted to only point to memory allocated or declared in the global memory space. Dereferencing a pointer either to global or shared memory in code that is executed on the host or to host memory in code that is executed on the device results in an undefined behavior, most often in a segmentation fault and application termination.

100

CUDA Programming Guide Version 3.0

float4. ushort1.Appendix A.3 B. uchar1. ushort4. uint2 int3. ushort1 short2. float3. __shared__ or __constant__ variable can only be used in device code. char3. uchar4 short1.3. uint4 long1.4 can only be used in host code. long4. ulong1. ushort4 int1. z. int4. double1. In host code. uint3. Table B-1. int y). uint4. They all come with a constructor function of the form make_<type name>. uint1 int2. uint1. float1. ushort3 short4. float2. ushort2 short3. double2 These are vector types derived from the basic integer and floating-point types. 3rd. char4.1 Built-in Vector Types char1. 2nd. uchar2 char3. The address of a __device__ or __constant__ variable obtained through cudaGetSymbolAddress() as described in Section 3. uchar3. short1. uchar3 char4. uchar4. uint3 int4. longlong2. This is not always the case in device code as detailed in Table B-1. They are structures and the 1st. long1. short2. y. uchar1 char2. ulong3. ulong1 Alignment 1 2 1 4 2 4 2 8 4 8 4 16 Same as int1 or longlong1 depending on platform CUDA Programming Guide Version 3. and w. ushort2. ushort3. ulong4. B. short4.3. long3. ulong2. int2. the alignment requirement of a vector type is equal to the alignment requirement of its base type. uint2. and 4th components are accessible through the fields x. int1. for example. short3. respectively. char2. C Language Extensions The address obtained by taking the address of a __device__. uchar2. longlong1. int3. long2.0 101 . y). Alignment Requirements in Device Code Type char1. int2 make_int2(int x. which creates a vector of type int2 with value (x.

any component left unspecified is initialized to 1. B.4.2) and contains the dimensions of the grid. B. They are only valid within functions that are executed on the device.3. ulong3 long4.1) and contains the block index within the grid.3. B. ulong4 longlong1 longlong2 float1 float2 float3 float4 double1 double2 Same as int2 or longlong2 depending on platform Same as int3 depending on platform Same as int4 depending on platform 8 16 4 8 4 16 8 16 B.3.4. C Language Extensions long2.1) and contains the thread index within the block.4.Appendix A.4 Built-in Variables Built-in variables specify the grid and block dimensions and the block and thread indices.3.2) and contains the dimensions of the block.2 dim3 This type is an integer vector type based on uint3 that is used to specify dimensions. B. When defining a variable of type dim3.1 gridDim This variable is of type dim3 (see Section B.2 blockIdx This variable is of type uint3 (see Section B. B. ulong2 long3.4 threadIdx This variable is of type uint3 (see Section B.3. 102 CUDA Programming Guide Version 3.3 blockDim This variable is of type dim3 (see Section B.4.0 .

the counter might increment before the partial sum is stored and therefore.4.1 for the definition of a warp). waits until all global and shared memory accesses made by the calling thread prior to __threadfence() are visible to:   All threads in the thread block for shared memory accesses. void __threadfence_system(). waits until all global and shared memory accesses made by the calling thread prior to __threadfence_system() are visible to: All threads in the thread block for shared memory accesses.3). Each block first sums a subset of the array and stores the result in global memory.  All threads in the device for global memory accesses.0 103 .  B. __threadfence_block().5. All threads in the device for global memory accesses. __threadfence().6 Restrictions It is not allowed to take the address of any of the built-in variables. waits until all global and shared memory accesses made by the calling thread prior to __threadfence_block() are visible to all threads in the thread block. when a thread issues a series of writes to memory in a particular order. each block atomically increments a counter to signal that it is done with computing and storing its partial sum (see Section B. C Language Extensions B.x-1. B. void __threadfence(). and __threadfence_system() can be used to enforce some ordering. the last block done reads each of these partial sums from global memory and sums them to obtain the final result. other threads may see the effects of these memory writes in a different order.0. might reach gridDim.  It is not allowed to assign values to any of the built-in variables.  In general. One use case is when threads consume some data produced by other threads as illustrated by the following code sample of a kernel that computes the sum of an array of N numbers in one call. The last block is the one that receives the counter value equal to gridDim. When all blocks are done.x-1 and let CUDA Programming Guide Version 3. __threadfence_system() is only supported by devices of compute capability 2.5 Memory Fence Functions void __threadfence_block().Appendix A.2.4. If no fence is placed between storing the partial sum and incrementing the counter. In order to determine which block is finished last.10 about atomic functions).5 warpSize This variable is of type int and contains the warp size in threads (see Section 4.  Host threads for page-locked host memory accesses (see Section 3.

x).x] = partialSum. if (threadIdx.Appendix A. N). } // Synchronize to make sure that each thread reads // the correct value of isLastBlockDone __syncthreads(). gridDim.. count = 0. float* result) { // Each block sums a subset of the input array float partialSum = calculatePartialSum(array. C Language Extensions the last block start reading partial sums before they have been actually updated in memory.x-1] float totalSum = calculateTotalSum(result).1)).x .6 Synchronization Functions void __syncthreads(). // Thread 0 of each block determines if its block is // the last block to be done isLastBlockDone = (value == (gridDim. } } } B.x == 0) { // Thread 0 of each block stores the partial sum // to global memory result[blockIdx. __shared__ bool isLastBlockDone. if (isLastBlockDone) { // The last block sums the partial sums // stored in result[0 . // Thread 0 makes sure its result is visible to // all other threads __threadfence(). __device__ unsigned int count = 0. unsigned int N.0 . __global__ void sum(const float* array.x == 0) { // Thread 0 of last block stores total sum // to global memory and resets count so that // next kernel call works properly result[0] = totalSum. gridDim. // Thread 0 of each block signals that it is done unsigned int value = atomicInc(&count. if (threadIdx. 104 CUDA Programming Guide Version 3.

int __syncthreads_or(int predicate). write-after-read. __syncthreads() is allowed in conditional code but only if the conditional evaluates identically across the entire thread block. it has the same name prefixed with __ (such as __sinf(x)). __syncthreads() is used to coordinate communication between the threads of the same block. along with their respective error bounds.0 105 . it may also cause some differences in special case handling. but faster version exists in the device runtime component. is identical to __syncthreads() with the additional feature that it evaluates predicate for all threads of the block and returns the number of threads for which predicate evaluates to non-zero. In addition to reduce accuracy of the affected functions. These intrinsic functions are listed in Section C. int __syncthreads_count(int predicate). When executed in host code. The compiler has an option (-use_fast_math) that forces each function in Table B-2 to compile to its intrinsic counterpart.0 support three variations of __syncthreads() described below.2. is identical to __syncthreads() with the additional feature that it evaluates predicate for all threads of the block and returns non-zero if and only if predicate evaluates to non-zero for any of them. int __syncthreads_and(int predicate). When some threads within a block access the same addresses in shared or global memory. along with their respective error bounds. a less accurate. For some of the functions of Section C. Devices of compute capability 2. otherwise the code execution is likely to hang or produce unintended side effects. B. C Language Extensions waits until all threads in the thread block have reached this point and all global and shared memory accesses made by these threads prior to __syncthreads() are visible to all threads in the block. is identical to __syncthreads() with the additional feature that it evaluates predicate for all threads of the block and returns non-zero if and only if predicate evaluates to non-zero for all of them. or write-after-write hazards for some of these memory accesses.Appendix A.1. a given function uses the C runtime implementation if available. Table B-2. Functions Affected by –use_fast_math CUDA Programming Guide Version 3.7 Mathematical Functions Section C.1 contains a comprehensive list of the C/C++ standard library mathematical functions that are currently supported in device code. A more robust approach is to selectively replace mathematical function calls by calls to intrinsic functions only where it is merited by the performance gains and where changed properties such as reduced accuracy and different special case handling can be tolerated. there are potential read-after-write. These data hazards can be avoided by synchronizing threads in-between these accesses.

Mutable attributes are described in Section 3. float tex1Dfetch( texture<signed char.y) Device Function __fdividef(x. runtime) attributes determine how the texture coordinates are interpreted. and depth depending on the compute capability of the device. int x).0 . compile-time) and mutable (i. 1. 1. height. a combination of the texture reference‟s immutable (i. float tex1Dfetch( texture<unsigned short. int x). For 106 CUDA Programming Guide Version 3.8.4.2. int x). C Language Extensions Operator/Function x/y sinf(x) cosf(x) tanf(x) sincosf(x. int x).8 Texture Functions For texture functions. cudaReadModeNormalizedFloat> texRef.4.y) __sinf(x) __cosf(x) __tanf(x) __sincosf(x.sptr. 1. fetch the region of linear memory bound to texture reference texRef using integer texture coordinate x. cudaReadModeNormalizedFloat> texRef. cudaReadModeNormalizedFloat> texRef.2. and the return value delivered by the texture fetch. cudaReadModeElementType> texRef. 1. Immutable attributes are described in Section 3. int x). 1.e. float tex1Dfetch( texture<signed short. what processing occurs during the texture fetch. No texture filtering and addressing modes are supported.1 tex1Dfetch() template<class Type> Type tex1Dfetch( texture<Type.Appendix A. Texture fetching is described in Appendix F. B.e. Appendix G lists the maximum texture width.1.2.y) B.sptr. float tex1Dfetch( texture<unsigned char. cudaReadModeNormalizedFloat> texRef.cptr) logf(x) log2f(x) log10f(x) expf(x) exp10f(x) powf(x.cptr) __logf(x) __log2f(x) __log10f(x) __expf(x) __exp10f(x) __powf(x.

