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

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

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

viii .

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

and for language statements in the text. September. x . Release 10. June 2010.1. multiple filenames are allowed. Geneva. examples.com/doc/pgiug. 7/21/2010.pdf. PGI provides a glossary of terms which you can access at www.pgroup. { item2 | item 3} braces indicate that a selection is required. NVIDIA. Version 3. directories. • ISO/IEC 1539-1:1997. Available online at http://www. including assembly language statements. NVIDIA. C/C++ C/C++ language statements are shown in the test of this guide using a reduced fixed point size. • PGI Users Guide. In this case item1 is optional.9. In the context of p/tsets.nvidia. options.com/support/definitions. Terminology If there are terms in this guide with which you are unfamiliar. • NVIDIA CUDA Compute Unified Device Architecture Reference Manual.com/cuda. 2009.. Bold is used for commands. Available online at http:// www. square brackets indicate optional items. In this example. Zero or more of the preceding item may occur.com/cuda.1. There are a wide variety of releases and distributions of each of these types of operating systems. • NVIDIA CUDA Programming Guide.. square brackets are required to specify a p/t-set. arguments. Information Technology – Programming Languages – FORTRAN. The Portland Group. 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. Version 3. 1997 (Fortran 95).nvidia.htm Related Publications The following documents contain additional information related to CUDA Fortran programming. Available online at http:// www.1. you must select either item2 or item3.Terminology Constant Width is used for filenames. ellipsis indicate a repetition. FORTRAN Fortran language statements are shown in the text of this guide using a reduced fixed point size. [ item1 ] in general. In this case.pgroup. filename .

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

2 .

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

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

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

If declared in a module. Datatypes in Device Subprograms The following intrinsic datatypes are allowed in device subprograms and device data: Table 2.Datatypes in Device Subprograms • If declared in a host subprogram.4. Attributes(constant) A variable with the constant attributes is called a device constant variable.8 equivalent to real(kind=8) 6 . though a write in one thread is only guaranteed to be visible to other threads after the next call to the SYNCTHREADS() intrinsic.8 4. Other device variables have a lifetime of the entire application. When a pinned variable is allocated. A device array may be an explicit-shape array. or an assumed-shape dummy array.8 1. and may only be accessed within that subprogram. Device constant variables are allocated in the device constant memory space. A pinned variable must be an allocatable array. or size of page-locked memory. A shared variable may not be data initialized. but may be modified in host subprograms. from when it is allocated until it is deallocated.4. Some operating systems or installations may restrict the use. 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. the variable will be allocated in the normal host paged memory.1. An allocatable device variable has a dynamic lifetime. Device constant data may not be assigned or modified in any device subprogram. A shared variable is allocated in the device shared memory for a thread block. or by other device subprograms to which it is passed as an argument. and have a lifetime of the entire application. availability. Device constant variables may not be allocatable. A shared variable may only be declared in a device subprogram. an allocatable array. Attributes(pinned) A variable with the pinned attributes is called a pinned variable. the variable may be accessed by any subprogram in that module and by any subprogram that uses the module. Attributes(shared) A variable with the shared attributed is called a device shared variable or a shared variable. Intrinsic Datatypes Type integer logical real double precision Type Kind 1.2. if the allocation in page-locked memory fails. It can be read or written by all threads in the block. it will be allocated in host pagelocked memory.2. and has a lifetime of the thread block. the variable may be accessed by that subprogram or subprograms contained in that subprogram.

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

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

m do i = 1. block >>> The compiler maps the launch configuration specified by the grid and block values onto the outermost n loops.j) +c(i. only the outer j loop is run in parallel with a thread block size of 256. Programming Guide Syntax The general form of the kernel directive is: !$cuf kernel do[(n)] <<< grid. (256. !$cuf kernel do(2) <<< *. You might consider if the code in Example 2 would perform better if the two loops were interchanged. Example 1 !$cuf kernel do(2) <<< (*.*).j) = b(i. that is.j) end do end do In this example.j) = b(i. the default value is 1. Alternatively.j) = b(i. m do i = 1.j) +c(i. so all iterations are computed. (32.j) +c(i. the directive defines a two-dimensional thread block of size 32x4. The grid and block values can be an integer scalar or a parenthesized list. ThreadIdx%y runs from 1 to 4 and is mapped onto the outer j loop. In this case. specified as (*. is computed by the compiler and runtime by dividing the loop trip counts n and m by the thread block size.j) end do end do Without an explicit n on the do. Alternatively. you could specify a configuration like the following in which the threads read and write the matrices in coalesced fashion. The body of the doubly-nested loop is turned in the kernel body: ThreadIdx%x runs from 1 to 32 and is mapped onto the inner i loop.*). starting at loop n and working out.4) >>> do j = 1.Chapter 2.1) >>> do j = 1. m do i = 1. The grid shape. The inner i dimension is run sequentially on each thread. n a(i. using asterisks tells the compiler to choose a thread block shape and/or compute the grid shape from the thread block shape and the loop limits.j) end do end do 9 . Example 2 !$cuf kernel do <<< *. Loops which are not mapped onto the grid and block values are run sequentially on each thread. the schedule applies just to the outermost loop. 256 >>> do j = 1. n a(i. n a(i.

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

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

12 .

