CUDA Fortran

Programming Guide and Reference

Release 2011

The Portland Group

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

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

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

ID: 103151117

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

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

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

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

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

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

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

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

viii .

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

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

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

2 .

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

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

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

it will be allocated in host pagelocked memory. Device constant variables are allocated in the device constant memory space. Attributes(pinned) A variable with the pinned attributes is called a pinned variable. The advantage of using pinned variables is that copies from page-locked memory to device memory are faster than copies from normal paged host memory. the variable will be allocated in the normal host paged memory. A device array may be an explicit-shape array. Attributes(shared) A variable with the shared attributed is called a device shared variable or a shared variable. or by other device subprograms to which it is passed as an argument. Intrinsic Datatypes Type integer logical real double precision Type Kind 1. A shared variable may only be declared in a device subprogram.4.4. availability. An allocatable device variable has a dynamic lifetime. if the allocation in page-locked memory fails.8 equivalent to real(kind=8) 6 . and has a lifetime of the thread block. though a write in one thread is only guaranteed to be visible to other threads after the next call to the SYNCTHREADS() intrinsic. and have a lifetime of the entire application. Datatypes in Device Subprograms The following intrinsic datatypes are allowed in device subprograms and device data: Table 2. or size of page-locked memory.8 1. Attributes(constant) A variable with the constant attributes is called a device constant variable. Some operating systems or installations may restrict the use. the variable may be accessed by any subprogram in that module and by any subprogram that uses the module.1. Other device variables have a lifetime of the entire application. but may be modified in host subprograms. or an assumed-shape dummy array. If declared in a module. A shared variable may not be data initialized. Device constant variables may not be allocatable. A shared variable is allocated in the device shared memory for a thread block.2. Device constant data may not be assigned or modified in any device subprogram. and may only be accessed within that subprogram. from when it is allocated until it is deallocated.2.Datatypes in Device Subprograms • If declared in a host subprogram. the variable may be accessed by that subprogram or subprograms contained in that subprogram. an allocatable array. When a pinned variable is allocated. A pinned variable must be an allocatable array.8 4. It can be read or written by all threads in the block.

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

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

the directive defines a two-dimensional thread block of size 32x4.j) +c(i.j) end do end do Without an explicit n on the do. n a(i. (256.j) +c(i. Alternatively. so all iterations are computed.4) >>> do j = 1. Example 1 !$cuf kernel do(2) <<< (*. n a(i. The grid shape. The body of the doubly-nested loop is turned in the kernel body: ThreadIdx%x runs from 1 to 32 and is mapped onto the inner i loop.j) end do end do 9 . In this case. only the outer j loop is run in parallel with a thread block size of 256. you could specify a configuration like the following in which the threads read and write the matrices in coalesced fashion.j) = b(i. The grid and block values can be an integer scalar or a parenthesized list. using asterisks tells the compiler to choose a thread block shape and/or compute the grid shape from the thread block shape and the loop limits. m do i = 1. is computed by the compiler and runtime by dividing the loop trip counts n and m by the thread block size. Programming Guide Syntax The general form of the kernel directive is: !$cuf kernel do[(n)] <<< grid.j) +c(i. the default value is 1. Loops which are not mapped onto the grid and block values are run sequentially on each thread. You might consider if the code in Example 2 would perform better if the two loops were interchanged. (32.Chapter 2. that is. n a(i. the schedule applies just to the outermost loop. ThreadIdx%y runs from 1 to 4 and is mapped onto the outer j loop. 256 >>> do j = 1.1) >>> do j = 1. starting at loop n and working out.*). specified as (*. Alternatively. m do i = 1. !$cuf kernel do(2) <<< *.*). The inner i dimension is run sequentially on each thread.j) = b(i.j) end do end do In this example.j) = b(i. block >>> The compiler maps the launch configuration specified by the grid and block values onto the outermost n loops. m do i = 1. Example 2 !$cuf kernel do <<< *.

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

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

12 .

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

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

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

the last dimension of a shared array may have an asterisk as its upper bound: real. variables and arrays with the constant attribute may not be assigned or modified.Variable Attributes • Members of a derived type may not have the constant attribute. • Shared variables may not be in a Common block or Equivalence statement. This memory is used for the 16 . Target. Shared arrays that are not dummy arguments may be declared as assumed-size arrays. as long as the interface is explicit and the matching dummy argument has the shared attribute. or Allocatable attributes. to be shared arrays of size 100. • Within device subprograms. variables and arrays with the constant attribute may be read and written. 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. • 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. The shared attribute can be specified with the attributes statement. • Members of a derived type may not have the shared attribute. the value of the bytes argument in the execution configuration is used to specify the number of bytes of shared memory that is dynamically allocated for each thread block. The following example declares two arrays. • Arrays with the constant attribute must have fixed size. real :: c(100) attributes(shared) :: c real. that is. or as an attribute on the type declaration statement. 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 :: x(*) Such an array has special significance. s and t. • Within host subprograms. In host subprograms. • Constant variables and arrays may be passed as actual arguments to host and device subprograms. as long as the subprogram interface is explicit. When the kernel is called. and the matching dummy argument also has the constant attribute.

