CUDA Fortran

Programming Guide and Reference

Release 2011

The Portland Group

While every precaution has been taken in the preparation of this document, The Portland Group® (PGI®), a wholly-owned subsidiary of STMicroelectronics, Inc., makes no warranty for the use of its products and assumes no responsibility for any errors that may appear, or for damages resulting from the use of the information contained herein. The Portland Group retains the right to make changes to this information at any time, without notice. The software described in this document is distributed under license from STMicroelectronics and/or The Portland Group and may be used or copied only in accordance with the terms of the license agreement ("EULA"). PGI Workstation, PGI Server, PGI Accelerator, PGF95, PGF90, PGFORTRAN, and PGI Unified Binary are trademarks; and PGI, PGHPF, PGF77, PGCC, PGC++, PGI Visual Fortran, PVF, PGI CDK, Cluster Development Kit, PGPROF, PGDBG, and The Portland Group are registered trademarks of The Portland Group Incorporated. No part of this document may be reproduced or transmitted in any form or by any means, for any purpose other than the purchaser's or the end user's personal use without the express written permission of STMicroelectronics and/or The Portland Group.

Copyright © 2010 The Portland Group® and STMicroelectronics, Inc.All rights reserved. Printed in the United States of America Release 2010, version 10.8, August 2010 Release 2010, version 10.9, September 2010 Release 2011, version 11.0, November 2010

Technical support: trs@pgroup.com Sales: sales@pgroup.com Web: www.pgroup.com

ID: 103151117

Contents
Preface ...................................................................................................................................... ix
Intended Audience ................................................................................................................. ix Organization .......................................................................................................................... ix Conventions ........................................................................................................................... ix Terminology ........................................................................................................................... x Related Publications ................................................................................................................ x

1. Introduction ........................................................................................................................ 1 2. Programming Guide .......................................................................................................... 3
CUDA Fortran Kernels .............................................................................................................. 3 Thread Blocks ........................................................................................................................ 4 Memory Hierarchy .................................................................................................................. 4 Subroutine / Function Qualifiers ................................................................................................ 4 Attributes(host) .............................................................................................................. 5 Attributes(global) ............................................................................................................ 5 Attributes(device) ........................................................................................................... 5 Restrictions .................................................................................................................... 5 Variable Qualifiers .................................................................................................................. 5 Attributes(device) ........................................................................................................... 5 Attributes(constant) ........................................................................................................ 6 Attributes(shared) ........................................................................................................... 6 Attributes(pinned) .......................................................................................................... 6 Datatypes in Device Subprograms ............................................................................................. 6 Predefined Variables in Device Subprograms .............................................................................. 7 Execution Configuration ........................................................................................................... 7 Asynchronous concurrent execution .......................................................................................... 8 Concurrent Host and Device Execution .............................................................................. 8 Concurrent Stream Execution ........................................................................................... 8 Kernel Loop Directive .............................................................................................................. 8 Restrictions on the CUF kernel directive ........................................................................... 10 Building a CUDA Fortran Program ........................................................................................... 11 Emulation Mode ................................................................................................................... 11 iii

3. Reference ........................................................................................................................... 13
New Subroutine and Function Attributes ................................................................................... Host Subroutines and Functions ...................................................................................... Global Subroutines ........................................................................................................ Device Subroutines and Functions ................................................................................... Restrictions on Device Subprograms ................................................................................ Variable Attributes ................................................................................................................. Device data .................................................................................................................. Constant data ................................................................................................................ Shared data .................................................................................................................. Value dummy arguments ................................................................................................ Pinned arrays ............................................................................................................... Allocating Device and Pinned Arrays ........................................................................................ Allocating Device Memory .............................................................................................. Allocating Device Memory Using Runtime Routines ............................................................ Allocating Pinned Memory .............................................................................................. Data transfer between host and device memory ......................................................................... Data Transfer Using Assignment Statements ....................................................................... Implicit Data Transfer in Expressions ............................................................................... Data Transfer Using Runtime Routines ............................................................................. Invoking a kernel subroutine .................................................................................................. Device code ......................................................................................................................... Datatypes allowed ......................................................................................................... Builtin variables ............................................................................................................ Fortran intrinsics .......................................................................................................... New Intrinsic Functions ................................................................................................. Warp-Vote Operations .................................................................................................... Atomic Functions .......................................................................................................... Restrictions .................................................................................................................. Host code ............................................................................................................................ SIZEOF Intrinsic ............................................................................................................ 13 13 13 14 14 14 14 15 16 17 17 17 17 18 18 19 19 19 20 20 21 21 22 22 24 26 27 29 29 29

4. Runtime APIs .................................................................................................................... 31
Initialization ......................................................................................................................... Device Management ............................................................................................................... cudaGetDeviceCount ...................................................................................................... cudaSetDevice ............................................................................................................... cudaSetDeviceFlags ........................................................................................................ cudaGetDevice .............................................................................................................. cudaGetDeviceProperties ................................................................................................ cudaChooseDevice ......................................................................................................... Thread Management .............................................................................................................. cudaThreadSynchronize ................................................................................................. cudaThreadExit ............................................................................................................. Memory Management ............................................................................................................ cudaMalloc .................................................................................................................. iv 31 31 31 31 31 32 32 32 32 32 32 32 33

cudaMallocPitch ............................................................................................................ cudaFree ...................................................................................................................... cudaMallocArray ........................................................................................................... cudaFreeArray .............................................................................................................. cudaMemset ................................................................................................................. cudaMemset2D ............................................................................................................. cudaMemcpy ................................................................................................................ cudaMemcpyAsync ........................................................................................................ cudaMemcpy2D ............................................................................................................ cudaMemcpy2DAsync .................................................................................................... cudaMemcpyToArray ...................................................................................................... cudaMemcpy2DToArray .................................................................................................. cudaMemcpyFromArray .................................................................................................. cudaMemcpy2DFromArray .............................................................................................. cudaMemcpyArrayToArray .............................................................................................. cudaMemcpy2DArrayToArray .......................................................................................... cudaMalloc3D .............................................................................................................. cudaMalloc3DArray ....................................................................................................... cudaMemset3D ............................................................................................................. cudaMemcpy3D ............................................................................................................ cudaMemcpy3DAsync .................................................................................................... cudaMemcpyToSymbol ................................................................................................... cudaMemcpyFromSymbol ............................................................................................... cudaMemcpyToSymbolAsync ........................................................................................... cudaMemcpyFromSymbolAsync ....................................................................................... cudaGetSymbolAddress ................................................................................................... cudaGetSymbolSize ........................................................................................................ cudaHostAlloc ............................................................................................................... cudaHostGetDevicePointer .............................................................................................. cudaMallocHost ............................................................................................................ cudaFreeHost ................................................................................................................ Stream Management .............................................................................................................. cudaStreamCreate .......................................................................................................... cudaStreamQuery .......................................................................................................... cudaStreamSynchronize .................................................................................................. cudaStreamDestroy ........................................................................................................ Event Management ................................................................................................................ cudaEventCreate ............................................................................................................ cudaEventRecord .......................................................................................................... cudaEventQuery ............................................................................................................ cudaEventSynchronize .................................................................................................... cudaEventDestroy .......................................................................................................... cudaEventElapsedTime ................................................................................................... Error Handling ..................................................................................................................... cudaGetLastError .......................................................................................................... cudaGetErrorString ........................................................................................................

33 33 33 33 34 34 34 34 34 35 35 35 35 35 36 36 36 36 36 36 37 37 37 37 37 38 38 38 38 39 39 39 39 39 39 39 40 40 40 40 40 40 40 41 41 41 v

.............................................. Examples ................................................................................. Contact Information ................................. 45 6..................................................................................................................Version Management ..................................................................................................................................................................................................... 43 Source Code Description .......................................... 41 cudaRuntimeGetVersion ........................ 43 Matrix Multiplication Example ...................................... 43 Source Code Listing ....... 41 5............. 47 vi ........................................................................................................... 41 cudaDriverGetVersion ..................................................................................................................................

.. Fortran Random Number Intrinsics ............... 6 3........ 23 3........................................................ 24 3.................................................................................................................1..........................Tables 2...................... Fortran Mathematical Intrinsics .................................2. 24 3.8....... Intrinsic Datatypes ........................ Fortran Numeric and Logical Intrinsics .........6..............................4......... Fortran Reduction Intrinsics .................................................................... 28 3.......7...............5............... 23 3..................................... Device Code Intrinsic Datatypes .........................................9................................................... Counting Atomic Functions ................................................... Fortran Bit Manipulation Intrinsics .... 22 3................. 23 3........................ Fortran Numeric Inquiry Intrinsics ....................................................... 21 3..................................................... Arithmetic and Bitwise Atomic Functions .. 27 3...................................................................................................................................... Compare and Swap Atomic Function .........................................10...................1.........................................................3....... 29 vii ..............

viii .

Intended Audience This guide is intended for application programmers. ix . “Runtime APIs” describes the interface between CUDA Fortran and the CUDA Runtime API Chapter 5. a small set of extensions to Fortran that supports and is built upon the CUDA computing architecture. scientists and engineers proficient in programming with the Fortran. Organization The organization of this document is as follows: Chapter 1. AMD64. The PGI tools are available on a variety of operating systems for the X86. C. Conventions This guide uses the following conventions: italic is used for emphasis. “Introduction” contains a general introduction Chapter 2. This guide assumes familiarity with basic operating system usage. and Intel 64 hardware platforms. and/or C++ languages. “Programming Guide” serves as a programming guide for CUDA Fortran Chapter 3.Preface This document describes CUDA Fortran. “Examples” provides sample code and an explanation of the simple example. “Reference” describes the CUDA Fortran language reference Chapter 4.