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

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

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

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

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

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

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

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

a kernel subroutine attributes(global) subroutine sub( a ) can be called like: call sub <<< DG. defined as type(dim3) integer(kind=4) :: x. • stream is optional.5 on page 41. the value zero is used. Device code Datatypes allowed Variables and arrays with the device.y. they may be of derived type. allowing for static shared memory declared in the kernel and for other dedicated uses. Table 3. or shared attributes. constant.8 1 (default) Additionally.Chapter 3. if present. or declared in device subprograms. are limited to the types described in this section.4(default). it must be an integer. If not specified. See Section 4. such as the function arguments and execution configuration arguments. They may have any of the intrinsic datatypes in the following table. DB >>> ( A ) The function call will fail if the grid or block arguments are greater than the maximum sizes allowed. It specifies the stream to which this call is enqueued.” on page 16. refer to “Shared data.4(default). The system module cudafor includes definitions of the derived type dim3.2. For more information. and specifies the number of bytes of shared memory to be allocated for each thread block to use for assumed-size shared memory arrays. or another allowed derived type.z end type 21 . and have a value of zero. if present. Device Code Intrinsic Datatypes Type integer logical real double precision complex character(len=1) Type Kind 1. Reference • bytes is optional. For instance. it must be a scalar integer. where the members of the derived type have one of the allowed intrinsic datatypes.1.8 1. or if bytes is greater than the shared memory available. or a value returned by a call to cudaStreamCreate.2.8 4(default).8 equivalent to real(kind=8) 4(default).

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

Fortran Numeric Inquiry Intrinsics Name bit_size digits epsilon huge maxexponent minexponent Argument Datatypes integer integer. real. complex real real. Fortran Mathematical Intrinsics Name acos asin atan atan2 cos cosh exp Argument Datatypes real real real (real. complex Name log log10 sin sinh sqrt tan tanh Argument Datatypes real. complex real real Fortran Numeric Inquiry Intrinsics Table 3.integer) real Fortran Bit Manipulation Intrinsics Table 3.Chapter 3. Reference Fortran Mathematical Intrinsics Table 3. real integer. complex integer.5.4. complex real real. real real integer. real real real Name precision radix range selected_int_kind selected_real_kind tiny Argument Datatypes real.real) real. complex real real. complex integer (integer.3. 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 .

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

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

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

and store the combined value back to the first argument location. They also combine that value with the value of the second argument. Example: if( ballot(int_value) ) allneg = .0.0) ) allneg = . BALLOT The ballot function is a warp-vote operation with a single integer argument: unsigned integer ballot(int_value) The function ballot evaluates the argument int_value for all threads of the warp and returns an integer whose Nth bit is set if and only if int_value evaluates to non-zero for the Nth thread of the warp.true. only if the value of the argument is . Note The return value for each of these functions is the first argument. These functions are: Table 3.false. Atomic Functions The atomic functions read and write the value of their first operand.false.true. This function is only supported by devices of compute capability 2.Chapter 3. The function anythread evaluates its argument for all threads in the current warp. 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. value ) atomicsub( mem.2 or higher is required if the first argument has the shared attribute. value ) Additional Atomic Update mem = mem + value mem = mem – value 27 . The value of the function is . Atomic functions are only supported by devices with compute capability 1. Arithmetic and Bitwise Atomic Functions These atomic functions read and return the value of the first argument.1 and higher. Compute capability 1. Arithmetic and Bitwise Atomic Functions Function atomicadd( mem. for all threads in the warp. which must be a variable or array element in shared memory (with the shared attribute) or in device global memory (with the device attribute). mem. Both arguments must be of type integer(kind=4). Reference ANYTHREAD The anythread function is a warp-vote operation with a single scalar logical argument: if( anythread(a(i)<0. depending on the function.8.

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

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

30 .

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

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

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

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

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

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

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

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

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

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

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

42 .

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

k) * Bsub(k. dimension(:. C ) real.:) = B(1:M. L/16.j) = Cij end subroutine mmul_kernel ! The host routine to drive the matrix multiplication subroutine mmul( A.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. 2 ) ! Allocate the device arrays allocate( Adev(N. 1 ) call mmul_kernel<<<dimGrid. B.j) = sum(A(i.ty) = B(kb+tx-1.j)) i = (blockidx%x-1) * 16 + tx j = (blockidx%y-1) * 16 + ty Cij = 0. device.j) do k = 1. dimBlock ! Get the array sizes N = size( A. 16.1:L) ! Create the grid and block dimensions dimGrid = dim3( N/16.M). 2 ) L = size( B. Bdev(M.Cdev ! dim3 variables to define the grid and block shapes type(dim3) :: dimGrid.:) * B(:. dimension(:. C ! allocatable device arrays real.L).0 ! Do the k loop in chunks of 16.dimBlock>>>( Adev. Cdev(N.1:M) Bdev(:. the block size do kb = 1. Cdev ) end subroutine mmul end module mmul_mod N 44 .Matrix Multiplication Example real :: Cij ! Get the thread indices tx = threadidx%x ty = threadidx%y ! This thread computes C(i. 1 ) dimBlock = dim3( 16.:) :: A. ! Copy the results back and free up memory C(1:N. Bdev. 16 ! Fill the submatrices ! Each of the 16x16 threads in the thread block ! loads one element of Asub and Bsub Asub(tx.ty) = A(i.L) ) ! Copy A and B to the device Adev = A(1:N. 1 ) M = size( A. allocatable.Bdev. M.kb+ty-1) Bsub(tx.:) :: Adev.1:L) = Cdev deallocate( Adev. B.16 Cij = Cij + Asub(tx. Cdev.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. Bdev.

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

46 .

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

48 .

Sign up to vote on this title
UsefulNot useful