Skip to content

Commit

Permalink
Merge pull request #212 from DrTimothyAldenDavis/master
Browse files Browse the repository at this point in the history 
v8.0.0
  • Loading branch information
@DrTimothyAldenDavis
DrTimothyAldenDavis authored May 18, 2023
2 parents 4910708 + f2d3a1f commit 3138f5a
Show file tree
Hide file tree
Showing 7,346 changed files with 949,709 additions and 1,314,547 deletions.
The diff you're trying to view is too large. We only load the first 3000 changed files.
9 changes: 2 additions & 7 deletions .gitignore
View file
Original file line number Diff line number Diff line change
Expand Up @@ -52,16 +52,11 @@ Demo/*.log
Demo/complex_demo_out.m
Demo/complex_demo_out2.m
Demo/import_demo.out
Demo/gauss_demo1.out
Demo/gauss_demo.out
Demo/t1.out
Demo/t2.out

alternative/*.out
alternative/*_out.m
alternative/*_out2.m
alternative/*_demo
alternative/*.so*
alternative/*.dylib*

Test/*.log
Test/errlog.txt
Test/errlog*.txt
Expand Down
261 changes: 110 additions & 151 deletions CMakeLists.txt
View file

Large diffs are not rendered by default.

110 changes: 57 additions & 53 deletions CUDA/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
@@ -1,72 +1,76 @@
#-------------------------------------------------------------------------------
# GraphBLAS/CUDA/CMakeLists.txt: cmake script for GraphBLAS/CUDA
#-------------------------------------------------------------------------------

# SuiteSparse:GraphBLAS, Timothy A. Davis, (c) 2017-2023, All Rights Reserved.

# Some files in this folder are (c) NVIDIA or (c) Google. Please refer
# to their individual licenses (Apache, BSD, or others).
# SPDX-License-Identifier: Apache-2.0

#-------------------------------------------------------------------------------

cmake_minimum_required ( VERSION 3.19 )

# CMake build for generating googletest c++ files that can be compiled and executed in parallel.
# Build can be customized to speed up development by allowing the targeting of specific
# specific parameters. The output of this build is an executable that can be used to
# run the gtests.
# CMake build for generating googletest c++ files that can be compiled and
# executed in parallel. Build can be customized to speed up development by
# allowing the targeting of specific specific parameters. The output of this
# build is an executable that can be used to run the gtests.

project ( GRAPHBLAS_CUDA
VERSION "${GraphBLAS_VERSION_MAJOR}.${GraphBLAS_VERSION_MINOR}.${GraphBLAS_VERSION_SUB}"
LANGUAGES CXX CUDA )

cmake_policy ( SET CMP0135 NEW ) # URL download timestamp policy

set(CMAKE_CUDA_FLAGS "-cudart=static -lineinfo ")
set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} --std=c++17 -fPIC ")
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DGBNCPUFEAT")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DGBNCPUFEAT")
set(CMAKE_C_STANDARD 11)
set ( CMAKE_C_STANDARD 11)
set ( CMAKE_CXX_STANDARD 17 )

message(STATUS "C++ flags for CUDA:" "${CMAKE_CXX_FLAGS}")
set ( CMAKE_CUDA_FLAGS "-cudart=static -lineinfo " )
set ( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} --std=c++17 -fPIC " )

file(GLOB GRAPHBLAS_CUDA_SOURCES "*.cu" "*.c" "*.cpp")
add_compile_definitions ( GBNCPUFEAT )

add_library(graphblascuda SHARED
${GRAPHBLAS_CUDA_SOURCES}
)
message ( STATUS "C++ flags for CUDA: ${CMAKE_CXX_FLAGS}" )

file ( GLOB GRAPHBLAS_CUDA_SOURCES "*.cu" "*.c" "*.cpp" )

add_library ( graphblascuda SHARED ${GRAPHBLAS_CUDA_SOURCES} )

set_target_properties ( graphblascuda PROPERTIES
VERSION ${GraphBLAS_VERSION_MAJOR}.${GraphBLAS_VERSION_MINOR}.${GraphBLAS_VERSION_SUB}
SOVERSION ${GraphBLAS_VERSION_MAJOR}
C_STANDARD_REQUIRED 11 )

set(RMM_WRAP_INCLUDES "../rmm_wrap")
# find rmm_wrap, the malloc/calloc/realloc/free wrapper for the Rapids
# memory manager
set ( RMM_WRAP_INCLUDES "../rmm_wrap" )

message(STATUS "RMM_WRAP_INCLUDES: ${RMM_WRAP_INCLUDES}")
set(GRAPHBLAS_CUDA_INCLUDES
message ( STATUS "RMM_WRAP_INCLUDES: ${RMM_WRAP_INCLUDES}" )
set ( GRAPHBLAS_CUDA_INCLUDES
${RMM_WRAP_INCLUDES}
../Source
../Source/Shared
../Source/Template
../Source/SharedTemplate
../Source/Factories
../Include
../CUDA)

message(STATUS "GraphBLAS CUDA includes: " "${GRAPHBLAS_CUDA_INCLUDES}")

set(EXTERNAL_INCLUDES_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/external_includes)