The PGI compilers and tools are supported on both 32-bit and 64-bit variants of the Linux and Windows operating systems on a variety of x86-compatible processors. • ISO/IEC 1539-1:1997.1.htm Related Publications The following documents contain additional information related to CUDA Fortran programming. June 2010.pdf. directories. 1997 (Fortran 95). PGI provides a glossary of terms which you can access at www. In the context of p/tsets. C/C++ C/C++ language statements are shown in the test of this guide using a reduced fixed point size. filename . square brackets are required to specify a p/t-set. • NVIDIA CUDA Compute Unified Device Architecture Reference Manual.com/support/definitions. 2009. Geneva. The Portland Group. In this case item1 is optional.com/doc/pgiug.1. NVIDIA.1. [ item1 ] in general. September. { item2 | item 3} braces indicate that a selection is required. options. Information Technology – Programming Languages – FORTRAN.nvidia. examples. Terminology If there are terms in this guide with which you are unfamiliar.. including assembly language statements. ellipsis indicate a repetition. Available online at http:// www. • PGI Users Guide. Bold is used for commands. Release 10.pgroup. • NVIDIA CUDA Programming Guide. Zero or more of the preceding item may occur. arguments.pgroup. x .Terminology Constant Width is used for filenames. multiple filenames are allowed. FORTRAN Fortran language statements are shown in the text of this guide using a reduced fixed point size. Available online at http:// www. 7/21/2010.nvidia. Version 3. Version 3. square brackets indicate optional items. NVIDIA. In this example. There are a wide variety of releases and distributions of each of these types of operating systems. In this case..com/cuda.9. you must select either item2 or item3.com/cuda. Available online at http://www. and for language statements in the text.

but are general enough to be useful in many data-parallel. a general purpose parallel programming architecture. allowing direct programming of the GPU from a high level language. and back • Writing subroutines and functions to execute on the GPU • Invoking GPU subroutines from the host • Allocating pinned memory on the host • Using asynchronous transfers between the host and GPU 1 . PGI 2011 includes support for CUDA Fortran on Linux. highly parallel computational units with very high memory bandwidth. and tremendous potential for many applications. NVIDIA introduced CUDA™. CUDA Fortran includes a Fortran 2003 compiler and tool chain for programming NVIDIA GPUs using Fortran. regardless of the number of available processor cores. Each block is partitioned into fine grain threads. here called CUDA C. Introduction Welcome to Release 2011 of PGI CUDA Fortran. CUDA comes with an extended C compiler. A CUDA programmer must partition the program into coarse grain blocks that can be executed in parallel. Graphic processing units or GPUs have evolved into programmable. compute-intensive programs. and coordinated independent thread groups organized into a grid. which can cooperate using shared memory and barrier synchronization. GPU designs are optimized for the computations found in graphics rendering.Chapter 1. Mac OS X and Windows. The CUDA Fortran extensions described in this document allow the following operations in a Fortran program: • Declaring variables that are allocated in the GPU device memory • Allocating dynamic memory in the GPU device memory • Copying data from the host memory to the GPU memory. A properly designed CUDA program will run on any CUDA-enabled GPU. a small set of extensions to Fortran that supports and is built upon the CUDA computing architecture. Compared to the PGI Accelerator directives-based model and compilers. CUDA Fortran is an analog to NVIDIA's CUDA C compiler. CUDA Fortran is a lower-level explicit programming model with substantial runtime library components that give expert programmers direct control of all aspects of GPGPU programming. The programming model supports four key abstractions: cooperating threads organized into thread groups. with compilers and libraries to support the programming of NVIDIA GPUs. shared memory and barrier synchronization within thread groups.

2 .

a kernel is called using special chevron syntax to specify the number of thread blocks and threads within each thread block: ! Kernel definition attributes(global) subroutine ksaxpy( n. a. y real :: a integer :: n ! call the kernel call ksaxpy<<<n/64. refer to Chapter 3. each instance will be executed by a different CUDA thread. the call to the kernel ksaxpy specifies n/64 thread blocks.y real. Programming Guide This chapter introduces the CUDA programming model through examples written in CUDA Fortran. y ) real. Such a subroutine is called a device kernel or kernel. dimension(*) :: x. a. value :: n. dimension(*) :: x. A call to a kernel specifies how many parallel instances of the kernel must be executed. x. 64>>>( n. “Reference. value :: a integer. and a thread index accessed through threadidx. and a local thread index within its thread block. 3 .” on page 13. and each thread has a global thread block index. each with 64 threads. i i = (blockidx%x-1) * blockdim%x + threadidx%x if( i <= n ) y(i) = a * x(i) + y(i) end subroutine ! Host subroutine subroutine solve( n. x. x. CUDA Fortran Kernels CUDA Fortran allows the definition of Fortran subroutines that execute in parallel on the GPU when called from the Fortran program which has been invoked and is running on the host. The CUDA threads are organized into thread blocks. In this example.Chapter 2. y ) end subroutine In this case. For a reference for CUDA Fortran. a. y ) real. each thread performs one iteration of the common SAXPY loop operation. device. Each thread is assigned a thread block index accessed through the built-in blockidx variable. A kernel is defined using the attributes(global) specifier on the subroutine statement.

or called from another device subprogram. For a two-dimensional thread block of size (Dx. high bandwidth software managed cache memory. computed from the thread index. It can also directly copy data to and from the device global memory. and a block index within the grid.Dy). data in global device memory can be read or written by all threads. and a thread index accessed through threadidx. In CUDA Fortran. designating whether it is executed on the host or on the device. Memory Hierarchy CUDA Fortran programs have access to several memory spaces. • A subprogram declared with attributes(host). The thread blocks are organized into a one. Thread blocks must be able to execute independently.or two-dimensional grid of blocks. and the second argument is the thread block size. The shared memory acts like a low-latency. all threads can read data in constant memory. A kernel may be invoked with many thread blocks. whether it is a kernel. the host program can directly access data in the host main memory. two-.Dy. Threads in the same thread block can access and share data in shared memory. the thread index is equal to the thread ID. Each thread can also have private local memory. again implemented using DMA access. On the host side. the thread ID is (x+Dx(y-1)+Dy(z-1)). data in thread local memory may be implemented as processor registers or may be allocated in the global device memory. but it is read-only to the threads and limited in size. Accesses to constant memory are typically faster than accesses to global memory. so each thread has a thread index within the block. or with the host attribute by default. For a one-dimensional thread block. each with the same thread block size. the maximum number of threads in a thread block is 512. A unique thread ID is assigned to each thread. Data in constant memory space is initialized by the host program. by the same core or by different cores. The host can also set the values in the device constant memory. such data copies require DMA access to the device. • A subprogram declared with attributes(global) or attributes(device) is called a device subprogram. The thread index may be a one-. best performance will often be obtained when thread local data Subroutine / Function Qualifiers A subroutine or function in CUDA Fortran has an additional attribute. 4 . Threads in the same thread block may cooperate by using shared memory. is called a host subprogram. Each thread in the block waits at the call to SYNCTHREADS() until all threads have reached that call. For a three-dimensional thread block of size (Dx. and by synchronizing at a barrier using the SYNCTHREADS() intrinsic. On the device side. the thread index for each dimension starts at one. or three-dimensional index. data in shared memory has a lifetime of the thread block. called from the host.Thread Blocks Thread Blocks Each thread is assigned a thread block index accessed through the built-in blockidx variable. so are slow relative to the host memory. When invoking a kernel. the first argument in the chevron <<<>>> syntax is the grid size. Currently. the thread ID is equal to (x+Dx(y-1)). and if the latter.Dz). two thread blocks may be executed in parallel or one after the other.

Restrictions The following restrictions apply to subprograms: • A device subprogram must not be recursive. or with data initialization. Variable Qualifiers Variables in CUDA Fortran have a new attribute. 5 . • Calls to a kernel subroutine must specify the execution configuration. or device attributes is specified.” on page 7. Programming Guide • A subroutine declared with attributes(global) is also called a kernel subroutine. such a routine must be called from a subprogram with the global or device attribute. which declares in which memory the data is allocated. and is allocated in the device global memory. Attributes(host) The host attribute. or pinned attributes may be specified for a variable. Such a call is asynchronous. as described in “Predefined Variables in Device Subprograms. if none of the host. global. • A kernel subroutine may not also have the device or host attribute. The default is attributes(host). At most one of the device. and may only be called from the host using a kernel call containing the chevron syntax and runtime mapping parameters. constant. Attributes(device) The device attribute. variables declared in modules or host subprograms are allocated in the host main memory. Attributes(global) The global attribute may only be specified on a subroutine statement. • A device subprogram must not contain variables with the SAVE attribute. to be executed on the device. Dummy arguments in a device subprogram must not have the pointer attribute. • If declared in a module. and may not contain any subroutines or functions. it declares that the subroutine is a kernel subroutine. Attributes(device) A variable with the device attribute is called a device variable. specified on the subroutine or function statement. By default.Chapter 2. the host routine making the call continues to execute before the device has completed its execution of the kernel subroutine. that is. specified on the subroutine or function statement. Such a subprogram can only be called from another host subprogram. declares that the subprogram is to be executed on the device. • Device subprograms may not be contained in a host subroutine or function. declares that the subroutine or function is to be executed on the host. • A device subprogram must not have optional arguments. the variable may be accessed by any subprogram in that module and by any subprogram that uses the module. shared.

and has a lifetime of the thread block. it will be allocated in host pagelocked memory.4. the variable may be accessed by that subprogram or subprograms contained in that subprogram. A shared variable may not be data initialized. and may only be accessed within that subprogram.2. the variable may be accessed by any subprogram in that module and by any subprogram that uses the module. an allocatable array. Some operating systems or installations may restrict the use. Device constant variables are allocated in the device constant memory space. The advantage of using pinned variables is that copies from page-locked memory to device memory are faster than copies from normal paged host memory. or size of page-locked memory. Attributes(constant) A variable with the constant attributes is called a device constant variable.8 equivalent to real(kind=8) 6 . When a pinned variable is allocated. Datatypes in Device Subprograms The following intrinsic datatypes are allowed in device subprograms and device data: Table 2. It can be read or written by all threads in the block. or an assumed-shape dummy array. Device constant variables may not be allocatable. A device array may be an explicit-shape array. Attributes(shared) A variable with the shared attributed is called a device shared variable or a shared variable.8 4. if the allocation in page-locked memory fails.2. Intrinsic Datatypes Type integer logical real double precision Type Kind 1. A shared variable may only be declared in a device subprogram. Device constant data may not be assigned or modified in any device subprogram. from when it is allocated until it is deallocated.8 1.Datatypes in Device Subprograms • If declared in a host subprogram.4. Other device variables have a lifetime of the entire application. If declared in a module. and have a lifetime of the entire application. or by other device subprograms to which it is passed as an argument. the variable will be allocated in the normal host paged memory.1. A shared variable is allocated in the device shared memory for a thread block. though a write in one thread is only guaranteed to be visible to other threads after the next call to the SYNCTHREADS() intrinsic. An allocatable device variable has a dynamic lifetime. Attributes(pinned) A variable with the pinned attributes is called a pinned variable. availability. but may be modified in host subprograms. A pinned variable must be an allocatable array.

