Merged in benoitsteiner/opencl (pull request PR-246) Improved support for OpenCL
diff --git a/Eigen/Core b/Eigen/Core index 985fea7..8255815 100644 --- a/Eigen/Core +++ b/Eigen/Core
@@ -14,16 +14,6 @@ // first thing Eigen does: stop the compiler from committing suicide #include "src/Core/util/DisableStupidWarnings.h" -/// This will no longer be needed after the next release of the computecppCE -#ifdef EIGEN_USE_SYCL -#undef min -#undef max -#undef isnan -#undef isinf -#undef isfinite -#include <SYCL/sycl.hpp> -#endif - // Handle NVCC/CUDA/SYCL #if defined(__CUDACC__) || defined(__SYCL_DEVICE_ONLY__) // Do not try asserts on CUDA and SYCL!
diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake index 9ebd06a..a92a297 100644 --- a/cmake/EigenTesting.cmake +++ b/cmake/EigenTesting.cmake
@@ -1,23 +1,23 @@ macro(ei_add_property prop value) - get_property(previous GLOBAL PROPERTY ${prop}) + get_property(previous GLOBAL PROPERTY ${prop}) if ((NOT previous) OR (previous STREQUAL "")) set_property(GLOBAL PROPERTY ${prop} "${value}") else() set_property(GLOBAL PROPERTY ${prop} "${previous} ${value}") - endif() + endif() endmacro(ei_add_property) #internal. See documentation of ei_add_test for details. macro(ei_add_test_internal testname testname_with_suffix) set(targetname ${testname_with_suffix}) - + if(EIGEN_ADD_TEST_FILENAME_EXTENSION) set(filename ${testname}.${EIGEN_ADD_TEST_FILENAME_EXTENSION}) else() set(filename ${testname}.cpp) endif() - + if(EIGEN_ADD_TEST_FILENAME_EXTENSION STREQUAL cu) if(EIGEN_TEST_CUDA_CLANG) set_source_files_properties(${filename} PROPERTIES LANGUAGE CXX) @@ -42,7 +42,7 @@ else() add_executable(${targetname} ${filename}) endif() - + if (targetname MATCHES "^eigen2_") add_dependencies(eigen2_buildtests ${targetname}) else() @@ -56,20 +56,20 @@ ei_add_target_property(${targetname} COMPILE_FLAGS "-DEIGEN_DEBUG_ASSERTS=1") endif(EIGEN_DEBUG_ASSERTS) endif(EIGEN_NO_ASSERTION_CHECKING) - + ei_add_target_property(${targetname} COMPILE_FLAGS "-DEIGEN_TEST_MAX_SIZE=${EIGEN_TEST_MAX_SIZE}") ei_add_target_property(${targetname} COMPILE_FLAGS "-DEIGEN_TEST_FUNC=${testname}") - + if(MSVC) ei_add_target_property(${targetname} COMPILE_FLAGS "/bigobj") - endif() + endif() # let the user pass flags. if(${ARGC} GREATER 2) ei_add_target_property(${targetname} COMPILE_FLAGS "${ARGV2}") endif(${ARGC} GREATER 2) - + if(EIGEN_TEST_CUSTOM_CXX_FLAGS) ei_add_target_property(${targetname} COMPILE_FLAGS "${EIGEN_TEST_CUSTOM_CXX_FLAGS}") endif() @@ -95,12 +95,12 @@ # notice: no double quotes around ${libs_to_link} here. It may be a list. target_link_libraries(${targetname} ${libs_to_link}) endif() - endif() + endif() add_test(${testname_with_suffix} "${targetname}") - + # Specify target and test labels accoirding to EIGEN_CURRENT_SUBPROJECT - get_property(current_subproject GLOBAL PROPERTY EIGEN_CURRENT_SUBPROJECT) + get_property(current_subproject GLOBAL PROPERTY EIGEN_CURRENT_SUBPROJECT) if ((current_subproject) AND (NOT (current_subproject STREQUAL ""))) set_property(TARGET ${targetname} PROPERTY LABELS "Build${current_subproject}") add_dependencies("Build${current_subproject}" ${targetname}) @@ -128,14 +128,14 @@ OUTPUT ${include_file} COMMAND ${CMAKE_COMMAND} -E echo "\\#include \\\"${host_file}\\\"" > ${include_file} COMMAND ${CMAKE_COMMAND} -E echo "\\#include \\\"${bc_file}.sycl\\\"" >> ${include_file} - DEPENDS ${filename} + DEPENDS ${filename} ${bc_file}.sycl COMMENT "Building ComputeCpp integration header file ${include_file}" ) # Add a custom target for the generated integration header - add_custom_target(${testname}_integration_header_woho DEPENDS ${include_file}) + add_custom_target(${testname}_integration_header_sycl DEPENDS ${include_file}) add_executable(${targetname} ${include_file}) - add_dependencies(${targetname} ${testname}_integration_header_woho) + add_dependencies(${targetname} ${testname}_integration_header_sycl) add_sycl_to_target(${targetname} ${filename} ${CMAKE_CURRENT_BINARY_DIR}) if (targetname MATCHES "^eigen2_") @@ -258,7 +258,7 @@ else() set(filename ${testname}.cpp) endif() - + file(READ "${filename}" test_source) set(parts 0) string(REGEX MATCHALL "CALL_SUBTEST_[0-9]+|EIGEN_TEST_PART_[0-9]+|EIGEN_SUFFIXES(;[0-9]+)+" @@ -379,7 +379,7 @@ elseif(EIGEN_TEST_NO_EXPLICIT_VECTORIZATION) message(STATUS "Explicit vectorization disabled (alignment kept enabled)") else() - + message(STATUS "Maximal matrix/vector size: ${EIGEN_TEST_MAX_SIZE}") if(EIGEN_TEST_SSE2) @@ -453,13 +453,13 @@ else() message(STATUS "ARMv8 NEON: Using architecture defaults") endif() - + if(EIGEN_TEST_ZVECTOR) message(STATUS "S390X ZVECTOR: ON") else() message(STATUS "S390X ZVECTOR: Using architecture defaults") endif() - + if(EIGEN_TEST_CXX11) message(STATUS "C++11: ON") else() @@ -505,7 +505,7 @@ set_property(GLOBAL PROPERTY EIGEN_FAILTEST_FAILURE_COUNT "0") set_property(GLOBAL PROPERTY EIGEN_FAILTEST_COUNT "0") - + # uncomment anytime you change the ei_get_compilerver_from_cxx_version_string macro # ei_test_get_compilerver_from_cxx_version_string() endmacro(ei_init_testing) @@ -514,22 +514,22 @@ # if the sitename is not yet set, try to set it if(NOT ${SITE} OR ${SITE} STREQUAL "") set(eigen_computername $ENV{COMPUTERNAME}) - set(eigen_hostname $ENV{HOSTNAME}) + set(eigen_hostname $ENV{HOSTNAME}) if(eigen_hostname) set(SITE ${eigen_hostname}) - elseif(eigen_computername) - set(SITE ${eigen_computername}) + elseif(eigen_computername) + set(SITE ${eigen_computername}) endif() endif() # in case it is already set, enforce lower case if(SITE) string(TOLOWER ${SITE} SITE) - endif() + endif() endmacro(ei_set_sitename) macro(ei_get_compilerver VAR) if(MSVC) - # on windows system, we use a modified CMake script + # on windows system, we use a modified CMake script include(EigenDetermineVSServicePack) EigenDetermineVSServicePack( my_service_pack ) @@ -541,20 +541,20 @@ else() # on all other system we rely on ${CMAKE_CXX_COMPILER} # supporting a "--version" or "/version" flag - + if(WIN32 AND ${CMAKE_CXX_COMPILER_ID} EQUAL "Intel") set(EIGEN_CXX_FLAG_VERSION "/version") else() set(EIGEN_CXX_FLAG_VERSION "--version") endif() - + execute_process(COMMAND ${CMAKE_CXX_COMPILER} ${EIGEN_CXX_FLAG_VERSION} OUTPUT_VARIABLE eigen_cxx_compiler_version_string OUTPUT_STRIP_TRAILING_WHITESPACE) string(REGEX REPLACE "[\n\r].*" "" eigen_cxx_compiler_version_string ${eigen_cxx_compiler_version_string}) - + ei_get_compilerver_from_cxx_version_string("${eigen_cxx_compiler_version_string}" CNAME CVER) set(${VAR} "${CNAME}-${CVER}") - + endif() endmacro(ei_get_compilerver) @@ -563,13 +563,13 @@ # the testing macro call in ei_init_testing() of the EigenTesting.cmake file. # See also the ei_test_get_compilerver_from_cxx_version_string macro at the end of the file macro(ei_get_compilerver_from_cxx_version_string VERSTRING CNAME CVER) - # extract possible compiler names + # extract possible compiler names string(REGEX MATCH "g\\+\\+" ei_has_gpp ${VERSTRING}) string(REGEX MATCH "llvm|LLVM" ei_has_llvm ${VERSTRING}) string(REGEX MATCH "gcc|GCC" ei_has_gcc ${VERSTRING}) string(REGEX MATCH "icpc|ICC" ei_has_icpc ${VERSTRING}) string(REGEX MATCH "clang|CLANG" ei_has_clang ${VERSTRING}) - + # combine them if((ei_has_llvm) AND (ei_has_gpp OR ei_has_gcc)) set(${CNAME} "llvm-g++") @@ -584,7 +584,7 @@ else() set(${CNAME} "_") endif() - + # extract possible version numbers # first try to extract 3 isolated numbers: string(REGEX MATCH " [0-9]+\\.[0-9]+\\.[0-9]+" eicver ${VERSTRING}) @@ -602,9 +602,9 @@ endif() endif() endif() - + string(REGEX REPLACE ".(.*)" "\\1" ${CVER} ${eicver}) - + endmacro(ei_get_compilerver_from_cxx_version_string) macro(ei_get_cxxflags VAR) @@ -633,30 +633,30 @@ elseif(EIGEN_TEST_SSE3) set(${VAR} SSE3) elseif(EIGEN_TEST_SSE2 OR IS_64BIT_ENV) - set(${VAR} SSE2) + set(${VAR} SSE2) endif() if(EIGEN_TEST_OPENMP) if (${VAR} STREQUAL "") - set(${VAR} OMP) - else() - set(${VAR} ${${VAR}}-OMP) - endif() + set(${VAR} OMP) + else() + set(${VAR} ${${VAR}}-OMP) + endif() endif() - + if(EIGEN_DEFAULT_TO_ROW_MAJOR) if (${VAR} STREQUAL "") - set(${VAR} ROW) - else() - set(${VAR} ${${VAR}}-ROWMAJ) - endif() + set(${VAR} ROW) + else() + set(${VAR} ${${VAR}}-ROWMAJ) + endif() endif() endmacro(ei_get_cxxflags) macro(ei_set_build_string) ei_get_compilerver(LOCAL_COMPILER_VERSION) ei_get_cxxflags(LOCAL_COMPILER_FLAGS) - + include(EigenDetermineOSVersion) DetermineOSVersion(OS_VERSION) @@ -672,11 +672,11 @@ else() set(TMP_BUILD_STRING ${TMP_BUILD_STRING}-64bit) endif() - + if(EIGEN_TEST_CXX11) set(TMP_BUILD_STRING ${TMP_BUILD_STRING}-cxx11) endif() - + if(EIGEN_BUILD_STRING_SUFFIX) set(TMP_BUILD_STRING ${TMP_BUILD_STRING}-${EIGEN_BUILD_STRING_SUFFIX}) endif()
