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

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

.................... 24 3...............................................8................ Device Code Intrinsic Datatypes .......................9...........................7................2....................................................1.. Fortran Random Number Intrinsics ......... Arithmetic and Bitwise Atomic Functions ..........Tables 2..................... 21 3.................................. Fortran Numeric and Logical Intrinsics ..................................................... 29 vii ............10......6....... 24 3...... Fortran Numeric Inquiry Intrinsics ............................................................................... Intrinsic Datatypes ................................. Fortran Bit Manipulation Intrinsics ........................... Fortran Mathematical Intrinsics . 23 3......... Fortran Reduction Intrinsics ...................................................................1...........................................................................................................3...........................................4.............. 23 3.5............................................................. 27 3................................... Compare and Swap Atomic Function ............................................................................ Counting Atomic Functions .................. 28 3................................... 6 3.................................. 22 3........ 23 3...........................................................

viii .

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

There are a wide variety of releases and distributions of each of these types of operating systems. Available online at http://www.pgroup. 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. filename . Bold is used for commands.. Available online at http:// www. C/C++ C/C++ language statements are shown in the test of this guide using a reduced fixed point size.pdf. you must select either item2 or item3.Terminology Constant Width is used for filenames. including assembly language statements.9.1. directories. Version 3. multiple filenames are allowed. Release 10.1. June 2010. and for language statements in the text. Version 3. Terminology If there are terms in this guide with which you are unfamiliar.. examples. 1997 (Fortran 95).nvidia. September. x . • NVIDIA CUDA Programming Guide. PGI provides a glossary of terms which you can access at www.nvidia. options.htm Related Publications The following documents contain additional information related to CUDA Fortran programming. • NVIDIA CUDA Compute Unified Device Architecture Reference Manual.com/doc/pgiug. Available online at http:// www.1. Geneva. [ item1 ] in general. Information Technology – Programming Languages – FORTRAN. FORTRAN Fortran language statements are shown in the text of this guide using a reduced fixed point size.pgroup. Zero or more of the preceding item may occur. square brackets are required to specify a p/t-set. In the context of p/tsets. • ISO/IEC 1539-1:1997. NVIDIA. In this example. In this case item1 is optional.com/cuda. 2009. square brackets indicate optional items.com/cuda. The Portland Group. 7/21/2010. { item2 | item 3} braces indicate that a selection is required. • PGI Users Guide.com/support/definitions. ellipsis indicate a repetition. In this case. NVIDIA. arguments.

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

2 .

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

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

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

availability. Device constant data may not be assigned or modified in any device subprogram.2. the variable will be allocated in the normal host paged memory. though a write in one thread is only guaranteed to be visible to other threads after the next call to the SYNCTHREADS() intrinsic.4.Datatypes in Device Subprograms • If declared in a host subprogram.2.1. Some operating systems or installations may restrict the use. Datatypes in Device Subprograms The following intrinsic datatypes are allowed in device subprograms and device data: Table 2. Attributes(pinned) A variable with the pinned attributes is called a pinned variable. An allocatable device variable has a dynamic lifetime. Device constant variables are allocated in the device constant memory space. Other device variables have a lifetime of the entire application. A shared variable may not be data initialized. the variable may be accessed by that subprogram or subprograms contained in that subprogram. 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. Intrinsic Datatypes Type integer logical real double precision Type Kind 1. A shared variable may only be declared in a device subprogram. an allocatable array. and has a lifetime of the thread block. Attributes(constant) A variable with the constant attributes is called a device constant variable.8 4. from when it is allocated until it is deallocated. It can be read or written by all threads in the block. and may only be accessed within that subprogram.8 equivalent to real(kind=8) 6 . or an assumed-shape dummy array. the variable may be accessed by any subprogram in that module and by any subprogram that uses the module. if the allocation in page-locked memory fails. When a pinned variable is allocated. but may be modified in host subprograms. A pinned variable must be an allocatable array. A shared variable is allocated in the device shared memory for a thread block.8 1. A device array may be an explicit-shape array. or by other device subprograms to which it is passed as an argument. it will be allocated in host pagelocked memory. and have a lifetime of the entire application. Attributes(shared) A variable with the shared attributed is called a device shared variable or a shared variable. If declared in a module.4. Device constant variables may not be allocatable. or size of page-locked memory.

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

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

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

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

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

12 .

