Skip to content

Latest commit

 

History

History
313 lines (263 loc) · 13.6 KB

README.md

File metadata and controls

313 lines (263 loc) · 13.6 KB

CLOC - V 1.3.2

CLOC: CL Offline Compiler Generate HSA code object from a cl (Kernel c Language) file. SNACK: Structured No API Compiled Kernels. Launch GPU kernels as host-callable functions with structured launch parameters.

Table of contents

# Copyright and Disclaimer ------------------------

Copyright (c) 2016 ADVANCED MICRO DEVICES, INC.

AMD is granting you permission to use this software and documentation (if any) (collectively, the �Materials�) pursuant to the terms and conditions of the Software License Agreement included with the Materials. If you do not have a copy of the Software License Agreement, contact your AMD representative for a copy.

You agree that you will not reverse engineer or decompile the Materials, in whole or in part, except for example code which is provided in source code form and as allowed by applicable law.

WARRANTY DISCLAIMER: THE SOFTWARE IS PROVIDED "AS IS" WITHOUT WARRANTY OF ANY KIND. AMD DISCLAIMS ALL WARRANTIES, EXPRESS, IMPLIED, OR STATUTORY, INCLUDING BUT NOT LIMITED TO THE IMPLIED WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE, NON-INFRINGEMENT, THAT THE SOFTWARE WILL RUN UNINTERRUPTED OR ERROR- FREE OR WARRANTIES ARISING FROM CUSTOM OF TRADE OR COURSE OF USAGE. THE ENTIRE RISK ASSOCIATED WITH THE USE OF THE SOFTWARE IS ASSUMED BY YOU. Some jurisdictions do not allow the exclusion of implied warranties, so the above exclusion may not apply to You.

LIMITATION OF LIABILITY AND INDEMNIFICATION: AMD AND ITS LICENSORS WILL NOT, UNDER ANY CIRCUMSTANCES BE LIABLE TO YOU FOR ANY PUNITIVE, DIRECT, INCIDENTAL, INDIRECT, SPECIAL OR CONSEQUENTIAL DAMAGES ARISING FROM USE OF THE SOFTWARE OR THIS AGREEMENT EVEN IF AMD AND ITS LICENSORS HAVE BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. In no event shall AMD's total liability to You for all damages, losses, and causes of action (whether in contract, tort (including negligence) or otherwise) exceed the amount of $100 USD. You agree to defend, indemnify and hold harmless AMD and its licensors, and any of their directors, officers, employees, affiliates or agents from and against any and all loss, damage, liability and other expenses (including reasonable attorneys' fees), resulting from Your use of the Software or violation of the terms and conditions of this Agreement.

U.S. GOVERNMENT RESTRICTED RIGHTS: The Materials are provided with "RESTRICTED RIGHTS." Use, duplication, or disclosure by the Government is subject to the restrictions as set forth in FAR 52.227-14 and DFAR252.227-7013, et seq., or its successor. Use of the Materials by the Government constitutes acknowledgement of AMD's proprietary rights in them.

EXPORT RESTRICTIONS: The Materials may be subject to export restrictions as stated in the Software License Agreement.

# Command Help --------- ## The cloc.sh command
   cloc.sh: Compile a cl or cu file into an HSA Code object file (.hsaco)  
            using the LLVM Ligntning Compiler. An hsaco file contains 
            the amdgpu isa that can be loaded by the HSA Runtime.
            As of amdcloc 1.3.1, use of cuda is only experimental.  
            Generated kernels from cuda will not execute. 

   Usage: cloc.sh [ options ] filename.cl

   Options without values:
    -ll       Generate IR for LLVM steps before generating hsaco
    -s        Generate dissassembled gcn from hsaco
    -g        Generate debug information
    -noqp     No quickpath, Use LLVM commands instead of clang driver
    -noshared Do not link hsaco as shared object, forces noqp
    -version  Display version of cloc then exit
    -v        Verbose messages
    -n        Dryrun, do nothing, show commands that would execute
    -h        Print this help message
    -k        Keep temporary files
    -brig     Generate brig instead of hsaco (deprecated)
    -hsail    Generate dissassembled hsail for brig  (deprecated)