…) where • grid and block are either integer expressions (for one-dimensional grids and thread blocks). two-. • If grid is type(dim3). This memory is used for the 7 . • If block is type(dim3). • The variable is warpsize is declared to be type integer. it specifies the number of bytes of shared memory to be allocated for each thread block. in bytes. to support concurrent stream execution on the device. for one. A kernel subroutine call looks like this: call kernel<<<grid. the threadidx%y and/or threadidx%z components have the value one. called warps. warpsize contains the number of threads in a warp. Execution Configuration A call to a kernel subroutine must specify an execution configuration.streamid]]>>>(arg1. the module cudafor defines the derived type dim3 as follows: type(dim3) integer(kind=4) :: x. and a stream identifier. are not accessible in host subprograms.arg2. or are type(dim3).block[. • The variable blockdim contains the dimensions of the thread block.y. Programming Guide complex character(len=1) 4. in addition to the statically allocated shared memory. for one. • The variable threadidx contains the thread index within its thread block.z end type These predefined variables. • The variable blockidx contains the block index within the grid. as with threadidx. Predefined Variables in Device Subprograms Device subprograms have access to block and grid indices and dimensions through several builtin read-only variables. These variables are of type dim3. The execution configuration defines the dimensionality and extent of the grid and thread blocks that execute the subroutine. blockdim has the same value for all thread blocks in the same grid. the value of each component must be equal to or greater than one. The value of blockidx%z is always one. and block%x and block%y must be equal to or greater than one. • The variable griddim contains the dimensions of the grid.or two-dimensional thread blocks. and the product of the component values must be less than or equal to 512. or three-dimensional thread blocks.8 1 Derived types may contain members with these intrinsic datatypes or other allowed derived types.Chapter 2. Threads are executed in groups of 32. • The value of bytes must be an integer. blockidx%y has the value one. except for warpsize. for one-dimensional grids.bytes[. the value of griddim%z is always one. It may also specify a dynamic shared memory extent. the value of grid%z must be one.or two-dimensional grids and one-.

In this way.2. inserting synchronization into the kernel as appropriate. are implemented using stream queues. An operation is placed at the end of the stream queue. If not specified. and no subsequent operations on any queue begin until the stream zero operation is complete. and will only be initiated when all previous operations on that queue have been completed. its value is treated as zero. The work in the loops specified by the directive is executed in parallel. Asynchronous concurrent execution There are two components to asynchronous concurrent execution with CUDA Fortran. stream zero is special in that operations on the stream zero queue will begin only after all preceding operations on all queues are complete.3. the call actually returns to the host program before the kernel subroutine begins execution. The program can use cudaThreadSynchronize() or CUDA Events to wait for the completion of the kernel. such as summing the values in a vector or matrix. used when no stream identifier is specified. the host program should synchronize with the device to avoid erroneous results. including calling or queueing more kernels for execution on the device. and may execute concurrently with each other. where the launch actually corresponds to placing the kernel on a queue for execution by the device. Kernel Loop Directive CUDA Fortran allows automatic kernel generation and invocation from a region of host code containing one or more tightly nested loops. the host program can synchronize and wait for all previously launched or queued kernels. the compiler handles the generation of the final reduction kernel. Each user-created stream manages its own queue. the launch is asynchronous. By calling the runtime routine cudaThreadSynchronize.Asynchronous concurrent execution assumed-size shared variables in the thread block. is stream zero. it is the programmer's responsibility to ensure that parallel execution is legal and produces the correct answer. Concurrent Host and Device Execution When a host subprogram calls a kernel subroutine. • The value of streamid must be an integer greater than or equal to zero. in cases where the host program reads or modifies device or constant data. Concurrent Stream Execution Operations involving the device. As with any kernel. the host can continue executing. An application can manage more concurrency by using multiple streams. For these operations. The default stream. including kernel execution and data copies to and from device memory. The one exception to this rule is a scalar reduction operation. The call can be treated as a kernel launch operation. across the thread blocks and grid. it specifies the stream to which this call is associated. 8 . Launch configuration and mapping of the loop iterations onto the hardware is controlled and specified as part of the directive body using the familiar CUDA chevron syntax. operations on different stream queues may execute out-of-order with respect to when they were placed on the queues. Programmers must be careful when using concurrent host and device execution. see Section 3.

The grid shape. the schedule applies just to the outermost loop. the default value is 1. m do i = 1.j) end do end do In this example. that is.j) +c(i.j) end do end do 9 . m do i = 1. Programming Guide Syntax The general form of the kernel directive is: !$cuf kernel do[(n)] <<< grid. the directive defines a two-dimensional thread block of size 32x4.j) +c(i. is computed by the compiler and runtime by dividing the loop trip counts n and m by the thread block size.j) = b(i.j) end do end do Without an explicit n on the do. Example 1 !$cuf kernel do(2) <<< (*. so all iterations are computed. n a(i.*). Loops which are not mapped onto the grid and block values are run sequentially on each thread. n a(i. The grid and block values can be an integer scalar or a parenthesized list. n a(i. (256. starting at loop n and working out. Alternatively.Chapter 2.j) = b(i. using asterisks tells the compiler to choose a thread block shape and/or compute the grid shape from the thread block shape and the loop limits. you could specify a configuration like the following in which the threads read and write the matrices in coalesced fashion.j) = b(i. You might consider if the code in Example 2 would perform better if the two loops were interchanged. 256 >>> do j = 1.j) +c(i. block >>> The compiler maps the launch configuration specified by the grid and block values onto the outermost n loops. Example 2 !$cuf kernel do <<< *. Alternatively. m do i = 1. !$cuf kernel do(2) <<< *. specified as (*. only the outer j loop is run in parallel with a thread block size of 256. (32. In this case.4) >>> do j = 1. ThreadIdx%y runs from 1 to 4 and is mapped onto the outer j loop.*).1) >>> do j = 1. The inner i dimension is run sequentially on each thread. The body of the doubly-nested loop is turned in the kernel body: ThreadIdx%x runs from 1 to 32 and is mapped onto the inner i loop.

j) = b(i. 10 .*). • Subroutine and function calls to attributes(device) subprograms are allowed if they are in the same module as the code containing the directive. atomic functions. IF statements. are not.j) end do end do Now the threads in a thread block handle all values in the i direction. and increment must be invariant with respect to any other loop in the kernel do. Further expansion of the work along the i direction and all work across the j dimension is handled by the mapping onto the grid dimensions. but the device-specific intrinsics such as syncthreads. • CUDA Fortran intrinsic functions are allowed. the j loop is mapped onto the grid x-dimension. upper limit. incrementing by 256. Specifically. it must be followed by at least that many tightly-nested DO loops. • Arrays used or assigned in the loop must have the device attribute. if they are allowed in device code. Except in the case of reductions. loops.j) +c(i. Summation Example The simplest directive form for performing a dot product on two device arrays takes advantage of the properties for scalar use outlined previously.1) >>> do j = 1. Restrictions on the CUF kernel directive The following restrictions apply to CUF kernel directives: • If the directive specifies n dimensions. • There can be no GOTO or EXIT statements within or between any loops that have been mapped onto the grid and block configuration values. gridDim%x is equal to m. n a(i. and GOTO statements. etc. in concert. One thread block is created for each j. or the compiler will make a device copy of that variable live for the duration of the loops. • Scalars used or assigned in the loop must either have the device attribute. To "unroll" more work into each thread. m do i = 1. the 256 threads in each block each do one element of the matrix addition.Kernel Loop Directive Example 3 In Example 2. • The body of the loops may contain assignment statements. (256. when a reduction has a scalar target. specify non-asterisk values for the grid. In CUDA built-in language. as illustrated here: !$cuf kernel do(2) <<< (1. • The tightly-nested DO loops must have invariant loop limits: the lower limit. one for each thread. • Implicit loops and F90 array syntax are not allowed within the directive loops. • Only CUDA Fortran dataypes are allowed within the loops. the compiler generates a correct sequence of synchronized operations to produce one copy either in device global memory or on the host. because the compiler skips over the constant 1 in the i loop grid size.

Emulation Mode PGI Fortran compilers support an emulation mode for program development on workstations or systems without a CUDA-enabled GPU and for debugging. n rsum = rsum + x(i) * y(i) end do For reductions. the compiler recognizes the use of the scalar and generates just one final result. compile and link with the –Mcuda=emu command line option. Building a CUDA Fortran Program CUDA Fortran is supported by the PGI Fortran compilers when the filename uses a CUDA Fortran extension. such as memory races. In emulation mode. emulation mode may execute a single thread block at a time. which may produce slightly different answers than the device units and intrinsics. in which case the program is processed by the preprocessor before being compiled. add the command line option –Mfixed. To compile a fixed-format program. use of rsum on the host following this loop is undefined. allowing the programmer to use a host debugger. the device code is compiled for and runs on the host. * >>> do i = 1. This will not expose certain errors. In emulation mode. The . This CUF kernel can be followed by another CUF kernel in the same subprogram: !cuf kernel do <<< *. the compiler recognizes rsum as a scalar temporary which should be allocated locally on every thread.Chapter 2.cuf extension specifies that the file is a free-format CUDA Fortran program. Programming Guide rsum = 0. the . It’s important to note that the emulation is far from exact. CUDA Fortran extensions can be enabled in any Fortran source file by adding the –Mcuda command line option.CUF extension may also be used.0 !cuf kernel do <<< *. * >>> do i = 1. n rsum = x(i) * y(i) z(i) = rsum end do In this CUF kernel. 11 . However. In particular. To build a program using emulation mode. the host floating point units and intrinsics are used.