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

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

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

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

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

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

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

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

constant. or shared attributes. The system module cudafor includes definitions of the derived type dim3.4(default). Reference • bytes is optional.2. defined as type(dim3) integer(kind=4) :: x. For more information. They may have any of the intrinsic datatypes in the following table.8 equivalent to real(kind=8) 4(default). and have a value of zero. If not specified.4(default). if present. are limited to the types described in this section.8 1 (default) Additionally.z end type 21 .5 on page 41. Device code Datatypes allowed Variables and arrays with the device.1. or if bytes is greater than the shared memory available. or a value returned by a call to cudaStreamCreate. or declared in device subprograms. or another allowed derived type. such as the function arguments and execution configuration arguments. allowing for static shared memory declared in the kernel and for other dedicated uses.y. refer to “Shared data.8 4(default). a kernel subroutine attributes(global) subroutine sub( a ) can be called like: call sub <<< DG. Device Code Intrinsic Datatypes Type integer logical real double precision complex character(len=1) Type Kind 1. it must be a scalar integer. It specifies the stream to which this call is enqueued. and specifies the number of bytes of shared memory to be allocated for each thread block to use for assumed-size shared memory arrays. it must be an integer. they may be of derived type.” on page 16. DB >>> ( A ) The function call will fail if the grid or block arguments are greater than the maximum sizes allowed. the value zero is used. For instance. See Section 4. if present.8 1. Table 3. where the members of the derived type have one of the allowed intrinsic datatypes.Chapter 3.2. • stream is optional.

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

complex real real Fortran Numeric Inquiry Intrinsics Table 3.3. complex real real.Chapter 3.real) real. complex real real. Fortran Bit Manipulation Intrinsics Name btest iand ibclr ibits ibset ieor ior Argument Datatypes integer integer integer integer integer integer integer Name ishft ishftc leadz mvbits not popcnt poppar Argument Datatypes integer integer integer integer integer integer integer 23 .integer) real Fortran Bit Manipulation Intrinsics Table 3. Reference Fortran Mathematical Intrinsics Table 3. complex Name log log10 sin sinh sqrt tan tanh Argument Datatypes real.4. real. complex integer. real real real Name precision radix range selected_int_kind selected_real_kind tiny Argument Datatypes real.5. complex real real. complex integer (integer. real real integer. Fortran Numeric Inquiry Intrinsics Name bit_size digits epsilon huge maxexponent minexponent Argument Datatypes integer integer. Fortran Mathematical Intrinsics Name acos asin atan atan2 cos cosh exp Argument Datatypes real real real (real. real integer.

real. To avoid these potential issues. write-after-read.7. or write-after-write hazards for some of these memory accesses. use syncthreads()to specify synchronization points in the kernel. Fortran Random Number Intrinsics Name random_number random_seed Argument Datatypes real integer New Intrinsic Functions This section describes the new intrinsic functions and subroutines supported in device subprograms. complex Fortran Random Number Intrinsics Table 3. refer to “Thread Management. real integer. This 24 . real integer. it has no arguments: void syncthreads() Sometimes threads within a block access the same addresses in shared or global memory. syncthreads syncthreads_count syncthreads_and syncthread_or threadfence threadfence_block threadfence_system For detailed information on these functions. Synchronization Functions The synchronization functions control the synchronization of various threads during execution of thread blocks. complex integer. .Device code Fortran Reduction Intrinsics Table 3.” on page 32.6. thus creating potential read-after-write. real Name minloc minval product sum Argument Datatypes integer. real integer. Fortran Reduction Intrinsics Name all any count maxloc maxval Argument Datatypes logical logical logical integer. real. SYNCTHREADS The syncthreads intrinsic subroutine acts as a barrier synchronization for all threads in a single thread block.

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

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

value ) Additional Atomic Update mem = mem + value mem = mem – value 27 . Example: if( ballot(int_value) ) allneg = . Note The return value for each of these functions is the first argument. for all threads in the warp.0.8.false.1 and higher. Atomic functions are only supported by devices with compute capability 1.false. only if the value of the argument is .2 or higher is required if the first argument has the shared attribute.0) ) allneg = . and store the combined value back to the first argument location. The value of the function is . depending on the function. They also combine that value with the value of the second argument. Arithmetic and Bitwise Atomic Functions Function atomicadd( mem. which must be a variable or array element in shared memory (with the shared attribute) or in device global memory (with the device attribute).Chapter 3. mem. Reference ANYTHREAD The anythread function is a warp-vote operation with a single scalar logical argument: if( anythread(a(i)<0. Atomic Functions The atomic functions read and write the value of their first operand. These functions are: Table 3.true.true. value ) atomicsub( mem. Compute capability 1. 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. Both arguments must be of type integer(kind=4). The function anythread evaluates its argument for all threads in the current warp. 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. Arithmetic and Bitwise Atomic Functions These atomic functions read and return the value of the first argument. This function is only supported by devices of compute capability 2.