IF(NOT EXISTS ${EXTERNAL_INCLUDES_DIRECTORY})
file(MAKE_DIRECTORY ${EXTERNAL_INCLUDES_DIRECTORY})
endif()
../CUDA )

IF(NOT EXISTS ${EXTERNAL_INCLUDES_DIRECTORY}/cuco)
execute_process(
COMMAND git clone "https://github.com/NVIDIA/cuCollections.git" --branch main --recursive cuco
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/external_includes)
endif()
message ( STATUS "GraphBLAS CUDA includes: ${GRAPHBLAS_CUDA_INCLUDES}" )

include_directories(${CMAKE_CURRENT_BINARY_DIR}/external_includes/cuco/include)
#-------------------------------------------------------------------------------
# graphblascuda properties
#-------------------------------------------------------------------------------

target_include_directories(graphblascuda PUBLIC
${CMAKE_CURRENT_BINARY_DIR}/external_includes/cuco/include
${CUDAToolkit_INCLUDE_DIRS}
${GRAPHBLAS_CUDA_INCLUDES})
set_target_properties(graphblascuda PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(graphblascuda PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
# FIXME: use SUITESPARSE_CUDA_ARCHITECTURES
set_target_properties(graphblascuda PROPERTIES CUDA_ARCHITECTURES "52;75;80" )

target_link_libraries(graphblascuda CUDA::nvrtc CUDA::cudart_static CUDA::nvToolsExt )
target_link_libraries(graphblascuda CUDA::nvrtc CUDA::cudart_static CUDA::cuda_driver CUDA::nvToolsExt )

#-------------------------------------------------------------------------------
# installation location
Expand All @@ -77,31 +81,29 @@ install ( TARGETS graphblascuda
ARCHIVE DESTINATION ${SUITESPARSE_LIBDIR}
RUNTIME DESTINATION ${SUITESPARSE_BINDIR}
PUBLIC_HEADER DESTINATION ${SUITESPARSE_INCLUDEDIR} )
# if ( BUILD_GRB_STATIC_LIBRARY )
# install ( TARGETS graphblas_static
# ARCHIVE DESTINATION ${SUITESPARSE_LIBDIR} )
# endif ( )

#-------------------------------------------------------------------------------
# test suite for the CUDA kernels
#-------------------------------------------------------------------------------

# 1. Execute enumify/stringify/jitify logic to compile ptx kernels and compile/link w/ relevant *.cu files.
# 1. Execute enumify/stringify/jitify logic to compile ptx kernels and
# compile/link w/ relevant *.cu files.

# TODO: Need to do this piece in cmake

# 2. Generate test .cu files named "{semiring_operation}_test_instances.hpp"
set(CUDA_TEST_SUITES
set ( CUDA_TEST_SUITES
AxB_dot3
# reduce_to_scalar
)

#
set(CUDA_TEST_MONOIDS PLUS MIN MAX) # TIMES ANY)
set(CUDA_TEST_BINOPS TIMES PLUS MIN MAX DIV) #MINUS RDIV RMINUS FIRST SECOND PAIR)
set(CUDA_TEST_SEMIRINGS PLUS_TIMES MIN_PLUS MAX_PLUS)
set(CUDA_TEST_DATATYPES int32_t int64_t uint32_t uint64_t float double)
set(CUDA_TEST_KERNELS vsvs) # mp vsvs dndn spdn vssp)
set(CUDA_TEST_FORMATS sparse dense sparse_dense reduce)

set ( CUDA_TEST_MONOIDS PLUS MIN MAX) # TIMES ANY )
set ( CUDA_TEST_BINOPS TIMES PLUS MIN MAX DIV ) #MINUS RDIV RMINUS FIRST SECOND PAIR )
set ( CUDA_TEST_SEMIRINGS PLUS_TIMES MIN_PLUS MAX_PLUS )
set ( CUDA_TEST_DATATYPES int32_t int64_t uint32_t uint64_t float double )
set ( CUDA_TEST_KERNELS vsvs) # mp vsvs dndn spdn vssp )
set ( CUDA_TEST_FORMATS sparse dense sparse_dense reduce )