12 .

so no other subroutine prefixes are allowed. A kernel subroutine is compiled as a kernel for execution on the device. as described in the following sections. Global Subroutines The global attribute may be explicitly specified on the Subroutine statement as follows: attributes(global) subroutine sub(…) Functions may not have the global attribute. This chapter describes how to specify the new attributes. to be called from a host routine using an execution configuration.Chapter 3. then default rules are used. and may not contain any other subprogram. their meaning and restrictions. or both. or device attribute. A host subprogram is compiled for execution on the host processor. if there is no attributes prefix on the subprogram statement. A kernel subroutine may not be contained in another subroutine or function. pure. A subroutine with the global attribute is called a kernel subroutine. A Subroutine may have the host. or may have both host and device attribute. pure. Reference This chapter is the CUDA Fortran Language Reference. elemental. New Subroutine and Function Attributes CUDA Fortran adds new attributes to subroutines and functions. A subroutine or function with the host attribute is called a host subroutine or function. Host Subroutines and Functions The host attribute may be explicitly specified on the Subroutine or Function statement as follows: attributes(host) subroutine sub(…) attributes(host) integer function func(…) integer attributes(host) function func(…) The host attributes prefix may be preceded or followed by any other allowable subroutine or function prefix specifiers (recursive. A Function may have the host or device attribute. 13 . or elemental. or a host subprogram. global. function return datatype). A subprogram with no attributes prefix has the host attribute by default. A kernel subroutine may not be recursive. These attributes are specified using the attributes(attr) prefix on the Subroutine or Function statement.

or as an attribute on the type declaration statement. and may only be called from device subprograms in the same module. or elemental. • It may not be contained in another subroutine or function. except for the function return datatype. • It may not contain another subprogram. For more information. refer to “Device code.” on page 21. Restrictions on Device Subprograms A subroutine or function with the device or global attribute must satisfy the following restrictions: • It may not be recursive. Device data A variable or array with the device attribute is defined to reside in the device global memory. to be device arrays of size 100. so no other subroutine or function prefixes are allowed. in the thread block shared memory. in page-locked host memory space. The advantage of using pinned memory is that transfers between the device and pinned memory are faster and can be asynchronous. nor have the recursive prefix on the subprogram statement. This section describes how to specify the new attributes and their meaning and restriction. Variable Attributes CUDA Fortran adds new attributes for variables and arrays. A subroutine or function with the device or kernel attribute is called a device subprogram. in constant memory space. Variables declared in a device subprogram may have one of four attributes: they may be declared to be in device global memory. a and b. or in thread local memory. pure. A subroutine or function with the device attribute must appear within a Fortran module. A device subprogram is compiled for execution on the device. The device attribute can be specified with the attributes statement. Variables in modules may be declared to be in device global memory or constant memory space. 14 . nor have the pure or elemental prefix on the subprogram statement. the array may be declared to be in pinned memory. The following example declares two arrays. that is. CUDA Fortran adds a new attribute for allocatable arrays in host memory. • It may not be pure or elemental.Variable Attributes Device Subroutines and Functions The device attribute may be explicitly specified on the Subroutine or Function statement as follows: attributes(device) subroutine sub(…) attributes(device) datatype function func(…) datatype attributes(device) function func(…) A subroutine or function with the device attribute may not be recursive.

• Members of a derived type may not have the device attribute unless they are allocatable. • Device variables and arrays may not have the Pointer or Target attributes. In host subprograms. Constant data A variable or array with the constant attribute is defined to reside in the device constant memory space. to be constant arrays of size 100. Reference real :: a(100) attributes(device) :: a real. • Device variables and arrays may be passed as actual arguments to host and device subprograms. constant :: d(100) These rules apply to constant data: • Constant data may not have the Pointer. The following example declares two arrays. or as an attribute on the type declaration statement. 15 . The constant attribute can be specified with the attributes statement. in that case. but may not be in a Common block or an Equivalence statement. • Device variables and arrays may appear in modules. • Device variables and arrays declared in a host subprogram cannot have the Save attribute. device :: b(100) These rules apply to device data: • An allocatable device array dynamically allocates device global memory. or may have adjustable extent. and the matching dummy argument must also have the device attribute. the subprogram interface must be explicit (in the Fortran sense). c and d. device data may only be used in the following manner: • In declaration statements • In Allocate and Deallocate statements • As an argument to the Allocated intrinsic function • As the source or destination in a data transfer assignment statement • As an actual argument to a kernel subroutine • As an actual argument to another host subprogram or runtime API call • As a dummy argument in a host subprogram A device array may have the allocatable attribute. or Allocatable attributes. Target. but may not be in a Common block or an Equivalence statement. • Constant variables and arrays may appear in modules. real :: c(100) attributes(constant) :: c real.Chapter 3.

• Shared variables may not be in a Common block or Equivalence statement. variables and arrays with the constant attribute may not be assigned or modified. that is. • Within host subprograms. the value of the bytes argument in the execution configuration is used to specify the number of bytes of shared memory that is dynamically allocated for each thread block. as long as the interface is explicit and the matching dummy argument has the shared attribute. and the matching dummy argument also has the constant attribute. or Allocatable attributes. The shared attribute can be specified with the attributes statement. shared :: d(100) These rules apply to shared data: • Shared data may not have the Pointer. Target. Its size is determined at run time by the call to the kernel. to be shared arrays of size 100. Shared arrays that are not dummy arguments may be declared as assumed-size arrays. • Constant variables and arrays may be passed as actual arguments to host and device subprograms. data with the constant attribute may only be used in the following manner: • In declaration statements • As the source or destination in a data transfer assignment statement • As an actual argument to another host subprogram • As a dummy argument in a host subprogram Shared data A variable or array with the shared attribute is defined to reside in the shared memory space of a thread block. A shared variable or array may only be declared and used inside a device subprogram. variables and arrays with the constant attribute may be read and written. This memory is used for the 16 . shared :: x(*) Such an array has special significance. the last dimension of a shared array may have an asterisk as its upper bound: real. When the kernel is called. The following example declares two arrays.Variable Attributes • Members of a derived type may not have the constant attribute. real :: c(100) attributes(shared) :: c real. as long as the subprogram interface is explicit. • Members of a derived type may not have the shared attribute. In host subprograms. • Within device subprograms. s and t. • Shared variables and arrays may be passed as actual arguments to from a device subprogram to another device subprogram. • Arrays with the constant attribute must have fixed size. or as an attribute on the type declaration statement.

as is done in C. Allocating Device and Pinned Arrays This section describes extensions to the Allocate statement. the declaration for the array must still have the pinned attribute. to be pinned allocatable arrays. dummy arguments are passed by default by reference. Allocating Device Memory Device arrays can have the allocatable attribute. it must be fixed size. Pinned arrays An allocatable array with the pinned attribute will be allocated in special page-locked host memory. Programmers must take this into account when coding. n ) real. or as an attribute on the type declaration statement. 17 .n) :: a. These arrays are dynamically allocated in host subprograms using the Allocate statement. An array with the pinned attribute may be declared in a module or in a host subprogram.Chapter 3. if there is more than one assumed-size shared memory array. Reference assumed-size shared memory arrays in that thread block. specifically for dynamically allocating device arrays and host pinned arrays. The following example declares two arrays. real :: p(:) allocatable :: p attributes(pinned) :: p real. starting at the same shared memory address. This means the actual argument must be stored in device global memory. or whether the dummy argument has the pinned and allocatable attributes. Where the array is deallocated. If a device array declared in a host subprogram does not have the Save attribute. p and q. they are all implicitly equivalenced. The variable arrays corresponding to the dummy arguments a and b must be set up before the call to reside on the device. and other supported methods for allocating device memory. pinned :: q(:) Pinned arrays may be passed as arguments to host subprograms regardless of whether the interface is explicit. attributes(global) subroutine madd( a. and the address of the argument is passed to the subprogram. the value of n can be passed from the host without needing to reside in device memory. or the deallocation may fail. Scalar arguments can be passed by value. allocatable. it will be automatically deallocated when the subprogram returns. following the rules of Fortran. dimension(n. b. when such memory is available. The pinned attribute can be specified with the attributes statement. Value dummy arguments In device subprograms. by adding the value attribute to the variable declaration. value :: n In this case. b integer. If a shared array is not a dummy argument and not assumed-size. and dynamically deallocated using the Deallocate statement.

allocatable. allocatable. 100) … istat = cudaFree(v) For a complete list of the memory management runtime routines. Allocating Pinned Memory Allocatable arrays with the pinned attribute are dynamically allocated using the Allocate statement. device :: v(:) istat = cudaMalloc(v. or if it is exhausted. device :: b(:) allocate(b(5024). device :: atmp(4) … end subroutine vfunc ! adev and atmp are deallocated Allocating Device Memory Using Runtime Routines For programmers comfortable with the CUDA C programming environment. To use these. These arrays will also have a lifetime of the subprogram as is usual with the Fortran language: subroutine vfunc(a. The cudaMalloc function can be used to allocate single-dimensional arrays of the supported intrinsic datatypes. declare and initialize the scalar on the host as: integer. device :: ndev allocate(ndev) ndev = 100 The language also supports the ability to create the equivalent of automatic and local device arrays without using the allocate statement. If no such memory space is available.c.” on page 32. real. These functions return memory which will bypass certain Fortran allocatable properties such as automatic deallocation. allocatable. the compiler allocates the array in normal paged host memory. pinned allocatable arrays work and act like any other allocatable array on the host. Mixing standard Fortran allocate/deallocate with the runtime Malloc/Free for a given array is not supported.stat=istat) … if(allocated(p)) deallocate(p) 18 . and cudaFree can be used to free it: real.n) real. refer to“Memory Management. Fortran interfaces to the CUDA memory management runtime routines are provided. The compiler will generate code to allocate the array in host page-locked memory. device :: adev(n) real. pinned :: p(:) allocate(p(5000). allocatable.stat=istat) … if(allocated(b)) deallocate(b) Scalar variables can be allocated on the device using the Fortran 2003 allocatable scalar feature. if available. Otherwise.Allocating Device and Pinned Arrays real. and thus the arrays are treated more like C malloc’ed areas.

Similarly. This means such data copies are synchronous. It returns a logical success value. copies data from the device global memory to the host memory. • An assignment statement where the left hand side is a host variable or host array or array section. and cdev are conforming device arrays. • An assignment statement where the left hand side is a device variable or device array or array section. which normally only allows one device array to appear on the righthand-side of an expression. Compiler-generated temporary arrays are generated to accommodate the host copies of device data as needed. Implicit Data Transfer in Expressions Some limited data transfer can be enclosed within expressions. and the right hand side is a device variable or device array or array section. pinned=plog) if (. bdev. you can use simple assignment statements to copy or assign variables or arrays with the constant attribute. the rule of thumb is all arithmetic or operations must occur on the host. and the right hand is a host variable or host array or array section. In general. . Note Using assignment statements to read or write device or constant data implicitly uses CUDA stream zero. stat=istat. and c are conforming host arrays. b. Data transfer between host and device memory This section provides methods to transfer data between the host and device memory. logical plog allocate(p(5000). and adev. an additional PINNED keyword is added to the allocate statement. meaning the data copy waits until all previous kernels and data copies complete. For instance. plog) then . .Chapter 3. • An assignment statement with a device variable or device array or array section on both sides of the assignment statement copies data between two device variables or arrays. Reference To determine whether or not the allocation from page-locked memory was successful.not. if a. the following expressions are legal: a = adev adev = a b = a + adev c = x * adev + b 19 . Data Transfer Using Assignment Statements You can copy variables and arrays from the host memory to the device memory by using simple assignment statements in host subprograms. copies data from the host memory to the device global memory.