   Options with values:
    -amdllvm   <path>           $AMDLLVM or /opt/amd/llvm
    -libgcn    <path>           $LIBGCN or /opt/rocm/libamdgcn  
    -hlcpath   <path>           $HLC_PATH or /opt/rocm/hlc3.2/bin  
    -cuda-path <path>           $CUDA_PATH or /usr/local/cuda
    -mcpu      <cputype>        Default= value returned by mymcpu
    -bclib     <bcfile>         Add a bc library for llvm-link
    -clopts    <compiler opts>  Addtional options for cl frontend
    -cuopts    <compiler opts>  Additonal options for cu frontend
    -I         <include dir>    Provide one directory per -I option
    -lkopts    <LLVM link opts> Default=$LIBGCN/lib/libamdgcn.$mcpu.bc
    -hsaillib  <fname>          Filename of hsail library. (deprecated)
    -opt       <LLVM opt>       LLVM optimization level
    -o         <outfilename>    Default=<filename>.<ft> ft=brig or hsail
    -t         <tdir>           Temporary directory or intermediate files
                                Default=/tmp/cloc-tmp-$$
   Examples:
    cloc.sh my.cl             /* creates my.hsaco                    */
    cloc.sh whybother.cu      /* creates whybother.hsaco             */

   Note: Instead of providing these command line options:
   -opt,-hlcpath,-amdllvm,-libgcn,-cuda-path,-mcpu,-clopts, or -lkopts 

   You may set these environment variables, respectively:
   LLVMOPT,HLC_PATH,AMDLLVM,LIBGCN,CUDA_PATH,LC_MCPU,CLOPTS, or LKOPTS 

   Command line options will take precedence over environment variables. 

   Copyright (c) 2016 ADVANCED MICRO DEVICES, INC.

   snack: Generate host-callable "snack" functions for GPU kernels.
          Snack generates the source code and headers for each kernel 
          in an OpenCL cl or CUDA cu file.  The -c option will compile 
          the source with gcc so you can link with your host application.
          Host applicaton requires no API to use snack functions.

   Usage: snack.sh [ options ] filename.cl

   Options without values:
    -c        Compile generated source code to create .o file
    -ll       Tell cloc.sh to generate IR for LLVM steps
    -noqp     Tell cloc.sh to not use quick path
    -noshared Tell cloc.sh to not create shared object.
    -s        Generate gcn assembly (.s) from lld output
    -version  Display version of snack then exit
    -v        Verbose messages
    -vv       Get additional verbose messages from cloc.sh
    -n        Dryrun, do nothing, show commands that would execute
    -h        Print this help message
    -k        Keep temporary files
    -fort     Generate fortran function names
    -noglobs  Do not generate global functions 
    -kstats   Print out code object kernel statistics 
    -m32      Generate snackwrap in 32-bit mode. If -c, also compile in 32
              bit mode

   Options with values:
    -path      <path>         $CLOC_PATH or <sdir> if CLOC_PATH not set
                              <sdir> is directory where snack.sh is found
    -mcpu      <cpu>          Default=`'mymcpu`, Options: kaveri,carrizo,fiji
    -amdllvm   <path>         Default=/opt/amd/llvm or env var AMDLLVM 
    -libgcn    <path>         Default=/opt/rocm/libamdgcn or env var LIBGCN 
    -cuda-path <path>         $CUDA_PATH or /usr/local/cuda
    -I        <include dir>  Provide one directory per -I option for cloc.sh
    -bclib    <bcfile>       Add a bc library for llvm-link
    -opt      <LLVM opt>     Default=2, passed to cloc.sh to build code object
    -gccopt   <gcc opt>      Default=2, gcc optimization for snack wrapper
    -t        <tempdir>      Default=/tmp/snk_$$, Temp dir for files
    -s        <symbolname>   Default=filename 
    -hsart    <HSA RT>       Default=CLOC_PATH/..
    -o        <outfilename>  Default=<filename>.<ft> 
    -foption  <fnlizer opts> Default=""  Finalizer options

   Examples:
    snack.sh my.cl              /* create my.snackwrap.c and my.h    */
    snack.sh -c my.cl           /* gcc compile to create  my.o       */
    snack.sh -c whybother.cl    /* compile to create  whybother.o    */
    snack.sh -t /tmp/foo my.cl  /* will automatically set -k         */

   Instead of providing these command line options:
   -path,-hsart,-amdllvm,-libgcn,-cuda-path,-mcpu

   You may set these environment variables respectively:
   CLOC_PATH,HSA_RT,AMDLLVM,LIBGCN,CUDA_PATH,LC_MCPU

   Command line options will take precedence over environment variables. 

   Copyright (c) 2016 ADVANCED MICRO DEVICES, INC.

# Examples --------

This version of cloc supports the SNACK method of writing accelerated kernels in c. With SNACK, the host program can directly call the accelerated function without an API such as OpenCL or CUDA.

Here is the c++ source code HelloWorld.cpp using SNACK.

