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

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

............. Fortran Reduction Intrinsics ........4...1................................................... Fortran Mathematical Intrinsics ............................ Fortran Bit Manipulation Intrinsics ............................... 6 3.... Fortran Numeric and Logical Intrinsics ........................ Device Code Intrinsic Datatypes ...................................... 23 3...................9.................................................... 24 3.10.......................................................... Compare and Swap Atomic Function ....................6................................... 27 3... 23 3. 23 3.......................................................................................... 24 3......2............................................................................................................8...................................................................................... Fortran Random Number Intrinsics ............ Counting Atomic Functions ................................3.....Tables 2............................................................................... Fortran Numeric Inquiry Intrinsics ..................................... 28 3................ 29 vii .....7.......... 22 3...........................................5...............1................................... Arithmetic and Bitwise Atomic Functions ............................................. Intrinsic Datatypes ............ 21 3...........................................

viii .

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

2009. Version 3. In the context of p/tsets. • ISO/IEC 1539-1:1997. • PGI Users Guide. filename . including assembly language statements. { item2 | item 3} braces indicate that a selection is required. square brackets indicate optional items.9. FORTRAN Fortran language statements are shown in the text of this guide using a reduced fixed point size. directories. arguments. There are a wide variety of releases and distributions of each of these types of operating systems. Terminology If there are terms in this guide with which you are unfamiliar. NVIDIA.pgroup. square brackets are required to specify a p/t-set. September. x . and for language statements in the text. • NVIDIA CUDA Programming Guide. Available online at http:// www. Version 3.pdf. Release 10. options.com/doc/pgiug. The Portland Group.htm Related Publications The following documents contain additional information related to CUDA Fortran programming.1.1.com/cuda. PGI provides a glossary of terms which you can access at www. ellipsis indicate a repetition. NVIDIA.nvidia.Terminology Constant Width is used for filenames. Available online at http:// www. 1997 (Fortran 95). In this case. Geneva. Bold is used for commands.com/cuda. • NVIDIA CUDA Compute Unified Device Architecture Reference Manual. Information Technology – Programming Languages – FORTRAN.nvidia. In this example.1. 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.pgroup. 7/21/2010. [ item1 ] in general..com/support/definitions. multiple filenames are allowed. In this case item1 is optional. June 2010. Available online at http://www.. examples. Zero or more of the preceding item may occur. you must select either item2 or item3. C/C++ C/C++ language statements are shown in the test of this guide using a reduced fixed point size.

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

2 .

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

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

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

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

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

The call can be treated as a kernel launch operation. the host can continue executing. and may execute concurrently with each other. and will only be initiated when all previous operations on that queue have been completed. Kernel Loop Directive CUDA Fortran allows automatic kernel generation and invocation from a region of host code containing one or more tightly nested loops.2.3. The default stream. As with any kernel. An operation is placed at the end of the stream queue. see Section 3. In this way. The one exception to this rule is a scalar reduction operation. inserting synchronization into the kernel as appropriate. Programmers must be careful when using concurrent host and device execution. and no subsequent operations on any queue begin until the stream zero operation is complete. 8 . are implemented using stream queues. across the thread blocks and grid. By calling the runtime routine cudaThreadSynchronize. An application can manage more concurrency by using multiple streams. Concurrent Host and Device Execution When a host subprogram calls a kernel subroutine. used when no stream identifier is specified. including kernel execution and data copies to and from device memory. If not specified. in cases where the host program reads or modifies device or constant data. The program can use cudaThreadSynchronize() or CUDA Events to wait for the completion of the kernel. the host program can synchronize and wait for all previously launched or queued kernels. • The value of streamid must be an integer greater than or equal to zero. Concurrent Stream Execution Operations involving the device. where the launch actually corresponds to placing the kernel on a queue for execution by the device. it specifies the stream to which this call is associated. 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. the host program should synchronize with the device to avoid erroneous results. its value is treated as zero. the compiler handles the generation of the final reduction kernel. the call actually returns to the host program before the kernel subroutine begins execution. For these operations. including calling or queueing more kernels for execution on the device. operations on different stream queues may execute out-of-order with respect to when they were placed on the queues. such as summing the values in a vector or matrix. 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. the launch is asynchronous. Asynchronous concurrent execution There are two components to asynchronous concurrent execution with CUDA Fortran. The work in the loops specified by the directive is executed in parallel. stream zero is special in that operations on the stream zero queue will begin only after all preceding operations on all queues are complete. is stream zero.Asynchronous concurrent execution assumed-size shared variables in the thread block.

