From 493909ddfe129b9297673f148b036d8abf8a7089 Mon Sep 17 00:00:00 2001 From: boufaras Date: Tue, 25 Oct 2011 12:53:56 +0000 Subject: [PATCH] add Cuda cmake git-svn-id: svn://scm.gforge.inria.fr/svnroot/paradiseo@2501 331e1502-861f-0410-8da2-ba01fb791d7f --- .../CMake/cuda/FindCUDA/make2cmake.cmake | 79 +++++ .../CMake/cuda/FindCUDA/parse_cubin.cmake | 112 +++++++ .../CMake/cuda/FindCUDA/run_nvcc.cmake | 280 ++++++++++++++++++ 3 files changed, 471 insertions(+) create mode 100644 trunk/paradiseo-gpu/CMake/cuda/FindCUDA/make2cmake.cmake create mode 100644 trunk/paradiseo-gpu/CMake/cuda/FindCUDA/parse_cubin.cmake create mode 100644 trunk/paradiseo-gpu/CMake/cuda/FindCUDA/run_nvcc.cmake diff --git a/trunk/paradiseo-gpu/CMake/cuda/FindCUDA/make2cmake.cmake b/trunk/paradiseo-gpu/CMake/cuda/FindCUDA/make2cmake.cmake new file mode 100644 index 000000000..7fce167c2 --- /dev/null +++ b/trunk/paradiseo-gpu/CMake/cuda/FindCUDA/make2cmake.cmake @@ -0,0 +1,79 @@ +# James Bigler, NVIDIA Corp (nvidia.com - jbigler) +# Abe Stephens, SCI Institute -- http://www.sci.utah.edu/~abe/FindCuda.html +# +# Copyright (c) 2008 - 2009 NVIDIA Corporation. All rights reserved. +# +# Copyright (c) 2007-2009 +# Scientific Computing and Imaging Institute, University of Utah +# +# This code is licensed under the MIT License. See the FindCUDA.cmake script +# for the text of the license. + +# The MIT License +# +# License for the specific language governing rights and limitations under +# Permission is hereby granted, free of charge, to any person obtaining a +# copy of this software and associated documentation files (the "Software"), +# to deal in the Software without restriction, including without limitation +# the rights to use, copy, modify, merge, publish, distribute, sublicense, +# and/or sell copies of the Software, and to permit persons to whom the +# Software is furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included +# in all copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS +# OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER +# DEALINGS IN THE SOFTWARE. +# + +####################################################################### +# This converts a file written in makefile syntax into one that can be included +# by CMake. + +file(READ ${input_file} depend_text) + +if (${depend_text} MATCHES ".+") + + # message("FOUND DEPENDS") + + # Remember, four backslashes is escaped to one backslash in the string. + string(REGEX REPLACE "\\\\ " " " depend_text ${depend_text}) + + # This works for the nvcc -M generated dependency files. + string(REGEX REPLACE "^.* : " "" depend_text ${depend_text}) + string(REGEX REPLACE "[ \\\\]*\n" ";" depend_text ${depend_text}) + + set(dependency_list "") + + foreach(file ${depend_text}) + + string(REGEX REPLACE "^ +" "" file ${file}) + + if(NOT IS_DIRECTORY ${file}) + # If softlinks start to matter, we should change this to REALPATH. For now we need + # to flatten paths, because nvcc can generate stuff like /bin/../include instead of + # just /include. + get_filename_component(file_absolute "${file}" ABSOLUTE) + list(APPEND dependency_list "${file_absolute}") + endif(NOT IS_DIRECTORY ${file}) + + endforeach(file) + +else() + # message("FOUND NO DEPENDS") +endif() + +# Remove the duplicate entries and sort them. +list(REMOVE_DUPLICATES dependency_list) +list(SORT dependency_list) + +foreach(file ${dependency_list}) + set(cuda_nvcc_depend "${cuda_nvcc_depend} \"${file}\"\n") +endforeach() + +file(WRITE ${output_file} "# Generated by: make2cmake.cmake\nSET(CUDA_NVCC_DEPEND\n ${cuda_nvcc_depend})\n\n") diff --git a/trunk/paradiseo-gpu/CMake/cuda/FindCUDA/parse_cubin.cmake b/trunk/paradiseo-gpu/CMake/cuda/FindCUDA/parse_cubin.cmake new file mode 100644 index 000000000..2518c6852 --- /dev/null +++ b/trunk/paradiseo-gpu/CMake/cuda/FindCUDA/parse_cubin.cmake @@ -0,0 +1,112 @@ +# James Bigler, NVIDIA Corp (nvidia.com - jbigler) +# Abe Stephens, SCI Institute -- http://www.sci.utah.edu/~abe/FindCuda.html +# +# Copyright (c) 2008 - 2009 NVIDIA Corporation. All rights reserved. +# +# Copyright (c) 2007-2009 +# Scientific Computing and Imaging Institute, University of Utah +# +# This code is licensed under the MIT License. See the FindCUDA.cmake script +# for the text of the license. + +# The MIT License +# +# License