3.8. returns the value of a per-multiprocessor counter that is incremented every clock cycle. 2. y. CUDA Programming Guide Version 3. float y. for example: float4 tex1Dfetch( texture<uchar4.8. 1. The former number is greater that the latter since threads are time sliced. these functions may optionally promote the integer to single-precision floating point. fetches the CUDA array or the region of linear memory bound to texture reference texRef using texture coordinates x and y. fetches the CUDA array bound to texture reference texRef using floating-point texture coordinates x. B. fetches the region of linear memory bound to texture reference texRef using texture coordinate x. when executed in device code. readMode> texRef. 2-. taking the difference of the two samples. but not of the number of clock cycles the device actually spent executing thread instructions. float z). readMode> texRef. and recording the result per thread provides a measure for each thread of the number of clock cycles taken by the device to completely execute the thread. and z. enum cudaTextureReadMode readMode> Type tex3D(texture<Type. fetches the CUDA array bound to texture reference texRef using texture coordinates x. Besides the functions shown above. float x).Appendix A. float y). C Language Extensions integer types. Sampling this counter at the beginning and at the end of a kernel. float x. cudaReadModeNormalizedFloat> texRef.2 tex1D() template<class Type. B. B.0 107 .3 tex2D() template<class Type. 1. enum cudaTextureReadMode readMode> Type tex1D(texture<Type. readMode> texRef.9 Time Function clock_t clock().4 tex3D() template<class Type. int x). enum cudaTextureReadMode readMode> Type tex2D(texture<Type.8. B. and 4-tuples are supported. float x.

0 and atomicExch() for all devices. int val). computes (old . unsigned int val). 108 CUDA Programming Guide Version 3. unsigned long long int atomicAdd(unsigned long long int* address. computes (old + val). The floating-point version of atomicAdd() is only supported by devices of compute capability 2. The function returns old. unsigned long long int val).val).1 Arithmetic Functions atomicAdd() int atomicAdd(int* address. B. int val). adds a number to it. reads the 32-bit or 64-bit word old located at the address address in global or shared memory.2 and above. The operation is atomic in the sense that it is guaranteed to be performed without interference from other threads. Atomic operations only work with signed and unsigned integers with the exception of atomicAdd() for devices of compute capability 2.Appendix A. and writes the result back to the same address.10. and stores the result back to memory at the same address. These three operations are performed in one atomic transaction.10. B. In other words. The function returns old.3) are not atomic from the point of view of the host or other devices. 64-bit words are only supported for global memory.1. C Language Extensions B. float val).2 atomicSub() int atomicSub(int* address. For example.1 and above.5. Note that atomic functions operating on mapped page-locked memory (Section 3.0. atomicAdd() reads a 32-bit word at some address in global or shared memory. which also work for single-precision floating-point numbers. unsigned int val).0 . unsigned int atomicSub(unsigned int* address. unsigned int atomicAdd(unsigned int* address. no other thread can access this address until the operation is complete. reads the 32-bit word old located at the address address in global or shared memory. Atomic functions operating on shared memory and atomic functions operating on 64-bit words are only available for devices of compute capability 1.2.1 B. Atomic functions can only be used in device functions and are only available for devices of compute capability 1. These three operations are performed in one atomic transaction.10. float atomicAdd(float* address.1. and stores the result back to memory at the same address.10 Atomic Functions An atomic function performs a read-modify-write atomic operation on one 32-bit or 64-bit word residing in global or shared memory.

reads the 32-bit word old located at the address address in global or shared memory.10. The function returns old.4 atomicMin() int atomicMin(int* address. unsigned int val).10.1. int val). and stores the result back to memory at the same address. These three operations are performed in one atomic transaction. computes ((old >= val) ? 0 : (old+1)). and stores the result back to memory at the same address. float atomicExch(float* address.5 atomicMax() int atomicMax(int* address. The function returns old. These three operations are performed in one atomic transaction. and stores the result back to memory at the same address. 64-bit words are only supported for global memory.1. The function returns old.3 atomicExch() int atomicExch(int* address. float val).1. These three operations are performed in one atomic transaction. int val). B. unsigned int val).Appendix A. unsigned int atomicMin(unsigned int* address. computes the minimum of old and val. B. B. These three operations are performed in one atomic transaction.7 atomicDec() unsigned int atomicDec(unsigned int* address. unsigned int atomicMax(unsigned int* address. The function returns old. reads the 32-bit word old located at the address address in global or shared memory. computes the maximum of old and val.1. int compare. reads the 32-bit or 64-bit word old located at the address address in global or shared memory and stores val back to memory at the same address.10. These two operations are performed in one atomic transaction.1.6 atomicInc() unsigned int atomicInc(unsigned int* address.0 109 . reads the 32-bit word old located at the address address in global or shared memory. unsigned int atomicExch(unsigned int* address. int val). and stores the result back to memory at the same address. B. C Language Extensions B.10. The function returns old. unsigned long long int val). int* address. unsigned int val). int val). unsigned long long int atomicExch(unsigned long long int* address.8 atomicCAS() int atomicCAS(int* address.1. reads the 32-bit word old located at the address address in global or shared memory. computes (((old == 0) | (old > val)) ? val : (old-1)). unsigned int val). int val). int unsigned int atomicCAS(unsigned unsigned unsigned compare. unsigned int val). CUDA Programming Guide Version 3.10.10. B.

computes (old == compare ? val : old). B. The function returns old (Compare And Swap). 64-bit words are only supported for global memory. computes (old | val). unsigned int val). int val). computes (old & val). int __all(int predicate). These three operations are performed in one atomic transaction. B. These three operations are performed in one atomic transaction.2 B. int val).0 . and stores the result back to memory at the same address.3 atomicXor() int atomicXor(int* address. reads the 32-bit word old located at the address address in global or shared memory. The function returns old. and stores the result back to memory at the same address. These three operations are performed in one atomic transaction.2 atomicOr() int atomicOr(int* address.11 Warp Vote Functions Warp vote functions are only supported by devices of compute capability 1. B. 110 CUDA Programming Guide Version 3.2.2. and stores the result back to memory at the same address. The function returns old.2.Appendix A. reads the 32-bit or 64-bit word old located at the address address in global or shared memory. and stores the result back to memory at the same address. unsigned int val). unsigned int atomicOr(unsigned int* address.10. unsigned int atomicXor(unsigned int* address. unsigned long long int compare. computes (old ^ val). These three operations are performed in one atomic transaction. reads the 32-bit word old located at the address address in global or shared memory. unsigned long long int val). C Language Extensions unsigned long long int atomicCAS(unsigned long long int* address. int val).10.10. unsigned int atomicAnd(unsigned int* address. reads the 32-bit word old located at the address address in global or shared memory.10.2 and higher (see Section 4. The function returns old. unsigned int val).1 for the definition of a warp).1 Bitwise Functions atomicAnd() int atomicAnd(int* address. B.

B. Db. etc.y equals the number of blocks being launched.z equals the number of threads  per block. prof_trigger_01. unsigned int __ballot(int predicate).3. When using the driver API. C Language Extensions evaluates predicate for all threads of the warp and returns non-zero if and only if predicate evaluates to non-zero for all of them.y * Db.Appendix A.x * Db. Dg.z must be equal to 1. When using the runtime API (Section 3. This function is only supported by devices of compute capability 2.2).3. the execution configuration is specified by inserting an expression of the form <<< Dg.13 Execution Configuration Any call to a __global__ function must specify the execution configuration for that call.3. Counters 8 to 15 are reserved and should not be used by applications. S >>> between the function name and the parenthesized argument list. The value of counters 0. Db is of type dim3 (see Section B. the execution configuration is specified through a series of driver function calls as detailed in Section 3. B. 7 for the first multiprocessor can be obtained via the CUDA profiler by listing prof_trigger_00. as well as the associated stream (see Section 3.3. this dynamically allocated memory is used by any of the CUDA Programming Guide Version 3.2) and specifies the dimension and size of each block. increments by one per warp the per-multiprocessor hardware counter of index counter. where:  Dg is of type dim3 (see Section B.2) and specifies the dimension and size of the grid.0 111 .12 Profiler Counter Function Each multiprocessor has a set of sixteen hardware counters that an application can increment with a single instruction by calling the __prof_trigger() function.1 for a description of streams).9. Ns. The execution configuration defines the dimension of the grid and blocks that will be used to execute the function on the device.conf file (see the profiler manual for more details). such that Db. such that Dg. evaluates predicate for all threads of the warp and returns an integer whose Nth bit is set if and only if predicate evaluates to non-zero for the Nth thread of the warp. 1. …. evaluates predicate for all threads of the warp and returns non-zero if and only if predicate evaluates to non-zero for any of them.0. int __any(int predicate).  Ns is of type size_t and specifies the number of bytes in shared memory that is dynamically allocated per block for this call in addition to the statically allocated memory. All counters are reset before each kernel call (note that when an application is run via the CUDA debugger or the CUDA profiler. all launches are synchronous). void __prof_trigger(int counter). …. in the profiler.3.x * Dg. prof_trigger_07.

functions arguments (for devices of compute capability 1. a function declared as __global__ void Func(float* parameter).. }  maxThreadsPerBlock specifies the maximum number of threads per block with which the application will ever launch MyKernel().Appendix A. minus the amount of shared memory required for static allocation. it compiles to the .0 . minBlocksPerMultiprocessor is optional and specifies the desired  minimum number of resident blocks per multiprocessor. the more threads and thread blocks are likely to reside on a multiprocessor. B.3.2 for the relationship between the number of registers used by a kernel and the number of registers 112 CUDA Programming Guide Version 3. Ns is an optional argument which defaults to 0. If launch bounds are specified. C Language Extensions variables declared as an external array as mentioned in Section B. Ns >>>(parameter).2.3..3. As an example.minnctapersm PTX directive.. and execution configuration.maxntid PTX directive. the compiler uses heuristics to minimize register usage while keeping register spilling (see Section 5. The function call will fail if Dg or Db are greater than the maximum sizes allowed for the device as specified in Appendix G. must be called like this: Func<<< Dg.2.2. Therefore.. S is an optional argument which defaults to 0. the compiler first derives from them the upper limit L on the number of registers the kernel should use to ensure that minBlocksPerMultiprocessor blocks (or a single block if minBlocksPerMultiprocessor is not specified) of maxThreadsPerBlock threads can reside on the multiprocessor (see Section 4. The arguments to the execution configuration are evaluated before the actual function arguments and like the function arguments. or if Ns is greater than the maximum amount of shared memory available on the device.2) and instruction count to a minimum.x). Db. the fewer registers a kernel uses.14 Launch Bounds As discussed in detail in Section 5. are currently passed via shared memory to the device. which can improve performance. it compiles to the .) { . minBlocksPerMultiprocessor) MyKernel(. An application can optionally aid these heuristics by providing additional information to the compiler in the form of launch bounds that are specified using the __launch_bounds__() qualifier in the definition of a __global__ function: __global__ void __launch_bounds__(maxThreadsPerBlock.  S is of type cudaStream_t and specifies the associated stream.