1).Invoking a kernel subroutine The following expressions are not legal as they either promote a false impression of where the actual computation occurs. If it is type(dim3). not bytes. Fortran interfaces to the CUDA memory management runtime routines are provided. If grid is an integer. it is converted to dim3(grid.1. device to host. it has the form: <<< grid. Array slices are also supported. or both: c = adev + bdev adev = adev + a b = sqrt(adev) Elemental transfers are supported by the language but will perform poorly. and the implementation. which must be less than the maximum supported by the device. Data Transfer Using Runtime Routines For programmers comfortable with the CUDA C programming environment. If it is type(dim3). The product grid%x*grid%y gives the number of thread blocks to launch. amount of contiguous data in the slices. The execution configuration is specified after the subroutine name in the call statement. or of type(dim3).” on page 32. For a complete list of memory management runtime routines. or would be more efficient written in another way. cur.1. 512) For those familiar with the CUDA C routines. it is converted to dim3(block. The execution configuration gives the size and shape of the grid and thread blocks that execute the function. refer to “Memory Management. These functions can transfer data either from the host to device. block. the value of grid%z must be one. stream >>> • grid is an integer.1). bl • block is an integer. or from one device array to another. and the associated stream. The cudaMemcpy function can be used to copy data between the host and the GPU: real. 20 . and their performance is dependent on the size of the slice. bytes. device :: wrk(1024) real cur(512) istat = cudaMemcpy(wrk. as well as the amount of shared memory to use for assumed-size shared memory arrays. Invoking a kernel subroutine A call to a kernel subroutine must give the execution configuration for the call. Counts expressed in arguments to the Fortran runtime routines are expressed in terms of data type elements. If block is an integer. or of type(dim3). the number of threads per thread block is block%x*block%y*block%z. the kind parameter to the Memcpy routines is optional in Fortran since the attributes of the arrays are explicitly declared.

it must be an integer. the value zero is used.2. If not specified. such as the function arguments and execution configuration arguments. and have a value of zero.4(default). where the members of the derived type have one of the allowed intrinsic datatypes. or shared attributes.z end type 21 . For more information. allowing for static shared memory declared in the kernel and for other dedicated uses. or another allowed derived type.Chapter 3. Reference • bytes is optional.” on page 16. if present. See Section 4. a kernel subroutine attributes(global) subroutine sub( a ) can be called like: call sub <<< DG. refer to “Shared data. Table 3. Device Code Intrinsic Datatypes Type integer logical real double precision complex character(len=1) Type Kind 1.1.8 4(default). if present. Device code Datatypes allowed Variables and arrays with the device. constant.8 equivalent to real(kind=8) 4(default). • stream is optional.5 on page 41. or a value returned by a call to cudaStreamCreate. They may have any of the intrinsic datatypes in the following table.8 1. For instance. and specifies the number of bytes of shared memory to be allocated for each thread block to use for assumed-size shared memory arrays.2. they may be of derived type.8 1 (default) Additionally. it must be a scalar integer.y.4(default). It specifies the stream to which this call is enqueued. The system module cudafor includes definitions of the derived type dim3. are limited to the types described in this section. or declared in device subprograms. or if bytes is greater than the shared memory available. defined as type(dim3) integer(kind=4) :: x. DB >>> ( A ) The function call will fail if the grid or block arguments are greater than the maximum sizes allowed.

These variables are read-only. real integer. Fortran intrinsics This section lists the Fortran intrinsic functions allowed in device subprograms. blockdim has the same value for all threads in the same grid. as with threadidx. blockidx%y has the value one. blockidx. • The variables threadidx. real. blockidx. real real Name int logical max min mod modulo nint real sign Argument Datatypes integer. real 22 . The value of griddim is the same for all threads in the same grid. It has constant value. the value of griddim%z is always one. blockdim. the threadidx%y and/or threadidx%z components have the value one. Fortran Numeric and Logical Intrinsics Name abs aimag aint anint ceiling cmplx conjg dim floor Argument Datatypes integer. real. griddim integer(4) :: warpsize • The variable threadidx contains the thread index within its thread block. real integer. currently defined to be 32. for one. real. • The variable blockidx contains the block index within the grid. • The variable griddim contains the dimensions of the grid. the blockdim%y and/or blockdim %z components have the value one. real integer. They are declared as follows: type(dim3) :: threadidx.2.or two-dimensional thread blocks. real real integer. blockdim. The value of blockidx%z is always one.real) complex integer.or two-dimensional thread blocks. complex logical integer. complex complex real real real real or (real. and griddim are available only in device subprograms. complex integer. the value of griddim%y is one for one-dimensional grids. for onedimensional grids. • The variable warpsize contains the number of threads in a warp. Fortran Numeric and Logical Intrinsics Table 3.Device code Builtin variables The system module cudafor declares several predefined variables. The value of blockidx is the same for all threads in the same thread block. • The variable blockdim contains the dimensions of the thread block. the value of griddim%z is always one. for one.

Fortran Mathematical Intrinsics Name acos asin atan atan2 cos cosh exp Argument Datatypes real real real (real.integer) real Fortran Bit Manipulation Intrinsics Table 3. complex real real. complex real real.real) real. Fortran Numeric Inquiry Intrinsics Name bit_size digits epsilon huge maxexponent minexponent Argument Datatypes integer integer. real integer. real real real Name precision radix range selected_int_kind selected_real_kind tiny Argument Datatypes real.Chapter 3.5. complex real real. real. complex integer. complex integer (integer. Reference Fortran Mathematical Intrinsics Table 3. complex Name log log10 sin sinh sqrt tan tanh Argument Datatypes real. Fortran Bit Manipulation Intrinsics Name btest iand ibclr ibits ibset ieor ior Argument Datatypes integer integer integer integer integer integer integer Name ishft ishftc leadz mvbits not popcnt poppar Argument Datatypes integer integer integer integer integer integer integer 23 .3. real real integer. complex real real Fortran Numeric Inquiry Intrinsics Table 3.4.

This 24 . Synchronization Functions The synchronization functions control the synchronization of various threads during execution of thread blocks. refer to “Thread Management. thus creating potential read-after-write. real integer. or write-after-write hazards for some of these memory accesses. syncthreads syncthreads_count syncthreads_and syncthread_or threadfence threadfence_block threadfence_system For detailed information on these functions.Device code Fortran Reduction Intrinsics Table 3. To avoid these potential issues.6. real Name minloc minval product sum Argument Datatypes integer. real integer. Fortran Random Number Intrinsics Name random_number random_seed Argument Datatypes real integer New Intrinsic Functions This section describes the new intrinsic functions and subroutines supported in device subprograms. it has no arguments: void syncthreads() Sometimes threads within a block access the same addresses in shared or global memory. . real. complex integer.7. use syncthreads()to specify synchronization points in the kernel. write-after-read. real. SYNCTHREADS The syncthreads intrinsic subroutine acts as a barrier synchronization for all threads in a single thread block.” on page 32. real integer. Fortran Reduction Intrinsics Name all any count maxloc maxval Argument Datatypes logical logical logical integer. complex Fortran Random Number Intrinsics Table 3.

SYNCTHREADS_OR integer syncthreads_or(int_value) syncthreads_or. If no fence is placed between storing the partial sum and incrementing the counter. or the kernel fails to complete correctly. suppose you use a kernel to compute the sum of an array of N numbers in one call. To determine which block is finished last. Threads within a block cooperate and share data by synchronizing their execution to coordinate memory accesses. If any thread in a thread block issues a call to syncthreads. Reference intrinsic acts as a barrier at which all threads in the block must wait before any thread is allowed to proceed. acts a as a barrier at which all threads in the block must wait before any thread is allowed to proceed. other threads may see the effects of these memory writes in a different order. Typically. When all blocks are done. like syncthreads. In addition. like syncthreads. the last block done reads each of these partial sums from global memory and sums them to obtain the final result. SYNCTHREADS_COUNT integer syncthreads_count(int_value) syncthreads_count. SYNCTHREADS_AND integer syncthreads_and(int_value) syncthreads_and. and threadfence_system() to create a memory fence to enforce ordering. acts as a barrier at which all threads in the block must wait before any thread is allowed to proceed. syncthreads_and evaluates the integer argument int_value for all threads of the block and returns non-zero if and only if int_value evaluates to non-zero for all of them. the counter might increment before the partial sum is stored. Memory Fences In general. THREADFENCE void threadfence() threadfence acts as a memory fence. acts as a barrier at which all threads in the block must wait before any thread is allowed to proceed. In addition. each block atomically increments a counter to signal that it is done with computing and storing its partial sum. creating a wait. all threads must also reach and execute the same call statement. You can use threadfence(). other threads may see the effects of these memory writes in a different order. Each block first sums a subset of the array and stores the result in global memory. 25 . syncthreads_or evaluates the integer argument int_value for all threads of the block and returns non-zero if and only if int_value evaluates to non-zero for any of them. In addition. Each thread in a thread block pauses at the syncthreads call until all threads have reached that call. syncthreads_count evaluates the integer argument int_value for all threads of the block and returns the number of threads for which int_value evaluates to non-zero. like syncthreads. when a thread issues a series of writes to memory in a particular order. threadfence_block(). For example. when a thread issues a series of writes to memory in a particular order.Chapter 3.