for the specific language governing rights and limitations under +# Permission is hereby granted, free of charge, to any person obtaining a +# copy of this software and associated documentation files (the "Software"), +# to deal in the Software without restriction, including without limitation +# the rights to use, copy, modify, merge, publish, distribute, sublicense, +# and/or sell copies of the Software, and to permit persons to whom the +# Software is furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included +# in all copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS +# OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER +# DEALINGS IN THE SOFTWARE. +# + +####################################################################### +# Parses a .cubin file produced by nvcc and reports statistics about the file. + + +file(READ ${input_file} file_text) + +if (${file_text} MATCHES ".+") + + # Remember, four backslashes is escaped to one backslash in the string. + string(REGEX REPLACE ";" "\\\\;" file_text ${file_text}) + string(REGEX REPLACE "\ncode" ";code" file_text ${file_text}) + + list(LENGTH file_text len) + + foreach(line ${file_text}) + + # Only look at "code { }" blocks. + if(line MATCHES "^code") + + # Break into individual lines. + string(REGEX REPLACE "\n" ";" line ${line}) + + foreach(entry ${line}) + + # Extract kernel names. + if (${entry} MATCHES "[^g]name = ([^ ]+)") + string(REGEX REPLACE ".* = ([^ ]+)" "\\1" entry ${entry}) + + # Check to see if the kernel name starts with "_" + set(skip FALSE) + # if (${entry} MATCHES "^_") + # Skip the rest of this block. + # message("Skipping ${entry}") + # set(skip TRUE) + # else (${entry} MATCHES "^_") + message("Kernel: ${entry}") + # endif (${entry} MATCHES "^_") + + endif(${entry} MATCHES "[^g]name = ([^ ]+)") + + # Skip the rest of the block if necessary + if(NOT skip) + + # Registers + if (${entry} MATCHES "reg([ ]+)=([ ]+)([^ ]+)") + string(REGEX REPLACE ".*([ ]+)=([ ]+)([^ ]+)" "\\3" entry ${entry}) + message("Registers: ${entry}") + endif() + + # Local memory + if (${entry} MATCHES "lmem([ ]+)=([ ]+)([^ ]+)") + string(REGEX REPLACE ".*([ ]+)=([ ]+)([^ ]+)" "\\3" entry ${entry}) + message("Local: ${entry}") + endif() + + # Shared memory + if (${entry} MATCHES "smem([ ]+)=([ ]+)([^ ]+)") + string(REGEX REPLACE ".*([ ]+)=([ ]+)([^ ]+)" "\\3" entry ${entry}) + message("Shared: ${entry}") + endif() + + if (${entry} MATCHES "^}") + message("") + endif() + + endif(NOT skip) + + + endforeach(entry) + + endif(line MATCHES "^code") + + endforeach(line) + +else() + # message("FOUND NO DEPENDS") +endif() + + diff --git a/trunk/paradiseo-gpu/CMake/cuda/FindCUDA/run_nvcc.cmake b/trunk/paradiseo-gpu/CMake/cuda/FindCUDA/run_nvcc.cmake new file mode 100644 index 000000000..7349da322 --- /dev/null +++ b/trunk/paradiseo-gpu/CMake/cuda/FindCUDA/run_nvcc.cmake @@ -0,0 +1,280 @@ +# James Bigler, NVIDIA Corp (nvidia.com - jbigler) +# +# Copyright (c) 2008 - 2009 NVIDIA Corporation. All rights reserved. +# +# This code is licensed under the MIT License. See the FindCUDA.cmake script +# for the text of the license. + +# The MIT License +# +# License for the specific language governing rights and limitations under +# Permission is hereby granted, free of charge, to any person obtaining a +# copy of this software and associated documentation files (the "Software"), +# to deal in the Software without restriction, including without limitation +# the rights to use, copy, modify, merge, publish, distribute, sublicense, +# and/or sell copies of the Software, and to permit persons to whom the +# Software is furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included +# in all copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS +# OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER +# DEALINGS IN THE SOFTWARE. + + +########################################################################## +# This file runs the nvcc commands to produce the desired output file along with +# the dependency file needed by CMake to compute dependencies. In addition the +# file checks the output of each command and if the command fails it deletes the +# output files. + +# Input variables +# +# verbose:BOOL=<> OFF: Be as quiet as possible (default) +# ON : Describe each step +# +# build_configuration:STRING=<> Typically one of Debug, MinSizeRel, Release, or +# RelWithDebInfo, but it should match one of the +# entries in CUDA_HOST_FLAGS. This is the build +# configuration used when compiling the code. If +# blank or unspecified Debug is assumed as this is +# what CMake does. +# +# generated_file:STRING=<> File to generate. This argument must be passed in. +# +# generated_cubin_file:STRING=<> File to generate. This argument must be passed +# in if build_cubin is true. + +if(NOT generated_file) + message(FATAL_ERROR "You must specify generated_file on the command line") +endif() + +# Set these up as variables to make reading the generated file easier +set(CMAKE_COMMAND "@CMAKE_COMMAND@") +set(source_file "@source_file@") +set(NVCC_generated_dependency_file "@NVCC_generated_dependency_file@") +set(cmake_dependency_file "@cmake_dependency_file@") +set(CUDA_make2cmake "@CUDA_make2cmake@") +set(CUDA_parse_cubin "@CUDA_parse_cubin@") +set(build_cubin @build_cubin@) +# We won't actually use these variables for now, but we need to set this, in +# order to force this file to be run again if it changes. +set(generated_file_path "@generated_file_path@") +set(generated_file_internal "@generated_file@") +set(generated_cubin_file_internal "@generated_cubin_file@") + +set(CUDA_NVCC_EXECUTABLE "@CUDA_NVCC_EXECUTABLE@") +set(CUDA_NVCC_FLAGS "@CUDA_NVCC_FLAGS@;;@CUDA_WRAP_OPTION_NVCC_FLAGS@") +@CUDA_NVCC_FLAGS_CONFIG@ +set(nvcc_flags "@nvcc_flags@") +set(CUDA_NVCC_INCLUDE_ARGS "@CUDA_NVCC_INCLUDE_ARGS@") +set(format_flag "@format_flag@") + +if(build_cubin AND NOT generated_cubin_file) + message(FATAL_ERROR "You must specify generated_cubin_file on the command line") +endif() + +# This is the list of host compilation flags. It C or CXX should already have +# been chosen by FindCUDA.cmake. +@CUDA_HOST_FLAGS@ + +# Take the compiler flags and package them up to be sent to the compiler via -Xcompiler +set(nvcc_host_compiler_flags "") +# If we weren't given a build_configuration, use Debug. +if(NOT build_configuration) + set(build_configuration Debug) +endif() +string(TOUPPER "${build_configuration}" build_configuration) +#message("CUDA_NVCC_HOST_COMPILER_FLAGS = ${CUDA_NVCC_HOST_COMPILER_FLAGS}") +foreach(flag ${CMAKE_HOST_FLAGS} ${CMAKE_HOST_FLAGS_${build_configuration}}) + # Extra quotes are added around each flag to help nvcc parse out flags with spaces. + set(nvcc_host_compiler_flags "${nvcc_host_compiler_flags},\"${flag}\"") +endforeach() +if (nvcc_host_compiler_flags) + set(nvcc_host_compiler_flags "-Xcompiler" ${nvcc_host_compiler_flags}) +endif() +#message("nvcc_host_compiler_flags = \"${nvcc_host_compiler_flags}\"") +# Add the build specific configuration flags +list(APPEND CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS_${build_configuration}}) + +if(DEFINED CCBIN) + set(CCBIN -ccbin "${CCBIN}") +endif() + +# cuda_execute_process - Executes a command with optional command echo and status message. +# +# status - Status message to print if verbose is true +# command - COMMAND argument from the usual execute_process argument structure +# ARGN - Remaining arguments are the command with arguments +# +# CUDA_result - return value from running the command +# +# Make this a macro instead of a function, so that things like RESULT_VARIABLE +# and other return variables are present after executing the process. +macro(cuda_execute_process status command) + set(_command ${command}) + if(NOT _command STREQUAL "COMMAND") + message(FATAL_ERROR "Malformed call to cuda_execute_process. Missing COMMAND as second argument. (command = ${command})") + endif() + if(verbose) + execute_process(COMMAND "${CMAKE_COMMAND}" -E echo -- ${status}) + # Now we need to build up our command string. We are accounting for quotes + # and spaces, anything else is left up to the user to fix if they want to + # copy and paste a runnable command line. + set(cuda_execute_process_string) + foreach(arg ${ARGN}) + # If there are quotes, excape them, so they come through. + string(REPLACE "\"" "\\\"" arg ${arg}) + # Args with spaces need quotes around them to get them to be parsed as a single argument. + if(arg MATCHES " ") + list(APPEND cuda_execute_process_string "\"${arg}\"") + else() + list(APPEND cuda_execute_process_string ${arg}) + endif() + endforeach() + # Echo the