). Instead the number of threads per block should be determined: CUDA Programming Guide Version 3.3) and then applies similar heuristics as when no launch bounds are specified.1. This will not work however since __CUDA_ARCH__ is undefined in host code as mentioned in Section 3. when using one less register makes room for an additional resident block as in the example of Section 5.  Optimal launch bounds for a given kernel will usually differ across major architecture revisions. The sample code below shows how this is typically handled in device code using the __CUDA_ARCH__ macro introduced in Section 3.0 113 . usually at the expense of more local memory usage and/or higher number of instructions. it is tempting to use MY_KERNEL_MAX_THREADS as the number of threads per block in the execution configuration: // Host code MyKernel<<<blocksPerGrid.e. so MyKernel will launch with 256 threads per block even when __CUDA_ARCH__ is greater or equal to 200.. #define THREADS_PER_BLOCK #if __CUDA_ARCH__ >= 200 #define MY_KERNEL_MAX_THREADS #define MY_KERNEL_MIN_BLOCKS #else #define MY_KERNEL_MAX_THREADS #define MY_KERNEL_MIN_BLOCKS #endif 256 (2 * THREADS_PER_BLOCK) 3 THREADS_PER_BLOCK 2 // Device code __global__ void __launch_bounds__(MY_KERNEL_MAX_THREADS.. the compiler may increase register usage as high as L to reduce the number of instructions and better hide single thread instruction latency. The compiler then optimizes register usage in the following way: If the initial register usage is higher than L. C Language Extensions allocated per block). A kernel will fail to launch if it is executed with more threads per block than its launch bound maxThreadsPerBlock.Appendix A...4.) { .. the compiler reduces it further until it becomes less or equal to L. } In the common case where MyKernel is invoked with the maximum number of threads per block (specified as the first parameter of __launch_bounds__()).  If both minBlocksPerMultiprocessor and maxThreadsPerBlock are specified.2.  If maxThreadsPerBlock is specified and minBlocksPerMultiprocessor is not.1. MY_KERNEL_MIN_BLOCKS) MyKernel(.4.. the compiler uses maxThreadsPerBlock to determine the register usage thresholds for the transitions between n and n+1 resident blocks (i.  If the initial register usage is lower than L. MY_KERNEL_MAX_THREADS>>>(.

The number of resident blocks can be derived from the occupancy reported by the CUDA profiler (see Section 5.2.. C Language Extensions  Either at compile time using a macro that does not depend on __CUDA_ARCH__. The value of -maxrregcount is ignored for functions with launch bounds. device).. 114 CUDA Programming Guide Version 3.Appendix A. for example // Host code MyKernel<<<blocksPerGrid.).0 . MyKernel<<<blocksPerGrid..). int threadsPerBlock = (deviceProp.major >= 2 ? 2 * THREADS_PER_BLOCK : THREADS_PER_BLOCK). Register usage can also be controlled for all __global__ functions in a file using the -maxrregcount compiler option. threadsPerBlock>>>(.  Or at runtime based on the compute capability // Host code cudaGetDeviceProperties(&deviceProp. THREADS_PER_BLOCK>>>(.3 for a definition of occupancy). Register usage is reported by the --ptxas-options=-v compiler option..

with the result being a single-precision floating-point number is rintf(). there are three prototypes for a given function <func-name>: (1) double <func-name>(double).2).g. Mathematical Functions Functions from Section C.2 can only be used in device code. CUDA Programming Guide Version 3. Note that floating-point functions are overloaded. This combination can be avoided by using the __fadd_rn() and __fmul_rn() intrinsic functions (see Section C.1 Single-Precision Floating-Point Functions Addition and multiplication are IEEE-compliant. truncf(). C.x. e. on the device. the compiler often combines them into a single multiply-add instruction (FMAD) and for devices of compute capability 1.5 ulp. and floorf() each map to a single instruction as well.1 can be used in both host and device code whereas functions from Section C. ceilf(). float log(float) (3) float <func-name>f(float). so they are not guaranteed bounds. whereas rintf() maps to a single instruction.Appendix C. not roundf().g. It also specifies the error bounds of each function when executed on the device. so that in general.1 Standard Functions This section lists all the mathematical standard library functions supported in device code.2. float logf(float) This means.g. However. C. e. e. The recommended way to round a single-precision floating-point operand to an integer. FMAD truncates the intermediate result of the multiplication as mentioned in Section G. double log(double) (2) float <func-name>(float). in particular. so have a maximum error of 0.0 115 . These error bounds also apply when the function is executed on the host in the case where the host does not supply the function. that passing a float argument always results in a float result (variants (2) and (3) above). The reason is that roundf() maps to an 8-instruction sequence on the device.1. They are generated from extensive but not exhaustive tests.

0 for compute capability ≥ 2 when compiled with -prec-sqrt=true 3 (full range). otherwise 1 (full range) 2 (full range) 3 (full range) 2 (full range) 2 (full range) 2 (full range) 1 (full range) 1 (full range) 3 (full range) 3 (full range) 2 (full range) 2 (full range) 2 (full range) 4 (full range) 2 (full range) 2 (full range) 4 (full range) 3 (full range) 2 (full range) 3 (full range) 3 (full range) 2 (full range) x*y x/y 1/x rsqrtf(x) 1/sqrtf(x) sqrtf(x) cbrtf(x) rcbrtf(x) hypotf(x.Appendix B.x) sinhf(x) coshf(x) 116 CUDA Programming Guide Version 3. Mathematical Standard Library Functions with Maximum ULP Error The maximum error is stated as the absolute value of the difference in ulps between a correctly rounded single-precision result and the result returned by the CUDA library function.y) expf(x) exp2f(x) exp10f(x) expm1f(x) logf(x) log2f(x) log10f(x) log1pf(x) sinf(x) cosf(x) tanf(x) sincosf(x.x when addition is merged into an FMAD) 0 (IEEE-754 round-to-nearest-even) (except for devices of compute capability 1. Mathematical Functions Table C-1. otherwise 2 (full range) Applies to 1/sqrtf(x) only when it is converted to rsqrtf(x) by the compiler. Function x+y Maximum ulp error 0 (IEEE-754 round-to-nearest-even) (except for devices of compute capability 1.cptr) sinpif(x) asinf(x) acosf(x) atanf(x) atan2f(y.sptr.0 . otherwise 0 for compute capability ≥ 2 when compiled with -prec-div=true 1 (full range).x when multiplication is merged into an FMAD) 0 for compute capability ≥ 2 when compiled with -prec-div=true 2 (full range).

y) fmaxf(x. -2.001 .z) frexpf(x.y) truncf(x) roundf(x) rintf(x) nearbyintf(x) ceilf(x) floorf(x) lrintf(x) lroundf(x) llrintf(x) llroundf(x) signbit(x) isinf(x) isnan(x) isfinite(x) copysignf(x.y) erff(x) erfcf(x) erfinvf(x) erfcinvf(x) lgammaf(x) tgammaf(x) fmaf(x..exp) scalbnf(x.exp) ldexpf(x.y) fabsf(x) nanf(cptr) Maximum ulp error 2 (full range) 3 (full range) 4 (full range) 3 (full range) 8 (full range) 3 (full range) 6 (full range) 5 (full range) 7 (full range) 6 (outside interval -10.y) remainderf(x. Mathematical Functions Function tanhf(x) asinhf(x) acoshf(x) atanhf(x) powf(x.264.n) scalblnf(x. larger inside) 11 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) N/A N/A N/A N/A N/A N/A N/A N/A N/A CUDA Programming Guide Version 3.l) logbf(x) ilogbf(x) fmodf(x.iptr) modff(x.0 117 .y.iptr) fdimf(x.y) remquof(x.y.Appendix B.y) fminf(x..