The function allthreads evaluates its argument for all threads in the current warp. only if the value of the argument is . for all threads in the warp.0) ) allneg = . The clock frequency can be determined by calling cudaGetDeviceProperties. Each of these functions has a single argument. 26 . creating a wait 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 threadfence_system() is only supported by devices of compute capability 2. as described in “cudaGetDeviceProperties. GPU_TIME The gpu_time intrinsic returns the value of the clock cycle counter on the GPU.true.2 and higher.true. ALLTHREADS The allthreads function is a warp-vote operation with a single scalar logical argument: if( allthreads(a(i)<0.0 or higher.true.” on page 32. 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 THREADFENCE_BLOCK void threadfence_block() threadfence_block acts as a memory fence. THREADFENCE_SYSTEM void threadfence_system() threadfence_system acts as a memory fence. It has a single argument: integer(8) clock call gpu_time(clock) The argument to gpu_time is set to the value of the clock cycle counter.Device code threadfence() is one method to enforce a specific order. Warp-Vote Operations Warp-vote operations are only supported by devices with compute capability 1. The value of the function is . creating a wait 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 for all accesses.

Atomic functions are only supported by devices with compute capability 1. This function is only supported by devices of compute capability 2. The function anythread evaluates its argument for all threads in the current warp. Note The return value for each of these functions is the first argument.8. value ) atomicsub( mem. Atomic Functions The atomic functions read and write the value of their first operand. The atomic functions return correct values even if multiple threads in the same or different thread blocks try to read and update the same location without any synchronization. and store the combined value back to the first argument location. for all threads in the warp. only if the value of the argument is . Arithmetic and Bitwise Atomic Functions These atomic functions read and return the value of the first argument.true. Arithmetic and Bitwise Atomic Functions Function atomicadd( mem. They also combine that value with the value of the second argument.2 or higher is required if the first argument has the shared attribute. Reference ANYTHREAD The anythread function is a warp-vote operation with a single scalar logical argument: if( anythread(a(i)<0.false. depending on the function.1 and higher. These functions are: Table 3.true.Chapter 3. The value of the function is . mem. value ) Additional Atomic Update mem = mem + value mem = mem – value 27 .0) ) allneg = . which must be a variable or array element in shared memory (with the shared attribute) or in device global memory (with the device attribute). Example: if( ballot(int_value) ) allneg = . BALLOT The ballot function is a warp-vote operation with a single integer argument: unsigned integer ballot(int_value) The function ballot evaluates the argument int_value for all threads of the warp and returns an integer whose Nth bit is set if and only if int_value evaluates to non-zero for the Nth thread of the warp. Compute capability 1.false.0. Both arguments must be of type integer(kind=4).

mem>0) then mem = mem-1 else mem = imax endif atomicdec( mem. mem. imax ) Additional Atomic Update if (mem<imax) then mem = mem+1 else mem = 0 endif if (mem<imax . Both arguments must be of type integer(kind=4). Counting Atomic Functions Function atomicinc( mem. value ) atomicand( mem.value) mem = ior(mem. Note The return value for this function is the first argument. value ) atomicmin( mem.value) mem = min(mem. All three arguments must be of type integer(kind=4). value ) atomicor( mem. These functions are intended to implement circular counters. These functions are: Table 3. value ) mem = max(mem. They also compare the first argument with the second argument.value) mem = iand(mem. mem. and atomically stores a new value back to the first argument location if the first and second argument are equal.Device code atomicmax( mem.9. depending on the result of the comparison.value) mem = ieor(mem. value ) atomicexch( mem.value) mem = value Counting Atomic Functions These atomic functions read and return the value of the first argument.and. It also compares the first argument with the second argument. imax ) Compare and Swap Atomic Function This atomic function reads and returns the value of the first argument. counting up to or down from a maximum value specified in the second argument. value ) atomicxor( mem. Note The return value for each of these functions is the first argument. and stores a new value back to the first argument location. The function is: 28 .

j j = sizeof(i) ! this assigns the value 4 to j 29 . Compare and Swap Atomic Function Function atomiccas(mem. SIZEOF Intrinsic A call to sizeof(A). ENDFILE. • Objects with character type must have LEN=1. • Automatic arrays must be fixed size. such as the new sizeof intrinsic function.6.comp. • Subroutine and function calls are supported only if they can be inlined. CLOSE. • Fortran intrinsic functions not listed in Section 3. where A is a variable or expression. • ENTRY statements are not allowed.Chapter 3. • Alternate return specifications are not allowed.10. character substrings are not supported. Reference Table 3. WRITE. REWIND.3 are not supported. • STOP and PAUSE statements are not allowed. • Objects with the Pointer and Allocatable attribute are not allowed. • Cray pointers are not supported. integer(kind=4) :: i. • Optional arguments are not allowed. • Floating point exception handling is not supported. Host code Host subprograms may use intrinsic functions. INQUIRE. OPEN. FORMAT. NAMELIST. BACKSPACE.val) Additional Atomic Update if (mem == comp) then mem = val endif Restrictions This section lists restrictions on statements and features that can appear in device subprograms. • Recursive subroutines and functions are not allowed. • Input/Output statements are not allowed: READ. PRINT. returns the number of bytes required to hold the value of A.

30 .

refer to “Error Handling. 31 . intent(out) :: numdev cudaGetDeviceCount assigns the number of available devices to its first argument. and a nonzero value if there was an error.” on page 41. Tip When doing timing runs. Initialization No explicit initialization is required. cudaGetDeviceCount integer function cudaGetDeviceCount( numdev ) integer. be aware that initialization can add some overhead. Device Management Use the functions in this section for device management. To interpret the error codes. Runtime APIs The system module cudafor defines the interfaces to the Runtime API routines. cudaSetDeviceFlags integer function cudaSetDevice( flags ) integer. intent(in) :: devnum cudaSetDevice selects the device to associate with this host thread. the runtime initializes and connects to the device the first time a runtime routine is called or a device array is allocated.Chapter 4. intent(in) :: flags cudaSetDeviceFlags records how the CUDA runtime interacts with this host thread. Most of the runtime API routines are integer functions that return an error code. cudaSetDevice integer function cudaSetDevice( devnum ) integer. they return a value of zero if the call was successful.

as TYPE(C_DEVPTR). write-after-read. intent(out) :: devnum type(cudadeviceprop). as arguments to simplify interfacing to existing CUDA C code. defined in the cudafor module. intent(out) :: prop integer. intent(in) :: devnum cudaGetDeviceProperties returns the properties of a given device. prop ) integer. use the functions in this section for thread management. It may return an error condition if one of the preceding operations fails. CUDA Fortran has extended the F2003 derived type TYPE(C_PTR) by providing a C device pointer. Thread Management Sometimes threads within a block access the same addresses in shared or global memory. cudaThreadExit integer function cudaThreadExit() cudaThreadExit explicitly cleans up all runtime-related CUDA resources associated with the host thread. devnum ) type(cudadeviceprop). Calling cudaThreadExit is optional. cudaThreadSynchronize integer function cudaThreadSynchronize() cudaThreadSynchronize blocks execution of the host subprogram until all preceding kernels and operations are complete. it is implicitly called when the host thread exits. or write-after-write hazards for some of these memory accesses. provided through the Fortran 2003 iso_c_binding module. cudaChooseDevice integer function cudaChooseDevice ( devnum. cudaGetDeviceProperties integer function cudaGetDeviceProperties( prop. Memory Management Many of the memory management routines can take device arrays as arguments. Any subsequent CUDA calls or operations will reinitialize the runtime.Thread Management cudaGetDevice integer function cudaGetDevice( devnum ) integer. intent(in) :: prop cudaChooseDevice assigns the device number that best matches the properties given in prop to its first argument. intent(out) :: devnum cudaGetDevice assigns the device number associated with this host thread to its first argument. To avoid these potential issues. thus creating potential read-after-write. Consistent use of TYPE(C_PTR) and 32 . Some can also take C types.

should be of benefit. cudaFree integer function cudaFree(devptr) cudaFree deallocates data on the device. “Device Code Intrinsic Datatypes. height) cudaMallocPitch allocates data on the device. devptr may be of TYPE(C_DEVPTR). cudaMallocPitch integer function cudaMallocPitch(devptr. an allocatable device array as the second argument. count) cudaMalloc allocates data on the device.” on page 21. height cudaMallocArray allocates a data array on the device. in which case the integer values are expressed in bytes.1. cdesc.” on page 21. The count is in terms of elements. proper Fortran pointers for device data are supported. as well as consistency checks between Fortran device arrays and host arrays.Chapter 4. The height is an integer. and the padded width is returned in the variable pitch. devptr may also be of TYPE(C_DEVPTR). devptr may be any allocatable. cudaFreeArray integer function cudaFreeArray(carray) type(cudaArrayPtr) :: carray 33 . it is possible to construct a Fortran device array out of a TYPE(C_DEVPTR) by using an extension of the iso_c_binding subroutine c_f_pointer. there is also a function C_DEVLOC() defined which will create a TYPE(C_DEVPTR) that holds the C address of the Fortran device array argument. in which case the count is in bytes. Both of these features are subject to change when.” on page 21. devptr may be any allocatable device array of a supported type specified in Table 3. cudaMallocArray integer function cudaMallocArray(carray. two-dimensional device array of a supported type specified in Table 3.1. Under CUDA Fortran. “Device Code Intrinsic Datatypes. height) type(cudaArrayPtr) :: carray type(cudaChannelFormatDesc) :: cdesc integer :: width. c_f_pointer will take a TYPE(C_DEVPTR) as the first argument. Currently. cudaMalloc integer function cudaMalloc(devptr. cudaMallocPitch may pad the data. The width is in terms of number of elements. width. Runtime APIs TYPE(C_DEVPTR). “Device Code Intrinsic Datatypes. width. Or. Similarly. pitch. devptr may be of TYPE(C_DEVPTR). Or. in the future. one-dimensional device array of a supported type specified in Table 3.1. and in effect transfer the allocation to the Fortran array. a shape as the third argument. Use the functions in this section for memory management. devptr may be any allocatable.