diff --git a/cmake/FindComputeCpp.cmake b/cmake/FindComputeCpp.cmake index 3aab5b8..07ebed6 100644 --- a/cmake/FindComputeCpp.cmake +++ b/cmake/FindComputeCpp.cmake
@@ -1,6 +1,21 @@ #.rst: # FindComputeCpp #--------------- +# +# Copyright 2016 Codeplay Software Ltd. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use these files except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. ######################### # FindComputeCpp.cmake @@ -8,6 +23,11 @@ # # Tools for finding and building with ComputeCpp. # +# User must define COMPUTECPP_PACKAGE_ROOT_DIR pointing to the ComputeCpp +# installation. +# +# Latest version of this file can be found at: +# https://github.com/codeplaysoftware/computecpp-sdk # Require CMake version 3.2.2 or higher cmake_minimum_required(VERSION 3.2.2) @@ -32,7 +52,6 @@ message(FATAL_ERROR "host compiler - Not found! (clang version must be at least 3.6)") else() - set(COMPUTECPP_DISABLE_GCC_DUAL_ABI "True") message(STATUS "host compiler - clang ${CMAKE_CXX_COMPILER_VERSION}") endif() else() @@ -48,11 +67,12 @@ # Find OpenCL package find_package(OpenCL REQUIRED) -# Find ComputeCpp package -if(EXISTS ${COMPUTECPP_PACKAGE_ROOT_DIR}) - message(STATUS "ComputeCpp package - Found (${COMPUTECPP_PACKAGE_ROOT_DIR})") +# Find ComputeCpp packagee +if(NOT COMPUTECPP_PACKAGE_ROOT_DIR) + message(FATAL_ERROR + "ComputeCpp package - Not found! (please set COMPUTECPP_PACKAGE_ROOT_DIR") else() - message(FATAL_ERROR "ComputeCpp package - Not found! (please set COMPUTECPP_PACKAGE_ROOT_DIR) (${COMPUTECPP_PACKAGE_ROOT_DIR})") + message(STATUS "ComputeCpp package - Found") endif() option(COMPUTECPP_PACKAGE_ROOT_DIR "Path to the ComputeCpp Package") @@ -61,9 +81,9 @@ ${COMPUTECPP_PACKAGE_ROOT_DIR} PATH_SUFFIXES bin) if (EXISTS ${COMPUTECPP_DEVICE_COMPILER}) mark_as_advanced(COMPUTECPP_DEVICE_COMPILER) - message(STATUS "compute++ - Found (${COMPUTECPP_PACKAGE_ROOT_DIR})") + message(STATUS "compute++ - Found") else() - message(FATAL_ERROR "compute++ - Not found! (${COMPUTECPP_DEVICE_COMPILER}) (${COMPUTECPP_PACKAGE_ROOT_DIR})") + message(FATAL_ERROR "compute++ - Not found! (${COMPUTECPP_DEVICE_COMPILER})") endif() # Obtain the path to computecpp_info @@ -71,9 +91,9 @@ ${COMPUTECPP_PACKAGE_ROOT_DIR} PATH_SUFFIXES bin) if (EXISTS ${COMPUTECPP_INFO_TOOL}) mark_as_advanced(${COMPUTECPP_INFO_TOOL}) - message(STATUS "computecpp_info - Found (${COMPUTECPP_PACKAGE_ROOT_DIR})") + message(STATUS "computecpp_info - Found") else() - message(FATAL_ERROR "computecpp_info - Not found! (${COMPUTECPP_INFO_TOOL}) (${COMPUTECPP_PACKAGE_ROOT_DIR})") + message(FATAL_ERROR "computecpp_info - Not found! (${COMPUTECPP_INFO_TOOL})") endif() # Obtain the path to the ComputeCpp runtime library @@ -85,15 +105,15 @@ mark_as_advanced(COMPUTECPP_RUNTIME_LIBRARY) message(STATUS "libComputeCpp.so - Found") else() - message(FATAL_ERROR "libComputeCpp.so - Not found! (${COMPUTECPP_PACKAGE_ROOT_DIR})") + message(FATAL_ERROR "libComputeCpp.so - Not found!") endif() # Obtain the ComputeCpp include directory set(COMPUTECPP_INCLUDE_DIRECTORY ${COMPUTECPP_PACKAGE_ROOT_DIR}/include/) if (NOT EXISTS ${COMPUTECPP_INCLUDE_DIRECTORY}) - message(FATAL_ERROR "ComputeCpp includes - Not found! (${COMPUTECPP_PACKAGE_ROOT_DIR}/include/)") + message(FATAL_ERROR "ComputeCpp includes - Not found!") else() - message(STATUS "ComputeCpp includes - Found (${COMPUTECPP_PACKAGE_ROOT_DIR})") + message(STATUS "ComputeCpp includes - Found") endif() # Obtain the package version @@ -144,7 +164,7 @@ # # targetName : Name of the target. # sourceFile : Source file to be compiled. -# binaryDir : Intermediate output directory for the integration header. +# binaryDir : Intermediate directory to output the integration header. # function(__build_spir targetName sourceFile binaryDir) @@ -176,12 +196,13 @@ OUTPUT ${outputSyclFile} COMMAND ${COMPUTECPP_DEVICE_COMPILER} ${COMPUTECPP_DEVICE_COMPILER_FLAGS} - -I${COMPUTECPP_INCLUDE_DIRECTORY} + -isystem ${COMPUTECPP_INCLUDE_DIRECTORY} ${COMPUTECPP_PLATFORM_SPECIFIC_ARGS} ${device_compiler_includes} -o ${outputSyclFile} -c ${CMAKE_CURRENT_SOURCE_DIR}/${sourceFile} DEPENDS ${sourceFile} + WORKING_DIRECTORY ${binaryDir} COMMENT "Building ComputeCpp integration header file ${outputSyclFile}") # Add a custom target for the generated integration header @@ -190,10 +211,6 @@ # Add a dependency on the integration header add_dependencies(${targetName} ${targetName}_integration_header) - # Force inclusion of the integration header for the host compiler - #set(compileFlags -include ${include_file} "-Wall") - target_compile_options(${targetName} PUBLIC ${compileFlags}) - # Set the host compiler C++ standard to C++11 set_property(TARGET ${targetName} PROPERTY CXX_STANDARD 11) @@ -210,11 +227,11 @@ ####################### # # Adds a SYCL compilation custom command associated with an existing -# target and sets a dependency on that new command. +# target and sets a dependancy on that new command. # # targetName : Name of the target to add a SYCL to. # sourceFile : Source file to be compiled for SYCL. -# binaryDir : Intermediate output directory for the integration header. +# binaryDir : Intermediate directory to output the integration header. # function(add_sycl_to_target targetName sourceFile binaryDir)
diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 388976d..7ecb4c7 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor
@@ -13,6 +13,18 @@ #include "../../../Eigen/Core" +#ifdef EIGEN_USE_SYCL +#undef min +#undef max +#undef isnan +#undef isinf +#undef isfinite +#include <SYCL/sycl.hpp> +#include <map> +#include <memory> +#include <utility> +#endif + #include <Eigen/src/Core/util/DisableStupidWarnings.h> #include "../SpecialFunctions" @@ -69,10 +81,6 @@ #endif #endif -#ifdef EIGEN_USE_SYCL -#include <SYCL/sycl.hpp> -#endif - #include "src/Tensor/TensorMacros.h" #include "src/Tensor/TensorForwardDeclarations.h" #include "src/Tensor/TensorMeta.h" @@ -81,7 +89,6 @@ #include "src/Tensor/TensorDeviceDefault.h" #include "src/Tensor/TensorDeviceThreadPool.h" #include "src/Tensor/TensorDeviceCuda.h" -#include "src/Tensor/TensorSycl.h" #include "src/Tensor/TensorDeviceSycl.h" #include "src/Tensor/TensorIndexList.h" #include "src/Tensor/TensorDimensionList.h" @@ -128,6 +135,7 @@ #include "src/Tensor/TensorAssign.h" #include "src/Tensor/TensorScan.h" +#include "src/Tensor/TensorSycl.h" #include "src/Tensor/TensorExecutor.h" #include "src/Tensor/TensorDevice.h"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h index bfd36f5..7c03989 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -1,12 +1,11 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com> // Mehdi Goli Codeplay Software Ltd. // Ralph Potter Codeplay Software Ltd. // Luke Iwanski Codeplay Software Ltd. -// Cummins Chris PhD student at The University of Edinburgh. // Contact: <eigen@codeplay.com> +// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com> // // This Source Code Form is subject to the terms of the Mozilla @@ -17,106 +16,107 @@ #define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H namespace Eigen { -/// \struct BufferT is used to specialise add_sycl_buffer function for -// two types of buffer we have. When the MapAllocator is true, we create the -// sycl buffer with MapAllocator. -/// We have to const_cast the input pointer in order to work around the fact -/// that sycl does not accept map allocator for const pointer. -template <typename T, bool MapAllocator> -struct BufferT { - using Type = cl::sycl::buffer<T, 1, cl::sycl::map_allocator<T>>; - static inline void add_sycl_buffer( - const T *ptr, size_t num_bytes, - std::map<const void *, std::shared_ptr<void>> &buffer_map) { - buffer_map.insert(std::pair<const void *, std::shared_ptr<void>>( - ptr, std::shared_ptr<void>(std::make_shared<Type>( - Type(const_cast<T *>(ptr), cl::sycl::range<1>(num_bytes)))))); - } -}; - -/// specialisation of the \ref BufferT when the MapAllocator is false. In this -/// case we only create the device-only buffer. -template <typename T> -struct BufferT<T, false> { - using Type = cl::sycl::buffer<T, 1>; - static inline void add_sycl_buffer( - const T *ptr, size_t num_bytes, - std::map<const void *, std::shared_ptr<void>> &buffer_map) { - buffer_map.insert(std::pair<const void *, std::shared_ptr<void>>( - ptr, std::shared_ptr<void>( - std::make_shared<Type>(Type(cl::sycl::range<1>(num_bytes)))))); - } -}; - struct SyclDevice { /// class members /// sycl