# TODO: Update testGen.py to accept the above CUDA_TEST_* params as arguments

Expand All @@ -111,9 +113,9 @@ set(CUDA_TEST_FORMATS sparse dense sparse_dense reduce)
# Separate individual kernels from larger "overview" test (e.g. 2-level testing structure)
# We want to test all the *_cuda versions

# FIXME
# TODO: make this a shorter test
set(CUDA_TEST_CPP_FILES "")
if ( FALSE )
if ( FALSE ) # TODO: use a cmake option
foreach(var ${CUDA_TEST_SUITES})
foreach(semiring ${CUDA_TEST_SEMIRINGS})
foreach(kernel ${CUDA_TEST_KERNELS})
Expand All @@ -125,8 +127,8 @@ if ( FALSE )
OUTPUT
${CMAKE_CURRENT_BINARY_DIR}/${var}_${semiring}_${format}_test_instances.hpp
${CMAKE_CURRENT_BINARY_DIR}/${var}_${semiring}_${format}_cuda_tests.cpp
DEPENDS
jitFactory.hpp
# DEPENDS
# jitFactory.hpp
COMMAND python3 ${CMAKE_CURRENT_SOURCE_DIR}/test/testGen_cmake.py "\"${CMAKE_CURRENT_SOURCE_DIR}\"" "\"${var}\"" "\"${CUDA_TEST_MONOIDS}\""
"\"${CUDA_TEST_BINOPS}\"" "\"${semiring}\"" "\"${CUDA_TEST_DATATYPES}\""
"\"${kernel}\""
Expand Down Expand Up @@ -187,6 +189,7 @@ target_link_libraries(graphblascuda_test
PUBLIC
graphblas
graphblascuda
rmm_wrap
CUDA::cudart_static
CUDA::nvrtc
${ADDITIONAL_DEPS}
Expand All @@ -195,6 +198,7 @@ target_link_libraries(graphblascuda_test

target_include_directories(graphblascuda_test
PUBLIC
rmm_wrap
${ADDITIONAL_INCLUDES}
${CUDAToolkit_INCLUDE_DIRS}
${GRAPHBLAS_CUDA_INCLUDES})
Expand Down
72 changes: 72 additions & 0 deletions CUDA/Config/GB_cuda_common_jitFactory.hpp.in
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,72 @@
//------------------------------------------------------------------------------
// GB_cuda_common_jitFactory.hpp: common defines for all jitFactory classes
//------------------------------------------------------------------------------

// (c) Nvidia Corp. 2020 All rights reserved
// SPDX-License-Identifier: Apache-2.0

//------------------------------------------------------------------------------

// Common defines for all jitFactory classes:
// iostream callback to deliver the buffer to jitify as if read from a file
// compiler flags
// Include this file along with any jitFactory you need.

// NOTE: do not edit the GB_cuda_common_jitFactory.hpp directly. It is
// configured by cmake from the following file:
// GraphBLAS/CUDA/Config/GB_cuda_common_jitFactory.hpp.in

#ifndef GB_COMMON_JITFACTORY_H
#define GB_COMMON_JITFACTORY_H

#pragma once

extern "C"
{
#include "GB.h"
#include "GraphBLAS.h"
#include "GB_stringify.h"
}

#include <iostream>
#include <cstdint>
#include "GB_jit_cache.h"
#include "GB_jit_launcher.h"
#include "GB_cuda_mxm_factory.hpp"
#include "GB_cuda_buckets.h"
#include "GB_cuda_type_wrap.hpp"
#include "GB_cuda_error.h"
#include "../rmm_wrap/rmm_wrap.h"
#include "GB_iceil.h"

// amount of shared memory to use in CUDA kernel launches
constexpr unsigned int SMEM = 0 ;

static const std::vector<std::string> GB_jit_cuda_compiler_flags{
"-std=c++17",
//"-G",
"-remove-unused-globals",
"-w",
"-D__CUDACC_RTC__",
// "-I.",
// "-I..",
// "-I../templates",
// "-I../CUDA",
// "-I../Source/Shared",

// Add includes relative to GRAPHBLAS_SOURCE_PATH variable
"-I" + jit::get_user_graphblas_source_path() + "/CUDA",
"-I" + jit::get_user_graphblas_source_path() + "/Source/Shared",
"-I" + jit::get_user_graphblas_source_path() + "/CUDA/templates",
"-I/usr/local/cuda/include",

// FIXME: add SUITESPARSE_CUDA_ARCHITECTURES here, via config
};

// FIXME: rename GB_jit_cuda_header_names or something
static const std::vector<std::string> header_names ={};

// FIXME: rename GB_jit_cuda_file_callback
inline std::istream* (*file_callback)(std::string, std::iostream&);

#endif
Loading

0 comments on commit 3138f5a

Please sign in to comment.