#include <string.h>
#include <stdlib.h>
#include <iostream>
using namespace std;
#include "hw.h"
int main(int argc, char* argv[]) {
	const char* input_const = "Gdkkn\x1FGR@\x1FVnqkc";
	size_t strlength = strlen(input_const);
	char *output = (char*) malloc_global(strlength + 1);
	char *input = (char*) malloc_global(strlength + 1);
        strncpy(input,input_const,strlength);
        SNK_INIT_LPARM(lparm,strlength);
        decode(input,output,lparm);
	output[strlength] = '\0';
	cout << output << endl;
	free_global(output);
	free_global(input);
	return 0;
}

The c source for the accelerated kernel is in file hw.cl.

/*
    File:  hw.cl 
*/
__kernel void decode(__global const char* in, __global char* out) {
	int num = get_global_id(0);
	out[num] = in[num] + 1;
}

The -c option of snack.sh will call the gcc compiler and create the object file and the header file "hw.h" required to build the host program. Without -c you get the generated c code hw.snackwrap.c and the header file. The header file has the function prototypes for all kernels declared in the .cl file. Use this command to compile the hw.cl file with snack.sh.

snack.sh -c hw.cl

You can now compile and build the binary "HelloWorld" with any c++ compiler. Here is the command to build HelloWorld with g++.

g++ -o HelloWorld hw.o HelloWorld.cpp -L/opt/rocm/lib -lhsa-runtime64 -lelf 

Then execute the program as follows.

$ ./HelloWorld
Hello HSA World

This example and other examples can be found in the cloc examples directory found in /opt/rocm/cloc/examples

Example 2: Use of -hsaillib by cloc.sh and snackhsail.sh

The snack.sh command creates and embeds HSA code object or isa. To create and embed the HSAIL intermediate language you can use the snackhsail.sh command in an similar way.

The snackhsail.sh command has an additional option to provide an HSAIL library. This options specifies a single file with the hsail for multiple functions that can be called directly by cl. One should have a corresponding header file for these functions.
An example to use the -hsaillib option is provided in the directory examples/snack/test_hsail_lib. This example uses the simple HSAIL library provided in the directory examples/mathdemo_hsaillib

Here you see the header file and the corresponding hsail file. This is the contents of the header file mathdemo_hsaillib.h.

float __sin(float in);
float __cos(float in);
float __exp(float in);

The example cl file (examples/snack/test_hsail_lib/test_hsail_lib.cl) includes the above header file.

#include "../../mathdemo_hsaillib/mathdemo_hsaillib.h"
__kernel void testkernel( __global float * outfval , __global const float  * fval) {
   int i = get_global_id(0);
   outfval[i] =  __sin(fval[i]);
}

If you want to compile and build the example, execute the buildrun.sh script found in the example directory. It calls snackhsail.sh as follows:

snackhsail.sh  -v -c  -hsaillib ../../mathdemo_hsaillib/mathdemo_hsaillib.hsail test_hsail_lib.cl

If you are just building the hsail for the above kernel, provide the -hsaillib option to cloc.sh as follows.

cloc.sh -hsail -hsaillib ../../mathdemo_hsaillib/mathdemo_hsaillib.hsail test_hsail_lib.cl

Your resulting hsail test_hsail_lib.hsail will include the hsaillib inserted into the correct position. The generated brig will also include the library.

Example 3: Creating code object file with GCN isa and loading it with HSA API.

In it's simplest form, cloc.sh compiles a .cl file into an HSA code object file. This is a standard ELF file that can be loaded by the HSA API without the need for finalization. However, the rest of the HSA API is required to launch this object code to the GPU. An example is provided in the examples/hsa directory that shows the necessary HSA API to launch an HSA code object. Use these commands to run this example:

cd $HOME
cp -rp /opt/rocm/cloc/examples/hsa/vector_copy_codeobject_rocm /tmp
cd /tmp/vector_copy_codeobject_rocm
make
make test

The make command will compile vector_copy_codeobject.cpp with these commands:

g++ -c -std=c++11 -I/opt/rocm/hsa/include -o obj/vector_copy_codeobject.o vector_copy_codeobject.cpp
g++ obj/vector_copy_codeobject.o -L/opt/rocm/lib -lhsa-runtime64 -o vector_copy_codeobject

It then calls cloc.sh to create the file vectory_copy_codeobject.hsaco with this command.

cloc.sh vector_copy_codeobject.cl

The program will load the HSA code object file and launch it with the HSA API. One could compare the minor difference in the HSA API between loading a code object and loading brig by comparing the source code and Makefile from the above example to the vector_copy example found in /opt/rocm/cloc/examples/hsa/vector_copy. In the vector_copy example, cloc.sh requires the -brig option and the API source code in vector_copy.c has the extra finalization step.