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

Chapter 3. such as the new sizeof intrinsic function.6. PRINT. • Automatic arrays must be fixed size. character substrings are not supported. • Alternate return specifications are not allowed. integer(kind=4) :: i. INQUIRE. Host code Host subprograms may use intrinsic functions. • Floating point exception handling is not supported. WRITE.10. • STOP and PAUSE statements are not allowed. OPEN. Compare and Swap Atomic Function Function atomiccas(mem. • Cray pointers are not supported. where A is a variable or expression. • Recursive subroutines and functions are not allowed. FORMAT. • Input/Output statements are not allowed: READ. BACKSPACE. SIZEOF Intrinsic A call to sizeof(A). Reference Table 3.3 are not supported. • Objects with the Pointer and Allocatable attribute are not allowed. j j = sizeof(i) ! this assigns the value 4 to j 29 . • Objects with character type must have LEN=1.comp. CLOSE. • ENTRY statements are not allowed. • Optional arguments are not allowed. returns the number of bytes required to hold the value of A. • Fortran intrinsic functions not listed in Section 3. ENDFILE. REWIND.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. NAMELIST. • Subroutine and function calls are supported only if they can be inlined.

30 .

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

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

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

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

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

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

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

cudaHostGetDevicePointer integer function cudaHostGetDevicePointer(devptr. flags cudaHostAlloc allocates pinned memory on the host. dst may be any host scalar or array of a supported type specified in “Datatypes allowed. hostptr. allocatable. hostptr is a pinned memory buffer that was allocated via cudaHostAlloc(). csvx) Call c_f_pointer(cdvx. vx.” on page 21. The count is in terms of elements. The copy can be associated with a stream by passing a non-zero stream argument. 100) Vx = 0.0 cudaGetSymbolSize integer function cudaGetSymbolSize(size. This function operates on page-locked host memory only. symbol) type(C_DEVPTR) :: devptr type(cudaSymbol) :: symbol cudaGetSymbolAddress returns in the devptr argument the address of symbol on the device. The following code sequence initializes a global device array “vx” from a CUDA C kernel: type(cudaSymbol) :: csvx type(c_devptr) :: cdvx real. 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. cudaHostAlloc integer function cudaHostAlloc(hostptr. cudaGetSymbolAddress integer function cudaGetSymbolAddress(devptr. device :: vx(:) csvx = “vx” Istat = cudaGetSymbolAddress(cdvx. It returns in devptr an address that can be passed to. Size is in bytes. A symbol can be set to an external device name via a character string. It returns in hostptr the address of the pagelocked allocation. The flags argument enables different options to be specified that affect the allocation. and read and written by. size. or returns an error if the memory is unavailable. The normal iso_c_binding subroutine c_f_pointer can be used to move the type(c_ptr) to a Fortran pointer. 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. flags) type(C_PTR) :: hostptr integer :: size. a kernel which runs on 38 .Memory Management cudaMemcpyFromSymbol copies data from a device area in global or constant memory space referenced by a symbol to a destination on the host.

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

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

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

the program assumes the matrix sizes are a multiple of 16. k.htm Matrix Multiplication Example This example shows a program to compute the product C of two matrices A and B. accumulates the submatrix product. M. C is then NxL attributes(global) subroutine mmul_kernel( A.com/support/index.j) being computed 43 . and has not been highly optimized for execution time. 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). L integer :: i. Each element of the result is the product of one row of A by one column of B. B(M. kb. then moves to the next submatrix of A rowwise and of B columnwise. j. shared :: Asub(16. N. M. C. The program computes the products by accumulating submatrix products. C(N. as follows: • Each thread block computes one 16x16 submatrix of C. L ) real :: A(N. For simplicity. tx. value :: N. • Each thread within the block computes one element of the submatrix. 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. Examples This chapter contains For up-to-date information about the state of the current release. The program caches the submatrices of A and B in the fast shared memory.Chapter 5. visit the frequently asked questions (FAQ) section of the pgroup. B. Bsub(16.16) ! the value of C(i. B is MxL.com web page at: www.L) integer.16). ty ! submatrices stored in shared memory real.pgroup. it reads a block submatrix of A and a block submatrix of B.M).L).

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

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

46 .

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

48 .