cudaMemcpy2D integer function cudaMemcpy2D(dst. in which case the count is in term of bytes. kdir) cudaMemcpy copies data from one location to another.” on page 21. value. it must be one of the defined enums cudaMemcpyHostToDevice. of a supported type specified in Table 3. dst and src may be any device or host. cudaMemset integer function cudaMemset(devptr. and height are in terms of elements. stream) cudaMemcpy copies data from one location to another. If kdir is specified. width. in which case the count is in term of bytes. dst and src may be any device or host array. “Device Code Intrinsic Datatypes. The width 34 .” on page 20. The count is in terms of elements. dpitch. kdir may be optional. for more information.” on page 20.” on page 21. dst and src may be of TYPE(C_DEVPTR) or TYPE(C_PTR). “Device Code Intrinsic Datatypes. src. This function operates on page-locked host memory only. of a supported type specified in Table 3. height. The value must match in type and kind. devptr may be of TYPE(C_DEVPTR).1. kdir may be optional. scalar or array. kdir) cudaMemcpy2D copies data from one location to another. count) cudaMemset sets a location or array to the specified value. The count is in terms of elements. pitch. “Device Code Intrinsic Datatypes. count. dst and src may be any device or host. spitch. width.1. devptr may be any device array of a supported type specified in Table 3. it must be one of the defined enums cudaMemcpyHostToDevice. cudaMemcpyDeviceToHost. src. kdir. devptr may be any device scalar or array of a supported type specified in Table 3.Memory Management cudaFreeArray frees an array that was allocated on the device. cudaMemcpyAsync integer function cudaMemcpyAsync(dst. The pitch. cudaMemcpy integer function cudaMemcpy(dst.” on page 21. devptr may be of TYPE(C_DEVPTR). and the lowest byte of value is used.” on page 21. in which case the pitch.1.1.1. width. refer to “Data Transfer Using Runtime Routines. scalar or array. in which case the count is in term of bytes. width. refer to “Data Transfer Using Runtime Routines. Or. “Device Code Intrinsic Datatypes. or cudaMemcpyDeviceToDevice. src. The copy can be associated with a stream by passing a non-zero stream argument. otherwise the stream argument is optional and defaults to zero. cudaMemcpyDeviceToHost. “Device Code Intrinsic Datatypes. The value must match in type and kind. height) cudaMemset2D sets an array to the specified value. Or. and the lowest byte of value is used. or cudaMemcpyDeviceToDevice. Alternatively. count. value. Alternatively. The count is in terms of elements. cudaMemset2D integer function cudaMemset2D(devptr. If kdir is specified. dst and src may be of TYPE(C_DEVPTR) or TYPE(C_PTR). and height are in terms of bytes. for more information. of a supported type specified in Table 3.” on page 21.

srca. kdir cudaMemcpy2DToArray copies array data to and from the device. spitch. spitch. height. dst and src may be of TYPE(C_DEVPTR) or TYPE(C_PTR). kdir) type(cudaArrayPtr) :: srca integer :: dpitch. count. it must be one of the defined enums cudaMemcpyHostToDevice. If kdir is specified. dsty. srcx.Chapter 4. stream) cudaMemcpy2D copies data from one location to another. dstx. dst and src may be of TYPE(C_DEVPTR) or TYPE(C_PTR). count. spitch. srcx. dsty.” on page 21. src. refer to “Data Transfer Using Runtime Routines. cudaMemcpyFromArray integer function cudaMemcpyFromArray(dst. otherwise the stream argument is optional and defaults to zero. cudaMemcpyDeviceToHost. cudaMemcpy2DFromArray integer function cudaMemcpy2DFromArray(dst. kdir may be optional. width. kdir. Alternatively. height.” on page 20. This function operates on page-locked host memory only. for more information. or cudaMemcpyDeviceToDevice. height. cudaMemcpy2DToArray integer function cudaMemcpy2DToArray(dsta. If kdir is specified. cudaMemcpy2DAsync integer function cudaMemcpy2DAsync(dst. width. srca. kdir cudaMemcpyFromArray copies array data to and from the device. srcy. kdir) type(cudaArrayPtr) :: dsta integer :: dstx. height. refer to “Data Transfer Using Runtime Routines. it must be one of the defined enums cudaMemcpyHostToDevice. Alternatively. dst and src may be any device or host array. src. kdir) type(cudaArrayPtr) :: srca integer :: dstx. width. kdir cudaMemcpy2DFromArray copies array data to and from the device. cudaMemcpyToArray integer function cudaMemcpyToArray(dsta. dsty. The copy can be associated with a stream by passing a non-zero stream argument. width. dsty. Runtime APIs and height are in terms of elements. for more information.” on page 20. in which case the width and height are in term of bytes. kdir may be optional. srcy. srcy. width. src. dsty. dpitch. count. cudaMemcpyDeviceToHost. dstx. in which case the width and height are in term of bytes.1. height. of a supported type specified in Table 3. “Device Code Intrinsic Datatypes. kdir cudaMemcpyToArray copies array data to and from the device. 35 . count. or cudaMemcpyDeviceToDevice. srcx. The width and height are in terms of elements. kdir) type(cudaArrayPtr) :: dsta integer :: dstx. dpitch.

intent(out) :: pitchptr type(cudaExtent). srca. srcy. value. cudaMemcpy2DArrayToArray integer function cudaMemcpy2DArrayToArray(dsta. srca. dsty. dstx. cudaMemcpy3D integer function cudaMemcpy3D(p) type(cudaMemcpy3DParms) :: p cudaMemcpy3D copies elements from one 3D array to another as specified by the data held in the derived type p. height. kdir) type(cudaArrayPtr) :: dsta. srcx. cdesc. height. cudaMalloc3DArray integer function cudaMalloc3DArray(carray. srcx. count. three-dimensional device array of a supported type specified in “Datatypes allowed. Alternatively. width. kdir) type(cudaArrayPtr) :: dsta. cext) type(cudaPitchedPtr). srcx. the extents in each dimension specified by cext. dsty. which was allocated with cudaMalloc3D to a specified value. srca integer :: dstx.Memory Management cudaMemcpyArrayToArray integer function cudaMemcpyArrayToArray(dsta. count. dsty. cudaMalloc3D integer function cudaMalloc3D(pitchptr. dstx. srca integer :: dstx. kdir cudaMemcpyArrayToArray copies array data to and from the device. srcy. width. cext) type(cudaPitchedPtr) :: pitchptr integer :: value type(cudaExtent) :: cext cudaMemset3D sets elements of an array. srcy. dsty. kdir cudaMemcpy2DArrayToArray copies array data to and from the device. srcy. intent(in) :: cext cudaMalloc3D allocates data on the device. cudaMemset3D integer function cudaMemset3D(pitchptr. srcx. pitchptr may be any allocatable.” on page 21. 36 . cext is also a derived type which holds the extents of the allocated array. pitchptr is a derived type defined in the cudafor module. cext) type(cudaArrayPtr) :: carray type(cudaChannelFormatDesc) :: cdesc type(cudaExtent) :: cext cudaMalloc3DArray allocates array data on the device.

kdir integer. src. count. The copy can be associated with a stream by passing a non-zero stream argument. src may be any host scalar or array of a supported type specified in “Datatypes allowed. dst may be any host scalar or array of a supported type specified in “Datatypes allowed. cudaMemcpyToSymbolAsync integer function cudaMemcpyToSymbolAsync(symbol. src.” on page 21. This function operates on page-locked host memory only. This function operates on page-locked host memory only. cudaMemcpyFromSymbolAsync integer function cudaMemcpyFromSymbolAsync(dst. kdir cudaMemcpyToSymbol copies data from the source to a device area in global or constant memory space referenced by a symbol. stream) type(cudaMemcpy3DParms) :: p integer :: stream cudaMemcpy3DAsync copies elements from one 3D array to another as specified by the data held in the derived type p. kdir cudaMemcpyFromSymbol copies data from a device area in global or constant memory space referenced by a symbol to a destination on the host. The copy can be associated with a stream by passing a non-zero stream argument. src may be any host scalar or array of a supported type as specified in “Datatypes allowed. offset. stream) type(cudaSymbol) :: symbol integer :: count. offset. kdir) type(cudaSymbol) :: symbol integer :: count.Chapter 4.” on page 21. kdir. symbol. optional :: stream cudaMemcpyToSymbol copies data from the source to a device area in global or constant memory space referenced by a symbol. offset. offset. The count is in terms of elements. count. count. offset. optional :: stream 37 . symbol. kdir) type(cudaSymbol) :: symbol integer :: count. The count is in terms of elements. kdir. kdir integer. stream) type(cudaSymbol) :: symbol integer :: count. offset. offset.” on page 21. count. offset. cudaMemcpyToSymbol integer function cudaMemcpyToSymbol(symbol. The count is in terms of elements. cudaMemcpyFromSymbol integer function cudaMemcpyFromSymbol(dst. Runtime APIs cudaMemcpy3DAsync integer function cudaMemcpy3D(p.

cudaGetSymbolAddress integer function cudaGetSymbolAddress(devptr. symbol) integer :: size type(cudaSymbol) :: symbol cudaGetSymbolSize sets the variable size to the size of a device area in global or constant memory space referenced by the symbol. a kernel which runs on 38 . hostptr. The following code sequence initializes a global device array “vx” from a CUDA C kernel: type(cudaSymbol) :: csvx type(c_devptr) :: cdvx real. allocatable.” on page 21. and read and written by.0 cudaGetSymbolSize integer function cudaGetSymbolSize(size. The copy can be associated with a stream by passing a non-zero stream argument. vx. Size is in bytes. This function operates on page-locked host memory only. flags) type(C_DEVPTR) :: devptr type(C_PTR) :: hostptr integer :: flags cudaHostGetDevicePointer returns a pointer to a device memory address corresponding to the pinned memory on the host. cudaHostGetDevicePointer integer function cudaHostGetDevicePointer(devptr. The normal iso_c_binding subroutine c_f_pointer can be used to move the type(c_ptr) to a Fortran pointer. hostptr is a pinned memory buffer that was allocated via cudaHostAlloc(). csvx) Call c_f_pointer(cdvx. dst may be any host scalar or array of a supported type specified in “Datatypes allowed. size. flags) type(C_PTR) :: hostptr integer :: size. or returns an error if the memory is unavailable.Memory Management cudaMemcpyFromSymbol copies data from a device area in global or constant memory space referenced by a symbol to a destination on the host. 100) Vx = 0. flags cudaHostAlloc allocates pinned memory on the host. It returns in devptr an address that can be passed to. The count is in terms of elements. device :: vx(:) csvx = “vx” Istat = cudaGetSymbolAddress(cdvx. symbol) type(C_DEVPTR) :: devptr type(cudaSymbol) :: symbol cudaGetSymbolAddress returns in the devptr argument the address of symbol on the device. A symbol can be set to an external device name via a character string. cudaHostAlloc integer function cudaHostAlloc(hostptr. The flags argument enables different options to be specified that affect the allocation. It returns in hostptr the address of the pagelocked allocation.

