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

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

................................................................. Counting Atomic Functions ...................... 23 3...................9.......10...................................... Arithmetic and Bitwise Atomic Functions .................................. Fortran Reduction Intrinsics ............................................. 23 3. 29 vii ............................................. 21 3.......................................................... Fortran Numeric and Logical Intrinsics ........................Tables 2.... Compare and Swap Atomic Function ...2........................ Intrinsic Datatypes ......................................................... 27 3.................................................. 6 3..................... 28 3............................................. Device Code Intrinsic Datatypes ......... Fortran Numeric Inquiry Intrinsics ..1..6...........1..........8.................................................................................. Fortran Mathematical Intrinsics .............7..3......................................................................................... 22 3........ Fortran Random Number Intrinsics .......................................... 24 3.......... 23 3................................................4........................................................... 24 3............................................................ Fortran Bit Manipulation Intrinsics .........................................................5..

viii .

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

Bold is used for commands. In this case.com/cuda. Version 3.1. Release 10. multiple filenames are allowed.com/support/definitions.Terminology Constant Width is used for filenames. In this case item1 is optional. 1997 (Fortran 95). • NVIDIA CUDA Programming Guide.com/cuda. [ item1 ] in general. and for language statements in the text. 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. { item2 | item 3} braces indicate that a selection is required. square brackets indicate optional items. In this example. Terminology If there are terms in this guide with which you are unfamiliar. square brackets are required to specify a p/t-set.htm Related Publications The following documents contain additional information related to CUDA Fortran programming. In the context of p/tsets. Available online at http:// www. x . examples.pgroup. 7/21/2010.pdf. There are a wide variety of releases and distributions of each of these types of operating systems. NVIDIA.nvidia. you must select either item2 or item3.com/doc/pgiug. Available online at http:// www. The Portland Group. • ISO/IEC 1539-1:1997. • PGI Users Guide. June 2010. September. 2009.. arguments.pgroup. Geneva. FORTRAN Fortran language statements are shown in the text of this guide using a reduced fixed point size. ellipsis indicate a repetition.. including assembly language statements. options. Information Technology – Programming Languages – FORTRAN. Available online at http://www. NVIDIA.9. Zero or more of the preceding item may occur.nvidia. C/C++ C/C++ language statements are shown in the test of this guide using a reduced fixed point size. Version 3. directories. • NVIDIA CUDA Compute Unified Device Architecture Reference Manual.1. PGI provides a glossary of terms which you can access at www.1. filename .

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. CUDA comes with an extended C compiler. highly parallel computational units with very high memory bandwidth. PGI 2011 includes support for CUDA Fortran on Linux. 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 . NVIDIA introduced CUDA™. but are general enough to be useful in many data-parallel. Each block is partitioned into fine grain threads. Graphic processing units or GPUs have evolved into programmable. regardless of the number of available processor cores. shared memory and barrier synchronization within thread groups. and tremendous potential for many applications.Chapter 1. here called CUDA C. CUDA Fortran is an analog to NVIDIA's CUDA C compiler. A CUDA programmer must partition the program into coarse grain blocks that can be executed in parallel. which can cooperate using shared memory and barrier synchronization. a small set of extensions to Fortran that supports and is built upon the CUDA computing architecture. Mac OS X and Windows. allowing direct programming of the GPU from a high level language. A properly designed CUDA program will run on any CUDA-enabled GPU. and coordinated independent thread groups organized into a grid. a general purpose parallel programming architecture. Introduction Welcome to Release 2011 of PGI CUDA Fortran. The programming model supports four key abstractions: cooperating threads organized into thread groups. CUDA Fortran includes a Fortran 2003 compiler and tool chain for programming NVIDIA GPUs using Fortran. compute-intensive programs. 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. Compared to the PGI Accelerator directives-based model and compilers. GPU designs are optimized for the computations found in graphics rendering. with compilers and libraries to support the programming of NVIDIA GPUs.