When compiling for devices without such support. with the result being a double-precision floating-point number is rint().sptr. and floor() each map to a single instruction as well. not round(). whereas rint() maps to a single instruction.2 Double-Precision Floating-Point Functions The errors listed below only apply when compiling for devices with native doubleprecision support.y) exp(x) exp2(x) exp10(x) expm1(x) log(x) log2(x) log10(x) log1p(x) sin(x) cos(x) tan(x) sincos(x. The recommended way to round a double-precision floating-point operand to an integer.Appendix B. The reason is that round() maps to an 8-instruction sequence on the device.0 .y) Maximum ulp error N/A C. Table C-2. trunc().2 and lower. the double type gets demoted to float by default and the double-precision math functions are mapped to their single-precision equivalents. such as devices of compute capability 1. Mathematical Standard Library Functions with Maximum ULP Error The maximum error is stated as the absolute value of the difference in ulps between a correctly rounded double-precision result and the result returned by the CUDA library function. Function x+y x*y x/y 1/x sqrt(x) rsqrt(x) cbrt(x) rcbrt(x) hypot(x.1.cptr) sinpi(x) Maximum ulp error 0 (IEEE-754 round-to-nearest-even) 0 (IEEE-754 round-to-nearest-even) 0 (IEEE-754 round-to-nearest-even) 0 (IEEE-754 round-to-nearest-even) 0 (IEEE-754 round-to-nearest-even) 1 (full range) 1 (full range) 1 (full range) 2 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 2 (full range) 2 (full range) 2 (full range) 2 (full range) 2 (full range) 118 CUDA Programming Guide Version 3. Mathematical Functions Function nextafterf(x. ceil().

y) trunc(x) round(x) rint(x) nearbyint(x) ceil(x) floor(x) lrint(x) lround(x) llrint(x) llround(x) signbit(x) isinf(x) isnan(x) Maximum ulp error 2 (full range) 2 (full range) 2 (full range) 2 (full range) 1 (full range) 1 (full range) 1 (full range) 2 (full range) 2 (full range) 2 (full range) 2 (full range) 2 (full range) 5 (full range) 8 (full range) 8 (full range) 4 (outside interval -11.exp) ldexp(x.0 119 .l) logb(x) ilogb(x) fmod(x.y) erf(x) erfc(x) erfinv(x) erfcinv(x) lgamma(x) tgamma(x) fma(x.x) sinh(x) cosh(x) tanh(x) asinh(x) acosh(x) atanh(x) pow(x.iptr) fdim(x.n) scalbln(x.Appendix B.y) remainder(x..y) remquo(x.0001 .z) frexp(x. Mathematical Functions Function asin(x) acos(x) atan(x) atan2(y.y. -2.iptr) modf(x. larger inside) 8 (full range) 0 (IEEE-754 round-to-nearest-even) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) N/A N/A N/A CUDA Programming Guide Version 3.exp) scalbn(x..2637.y.

0 . Functions suffixed with _rd operate using the round-down (to negative infinity) rounding mode. __int_as_float(0xC0000000) is equal to -2.y) delivers a result of zero.y) fmin(x. they have the same name prefixed with __ (such as __sinf(x)).y) are supported and map to a single instruction on the device. Functions suffixed with _rn operate using the round-to-nearest-even rounding mode. Both the regular floating-point division and __fdividef(x.2 Intrinsic Functions This section lists the intrinsic functions that are only supported in device code.3 Integer Functions Integer min(x.y) Maximum ulp error N/A N/A N/A N/A N/A N/A N/A C. For example.1.y) fmax(x. Among these functions are the less accurate.0f) is equal to 0x3f800000. but for 2126 < y < 2128.y) and max(x. Unlike type conversion functions (such as __int2float_rn) that convert from one type to another.1.Appendix B.2. whereas the regular division delivers the correct result to within the accuracy stated in Table C-3.y) delivers 120 CUDA Programming Guide Version 3. if x is infinity. C. additions and multiplications generated from the '*' and '+' operators will frequently be combined into FMADs. type casting functions simply perform a type cast on the argument. Mathematical Functions Function isfinite(x) copysign(x. Functions suffixed with _rz operate using the round-towards-zero rounding mode. leaving the value unchanged. for 2126 < y < 2128. By contrast. __float_as_int(1. Functions suffixed with _ru operate using the round-up (to positive infinity) rounding mode. __fdividef(x.y) fabs(x) nan(cptr) nextafter(x. __fdividef(x.1 Single-Precision Floating-Point Functions __fadd_rn() and __fmul_rn() map to addition and multiplication operations that the compiler never merges into FMADs. Also.y) have the same accuracy. C. but faster versions of some of the functions of Section C.

the maximum ulp error is 2.ru. __float2ll_[rn. IEEE-compliant. y) __int_as_float(x) __float_as_int(x) __saturate(x) __float2int_[rn. 2]. 2]. otherwise. 1 if x is more than 1. Table C-3. the maximum absolute error is 2-21.y) __expf(x) __exp10f(x) __logf(x) Error bounds IEEE-compliant.y) __fdividef(x.sptr. the maximum absolute error is 2-22.rz.rd](x) __fdiv_[rn.y) __fmaf_[rn. For x in [0.y) __fmul_[rn. N/A N/A N/A N/A N/A __log2f(x) __log10f(x) __sinf(x) __cosf(x) __sincosf(x.5.rd](x. and x otherwise. __saturate(x) returns 0 if x is less than 0. IEEE-compliant.41.cptr) __tanf(x) __powf(x.rz. Mathematical Functions a NaN (as a result of multiplying infinity by zero). ]. while the regular division returns infinity.rz. IEEE-compliant.ru. Derived from its implementation as Derived from its implementation as exp2f(y * __log2f(x)).ru. For x in [-.ru. otherwise.rz.rz.95 * x)).y.rd](x.rd](x.rz. For y in [2-126.41. The maximum ulp error is The maximum ulp error is 2 + floor(abs(1.rz. the maximum ulp error is 3. 2]. ].rz.ru.0 121 . CUDA Programming Guide Version 3. For x in [-.16 * x)). and larger otherwise.ru. and larger otherwise. Same as sinf(x) and cosf(x).rd](x. the maximum absolute error is 2-21.ru. the maximum ulp error is 2. the maximum absolute error is 2-21.z) __frcp_[rn.rz. IEEE-compliant.5.rd](x) __float2uint_[rn. For x in [0. the maximum absolute error is 2-24. IEEE-compliant.ru.rd](x) __fsqrt_[rn.rd](x) (respectively __float2ull_[rn. otherwise. Single-Precision Floating-Point Intrinsic Functions Supported by the CUDA Runtime Library with Respective Error Bounds Function __fadd_[rn.rd](x)) converts single-precision floating-point parameter x to 64-bit signed (respectively unsigned) integer with specified IEEE- 754 rounding modes.ru. For x in [0. 2126]. 2 + floor(abs(2.5.19.ru.rd](x) __sinf(x) * (1 / __cosf(x)).rz. the maximum ulp error is 3.Appendix B.

rd](x) __ull2float_[rn.y.ru.rd](x) __double2uint_[rn.rz. IEEE-compliant. Double-Precision Floating-Point Intrinsic Functions Supported by the CUDA Runtime Library with Respective Error Bounds Function __dadd_[rn.rd](x.rz.rz. IEEE-compliant.rz.rz. Requires compute capability ≥ 2 N/A 122 CUDA Programming Guide Version 3.rz.ru.z) __double2float_[rn. ys) __ddiv_[rn.ru.ru.rz. Table C-4.rd](x.rz.rd](x) Error bounds IEEE-compliant.Appendix B.y) __dmul_[rn. additions and multiplications generated from the '*' and '+' operators will frequently be combined into FMADs.rd](x) __double2float_[ru.y)(x.rz.ru.ru.y) __drcp_[rn. N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A IEEE-compliant.rz.ru.rz.rz](x) __double2int_[rn.rz.rd](x) N/A N/A N/A N/A N/A N/A C.ru.rd](x) __float2ull_[rn.rz.2 Double-Precision Floating-Point Functions __dadd_rn() and __dmul_rn() map to addition and multiplication operations that the compiler never merges into FMADs.rz.rd](x) __uint2float_[rn.rd](x) __double_as_longlong(x) __longlong_as_double(x) __double2hiint(x) __double2loint(x) __hiloint2double(x.ru.ru.rd](x.2.ru.rd](x) __int2double_rn(x) __uint2double_rn(x) __ll2double_[rn.ru. IEEE-compliant.ru.rz.ru.ru.rz.y) __fma_[rn. Requires compute capability ≥ 2.rd](x) __dsqrt_[rn.ru.rz.rd](x) __float2ll_[rn.rd](x) __double2ull_[rn. Requires compute capability ≥ 2 IEEE-compliant.rz.rd](x.ru.0 .rd](x) __double2ll_[rn. By contrast.rd](x) __ull2double_[rn. Mathematical Functions __int2float_[rn.ru.rd](x) __ll2float_[rn.

Note that this is identical to the Linux function ffs. between 0 and 64 inclusive.Appendix B.y. Mathematical Functions C.e. __ffs(x) returns the position of the first (least significant) bit set in integer parameter x. CUDA Programming Guide Version 3. i. i. __[u]sad(x.y) computes the product of the 64-bit integer parameters x and y and delivers the 64 most significant bits of the 128-bit result. __ffsll(x) returns the position of the first (least significant) bit set in 64-bit integer parameter x. of consecutive zero bits starting at the most significant bit (i. __clzll(x) returns the number.e. If x is 0.z) (Sum of Absolute Difference) returns the sum of integer parameter z and the absolute value of the difference between integer parameters x and y.3 Integer Functions __[u]mul24(x. bit N of the result corresponds to bit 63-N of x. The least significant bit is position 1. The least significant bit is position 1.e. between 0 and 32 inclusive. __popc(x) returns the number of bits that are set to 1 in the binary representation of 32-bit integer parameter x.0 123 . __popcll(x) returns the number of bits that are set to 1 in the binary representation of 64-bit integer parameter x. __clz(x) returns the number. bit N of the result corresponds to bit 31-N of x. __[u]mul64hi(x. __ffs() returns 0. __brev(x) reverses the bits of 32-bit unsigned integer parameter x.y) computes the product of the integer parameters x and y and delivers the 32 most significant bits of the 64-bit result. If x is 0. of consecutive zero bits starting at the most significant bit (i. bit 31) of integer parameter x.2.y) computes the product of the 24 least significant bits of the integer parameters x and y and delivers the 32 least significant bits of the result. The 8 most significant bits of x or y are ignored. __ffsll() returns 0. Note that this is identical to the Linux function ffsll.e. bit 63) of 64-bit integer parameter x. __[u]mulhi(x. __brevll(x) reverses the bits of 64-bit unsigned long long parameter x.

.