command + execute_process(COMMAND ${CMAKE_COMMAND} -E echo ${cuda_execute_process_string}) + endif(verbose) + # Run the command + execute_process(COMMAND ${ARGN} RESULT_VARIABLE CUDA_result ) +endmacro() + +# Delete the target file +cuda_execute_process( + "Removing ${generated_file}" + COMMAND "${CMAKE_COMMAND}" -E remove "${generated_file}" + ) + +# For CUDA 2.3 and below, -G -M doesn't work, so remove the -G flag +# for dependency generation and hope for the best. +set(depends_CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS}") +set(CUDA_VERSION @CUDA_VERSION@) +if(CUDA_VERSION VERSION_LESS "3.0") + cmake_policy(PUSH) + # CMake policy 0007 NEW states that empty list elements are not + # ignored. I'm just setting it to avoid the warning that's printed. + cmake_policy(SET CMP0007 NEW) + # Note that this will remove all occurances of -G. + list(REMOVE_ITEM depends_CUDA_NVCC_FLAGS "-G") + cmake_policy(POP) +endif() + +# nvcc doesn't define __CUDACC__ for some reason when generating dependency files. This +# can cause incorrect dependencies when #including files based on this macro which is +# defined in the generating passes of nvcc invokation. We will go ahead and manually +# define this for now until a future version fixes this bug. +set(CUDACC_DEFINE -D__CUDACC__) + +# Generate the dependency file +cuda_execute_process( + "Generating dependency file: ${NVCC_generated_dependency_file}" + COMMAND "${CUDA_NVCC_EXECUTABLE}" + -M + ${CUDACC_DEFINE} + "${source_file}" + -o "${NVCC_generated_dependency_file}" + ${CCBIN} + ${nvcc_flags} + ${nvcc_host_compiler_flags} + ${depends_CUDA_NVCC_FLAGS} + -DNVCC + ${CUDA_NVCC_INCLUDE_ARGS} + ) + +if(CUDA_result) + message(FATAL_ERROR "Error generating ${generated_file}") +endif() + +# Generate the cmake readable dependency file to a temp file. Don't put the +# quotes just around the filenames for the input_file and output_file variables. +# CMake will pass the quotes through and not be able to find the file. +cuda_execute_process( + "Generating temporary cmake readable file: ${cmake_dependency_file}.tmp" + COMMAND "${CMAKE_COMMAND}" + -D "input_file:FILEPATH=${NVCC_generated_dependency_file}" + -D "output_file:FILEPATH=${cmake_dependency_file}.tmp" + -P "${CUDA_make2cmake}" + ) + +if(CUDA_result) + message(FATAL_ERROR "Error generating ${generated_file}") +endif() + +# Copy the file if it is different +cuda_execute_process( + "Copy if different ${cmake_dependency_file}.tmp to ${cmake_dependency_file}" + COMMAND "${CMAKE_COMMAND}" -E copy_if_different "${cmake_dependency_file}.tmp" "${cmake_dependency_file}" + ) + +if(CUDA_result) + message(FATAL_ERROR "Error generating ${generated_file}") +endif() + +# Delete the temporary file +cuda_execute_process( + "Removing ${cmake_dependency_file}.tmp and ${NVCC_generated_dependency_file}" + COMMAND "${CMAKE_COMMAND}" -E remove "${cmake_dependency_file}.tmp" "${NVCC_generated_dependency_file}" + ) + +if(CUDA_result) + message(FATAL_ERROR "Error generating ${generated_file}") +endif() + +# Generate the code +cuda_execute_process( + "Generating ${generated_file}" + COMMAND "${CUDA_NVCC_EXECUTABLE}" + "${source_file}" + ${format_flag} -o "${generated_file}" + ${CCBIN} + ${nvcc_flags} + ${nvcc_host_compiler_flags} + ${CUDA_NVCC_FLAGS} + -DNVCC + ${CUDA_NVCC_INCLUDE_ARGS} + ) + +if(CUDA_result) + # Since nvcc can sometimes leave half done files make sure that we delete the output file. + cuda_execute_process( + "Removing ${generated_file}" + COMMAND "${CMAKE_COMMAND}" -E remove "${generated_file}" + ) + message(FATAL_ERROR "Error generating file ${generated_file}") +else() + if(verbose) + message("Generated ${generated_file} successfully.") + endif() +endif() + +# Cubin resource report commands. +if( build_cubin ) + # Run with -cubin to produce resource usage report. + cuda_execute_process( + "Generating ${generated_cubin_file}" + COMMAND "${CUDA_NVCC_EXECUTABLE}" + "${source_file}" + ${CUDA_NVCC_FLAGS} + ${nvcc_flags} + ${CCBIN} + ${nvcc_host_compiler_flags} + -DNVCC + -cubin + -o "${generated_cubin_file}" + ${CUDA_NVCC_INCLUDE_ARGS} + ) + + # Execute the parser script. + cuda_execute_process( + "Executing the parser script" + COMMAND "${CMAKE_COMMAND}" + -D "input_file:STRING=${generated_cubin_file}" + -P "${CUDA_parse_cubin}" + ) + +endif( build_cubin )