2 .

x. dimension(*) :: x. and a thread index accessed through threadidx. A kernel is defined using the attributes(global) specifier on the subroutine statement. y ) end subroutine In this case. and a local thread index within its thread block. In this example. A call to a kernel specifies how many parallel instances of the kernel must be executed. a. 3 . y real :: a integer :: n ! call the kernel call ksaxpy<<<n/64. The CUDA threads are organized into thread blocks. refer to Chapter 3. For a reference for CUDA Fortran. Each thread is assigned a thread block index accessed through the built-in blockidx variable. and each thread has a global thread block index.Chapter 2. 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. Programming Guide This chapter introduces the CUDA programming model through examples written in CUDA Fortran. value :: n. 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. a. Such a subroutine is called a device kernel or kernel. y ) real.” on page 13. value :: a integer. x. device. a. each thread performs one iteration of the common SAXPY loop operation. 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. each instance will be executed by a different CUDA thread. 64>>>( n. x. dimension(*) :: x.y real. “Reference. each with 64 threads. the call to the kernel ksaxpy specifies n/64 thread blocks. y ) real.

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

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

If declared in a module.Datatypes in Device Subprograms • If declared in a host subprogram.2. or size of page-locked memory. and may only be accessed within that subprogram. it will be allocated in host pagelocked memory. A shared variable may not be data initialized. availability.8 1. An allocatable device variable has a dynamic lifetime. Attributes(shared) A variable with the shared attributed is called a device shared variable or a shared variable. Device constant variables are allocated in the device constant memory space. Other device variables have a lifetime of the entire application. Some operating systems or installations may restrict the use.1. A shared variable may only be declared in a device subprogram. Intrinsic Datatypes Type integer logical real double precision Type Kind 1.4. A device array may be an explicit-shape array. the variable will be allocated in the normal host paged memory. though a write in one thread is only guaranteed to be visible to other threads after the next call to the SYNCTHREADS() intrinsic.8 4. or by other device subprograms to which it is passed as an argument. Datatypes in Device Subprograms The following intrinsic datatypes are allowed in device subprograms and device data: Table 2. 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. from when it is allocated until it is deallocated. but may be modified in host subprograms. Device constant variables may not be allocatable. When a pinned variable is allocated. A pinned variable must be an allocatable array.4.2. the variable may be accessed by that subprogram or subprograms contained in that subprogram. the variable may be accessed by any subprogram in that module and by any subprogram that uses the module. It can be read or written by all threads in the block. an allocatable array. and have a lifetime of the entire application. if the allocation in page-locked memory fails. Device constant data may not be assigned or modified in any device subprogram. and has a lifetime of the thread block. A shared variable is allocated in the device shared memory for a thread block. Attributes(constant) A variable with the constant attributes is called a device constant variable. or an assumed-shape dummy array. Attributes(pinned) A variable with the pinned attributes is called a pinned variable.8 equivalent to real(kind=8) 6 .

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

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

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

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

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

12 .

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

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

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

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

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

If no such memory space is available. declare and initialize the scalar on the host as: integer. Otherwise. or if it is exhausted. the compiler allocates the array in normal paged host memory. allocatable. Allocating Pinned Memory Allocatable arrays with the pinned attribute are dynamically allocated using the Allocate statement. and cudaFree can be used to free it: real. refer to“Memory Management.stat=istat) … if(allocated(b)) deallocate(b) Scalar variables can be allocated on the device using the Fortran 2003 allocatable scalar feature. The compiler will generate code to allocate the array in host page-locked memory. device :: b(:) allocate(b(5024). allocatable. pinned :: p(:) allocate(p(5000). 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. allocatable. allocatable. device :: adev(n) real. pinned allocatable arrays work and act like any other allocatable array on the host. Fortran interfaces to the CUDA memory management runtime routines are provided. To use these.stat=istat) … if(allocated(p)) deallocate(p) 18 . 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. 100) … istat = cudaFree(v) For a complete list of the memory management runtime routines.” on page 32. real. if available. Mixing standard Fortran allocate/deallocate with the runtime Malloc/Free for a given array is not supported. These functions return memory which will bypass certain Fortran allocatable properties such as automatic deallocation.Allocating Device and Pinned Arrays real. 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. and thus the arrays are treated more like C malloc’ed areas. device :: v(:) istat = cudaMalloc(v.n) real.c.