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

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

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

12 .

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

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

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

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

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

The cudaMalloc function can be used to allocate single-dimensional arrays of the supported intrinsic datatypes. pinned allocatable arrays work and act like any other allocatable array on the host. allocatable.” on page 32.stat=istat) … if(allocated(p)) deallocate(p) 18 . declare and initialize the scalar on the host as: integer. allocatable. Allocating Pinned Memory Allocatable arrays with the pinned attribute are dynamically allocated using the Allocate statement.stat=istat) … if(allocated(b)) deallocate(b) Scalar variables can be allocated on the device using the Fortran 2003 allocatable scalar feature. 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. These arrays will also have a lifetime of the subprogram as is usual with the Fortran language: subroutine vfunc(a.c. device :: v(:) istat = cudaMalloc(v. 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. device :: adev(n) real. Otherwise.n) real. To use these. Fortran interfaces to the CUDA memory management runtime routines are provided. refer to“Memory Management. and cudaFree can be used to free it: real. pinned :: p(:) allocate(p(5000). If no such memory space is available. real. the compiler allocates the array in normal paged host memory.Allocating Device and Pinned Arrays real. or if it is exhausted. device :: b(:) allocate(b(5024). and thus the arrays are treated more like C malloc’ed areas. The compiler will generate code to allocate the array in host page-locked memory. allocatable. allocatable. if available. 100) … istat = cudaFree(v) For a complete list of the memory management runtime routines. 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.

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

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

constant. and have a value of zero. For more information. They may have any of the intrinsic datatypes in the following table. and specifies the number of bytes of shared memory to be allocated for each thread block to use for assumed-size shared memory arrays. if present. allowing for static shared memory declared in the kernel and for other dedicated uses.4(default).Chapter 3. or another allowed derived type. Reference • bytes is optional. where the members of the derived type have one of the allowed intrinsic datatypes.8 equivalent to real(kind=8) 4(default).8 1. a kernel subroutine attributes(global) subroutine sub( a ) can be called like: call sub <<< DG. the value zero is used. are limited to the types described in this section. Device Code Intrinsic Datatypes Type integer logical real double precision complex character(len=1) Type Kind 1. DB >>> ( A ) The function call will fail if the grid or block arguments are greater than the maximum sizes allowed. Device code Datatypes allowed Variables and arrays with the device. • stream is optional. refer to “Shared data. if present. defined as type(dim3) integer(kind=4) :: x.4(default).z end type 21 .5 on page 41.8 1 (default) Additionally. If not specified. It specifies the stream to which this call is enqueued. it must be a scalar integer. such as the function arguments and execution configuration arguments. they may be of derived type. For instance. it must be an integer. or shared attributes.y.1.” on page 16. See Section 4. or a value returned by a call to cudaStreamCreate. or if bytes is greater than the shared memory available.8 4(default). The system module cudafor includes definitions of the derived type dim3. or declared in device subprograms.2.2. Table 3.

real integer. blockidx. complex complex real real real real or (real. the value of griddim%y is one for one-dimensional grids. the value of griddim%z is always one. the blockdim%y and/or blockdim %z components have the value one. griddim integer(4) :: warpsize • The variable threadidx contains the thread index within its thread block. real real Name int logical max min mod modulo nint real sign Argument Datatypes integer. and griddim are available only in device subprograms. Fortran Numeric and Logical Intrinsics Name abs aimag aint anint ceiling cmplx conjg dim floor Argument Datatypes integer. blockdim. Fortran Numeric and Logical Intrinsics Table 3. complex integer. real. blockidx. The value of blockidx is the same for all threads in the same thread block. real. real integer. • The variables threadidx. real 22 . blockdim. the threadidx%y and/or threadidx%z components have the value one. blockidx%y has the value one. real real integer. • The variable warpsize contains the number of threads in a warp.real) complex integer.or two-dimensional thread blocks. for one.Device code Builtin variables The system module cudafor declares several predefined variables. They are declared as follows: type(dim3) :: threadidx. • The variable blockdim contains the dimensions of the thread block. for onedimensional grids. the value of griddim%z is always one. real. currently defined to be 32.2. The value of blockidx%z is always one. The value of griddim is the same for all threads in the same grid. • The variable blockidx contains the block index within the grid. as with threadidx. • The variable griddim contains the dimensions of the grid. It has constant value. These variables are read-only. complex logical integer.or two-dimensional thread blocks. for one. real integer. blockdim has the same value for all threads in the same grid. Fortran intrinsics This section lists the Fortran intrinsic functions allowed in device subprograms.