intent(in) :: stream cudaStreamQuery tests whether all operations enqueued to the selected stream are complete. Stream Management Use the functions in this section for stream management.Chapter 4. intent(in) :: stream cudaStreamSynchronize blocks execution of the host subprogram until all preceding kernels and operations associated with the given stream are complete. cudaStreamCreate integer function cudaStreamCreate( stream ) integer. it returns zero (success) if all operations are complete. cudaStreamSynchronize integer function cudaStreamSynchronize( stream ) integer. The flags argument is provided for future releases. intent(out) :: stream cudaStreamCreate creates an asynchronous stream and assigns its identifier to its first argument. cudaMallocHost integer function cudaMallocHost(hostptr. It returns in hostptr the address of the page-locked allocation. size is in bytes. 39 . or returns an error if the memory is unavailable. asynchronous operations. The normal iso_c_binding subroutine c_f_pointer can be used to move the type(c_ptr) to a Fortran pointer. It may also return another error condition if some asynchronous operations failed. The normal iso_c_binding subroutine c_f_pointer can be used to move the type(c_devptr)to a device array. size) type(C_PTR) :: hostptr integer :: size cudaMallocHost allocates pinned memory on the host. cudaStreamDestroy integer function cudaStreamDestroy( stream ) integer. Runtime APIs the device. cudaStreamQuery integer function cudaStreamQuery( stream ) integer. intent(in) :: stream cudaStreamDestroy releases any resources associated with the given stream. and the value cudaErrorNotReady if not. It may return error codes from previous. cudaFreeHost integer function cudaFreeHost(hostptr) type(C_PTR) :: hostptr cudaFreeHost deallocates pinned memory on the host allocated with cudaMalloHost.

start. intent(in) :: event cudaEventDestroy destroys the resources associated with an event object. cudaEventDestroy integer function cudaEventDestroy( event ) type(cudaEvent). intent(in) :: event cudaEventQuery tests whether an event has been recorded. cudaEventCreate integer function cudaEventCreate( event ) type(cudaEvent). cudaEventElapsedTime integer function cudaEventElapsedTime( time. end) float :: time type(cudaEvent). It returns cudaErrorInvalidValue if cudaEventRecord has not been called for this event. It returns cudaErrorInvalidValue if either event has not yet been recorded. intent() :: start. This function is only valid with events recorded on stream zero. the event is recorded after all preceding operations in all streams are complete. intent(in) :: event cudaEventSynchronize blocks until the event has been recorded. end cudaEventElapsedTime computes the elapsed time between two events (in milliseconds). intent(in) :: stream cudaEventRecord issues an operation to the given stream to record an event. intent(out) :: event cudaEventCreate creates an event object and assigns the event identifier to its first argument cudaEventRecord integer function cudaEventRecord( event. stream ) type(cudaEvent). intent(in) :: event integer. and cudaErrorNotReady if it has not. cudaEventQuery integer function cudaEventQuery( event ) type(cudaEvent). The event is recorded after all preceding operations in the stream are complete. If stream is zero. It returns success (zero) if the event has been recorded. cudaEventSynchronize integer function cudaEventSynchronize( event ) type(cudaEvent). It returns a value of cudaErrorInvalidValue if cudaEventRecord has not been called for this event. 40 .Event Management Event Management Use the functions in this section to manage events.

41 . then it returns 0 as iversion. cudaGetErrorString function cudaGetErrorString( errcode ) integer. intent(in) :: errcode character*(*) :: cudaGetErrorString cudaGetErrorString returns the message string associated with the given error code. cudaRuntimeGetVersion integer function cudaRuntimeGetVersion(iversion) integer :: iversion cudaRuntimeGetVersion returns the version number of the installed CUDA Runtime as iversion. cudaDriverGetVersion integer function cudaDriverGetVersion(iversion) integer :: iversion cudaDriverGetVersion returns the version number of the installed CUDA driver as iversion. Version Management Use the functions in this section for version management. This function automatically returns cudaErrorInvalidValue if the iversion argument is NULL. If no driver is installed.Chapter 4. cudaGetLastError integer function cudaGetLastError() cudaGetLastError returns the error code that was most recently returned from any runtime call in this host thread. This function automatically returns cudaErrorInvalidValue if the iversion argument is NULL. Runtime APIs Error Handling Use the functions in this section for error handling.

42 .

The program caches the submatrices of A and B in the fast shared memory.j) being computed 43 . N. it reads a block submatrix of A and a block submatrix of B. j. C is then NxL attributes(global) subroutine mmul_kernel( A. kb. Bsub(16.M).Chapter 5. as follows: • Each thread block computes one 16x16 submatrix of C.com web page at: www. and has not been highly optimized for execution time. visit the frequently asked questions (FAQ) section of the pgroup. k.16). shared :: Asub(16. Each element of the result is the product of one row of A by one column of B.com/support/index. M. Examples This chapter contains For up-to-date information about the state of the current release. B(M.L) integer. C. The submatrix size is chosen so the number of threads in a block is a multiple of the warp size (32) and is less than the maximum number of threads per thread block (512). accumulates the submatrix product. tx. L integer :: i. B is MxL.16) ! the value of C(i. Source Code Listing ! start the module containing the matmul kernel module mmul_mod use cudafor contains ! mmul_kernel computes A*B into C where ! A is NxM. ty ! submatrices stored in shared memory real. C(N. M. B.pgroup. • Each thread within the block computes one element of the submatrix.htm Matrix Multiplication Example This example shows a program to compute the product C of two matrices A and B. For simplicity. the program assumes the matrix sizes are a multiple of 16. value :: N. The program computes the products by accumulating submatrix products. L ) real :: A(N. then moves to the next submatrix of A rowwise and of B columnwise.L).

ty) = B(kb+tx-1. B. C ) real.M).:) = B(1:M. M. 16.L) ) ! Copy A and B to the device Adev = A(1:N.Bdev. allocatable.:) :: Adev.:) :: A. device.j)) i = (blockidx%x-1) * 16 + tx j = (blockidx%y-1) * 16 + ty Cij = 0. 2 ) ! Allocate the device arrays allocate( Adev(N. L/16. 1 ) call mmul_kernel<<<dimGrid.Cdev ! dim3 variables to define the grid and block shapes type(dim3) :: dimGrid. Cdev.L). Bdev(M. 2 ) L = size( B.1:L) = Cdev deallocate( Adev. ! Copy the results back and free up memory C(1:N.16 Cij = Cij + Asub(tx. Bdev.j) ! Wait until all elements are filled call syncthreads() ! Multiply the two submatrices ! Each of the 16x16 threads accumulates the ! dot product for its element of C(i. 1 ) M = size( A.1:L) ! Create the grid and block dimensions dimGrid = dim3( N/16.0 ! Do the k loop in chunks of 16.ty) = A(i.dimBlock>>>( Adev. B. dimension(:.1:M) Bdev(:. 1 ) dimBlock = dim3( 16.ty) enddo ! Synchronize to make sure all threads are done ! reading the submatrices before overwriting them ! in the next iteration of the kb loop call syncthreads() enddo ! Each of the 16x16 threads stores its element ! to the global C array C(i. C ! allocatable device arrays real.j) do k = 1. dimBlock ! Get the array sizes N = size( A.kb+ty-1) Bsub(tx. Cdev(N. the block size do kb = 1.j) = sum(A(i.:) * B(:.Matrix Multiplication Example real :: Cij ! Get the thread indices tx = threadidx%x ty = threadidx%y ! This thread computes C(i. dimension(:. Cdev ) end subroutine mmul end module mmul_mod N 44 .k) * Bsub(k.j) = Cij end subroutine mmul_kernel ! The host routine to drive the matrix multiplication subroutine mmul( A. 16 ! Fill the submatrices ! Each of the 16x16 threads in the thread block ! loads one element of Asub and Bsub Asub(tx. Bdev.

C. it stores the computed value into the correct element of C. no extra synchronization is needed between the copy operations and the kernel launch. A and B. 45 .j) for a particular value of i and j. Because the data copy operations are synchronous. and one output array. • It determines the i and j indices. and Cdev.Chapter 5. and three scalars giving the array sizes. for which element of C(i. • It copies Cdev back from device memory to C. • It initializes a scalar in which it will accumulate the dot product. • It fills dimGrid and dimBlock to hold the grid and thread block sizes. • It calls mmul_kernel to compute Cdev on the device. it does the following steps: • It loads one element of the submatrices of A and B into shared memory. depending on the block and thread index. • It frees the device memory arrays.j) it is computing. • It accumulates the dot product of its row and column of the submatrices. The thread executing this routine is one of 16x16 threads cooperating in a thread block. • It synchronizes to make sure both submatrices are loaded by all threads in the block. C. • For each block. • Finally. • It allocates device memory arrays Adev. The routine performs the following operations: • It determines the size of the matrices in N. • It steps through the arrays A and B in blocks of size 16. The host subroutine mmul is a wrapper for the kernel routine mmul_kernel. Examples Source Code Description This source code module mmul_mod has two subroutines. passed as assumed-shape arrays. • It synchronizes again to make sure all threads are done reading the submatrices before starting the next block.:)*B(:. It performs the following operations: • It determines the thread indices for this thread. M. MMUL This host subroutine has two input arrays. This routine computes the dot product of A(i. MMUL_KERNEL This kernel subroutine has two device memory input arrays. one device memory output array. Bdev. and L. • It copies the arrays A and B to Adev and Bdev using array assignments. A and B.

46 .

pgroup. Two Centerpointe Drive Lake Oswego. 47 . Contact Information You can contact The Portland Group at: The Portland Group STMicroelectronics.com/resources/docs.com All technical support is by email or submissions using an online form at www.pgroup.pgroup.htm or in your local copy of the documentation in the release directory doc/index. Many questions and problems can be resolved at our frequently asked questions (FAQ) site at www.htm. Inc.php Or contact us electronically using any of the following means: Fax Sales Support WWW +1-503-682-2637 sales@pgroup. Phone support is not currently available.htm.pgroup. Log in to the PGI website to access the forum: www.com/userforum/index.com/support. PGI documentation is available at www. The forum newsgroups may contain answers to commonly asked questions.com trs@pgroup.Chapter 6.com/support/faq.pgroup. OR 97035 USA The PGI User Forum is monitored by members of the PGI engineering and support teams as well as other PGI customers.com www.

48 .

Sign up to vote on this title
UsefulNot useful