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

the number of threads per thread block is block%x*block%y*block%z. The execution configuration gives the size and shape of the grid and thread blocks that execute the function. Counts expressed in arguments to the Fortran runtime routines are expressed in terms of data type elements. it has the form: <<< grid. refer to “Memory Management. device :: wrk(1024) real cur(512) istat = cudaMemcpy(wrk. or from one device array to another. 512) For those familiar with the CUDA C routines. device to host. as well as the amount of shared memory to use for assumed-size shared memory arrays. and their performance is dependent on the size of the slice. block. or of type(dim3). amount of contiguous data in the slices. If it is type(dim3). These functions can transfer data either from the host to device. The product grid%x*grid%y gives the number of thread blocks to launch. If block is an integer. Array slices are also supported. or would be more efficient written in another way. If grid is an integer. the kind parameter to the Memcpy routines is optional in Fortran since the attributes of the arrays are explicitly declared. cur.Invoking a kernel subroutine The following expressions are not legal as they either promote a false impression of where the actual computation occurs. the value of grid%z must be one. stream >>> • grid is an integer. it is converted to dim3(grid. Data Transfer Using Runtime Routines For programmers comfortable with the CUDA C programming environment. and the associated stream. The cudaMemcpy function can be used to copy data between the host and the GPU: real. If it is type(dim3). Invoking a kernel subroutine A call to a kernel subroutine must give the execution configuration for the call.1). it is converted to dim3(block.1. which must be less than the maximum supported by the device. For a complete list of memory management runtime routines. bl • block is an integer. or both: c = adev + bdev adev = adev + a b = sqrt(adev) Elemental transfers are supported by the language but will perform poorly. bytes. not bytes. or of type(dim3). 20 .1. The execution configuration is specified after the subroutine name in the call statement. and the implementation. Fortran interfaces to the CUDA memory management runtime routines are provided.” on page 32.1).

it must be a scalar integer.1. Device code Datatypes allowed Variables and arrays with the device. such as the function arguments and execution configuration arguments. Table 3.8 1 (default) Additionally. For more information. or shared attributes. See Section 4. or a value returned by a call to cudaStreamCreate. It specifies the stream to which this call is enqueued. or another allowed derived type. or if bytes is greater than the shared memory available. they may be of derived type.y. if present. allowing for static shared memory declared in the kernel and for other dedicated uses. 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. or declared in device subprograms.4(default).2.8 equivalent to real(kind=8) 4(default). are limited to the types described in this section. where the members of the derived type have one of the allowed intrinsic datatypes. if present. The system module cudafor includes definitions of the derived type dim3. refer to “Shared data.8 4(default).” on page 16. constant. and specifies the number of bytes of shared memory to be allocated for each thread block to use for assumed-size shared memory arrays. They may have any of the intrinsic datatypes in the following table.4(default). For instance.2.z end type 21 . Reference • bytes is optional. a kernel subroutine attributes(global) subroutine sub( a ) can be called like: call sub <<< DG. If not specified. • stream is optional.Chapter 3. it must be an integer.5 on page 41.8 1. the value zero is used. Device Code Intrinsic Datatypes Type integer logical real double precision complex character(len=1) Type Kind 1. and have a value of zero.

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

real real real Name precision radix range selected_int_kind selected_real_kind tiny Argument Datatypes real. complex real real. complex real real Fortran Numeric Inquiry Intrinsics Table 3. complex real real. real real integer. complex real real.Chapter 3. Reference Fortran Mathematical Intrinsics Table 3.4.integer) real Fortran Bit Manipulation Intrinsics Table 3. complex integer (integer. 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 . complex Name log log10 sin sinh sqrt tan tanh Argument Datatypes real.3.real) real.5. Fortran Numeric Inquiry Intrinsics Name bit_size digits epsilon huge maxexponent minexponent Argument Datatypes integer integer. complex integer. Fortran Mathematical Intrinsics Name acos asin atan atan2 cos cosh exp Argument Datatypes real real real (real. real integer.

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

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

true. only if the value of the argument is .” on page 32.2 and higher. Warp-Vote Operations Warp-vote operations are only supported by devices with compute capability 1.0 or higher. Each of these functions has a single argument.true. 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. The clock frequency can be determined by calling cudaGetDeviceProperties. The value of the function is . 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 2.0) ) allneg = . ALLTHREADS The allthreads function is a warp-vote operation with a single scalar logical argument: if( allthreads(a(i)<0.Device code threadfence() is one method to enforce a specific order. 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. 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. The function allthreads evaluates its argument for all threads in the current warp. as described in “cudaGetDeviceProperties. GPU_TIME The gpu_time intrinsic returns the value of the clock cycle counter on the GPU. for all threads in the warp.true. 26 . THREADFENCE_SYSTEM void threadfence_system() threadfence_system acts as a memory fence.