1 Polymorphism Generally.0 125 . C++ Language Constructs CUDA supports the following C++ language constructs for device code: Polymorphism  Default Parameters  Operator Overloading  Namespaces  Function Templates  Classes for devices of compute capability 2.Appendix D. It is valid to use any of these constructs in . polymorphism is the ability to define that functions or operators behave differently in different contexts. a compiler may encounter multiple potential matches for a function invocation and in that case the matching rules as described in the C++ Language Standard apply.  The following subsections provide examples of the various constructs. Example: The following is valid CUDA code: __device__ void f(float x) { CUDA Programming Guide Version 3. In practical terms. This is also referred to as function (and operator. Because of implicit typecasting. device. this means that it is permissible to define two different functions within the same scope (namespace) as long as they have a distinguishable function signature. That means that the two functions either consume a different number of parameters or parameters of different types.cu CUDA files for host. When either of the multiple functions gets invoked the compiler resolves to the function‟s implementation that matches the function signature. like the lack of support for recursion. In practice this means that the compiler will pick the closest match in case of multiple potential matches. D. see below) overloading. Any restrictions detailed in previous parts of this programming guide. still apply.0 These C++ constructs are implemented as specified in “The C++ Programming Langue” reference. and kernel (__global__) functions.

[]. double y) { // do something with x and y } D. /. *. b.x = a. D. f(x).2 Default Parameters With support for polymorphism as described in the previous subsection and the function signature matching rules in place it becomes possible to provide support for default values for function parameters. etc. c. r. -. Example: The following is valid CUDA code.0f) { // do something with x } Kernel or other device functions can now invoke this version of f in one of two ways: f(). implementing the + operation between two uchar4 vectors: __device__ uchar4 operator+ (const uchar4 & a. 126 CUDA Programming Guide Version 3.x. .Appendix C. +=. Default parameters can only be given for the last n parameters of a function.0 .. Examples of overloadable operators in C++ are: +. return r. Example: __device__ void f(float x = 0. a = b = /* some initial value */. // or float x = /* some value */..x + b. const uchar4 & b) { uchar4 r. C++ Language Constructs // do something with x } __device__ void f(int i) { // do something with i } __device__ void f(double x.3 Operator Overloading Operator overloading allows programmers to define operators for new data-types. &. } This new operator can now be used like this: uchar4 a.

 Template specialization. Explicit instantiation. CUDA supports function templates to the full extent of the C++ standard.5f). D. The use of namespaces can be used to solve the problem of name-clashes (two different symbols using identical names).   Example: template <T> __device__ bool f(T x) { return /* some clever code that turns x into a bool here */ } This function will convert x of any data-type to a bool as long as the code in the function‟s body can be compiled for the actually type (T) of the variable x.4 Namespaces Namespaces in C++ allow for the creation of a hierarchy of scopes of visibility. f() can be invoked in two ways: int x = 1. All the symbols inside a namespace can be used within this namespaces without additional syntax.} } The functions can now be used anywhere via fully qualified names: nvidia::f(0.5 Function Templates Function templates are a form of meta-programming that allows writing a generic function in a data-type independent fashion. which commonly occurs when using multiple function libraries from different sources. CUDA Programming Guide Version 3.0 127 . Example: The following code defines two functions “f()” in two separate namespaces (“nvidia” and “other”): namespace nvidia { __device__ void f(float x) { /* do something with x */ . All the symbols in a namespace can be imported into another namespace (scope) like this: using namespace nvidia. D. C++ Language Constructs c = a + b. including the following concepts: Implicit template parameter deduction.5f).} } namespace other { __device__ void f(float x) { /* do something with x */ . f(0.Appendix C.

This first type of invocation relies on the compiler‟s ability to implicitly deduce the correct function type for T. g. 2D and 3D points. g_(0).1 Example 1 Pixel Data Type The following is an example of a data type for RGBA pixels with 8 bit per channel depth: class PixelRGBA { public: __device__ PixelRGBA(): r_(0).Appendix C. a_(0) { . a). all other types will be caught by the more general template and return false.  D. } In this case the implementation for T representing the int type are specialized to return true.0 and greater may make use of C++ classes. C++ Language Constructs bool result = f(x). vectors. b.6 Classes Code compiled for devices with compute capability 2. etc. } template <> __device__ bool f<int>(T x) { return true. E.  Functor classes. Function templates may be specialized: template <T> __device__ bool f(T x) { return false. The use of functors is necessitated by the fact that devicefunction pointers are not supported and thus it is not possible to pass functions as template parameters. In this case the compiler would deduce T to be int and instantiate f<int>(x).0 128 .5). as long as none of the member functions are virtual (this restriction will be removed in some future release).6. The second type of invoking the template function is via explicit instantiation like this: bool result = f<double>(0. data types like pixels (r.g. A workaround for this restriction is the use of functor classes (see code sample below). There are two common use cases for classes without virtual member functions: Small-data aggregations. The complete set of matching rules (for implicitly deducing template parameters) and matching polymorphous functions apply as specified in the C++ standard. } CUDA Programming Guide Version 3. D. b_(0).

p1. const PixelRGBA & p2) { return PixelRGBA(p1.b_.. a_. // [. } Other device code can now make use of this new data type as one would expect: PixelRGBA p1. unsigned char a = 255): r_(r).] initialization of p1 and p2 here PixelRGBA p3 = p1 + p2. } // other methods and operators left out for sake of brevity private: unsigned char r_.6. class Sub { public: __device__ float CUDA Programming Guide Version 3. C++ Language Constructs __device__ PixelRGBA(unsigned char r. __device__ PixelRGBA operator+(const PixelRGBA & p1.0 129 . Here are two functors for float addition and subtraction: class Add { public: __device__ float operator() (float a. const PixelRGBA &). p2.g_ + p2. unsigned char b. } }.2 Example 2 Functor Class The following example shows how functors may be used as function template parameters to implement a set of vector arithmetic operations. b_. p1.b_ + p2.a_). friend PixelRGBA operator+(const PixelRGBA &. g_(g).. g_. D.a_ + p2. float b) const { return a + b.g_. unsigned char g. }.r_ + p2.r_. p1. a_(a) { .Appendix C. b_(b).

if (iElement < N) { C[iElement] = op(A[iElement]. const float * B. float * C.Appendix C.x.0 .b. } }. 130 CUDA Programming Guide Version 3. v3. C++ Language Constructs operator() (float a. float b) const { return a . } } The VectorOperation kernel may now be launched like this in order to get a vector addition: // Host code VectorOperation<<<blocks. v2.x * blockIdx. O op) { unsigned int iElement = blockDim. B[iElement]).x + threadIdx. threads>>>(v1. unsigned int N. Add()). N. The following templatized kernel makes use of the functors like the ones above in order to implement operations on vectors of floats: // Device code template<class O> __global__ void VectorOperation(const float * A.

NVCC Specifics E. The compiler will also insert code to ensure correctness (in the example above.1 __noinline__ By default. It is up to the programmer to make sure that the specified unroll number gives the best performance. the loop is completely unrolled if its trip count is constant. For example. ++i) the loop will be unrolled 5 times.x. otherwise it is not unrolled at all. The #pragma unroll directive however can be used to control unrolling of any given loop. E.0 131 . It must be placed immediately before the loop and only applies to that loop. i < n. in this code sample: #pragma unroll 5 for (int i = 0. a __device__ function is always inlined. If no number is specified after #pragma unroll. for example). to ensure that there will only be n iterations if n is less than 5. For devices of compute capability 1. The __noinline__ function qualifier however can be used as a hint for the compiler not to inline the function if possible. the compiler unrolls small loops with a known trip count.2 #pragma unroll By default. the compiler will not honor the __noinline__ qualifier for functions with pointer parameters and for functions with large parameter lists. CUDA Programming Guide Version 3. #pragma unroll 1 will prevent the compiler from ever unrolling a loop. The function body must still be in the same file where it is called. It is optionally followed by a number that specifies how many times the loop must be unrolled.Appendix E.

float* __restrict__ c). Note that all pointer arguments need to be made restricted for the compiler optimizer to derive any benefit. and c restricted pointers. So the compiler cannot take advantage of the common sub-expression.3 __restrict__ nvcc supports restricted pointers via the __restrict__ keyword. Restricted pointers were introduced in C99 to alleviate the aliasing problem that exists in C-type languages. float t2 = t0 * t2. multiply them. c[0] = t2. This changes the function prototype as follows: void foo(const float* __restrict__ a. NVCC Specifics E.. the compiler can now reorder and do common sub-expression elimination at will. b. const float* b. float t3 = a[1].Appendix D. This means that to guarantee functional correctness. float* __restrict__ c) { float t0 = a[0]. . c[1] = a[0] * b[0]. and which inhibits all kind of optimization from code reordering to common sub-expression elimination. Likewise. By making a. b. which in this case means writes through c would never overwrite elements of a or b. c[5] = b[0]. c[2] = a[0] * b[0] * a[1]. c[3] = a[0] * a[1]. Here is an example subject to the aliasing issue. the compiler cannot load a[0] and b[0] into registers. a[0] is really the same location as c[0]. so any write through c could modify elements of a or b. the pointers a. With the __restrict keywords added. float t1 = b[0]. while retaining functionality identical with the abstract execution model: void foo(const float* __restrict__ a. because the results would differ from the abstract execution model if. } In C-type languages. where use of restricted pointer can help the compiler to reduce the number of instructions: void foo(const float* a.0 .. and c may be aliased. the programmer asserts to the compiler that the pointers are in fact not aliased. the compiler cannot just reorder the computation of c[4] into the proximity of the computation of c[0] and c[1] because the preceding write to c[3] could change the inputs to the computation of c[4]. const float* __restrict__ b. const float* __restrict__ b. say. and store the result to both c[0] and c[1]. 132 CUDA Programming Guide Version 3. float* c) { c[0] = a[0] * b[0]. c[4] = a[0] * b[0].

0 133 . t1. NVCC Specifics c[1] c[4] c[2] c[3] c[5] . t2 * t3.. The effects here are a reduced number of memory accesses and reduced number of computations. CUDA Programming Guide Version 3. This is balanced by an increase in register pressure due to "cached" loads and common sub-expressions. Since register pressure is a critical issue in many CUDA codes. } = = = = = t2. due to reduced occupancy.. t2.Appendix D. use of restricted pointers can have negative performance impact on CUDA code. t0 * t3.