complex Name log log10 sin sinh sqrt tan tanh Argument Datatypes real. real integer. 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 . Reference Fortran Mathematical Intrinsics Table 3.3. Fortran Mathematical Intrinsics Name acos asin atan atan2 cos cosh exp Argument Datatypes real real real (real. complex real real.4. complex real real. Fortran Numeric Inquiry Intrinsics Name bit_size digits epsilon huge maxexponent minexponent Argument Datatypes integer integer. real real real Name precision radix range selected_int_kind selected_real_kind tiny Argument Datatypes real.Chapter 3. complex integer. real.5. complex real real. complex real real Fortran Numeric Inquiry Intrinsics Table 3.integer) real Fortran Bit Manipulation Intrinsics Table 3. real real integer.real) real. complex integer (integer.

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

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

” on page 32. 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.true. 26 . GPU_TIME The gpu_time intrinsic returns the value of the clock cycle counter on the GPU. for all threads in the warp. 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. Each of these functions has a single argument. Warp-Vote Operations Warp-vote operations are only supported by devices with compute capability 1.2 and higher.Device code threadfence() is one method to enforce a specific order. 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. The function allthreads evaluates its argument for all threads in the current warp.0 or higher. as described in “cudaGetDeviceProperties. The value of the function is .true. 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. The clock frequency can be determined by calling cudaGetDeviceProperties.true. ALLTHREADS The allthreads function is a warp-vote operation with a single scalar logical argument: if( allthreads(a(i)<0. only if the value of the argument is .0) ) allneg = . THREADFENCE_SYSTEM void threadfence_system() threadfence_system acts as a memory fence.

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

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

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

30 .

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

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

“Device Code Intrinsic Datatypes. devptr may be any allocatable.” on page 21. cdesc. width. cudaFreeArray integer function cudaFreeArray(carray) type(cudaArrayPtr) :: carray 33 . one-dimensional device array of a supported type specified in Table 3. 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. and the padded width is returned in the variable pitch. pitch. should be of benefit. as well as consistency checks between Fortran device arrays and host arrays. proper Fortran pointers for device data are supported. height) cudaMallocPitch allocates data on the device. Under CUDA Fortran. cudaMallocPitch may pad the data. 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. “Device Code Intrinsic Datatypes. height cudaMallocArray allocates a data array on the device. in which case the count is in bytes. The count is in terms of elements. an allocatable device array as the second argument. Or. c_f_pointer will take a TYPE(C_DEVPTR) as the first argument. cudaMallocArray integer function cudaMallocArray(carray. “Device Code Intrinsic Datatypes. Both of these features are subject to change when. two-dimensional device array of a supported type specified in Table 3. cudaMalloc integer function cudaMalloc(devptr. in the future. The height is an integer.” on page 21. height) type(cudaArrayPtr) :: carray type(cudaChannelFormatDesc) :: cdesc integer :: width. and in effect transfer the allocation to the Fortran array.Chapter 4. devptr may be any allocatable device array of a supported type specified in Table 3. Or.1.1. devptr may also be of TYPE(C_DEVPTR).” on page 21. Use the functions in this section for memory management. The width is in terms of number of elements. a shape as the third argument. devptr may be of TYPE(C_DEVPTR). in which case the integer values are expressed in bytes. width. Currently. cudaMallocPitch integer function cudaMallocPitch(devptr. Runtime APIs TYPE(C_DEVPTR). cudaFree integer function cudaFree(devptr) cudaFree deallocates data on the device. count) cudaMalloc allocates data on the device.1. devptr may be any allocatable. devptr may be of TYPE(C_DEVPTR). Similarly.

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

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

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

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

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

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

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

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

42 .

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

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

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

46 .

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

48 .

Sign up to vote on this title
UsefulNot useful