0. Arithmetic and Bitwise Atomic Functions Function atomicadd( mem. Compute capability 1.true. Atomic functions are only supported by devices with compute capability 1. only if the value of the argument is . for all threads in the warp.1 and higher.2 or higher is required if the first argument has the shared attribute. value ) atomicsub( mem. The function anythread evaluates its argument for all threads in the current warp. and store the combined value back to the first argument location. value ) Additional Atomic Update mem = mem + value mem = mem – value 27 .true. which must be a variable or array element in shared memory (with the shared attribute) or in device global memory (with the device attribute). This function is only supported by devices of compute capability 2.0) ) allneg = . The value of the function is .false. Reference ANYTHREAD The anythread function is a warp-vote operation with a single scalar logical argument: if( anythread(a(i)<0. mem. Arithmetic and Bitwise Atomic Functions These atomic functions read and return the value of the first argument. 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. Example: if( ballot(int_value) ) allneg = .8. Both arguments must be of type integer(kind=4). Note The return value for each of these functions is the first argument. They also combine that value with the value of the second argument. Atomic Functions The atomic functions read and write the value of their first operand.false. 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.Chapter 3. These functions are: Table 3. depending on the function.

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

SIZEOF Intrinsic A call to sizeof(A). CLOSE. • Recursive subroutines and functions are not allowed. NAMELIST.10. • STOP and PAUSE statements are not allowed. • Objects with the Pointer and Allocatable attribute are not allowed. FORMAT. PRINT. OPEN.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. • Alternate return specifications are not allowed. REWIND. Reference Table 3.Chapter 3. where A is a variable or expression. • Objects with character type must have LEN=1. character substrings are not supported. such as the new sizeof intrinsic function. ENDFILE. j j = sizeof(i) ! this assigns the value 4 to j 29 . Compare and Swap Atomic Function Function atomiccas(mem. • Input/Output statements are not allowed: READ.comp. • Subroutine and function calls are supported only if they can be inlined. Host code Host subprograms may use intrinsic functions. WRITE. INQUIRE. • Fortran intrinsic functions not listed in Section 3.6. BACKSPACE. • Automatic arrays must be fixed size. • ENTRY statements are not allowed. returns the number of bytes required to hold the value of A. • Cray pointers are not supported. integer(kind=4) :: i.3 are not supported. • Floating point exception handling is not supported. • Optional arguments are not allowed.

30 .

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

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

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

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

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

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

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

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

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

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

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

42 .

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

j) do k = 1. dimBlock ! Get the array sizes N = size( A.Bdev. Bdev. 1 ) dimBlock = dim3( 16.ty) = A(i. B. Cdev ) end subroutine mmul end module mmul_mod N 44 . ! Copy the results back and free up memory C(1:N. Bdev. C ! allocatable device arrays real.1:M) Bdev(:.Matrix Multiplication Example real :: Cij ! Get the thread indices tx = threadidx%x ty = threadidx%y ! This thread computes C(i. B. 16 ! Fill the submatrices ! Each of the 16x16 threads in the thread block ! loads one element of Asub and Bsub Asub(tx. dimension(:.:) * B(:. 1 ) M = size( A.:) = B(1:M.kb+ty-1) Bsub(tx. device.1:L) ! Create the grid and block dimensions dimGrid = dim3( N/16. Cdev.M).16 Cij = Cij + Asub(tx. 2 ) L = size( B. L/16.j) = sum(A(i. dimension(:. allocatable. the block size do kb = 1.0 ! Do the k loop in chunks of 16.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:L) = Cdev deallocate( Adev.j)) i = (blockidx%x-1) * 16 + tx j = (blockidx%y-1) * 16 + ty Cij = 0. 2 ) ! Allocate the device arrays allocate( Adev(N.:) :: A.:) :: Adev.Cdev ! dim3 variables to define the grid and block shapes type(dim3) :: dimGrid.k) * Bsub(k. 1 ) call mmul_kernel<<<dimGrid.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.dimBlock>>>( Adev. 16.j) = Cij end subroutine mmul_kernel ! The host routine to drive the matrix multiplication subroutine mmul( A. Bdev(M.ty) = B(kb+tx-1. C ) real. Cdev(N.L). M.L) ) ! Copy A and B to the device Adev = A(1:N.

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

46 .

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

48 .