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

device :: v(:) istat = cudaMalloc(v. device :: b(:) allocate(b(5024). These arrays will also have a lifetime of the subprogram as is usual with the Fortran language: subroutine vfunc(a. allocatable.c. The compiler will generate code to allocate the array in host page-locked memory. The cudaMalloc function can be used to allocate single-dimensional arrays of the supported intrinsic datatypes. 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. declare and initialize the scalar on the host as: integer. and thus the arrays are treated more like C malloc’ed areas. These functions return memory which will bypass certain Fortran allocatable properties such as automatic deallocation. To use these. device :: adev(n) real. pinned :: p(:) allocate(p(5000).Allocating Device and Pinned Arrays real. or if it is exhausted. allocatable. if available. Fortran interfaces to the CUDA memory management runtime routines are provided. If no such memory space is available. refer to“Memory Management.stat=istat) … if(allocated(b)) deallocate(b) Scalar variables can be allocated on the device using the Fortran 2003 allocatable scalar feature. allocatable. Mixing standard Fortran allocate/deallocate with the runtime Malloc/Free for a given array is not supported.n) real. 100) … istat = cudaFree(v) For a complete list of the memory management runtime routines. allocatable. Otherwise. real. and cudaFree can be used to free it: real. device :: atmp(4) … end subroutine vfunc ! adev and atmp are deallocated Allocating Device Memory Using Runtime Routines For programmers comfortable with the CUDA C programming environment. pinned allocatable arrays work and act like any other allocatable array on the host. the compiler allocates the array in normal paged host memory.stat=istat) … if(allocated(p)) deallocate(p) 18 . Allocating Pinned Memory Allocatable arrays with the pinned attribute are dynamically allocated using the Allocate statement.” on page 32.

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

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

They may have any of the intrinsic datatypes in the following table. it must be a scalar integer. or shared attributes.y. or a value returned by a call to cudaStreamCreate. or declared in device subprograms. refer to “Shared data. a kernel subroutine attributes(global) subroutine sub( a ) can be called like: call sub <<< DG. if present.2. it must be an integer. are limited to the types described in this section. and specifies the number of bytes of shared memory to be allocated for each thread block to use for assumed-size shared memory arrays. or if bytes is greater than the shared memory available.8 4(default). such as the function arguments and execution configuration arguments. It specifies the stream to which this call is enqueued.1. if present. where the members of the derived type have one of the allowed intrinsic datatypes. and have a value of zero. If not specified. Reference • bytes is optional. • stream is optional. For more information. DB >>> ( A ) The function call will fail if the grid or block arguments are greater than the maximum sizes allowed.5 on page 41. or another allowed derived type.Chapter 3.2. For instance. defined as type(dim3) integer(kind=4) :: x.8 1 (default) Additionally. they may be of derived type. Device code Datatypes allowed Variables and arrays with the device. Table 3. allowing for static shared memory declared in the kernel and for other dedicated uses.z end type 21 . The system module cudafor includes definitions of the derived type dim3. Device Code Intrinsic Datatypes Type integer logical real double precision complex character(len=1) Type Kind 1.4(default).4(default). constant. the value zero is used.8 equivalent to real(kind=8) 4(default).” on page 16.8 1. See Section 4.

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

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

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

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

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

only if the value of the argument is .8. Compute capability 1. Arithmetic and Bitwise Atomic Functions Function atomicadd( mem. They also combine that value with the value of the second argument. The value of the function is .true. mem. Atomic functions are only supported by devices with compute capability 1.0) ) allneg = .true. Arithmetic and Bitwise Atomic Functions These atomic functions read and return the value of the first argument. BALLOT The ballot function is a warp-vote operation with a single integer argument: unsigned integer ballot(int_value) The function ballot evaluates the argument int_value for all threads of the warp and returns an integer whose Nth bit is set if and only if int_value evaluates to non-zero for the Nth thread of the warp. for all threads in the warp. Example: if( ballot(int_value) ) allneg = . The function anythread evaluates its argument for all threads in the current warp.1 and higher. which must be a variable or array element in shared memory (with the shared attribute) or in device global memory (with the device attribute). value ) Additional Atomic Update mem = mem + value mem = mem – value 27 . value ) atomicsub( mem.Chapter 3. Note The return value for each of these functions is the first argument. These functions are: Table 3. Reference ANYTHREAD The anythread function is a warp-vote operation with a single scalar logical argument: if( anythread(a(i)<0.2 or higher is required if the first argument has the shared attribute. and store the combined value back to the first argument location. depending on the function.0. Both arguments must be of type integer(kind=4).false. This function is only supported by devices of compute capability 2. 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. Atomic Functions The atomic functions read and write the value of their first operand.false.