.

x is replaced by 0 if x  0 and 1  1 N if 1  x . and ˆ. y . If x is normalized:   In clamp addressing mode. x is replaced by frac(x ) . and z as such: x  Nx .8 depending on the various attributes of the texture reference (see Section 3.0 135 . In wrap addressing mode. y . N  M texels for a two-dimensional texture. x .4). A texture coordinate must fall within T ‟s valid addressing range before it can be used to address T . z  Lz CUDA Programming Guide Version 3. y  My . It is fetched using texture coordinates x . where frac( x)  x  floor( x) and floor(x) is the largest integer not greater than x .2. and z are derived ˆ ˆ ˆ ˆ ˆ from the normalized texture coordinates x .Appendix F. x . y . and z are the non-normalized texture coordinates remapped to T ‟s valid addressing range. only the clamp addressing mode is supported and x is replaced by 0 if x  0 and N  1 if N  x . If x is non-normalized. and z . The texture bound to the texture reference is represented as an array T of N texels for a one-dimensional texture. or N  M  L texels for a three-dimensional texture. The addressing mode specifies how an out-of-range texture coordinate x is remapped to the valid range. In the remaining of the appendix. Texture Fetching This appendix gives the formula used to compute the value returned by the texture functions of Section B. y .

4.75 4 1 Non-Normalized Normalized Figure F-1. 136 CUDA Programming Guide Version 3.Appendix F. which is only available for floating-point textures.0] (see Section 3. where i  floor(x) .1 Nearest-Point Sampling In this filtering mode. For integer textures. z )  T [i. y )  T [i. j .5 3 0. k ] for a three-dimensional texture.2. CUDA-Enabled GPUs F.0 .0. j ] for a two-dimensional texture. j  floor(y) . Figure D-1 illustrates nearest-point sampling for a one-dimensional texture with N 4. the value returned by the texture fetch is    tex( x)  T [i ] for a one-dimensional texture.2 Linear Filtering In this filtering mode. the value returned by the texture fetch is  tex( x)  (1   )T [i]  T [i  1] for a one-dimensional texture. the value returned by the texture fetch can be optionally remapped to [0. tex(x) T[3] T[0] T[2] T[1] x 0 0 1 0. 1. y. and k  floor(z) . Nearest-Point Sampling of a One-Dimensional Texture of Four Texels F.25 2 0. tex( x.1). tex( x.

5 3 0.5 . j  1. j  floor( y B ) . y.   frac( y B ) . y )  (1   )(1   )T [i. z )  (1   )(1   )(1   )T [i. j . j  1.75 4 1 Non-Normalized Normalized Figure F-2. j . where:    i  floor( x B ) . C++ Language Constructs   tex( x. x B  x  0. tex( x.Appendix D. j  1]  T [i  1. y B  y  0. k ]  (1   )(1   )T [i. k  1]  for a three-dimensional texture. j . Figure F-2 illustrates nearest-point sampling for a one-dimensional texture with N 4. Linear Filtering of a One-Dimensional Texture of Four Texels in Clamp Addressing Mode CUDA Programming Guide Version 3. k  1]  T [i  1. j ]   (1   )T [i  1. z B  z  0.  .25 2 0. k  1]  (1   ) T [i. k  floor( z B ) . k ]  (1   )  (1   )T [i.   frac( z B ) .  . and  are stored in 9-bit fixed point format with 8 bits of fractional value (so 1.0 137 . j  1] for a two-dimensional texture.5 .   frac( x B ) . tex(x) T[3] T[0] T[2] T[1] x 0 0 1 0.0 is exactly represented). k  1]   (1   )T [i  1.5 . j . k ]   (1   )T [i  1. k ]   (1   )(1   )T [i  1. j ]  (1   ) T [i. j  1. j  1.

5) in order to ensure that TL (0)  T [0] and TL ( R)  T [ N  1] . CUDA-Enabled GPUs F.3 Table Lookup A table lookup TL (x ) where x spans the interval [0. One-Dimensional Table Lookup Using Linear Filtering 138 CUDA Programming Guide Version 3. TL(x) T[3] T[0] T[2] T[1] x 0 0 4/3 1/3 8/3 2/3 4 1 Figure F-3. R Figure F-3 illustrates the use of texture filtering to implement a table lookup with R  4 or R  1 from a one-dimensional texture with N  4 .Appendix F. R] can be implemented as TL( x)  tex( N 1 x  0.0 .

0 1.10) Integer atomic functions operating on 64-bit words in global memory (Section B.11) Double-precision floating-point numbers Floating-point atomic addition operating on 32-bit words in global and shared memory (Section B.1 1. __syncthreads_and(). Section G.6) No Yes No Yes No Yes No yes 1. __syncthreads_or() (Section B.11) __threadfence_system() (Section B.3 2.10) __ballot() (Section B.5) __syncthreads_count().2 reviews the compliance with the IEEE floating-point standard. respectively. Section G.10) Integer atomic functions operating on 32-bit words in shared memory (Section B.2 1.10) Warp vote functions (Section B.3 and 0 give more details on the architecture of devices of compute capability 1. Section G.0. G.x and 2.5). Compute Capabilities The general specifications and features of a compute device depend on its compute capability (see Section 2.1 gives the features and technical specifications associated to each compute capability.1 Features and Technical Specifications Compute Capability Feature Support (Unlisted features are supported for all compute capabilities) Integer atomic functions operating on 32-bit words in global memory (Section B.Appendix G.0 139 .0 CUDA Programming Guide Version 3.

Compute Capabilities Compute Capability Technical Specifications Maximum x.Appendix G.2 Floating-Point Standard All compute devices follow the IEEE 754-2008 standard for binary floating-point arithmetic with the following deviations: 140 CUDA Programming Guide Version 3.0 1.0 .0 1024 1024 2048 x 2048 x 2048 2 million G. between 6 KB and 8 KB 8192 227 65536 x 32768 65536 x 65536 4096 x 4096 x 4096 32768 512 512 64 32 8 32 1024 16 K 48 1536 32 K 48 KB 32 512 KB 1. height. and depth for a 3D texture reference bound to linear memory or a CUDA array Maximum number of instructions per kernel 24 768 8K 16 KB 16 16 KB 64 KB 8 KB Device dependent.1 1.or y-dimension of a grid of thread blocks Maximum number of threads per block Maximum x.2 65535 1.or y-dimension of a block Maximum z-dimension of a block Warp size Maximum number of resident blocks per multiprocessor Maximum number of resident warps per multiprocessor Maximum number of resident threads per multiprocessor Number of 32-bit registers per multiprocessor Maximum amount of shared memory per multiprocessor Number of shared memory banks Amount of local memory per thread Constant memory size Cache working set per multiprocessor for constant memory Cache working set per multiprocessor for texture memory Maximum width for a 1D texture reference bound to a CUDA array Maximum width for a 1D texture reference bound to linear memory Maximum width and height for a 2D texture reference bound to linear memory or a CUDA array Maximum width.3 2.

z. __frcp_r[n.u.u.infinity is not supported. __fadd_r[u. while SNaN encodings are supported. __fdiv_r[n. float): single-precision division with IEEE rounding modes.d](float): single-precision reciprocal with IEEE rounding modes.1):       __fmaf_r{n. these are passed through unchanged. floating-point arithmetic and comparison instructions convert denormalized operands to zero prior to the floating-point operation.z.  For addition and multiplication. There is no mechanism for detecting that a floating-point exception has occurred and all operations behave as if the IEEE-754 exceptions are always masked. float): single-precision multiplication with IEEE directed rounding. most of the operations support multiple IEEE rounding modes.d}(float. float): single-precision fused multiply-add with IEEE rounding modes.d](float): single-precision square root with IEEE rounding modes.x:  Denormalized numbers are not supported. which truncates (i.z. and deliver the masked response as defined by IEEE-754 if there is an exceptional event. for the same reason.u.2. directed rounding towards +/.e. only round-to-nearest-even and round-towards-zero are supported via static rounding modes. without rounding) the intermediate mantissa of the multiplication. they are not signaling and are handled as quiet.u. Section C.  Underflowed results are flushed to zero.Appendix G. float): single-precision addition with IEEE directed rounding.x: CUDA Programming Guide Version 3.d](float. For single-precision floating-point numbers on devices of compute capability 1. float.  Division is implemented via the reciprocal in a non-standard-compliant way.f.  Some instructions are not IEEE-compliant:  Addition and multiplication are often combined into a single multiplyadd instruction (FMAD).  For double-precision floating-point numbers on devices of compute capability 1.  Square root is implemented via the reciprocal square root in a nonstandard-compliant way. Double-precision floating-point absolute value and negation are not compliant with IEEE-754 with respect to NaNs. __fmul_r[u. IEEE-compliant software (and therefore slower) implementations are provided through the following intrinsics (c. __fsqrt_r[n. however. Compute Capabilities      There is no dynamically configurable rounding mode. To mitigate the impact of these restrictions. The result of a single-precision floating-point operation involving one or more input NaNs is the quiet NaN of bit pattern 0x7fffffff.0 141 .d](float. exposed via device intrinsics.d](float.z.

G.Appendix G.x.3. if one of the input parameters to fminf().  142 CUDA Programming Guide Version 3. When compiling for devices without native double-precision floating-point support. -prec-div=false. This is unlike the x86 architecture behavior. devices of compute capability 1. -prec-div=true.  2 special function units for single-precision floating-point transcendental functions (these units can also handle single-precision floating-point multiplications).x. the behavior is to clamp to the end of the supported range.  FFMA for single precision on devices of compute capability 2. i.0 and higher. FFMA.0 . code must be compiled with -ftz=false. As mentioned above. but not the other. division. Addition and multiplication are often combined into a single multiply-add instruction: FMAD for single precision on devices of compute capability 1.e.1 Compute Capability 1.  In accordance to the IEEE-754R standard. is an IEEE-754(2008) compliant fused multiply-add instruction. so the full-width product is being used in the addition and a single rounding occurs during generation of the final result. see the nvcc user manual for description of these compilation flags). the result is the non-NaN parameter. on the other hand. For compute devices.0. a multiprocessor consists of: 8 CUDA cores for integer and single-precision floating-point arithmetic operations. or fmax() is NaN. FMAD truncates the mantissa prior to use it in the addition. Compute Capabilities  Round-to-nearest-even is the only supported IEEE rounding mode for reciprocal. fmin(). each double variable is converted to single-precision floating-point format (but retains its size of 64 bits) and doubleprecision floating-point arithmetic gets demoted to single-precision floating-point arithmetic. code compiled with -ftz=true. While FFMA in general has superior numerical properties compared to FMAD. the switch from FMAD to FFMA can cause slight changes in numeric results and can in rare circumstances lead to slighty larger error in final results.x.  1 double-precision floating-point unit for double-precision floating-point arithmetic operations. fmaxf().2 and lower.3 G. and -prec-sqrt=false comes closest to the code generated for devices of compute capability 1. For devices of compute capability 2.x Architecture For devices of compute capability 1. and square root. The conversion of a floating-point value to an integer value in the case where the floating-point value falls outside the range of the integer format is left undefined by IEEE-754. and -prec-sqrt=true to ensure IEEE compliance (this is the default setting.

that are issued independently. The resulting memory transactions are serviced at the throughput of device memory.1 To coalesce.0 143 .  16. which resides in device memory.3.2. the first 8 words must lie in the same 128-byte segment and the last 8 words in the following 128-byte segment.Appendix G. If the half-warp meets these requirements.3. one for each half-warp.  16 clock cycles for a single-precision floating-point transcendental instruction. Each TPC has a read-only texture cache that is shared by all multiprocessors and speeds up reads from the texture memory space. the warp scheduler must therefore issue the instruction over:  4 clock cycles for an integer or single-precision floating-point arithmetic instruction.  32 clock cycles for a double-precision floating-point arithmetic instruction. Each multiprocessor accesses the texture cache via a texture unit that implements the various addressing modes and data filtering mentioned in Section 3.  Threads must access the words in sequence: The kth thread in the half-warp must access the kth word. all 16 words must lie in the same 64-byte segment.2 and 1.3. G. A multiprocessor also has a read-only constant cache that is shared by all functional units and speeds up reads from the constant memory space. Sections G.0 and 1.  If this size is:  4.3. G.4.2 describe how the memory accesses of threads within a half-warp are coalesced into one or more memory transactions depending on the compute capability of the device. the memory request for a half-warp must satisfy the following conditions: The size of the words accessed by the threads must be 4. a 64-byte memory transaction. which resides in device memory.0 and 1.2. Figure G-1 shows some examples of global memory accesses and corresponding memory transactions based on compute capability.   The local and global memory spaces reside in device memory and are not cached.1. all 16 words must lie in the same 128-byte segment.2. 3 for devices of compute capabilities 1.1 Devices of Compute Capability 1. To execute an instruction for all threads of a warp.2.3. or 16 bytes. 8.1 and G. a 128-byte memory transaction.  8.2 Global Memory A global memory request for a warp is split into two memory requests. The number of multiprocessors per TPC is: 2 for devices of compute capabilities 1. or two 128-byte memory transactions are issued if the size of  CUDA Programming Guide Version 3.  Multiprocessors are grouped into Texture Processor Clusters (TPCs). Compute Capabilities 1 warp scheduler.

A shared memory request for a warp is split into two memory requests. reduce the transaction size to 64 bytes. G.3 Shared Memory Shared memory has 16 banks that are organized such that successive 32-bit words are assigned to successive banks. Carry out the transaction and mark the serviced threads as inactive. G. reduce the transaction size to 32 bytes. This is in contrast with devices of compute capabilities 1. if possible:  If the transaction size is 128 bytes and only the lower or upper half is used.1 where threads need to access words in sequence and coalescing only happens if the half-warp addresses a single segment.Appendix G.  128 bytes for 4-. 16 separate 32-byte memory transactions are issued.2 Devices of Compute Capability 1. there can be no bank conflict between a thread belonging to the first half of a warp and a thread belonging to the second half of the same warp. and a single memory transaction for each segment addressed by the half-warp is issued. 8. 8. that are issued independently. Repeat until all threads in the half-warp are serviced. i.3.e. or 16. The segment size depends on the size of the words accessed by the threads:  32 bytes for 1-byte words.  If the transaction size is 64 bytes (originally or after reduction from 128 bytes) and only the lower or upper half is used.3 Threads can access any words in any order. If the half-warp does not meet these requirements.2. i. As a consequence. 144 CUDA Programming Guide Version 3. Find all other active threads whose requested address lies in the same segment. respectively.e. Coalescing is achieved even if the warp is divergent. only one thread per halfwarp performs a write and which thread performs the final write is undefined.and 16-byte words. one for each half-warp.2 and 1. there are some inactive threads that do not actually access memory. the following protocol is used to determine the memory transactions necessary to service all threads in a half-warp:      Find the memory segment that contains the address requested by the lowest numbered active thread. including the same words.3. Reduce the transaction size. Compute Capabilities the words accessed by the threads is 4.  64 bytes for 2-byte words. More precisely. If a non-atomic instruction executed by a warp writes to the same location in shared memory for more than one of the threads of the warp. interleaved. Each bank has a bandwidth of 32 bits per two clock cycles.0 .0 and 1.

for example. Which word is selected as the broadcast word and which address is picked up for each bank at each cycle are unspecified. There are no bank conflicts however. G. float data = shared[BaseIndex + s * tid]. For example. G.3. there will be no bank conflict only if half the warp size (i.3 8-Bit and 16-Bit Access 8-bit and 16-bit accesses typically generate bank conflicts. the subset is built from the remaining addresses that have yet to be serviced using the following procedure: Select one of the words pointed to by the remaining addresses as the broadcast word. i. CUDA Programming Guide Version 3.3.  One address for each bank (other than the broadcasting bank) pointed to by the remaining addresses. a memory read request made of several addresses is serviced in several steps over time by servicing one conflict-free subset of these addresses per step until all addresses have been serviced. 16) or. char data = shared[BaseIndex + tid]. belong to the same bank.3.1 32-Bit Strided Access A common access pattern is for each thread to access a 32-bit word from an array indexed by the thread ID tid and with some stride s: __shared__ float shared[32]. s is odd.0 145 . if the same array is accessed the following way: char data = shared[BaseIndex + 4 * tid]. shared[2]. at each step.e. In this case.e. threads tid and tid+n access the same bank whenever s*n is a multiple of the number of banks (i.3. because shared[0]. The same examples apply for devices of compute capability 1.3. that is only if d is equal to 1. Figure G-3 shows some examples of memory read accesses that involve the broadcast mechanism. Compute Capabilities G.  A common conflict-free case is when all threads of a half-warp read from an address within the same 32-bit word. shared[1].3. but with 16 banks instead of 32.2 32-Bit Broadcast Access Shared memory features a broadcast mechanism whereby a 32-bit word can be read and broadcast to several threads simultaneously when servicing one memory read request. 16) is less than or equal to 16/d.  Include in the subset:  All addresses that are within the broadcast word. equivalently. and shared[3]. Figure G-2 shows some examples of strided access for devices of compute capability 2.x.. As a consequence. More precisely.0. but with 16 banks instead of 32. This reduces the number of bank conflicts when several threads read from an address within the same 32-bit word. The same examples apply for devices of compute capability 1. there are bank conflicts if an array of char is accessed the following way: __shared__ char shared[32]. whenever n is a multiple of 16/d where d is the greatest common divisor of 16 and s.Appendix G.x.e.

0. since each member is accessed with an odd stride of three 32-bit words. The following code. as the memory request is compiled into two separate 32-bit requests with a stride of two.0 . double dataOut = __hiloint2double(shared_hi[BaseIndex + tid]. there are 2-way bank conflicts for arrays of doubles accessed as follows: __shared__ double shared[32]. shared_hi[BaseIndex + tid] = __double2hiint(dataIn). struct type data = shared[BaseIndex + tid].4 Larger Than 32-Bit Access Accesses that are larger than 32-bit per thread are split into 32-bit accesses that typically generate bank conflicts. y. results in:  Three separate reads without bank conflicts if type is defined as struct type { float x.0. For example. z.4 G. shared_lo[BaseIndex + tid] = __double2loint(dataIn). since each member is accessed with an even stride of two 32-bit words. One way to avoid bank conflicts in this case is two split the double operands like in the following sample code: __shared__ int shared_lo[32].Appendix G.0 Architecture For devices of compute capability 2. y. }.  146 CUDA Programming Guide Version 3. The same applies to structure assignments.1 Compute Capability 2.  Two separate reads with bank conflicts if type is defined as struct type { float x.4. double data = shared[BaseIndex + tid]. }. __shared__ int shared_hi[32].  4 special function units for single-precision floating-point transcendental functions. double dataIn. G.3.3. This might not always improve performance however and does perform worse on devices of compute capabilities 2. a multiprocessor consists of: 32 CUDA cores for integer and floating-point arithmetic operations. Compute Capabilities G. shared_lo[BaseIndex + tid]). for example: __shared__ struct type shared[32].