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

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

30 .

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

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

devptr may be any allocatable. in the future. Use the functions in this section for memory management. devptr may be of TYPE(C_DEVPTR). devptr may be of TYPE(C_DEVPTR). 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. “Device Code Intrinsic Datatypes.” on page 21. Or.1. Currently. height cudaMallocArray allocates a data array on the device. pitch.” on page 21.1. devptr may also be of TYPE(C_DEVPTR). Under CUDA Fortran. proper Fortran pointers for device data are supported. two-dimensional device array of a supported type specified in Table 3. width. cudaMallocPitch may pad the data. and in effect transfer the allocation to the Fortran array. “Device Code Intrinsic Datatypes. The width is in terms of number of elements.” on page 21. devptr may be any allocatable. c_f_pointer will take a TYPE(C_DEVPTR) as the first argument. Both of these features are subject to change when. a shape as the third argument. and the padded width is returned in the variable pitch. cudaFreeArray integer function cudaFreeArray(carray) type(cudaArrayPtr) :: carray 33 . “Device Code Intrinsic Datatypes. cudaMalloc integer function cudaMalloc(devptr. Or. height) cudaMallocPitch allocates data on the device. count) cudaMalloc allocates data on the device. Runtime APIs TYPE(C_DEVPTR). 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. Similarly.Chapter 4. cudaFree integer function cudaFree(devptr) cudaFree deallocates data on the device. height) type(cudaArrayPtr) :: carray type(cudaChannelFormatDesc) :: cdesc integer :: width. one-dimensional device array of a supported type specified in Table 3. cudaMallocPitch integer function cudaMallocPitch(devptr. as well as consistency checks between Fortran device arrays and host arrays. in which case the integer values are expressed in bytes. The count is in terms of elements. cudaMallocArray integer function cudaMallocArray(carray.1. The height is an integer. width. an allocatable device array as the second argument. in which case the count is in bytes. devptr may be any allocatable device array of a supported type specified in Table 3. should be of benefit. cdesc.

cudaMemset integer function cudaMemset(devptr. in which case the count is in term of bytes. Or.Memory Management cudaFreeArray frees an array that was allocated on the device. dpitch. The count is in terms of elements. dst and src may be of TYPE(C_DEVPTR) or TYPE(C_PTR). “Device Code Intrinsic Datatypes. kdir may be optional. cudaMemcpy2D integer function cudaMemcpy2D(dst. This function operates on page-locked host memory only. of a supported type specified in Table 3. src. “Device Code Intrinsic Datatypes.” on page 20. of a supported type specified in Table 3. The count is in terms of elements. dst and src may be any device or host array. The value must match in type and kind. kdir may be optional. kdir) cudaMemcpy copies data from one location to another. and the lowest byte of value is used. cudaMemcpy integer function cudaMemcpy(dst. value. scalar or array. in which case the count is in term of bytes. otherwise the stream argument is optional and defaults to zero. cudaMemcpyAsync integer function cudaMemcpyAsync(dst. dst and src may be any device or host. width. devptr may be of TYPE(C_DEVPTR).” on page 21. cudaMemcpyDeviceToHost. devptr may be of TYPE(C_DEVPTR). “Device Code Intrinsic Datatypes. and height are in terms of elements. and the lowest byte of value is used. cudaMemcpyDeviceToHost. cudaMemset2D integer function cudaMemset2D(devptr. src. The pitch.1. and height are in terms of bytes. value. Or. kdir) cudaMemcpy2D copies data from one location to another. devptr may be any device scalar or array of a supported type specified in Table 3.” on page 21. it must be one of the defined enums cudaMemcpyHostToDevice.” on page 21. count) cudaMemset sets a location or array to the specified value. pitch. in which case the pitch. Alternatively. of a supported type specified in Table 3. or cudaMemcpyDeviceToDevice.” on page 20. The copy can be associated with a stream by passing a non-zero stream argument. width. dst and src may be any device or host. dst and src may be of TYPE(C_DEVPTR) or TYPE(C_PTR).” on page 21. spitch. in which case the count is in term of bytes. for more information. kdir. it must be one of the defined enums cudaMemcpyHostToDevice. The value must match in type and kind. height. “Device Code Intrinsic Datatypes.” on page 21. The width 34 . scalar or array. width. The count is in terms of elements.1. “Device Code Intrinsic Datatypes. or cudaMemcpyDeviceToDevice. devptr may be any device array of a supported type specified in Table 3.1.1. If kdir is specified. refer to “Data Transfer Using Runtime Routines. stream) cudaMemcpy copies data from one location to another. width. refer to “Data Transfer Using Runtime Routines. for more information. src. Alternatively. height) cudaMemset2D sets an array to the specified value. count. count. If kdir is specified.1.

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

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

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

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

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

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

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

42 .

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

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

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

46 .

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

48 .