the other scheduler cannot issue any instruction. The same on-chip memory is used for both L1 and shared memory: It can be configured as 48 KB of shared memory with 16 KB of L1 cache (default setting) or as 16 KB of shared memory with 48 KB of L1 cache using cudaFuncSetCacheConfig() or cuFuncSetCacheConfig(): // Device code __global__ void MyKernel() { . The only exception is when a scheduler issues a double-precision floating-point instruction: In this case.. whether reads are cached in both L1 and L2 or in L2 only) can be partially configured on a peraccess basis using modifiers to the load or store instruction. the first scheduler issues an instruction for some warp with an odd ID and the second scheduler issues an instruction for some warp with an even ID. a warp scheduler must therefore issue the instruction over: 2 clock cycles for an integer or floating-point arithmetic instruction. To execute an instruction for all threads of a warp.  2 clock cycles for a double-precision floating-point arithmetic instruction. Compute Capabilities 2 warp schedulers. Each multiprocessor has a read-only texture cache to speed up reads from the texture memory space. A GPC includes four multiprocessors. The cache behavior (e.g. It accesses the texture cache CUDA Programming Guide Version 3. CU_FUNC_CACHE_PREFER_SHARED) Multiprocessors are grouped into Graphics Processor Clusters (GPCs).  A warp scheduler can issue an instruction to only half of the CUDA cores. At every instruction issue time. A multiprocessor also has a read-only uniform cache that is shared by all functional units and speeds up reads from the constant memory space. cuFuncSetCacheConfig(myKernel.0 147 . including temporary register spills.. which resides in device memory. both of which are used to cache accesses to local or global memory. } // Host code // Runtime API // cudaFuncCachePreferShared: shared memory is 48 KB // cudaFuncCachePreferL1: shared memory is 16 KB // cudaFuncCachePreferNone: no preference cudaFuncSetCacheConfig(MyKernel. cudaFuncCachePreferShared) // Driver API // CU_FUNC_CACHE_PREFER_SHARED: shared memory is 48 KB // CU_FUNC_CACHE_PREFER_L1: shared memory is 16 KB // CU_FUNC_CACHE_PREFER_NONE: no preference CUfunction myKernel.Appendix G.  There is an L1 cache for each multiprocessor and an L2 cache shared by all multiprocessors. which resides in device memory.  8 clock cycles for a single-precision floating-point transcendental instruction.

a memory request by a warp is first split into separate 128-byte memory requests that are issued independently: Two memory requests. only one thread performs a write and which thread does it is undefined. they can be configured at compile time to be cached in both L1 and L2 (-Xptxas -dlcm=ca) or in L2 only (-Xptxas -dlcm=cg). Each memory request is then broken down into cache line requests that are issued independently. otherwise.Appendix G.  Four memory requests. Compute Capabilities via a texture unit that implements the various addressing modes and data filtering mentioned in Section 3. 148 CUDA Programming Guide Version 3.0 . or at the throughput of device memory. one for each half-warp.2 Global Memory Global memory accesses are cached. if the size is 8 bytes. G.2. Using the –dlcm compilation flag. A cache line request is serviced at the throughput of L1 or L2 cache in case of a cache hit.  If a non-atomic instruction executed by a warp writes to the same location in global memory for more than one of the threads of the warp. one for each quarter-warp. If the size of the words accessed by each thread is more than 4 bytes. A cache line in L1 or L2 is 128 bytes and maps to a 128-byte aligned segment in device memory. if the size is 16 bytes.4.4.

0 Cached 1 x 128B at 128 1 x 128B at 256 Uncached 8x 8x 8x 8x 32B at 128 32B at 160 32B at 192 32B at 224 1 x 128B at 128 1 x 64B at 192 1 x 32B at 256 Figure G-1.0 Cached 1 x 128B at 128 Uncached 1 x 64B at 128 1 x 64B at 192 1 x 64B at 128 1 x 64B at 192 Aligned and non-sequential Addresses: 96 128 160 192 224 256 288 Threads: Compute capability: Memory transactions: 0 1. Compute Capabilities Aligned and sequential Addresses: 96 128 160 192 224 256 288 Threads: Compute capability: Memory transactions: 0 1.2 and 1.3 31 2.1 … 1.0 Cached 1 x 128B at 128 Uncached 8x 8x 8x 8x 32B at 128 32B at 160 32B at 192 32B at 224 1 x 64B at 128 1 x 64B at 192 Misaligned and sequential Addresses: 96 128 160 192 224 256 288 Threads: Compute capability: Memory transactions: 0 1.0 and 1.3 31 2.0 149 .2 and 1.2 and 1. and Associated Memory Transactions Based on Compute Capability CUDA Programming Guide Version 3.1 … 1.0 and 1.3 31 2.1 … 1.Appendix G. 4-Byte Word per Thread. Examples of Global Memory Accesses by a Warp.0 and 1.

for example: struct type { float x. A bank conflict only occurs if two or more threads access any bytes within different 32-bit words belonging to the same bank.2 Larger Than 32-Bit Access 64-bit and 128-bit accesses are specifically handled to minimize bank conflicts as described below. }. unlike for devices of lower compute capability. threads tid and tid+n access the same bank whenever s*n is a multiple of the number of banks (i.x. 150 CUDA Programming Guide Version 3.3. in particular.4.. struct type data = shared[BaseIndex + tid]. Therefore. a bank conflict only occurs if two or more threads in either of the half-warps access different addresses belonging to the same bank. for write accesses.e. 32) or.3. interleaved. __shared__ struct type shared[32]. Each bank has a bandwidth of 32 bits per two clock cycles. char data = shared[BaseIndex + tid].e. there is no bank conflict between these threads: For read accesses. each byte is written by only one of the threads (which thread performs the write is undefined). G. Compute Capabilities G. As a consequence.4. equivalently. the word is broadcast to the requesting threads (unlike for devices of compute capability 1. there are no bank conflicts if an array of char is accessed as follows. that is only if d is equal to 1. 64-bit. 64-Bit Accesses For 64-bit accesses. multiple words can be broadcast in a single transaction).x.3 Shared Memory Shared memory has 32 banks that are organized such that successive 32-bit words are assigned to successive banks. there will be no bank conflict only if the warp size (i. for example: __shared__ char shared[32].1 32-Bit Strided Access A common access pattern is for each thread to access a 32-bit word from an array indexed by the thread ID tid and with some stride s: __shared__ float shared[32]. float data = shared[BaseIndex + s * tid]. 32) is less than or equal to 32/d. whenever n is a multiple of 32/d where d is the greatest common divisor of 32 and s. Other accesses larger than 32-bit are split into 32-bit.Appendix G. that unlike for devices of compute capability 1. In this case. The following code. Figure G-2 shows some examples of strided access. s is odd. z. i. results in three separate 32-bit reads without bank conflicts since each member is accessed with a stride of three 32-bit words. or 128-bit accesses.e.4. This means. there may be bank conflicts between a thread belonging to the first half of a warp and a thread belonging to the second half of the same warp. If two or more threads access any bytes within the same 32-bit word. G. i.e.0 . y.

Appendix G.

Compute Capabilities

Unlike for devices of compute capability 1.x, there are no bank conflicts for arrays of doubles accessed as follows, for example:
__shared__ double shared[32]; double data = shared[BaseIndex + tid];

128-Bit Accesses
The majority of 128-bit accesses will cause 2-way bank conflicts, even if no two threads in a quarter-warp access different addresses belonging to the same bank. Therefore, to determine the ways of bank conflicts, one must add 1 to the maximum number of threads in a quarter-warp that access different addresses belonging to the same bank.

G.4.4

Constant Memory
In addition to the constant memory space supported by devices of all compute capabilities (where __constant__ variables reside), devices of compute capability 2.0 support the LDU (LoaD Uniform) instruction that the compiler use to load any variable that is: pointing to global memory,  read-only in the kernel (programmer can enforce this using the const keyword),  not dependent on thread ID.

CUDA Programming Guide Version 3.0

151

Appendix G.

Compute Capabilities

Threads: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Banks: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Threads: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Banks: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Threads: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Banks: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Left: Linear addressing with a stride of one 32-bit word (no bank conflict). Middle: Linear addressing with a stride of two 32-bit words (2-way bank conflicts). Right: Linear addressing with a stride of three 32-bit words (no bank conflict).

Figure G-2 Examples of Strided Shared Memory Accesses for Devices of Compute Capability 2.0

152

CUDA Programming Guide Version 3.0

Appendix G.

Compute Capabilities
Banks: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Threads: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Banks: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Threads: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Banks: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Threads: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Left: Conflict-free access via random permutation. Middle: Conflict-free access since threads 3, 4, 6, 7, and 9 access the same word within bank 5. Right: Conflict-free broadcast access (all threads access the same word).

Figure G-3 Examples of Irregular and Colliding Shared Memory Accesses for Devices of Compute Capability 2.0
CUDA Programming Guide Version 3.0
153

.

2 (March/April 2008). LISTS. http://mags. Other company and product names may be trademarks of the respective companies with which they are associated.acm. Copyright © 2006-2010 NVIDIA Corporation. DRAWINGS. and Quadro are trademarks or registered trademarks of NVIDIA Corporation. FILES. This publication supersedes and replaces all information previously supplied. Trademarks NVIDIA. MERCHANTABILITY. NVIDIA Corporation assumes no responsibility for the consequences of use of such information or for any infringement of patents or other rights of third parties that may result from its use. EXPRESSED.nvidia. This work incorporates portions of on an earlier work: Scalable Parallel Programming with CUDA. CA 95050 www. All rights reserved. in ACM Queue. the NVIDIA logo. No license is granted by implication or otherwise under any patent or patent rights of NVIDIA Corporation. Tesla. used under license to the Khronos Group Inc. IMPLIED.Notice ALL NVIDIA DESIGN SPECIFICATIONS. © ACM. Information furnished is believed to be accurate and reliable.” NVIDIA MAKES NO WARRANTIES. DIAGNOSTICS. AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT. AND FITNESS FOR A PARTICULAR PURPOSE.org/queue/20080304/?u1=texterity" NVIDIA Corporation 2701 San Tomas Expressway Santa Clara. No. VOL 6. AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY.com . Specifications mentioned in this publication are subject to change without notice. However. REFERENCE BOARDS. NVIDIA Corporation products are not authorized for use as critical components in life support devices or systems without express written approval of NVIDIA Corporation. GeForce. 2008. OR OTHERWISE WITH RESPECT TO THE MATERIALS. OpenCL is trademark of Apple Inc. “MATERIALS”) ARE BEING PROVIDED “AS IS. STATUTORY.

Sign up to vote on this title
UsefulNot useful

Master Your Semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master Your Semester with a Special Offer from Scribd & The New York Times

Cancel anytime.