queue - cl::sycl::queue &m_queue; + mutable cl::sycl::queue m_queue; /// std::map is the container used to make sure that we create only one buffer - /// per pointer. The lifespan of the buffer - /// now depends on the lifespan of SyclDevice. If a non-read-only pointer is - /// needed to be accessed on the host we should manually deallocate it. + /// per pointer. The lifespan of the buffer now depends on the lifespan of SyclDevice. + /// If a non-read-only pointer is needed to be accessed on the host we should manually deallocate it. mutable std::map<const void *, std::shared_ptr<void>> buffer_map; - - SyclDevice(cl::sycl::queue &q) : m_queue(q) {} + /// creating device by using selector + template<typename dev_Selector> SyclDevice(dev_Selector s) + : +#ifdef EIGEN_EXCEPTIONS + m_queue(cl::sycl::queue(s, [=](cl::sycl::exception_list l) { + for (const auto& e : l) { + try { + std::rethrow_exception(e); + } catch (cl::sycl::exception e) { + std::cout << e.what() << std::endl; + } + } + })) +#else + m_queue(cl::sycl::queue(s)) +#endif + {} // destructor ~SyclDevice() { deallocate_all(); } - template <typename T> - void deallocate(const T *p) const { + template <typename T> void deallocate(T *p) const { auto it = buffer_map.find(p); if (it != buffer_map.end()) { buffer_map.erase(it); + internal::aligned_free(p); } } - void deallocate_all() const { buffer_map.clear(); } + void deallocate_all() const { + std::map<const void *, std::shared_ptr<void>>::iterator it=buffer_map.begin(); + while (it!=buffer_map.end()) { + auto p=it->first; + buffer_map.erase(it); + internal::aligned_free(const_cast<void*>(p)); + it=buffer_map.begin(); + } + buffer_map.clear(); + } /// creation of sycl accessor for a buffer. This function first tries to find - /// the buffer in the buffer_map. - /// If found it gets the accessor from it, if not, the function then adds an - /// entry by creating a sycl buffer - /// for that particular pointer. - template <cl::sycl::access::mode AcMd, bool MapAllocator, typename T> - inline cl::sycl::accessor<T, 1, AcMd, cl::sycl::access::target::global_buffer> - get_sycl_accessor(size_t num_bytes, cl::sycl::handler &cgh, - const T *ptr) const { - auto it = buffer_map.find(ptr); - if (it == buffer_map.end()) { - BufferT<T, MapAllocator>::add_sycl_buffer(ptr, num_bytes, buffer_map); - } - return ( - ((typename BufferT<T, MapAllocator>::Type *)(buffer_map.at(ptr).get())) - ->template get_access<AcMd>(cgh)); + /// the buffer in the buffer_map. If found it gets the accessor from it, if not, + ///the function then adds an entry by creating a sycl buffer for that particular pointer. + template <cl::sycl::access::mode AcMd, typename T> inline cl::sycl::accessor<T, 1, AcMd, cl::sycl::access::target::global_buffer> + get_sycl_accessor(size_t num_bytes, cl::sycl::handler &cgh, const T * ptr) const { + return (get_sycl_buffer<T>(num_bytes, ptr)->template get_access<AcMd, cl::sycl::access::target::global_buffer>(cgh)); + } + + template<typename T> inline std::pair<std::map<const void *, std::shared_ptr<void>>::iterator,bool> add_sycl_buffer(const T *ptr, size_t num_bytes) const { + using Type = cl::sycl::buffer<T, 1>; + std::pair<std::map<const void *, std::shared_ptr<void>>::iterator,bool> ret = buffer_map.insert(std::pair<const void *, std::shared_ptr<void>>(ptr, std::shared_ptr<void>(new Type(cl::sycl::range<1>(num_bytes)), + [](void *dataMem) { delete static_cast<Type*>(dataMem); }))); + (static_cast<Type*>(buffer_map.at(ptr).get()))->set_final_data(nullptr); + return ret; + } + + template <typename T> inline cl::sycl::buffer<T, 1>* get_sycl_buffer(size_t num_bytes,const T * ptr) const { + return static_cast<cl::sycl::buffer<T, 1>*>(add_sycl_buffer(ptr, num_bytes).first->second.get()); } /// allocating memory on the cpu - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void *allocate(size_t num_bytes) const { - return internal::aligned_malloc(num_bytes); + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void *allocate(size_t) const { + return internal::aligned_malloc(8); } // some runtime conditions that can be applied here bool isDeviceSuitable() const { return true; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate(void *buffer) const { - internal::aligned_free(buffer); - } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void *dst, const void *src, - size_t n) const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void *dst, const void *src, size_t n) const { ::memcpy(dst, src, n); } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyHostToDevice( - void *dst, const void *src, size_t n) const { - memcpy(dst, src, n); + + template<typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyHostToDevice(T *dst, const T *src, size_t n) const { + auto host_acc= (static_cast<cl::sycl::buffer<T, 1>*>(add_sycl_buffer(dst, n).first->second.get()))-> template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::host_buffer>(); + memcpy(host_acc.get_pointer(), src, n); } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyDeviceToHost( - void *dst, const void *src, size_t n) const { - memcpy(dst, src, n); + /// whith the current implementation of sycl, the data is copied twice from device to host. This will be fixed soon. + template<typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyDeviceToHost(T *dst, const T *src, size_t n) const { + auto it = buffer_map.find(src); + if (it != buffer_map.end()) { + auto host_acc= (static_cast<cl::sycl::buffer<T, 1>*>(it->second.get()))-> template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::host_buffer>(); + memcpy(dst,host_acc.get_pointer(), n); + } else{ + eigen_assert("no device memory found. The memory might be destroyed before creation"); + } } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void *buffer, int c, - size_t n) const { + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void *buffer, int c, size_t n) const { ::memset(buffer, c, n); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int majorDeviceVersion() const { + return 1; + } }; + } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h index 68d14a7..0698713 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
@@ -47,13 +47,13 @@ template<typename XprType, template <class> class MakePointer_> struct eval<TensorEvalToOp<XprType, MakePointer_>, Eigen::Dense> { - typedef const TensorEvalToOp<XprType>& type; + typedef const TensorEvalToOp<XprType, MakePointer_>& type; }; template<typename XprType, template <class> class MakePointer_> struct nested<TensorEvalToOp<XprType, MakePointer_>, 1, typename eval<TensorEvalToOp<XprType, MakePointer_> >::type> { - typedef TensorEvalToOp<XprType> type; + typedef TensorEvalToOp<XprType, MakePointer_> type; }; } // end namespace internal
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index 6497b18..52b803d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -33,7 +33,7 @@ template<typename BinaryOp, typename LeftXprType, typename RightXprType> class TensorCwiseBinaryOp; template<typename TernaryOp, typename Arg1XprType, typename Arg2XprType, typename Arg3XprType> class TensorCwiseTernaryOp; template<typename IfXprType, typename ThenXprType, typename ElseXprType> class TensorSelectOp; -template<typename Op, typename Dims, typename XprType> class TensorReductionOp; +template<typename Op, typename Dims, typename XprType, template <class> class MakePointer_ = MakePointer > class TensorReductionOp; template<typename XprType> class TensorIndexTupleOp; template<typename ReduceOp, typename Dims, typename XprType> class TensorTupleReducerOp; template<typename Axis, typename LeftXprType, typename RightXprType> class TensorConcatenationOp;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h index d34ff98..41d0d00 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -2,6 +2,7 @@ // for linear algebra. // // Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com> +// Copyright (C) 2016 Mehdi Goli, Codeplay Software Ltd <eigen@codeplay.com> // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -20,8 +21,8 @@ */ namespace internal { -template<typename Op, typename Dims, typename XprType> -struct traits<TensorReductionOp<Op, Dims, XprType> > + template<typename Op, typename Dims, typename XprType,template <class> class MakePointer_ > + struct traits<TensorReductionOp<Op, Dims, XprType, MakePointer_> > : traits<XprType> { typedef traits<XprType> XprTraits; @@ -31,18 +32,24 @@ typedef typename XprType::Nested Nested; static const int NumDimensions = XprTraits::NumDimensions - array_size<Dims>::value; static const int Layout = XprTraits::Layout; + + template <class T> struct MakePointer { + // Intermediate typedef to workaround MSVC issue. + typedef MakePointer_<T> MakePointerT; + typedef typename MakePointerT::Type Type; + }; }; -template<typename Op, typename Dims, typename XprType> -struct eval<TensorReductionOp<Op, Dims, XprType>, Eigen::Dense> +template<typename Op, typename Dims, typename XprType, template <class> class MakePointer_> +struct eval<TensorReductionOp<Op, Dims, XprType, MakePointer_>, Eigen::Dense> { - typedef const TensorReductionOp<Op, Dims, XprType>& type; + typedef const TensorReductionOp<Op, Dims, XprType, MakePointer_>& type; }; -template<typename Op, typename Dims, typename XprType> -struct nested<TensorReductionOp<Op, Dims, XprType>, 1, typename eval<TensorReductionOp<Op, Dims, XprType> >::type> +template<typename Op, typename Dims, typename XprType, template <class> class MakePointer_> +struct nested<TensorReductionOp<Op, Dims, XprType, MakePointer_>, 1, typename eval<TensorReductionOp<Op, Dims, XprType, MakePointer_> >::type> { - typedef TensorReductionOp<Op, Dims, XprType> type; + typedef TensorReductionOp<Op, Dims, XprType, MakePointer_> type; }; @@ -339,8 +346,8 @@ } // end namespace internal -template <typename Op, typename Dims, typename XprType> -class TensorReductionOp : public TensorBase<TensorReductionOp<Op, Dims, XprType>, ReadOnlyAccessors> { +template <typename Op, typename Dims, typename XprType, template <class> class MakePointer_> +class TensorReductionOp : public TensorBase<TensorReductionOp<Op, Dims, XprType, MakePointer_>, ReadOnlyAccessors> { public: typedef typename Eigen::internal::traits<TensorReductionOp>::Scalar Scalar; typedef typename Eigen::NumTraits<Scalar>::Real RealScalar; @@ -371,18 +378,19 @@ // Eval as rvalue -template<typename Op, typename Dims, typename ArgType, typename Device> -struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device> +template<typename Op, typename Dims, typename ArgType, template <class> class MakePointer_, typename Device> +struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device> { - typedef TensorReductionOp<Op, Dims, ArgType> XprType; + typedef TensorReductionOp<Op, Dims, ArgType, MakePointer_> XprType; typedef typename XprType::Index Index; + typedef ArgType ChildType; typedef typename TensorEvaluator<ArgType, Device>::Dimensions InputDimensions; static const int NumInputDims = internal::array_size<InputDimensions>::value; static const int NumReducedDims = internal::array_size<Dims>::value; static const int NumOutputDims = NumInputDims - NumReducedDims; typedef typename internal::conditional<NumOutputDims==0, Sizes<>, DSizes<Index, NumOutputDims> >::type Dimensions; typedef typename XprType::Scalar Scalar; - typedef TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device> Self; + typedef TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device> Self; static const bool InputPacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess; typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType; typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType; @@ -401,7 +409,7 @@ static const bool RunningFullReduction = (NumOutputDims==0); EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) - : m_impl(op.expression(), device), m_reducer(op.reducer()), m_result(NULL), m_device(device) + : m_impl(op.expression(), device), m_reducer(op.reducer()), m_result(NULL), m_device(device), m_xpr_dims(op.dims()) { EIGEN_STATIC_ASSERT((NumInputDims >= NumReducedDims), YOU_MADE_A_PROGRAMMING_MISTAKE); EIGEN_STATIC_ASSERT((!ReducingInnerMostDims | !PreservingInnerMostDims | (NumReducedDims == NumInputDims)), @@ -471,25 +479,35 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool evalSubExprsIfNeeded(CoeffReturnType* data) { + EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool evalSubExprsIfNeeded(typename MakePointer_<CoeffReturnType>::Type data) { m_impl.evalSubExprsIfNeeded(NULL); // Use the FullReducer if possible. - if (RunningFullReduction && + if ((RunningFullReduction && RunningOnSycl) ||(RunningFullReduction && internal::FullReducer<Self, Op, Device>::HasOptimizedImplementation && ((RunningOnGPU && (m_device.majorDeviceVersion() >= 3)) || - !RunningOnGPU)) { + !RunningOnGPU))) { bool need_assign = false; if (!data) { m_result = static_cast<CoeffReturnType*>(m_device.allocate(sizeof(CoeffReturnType))); data = m_result; need_assign = true; } - Op reducer(m_reducer); internal::FullReducer<Self, Op, Device>::run(*this, reducer, m_device, data); return need_assign; } + else if(RunningOnSycl){ + const Index num_values_to_reduce = internal::array_prod(m_reducedDims); + const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions); + if (!data) { + data = static_cast<CoeffReturnType*>(m_device.allocate(sizeof(CoeffReturnType) * num_coeffs_to_preserve)); + m_result = data; + } + Op reducer(m_reducer); + internal::InnerReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve); + return (m_result != NULL); + } // Attempt to use an optimized reduction. else if (RunningOnGPU && (m_device.majorDeviceVersion() >= 3)) { @@ -572,7 +590,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { - if ((RunningFullReduction || RunningOnGPU) && m_result) { + if ((RunningOnSycl || RunningFullReduction || RunningOnGPU) && m_result) { return *(m_result + index); } Op reducer(m_reducer); @@ -644,7 +662,14 @@ } } - EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC typename MakePointer_<Scalar>::Type data() const { return m_result; } + /// required by sycl in order to extract the accessor + const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; } + /// added for sycl in order to construct the buffer from the sycl device + const Device& device() const{return m_device;} + /// added for sycl in order to re-construct the reduction eval on the device for the sub-kernel + const Dims& xprDims() const {return m_xpr_dims;} + private: template <int, typename, typename> friend struct internal::GenericDimReducer; @@ -737,12 +762,18 @@ // For full reductions #if defined(EIGEN_USE_GPU) && defined(__CUDACC__) static const bool RunningOnGPU = internal::is_same<Device, Eigen::GpuDevice>::value; + static const bool RunningOnSycl = false; +#elif defined(EIGEN_USE_SYCL) +static const bool RunningOnSycl = internal::is_same<typename internal::remove_all<Device>::type, Eigen::SyclDevice>::value; +static const bool RunningOnGPU = false; #else static const bool RunningOnGPU = false; + static const bool RunningOnSycl = false; #endif - CoeffReturnType* m_result; + typename MakePointer_<CoeffReturnType>::Type m_result; const Device& m_device; + const Dims& m_xpr_dims; }; } // end namespace Eigen
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h new file mode 100644 index 0000000..3daecb0 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
@@ -0,0 +1,242 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Mehdi Goli Codeplay Software Ltd. +// Ralph Potter Codeplay Software Ltd. +// Luke Iwanski Codeplay Software Ltd. +// Contact: <eigen@codeplay.com> +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +/***************************************************************** + * TensorSyclPlaceHolderExpr.h + * + * \brief: + * This is the specialisation of the placeholder expression based on the + * operation type + * +*****************************************************************/ + +#ifndef UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_REDUCTION_SYCL_HPP +#define UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_REDUCTION_SYCL_HPP + +namespace Eigen { +namespace internal { + +template<typename CoeffReturnType, typename KernelName> struct syclGenericBufferReducer{ +template<typename BufferTOut, typename BufferTIn> +static void run(BufferTOut* bufOut, BufferTIn& bufI, const Eigen::SyclDevice& dev, size_t length, size_t local){ + do { + auto f = [length, local, bufOut, &bufI](cl::sycl::handler& h) mutable { + cl::sycl::nd_range<1> r{cl::sycl::range<1>{std::max(length, local)}, + cl::sycl::range<1>{std::min(length, local)}}; + /* Two accessors are used: one to the buffer that is being reduced, + * and a second to local memory, used to store intermediate data. */ + auto aI = + bufI.template get_access<cl::sycl::access::mode::read_write>(h); + auto aOut = + bufOut->template get_access<cl::sycl::access::mode::discard_write>(h); + cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write, + cl::sycl::access::target::local> + scratch(cl::sycl::range<1>(local), h); + + /* The parallel_for invocation chosen is the variant with an nd_item + * parameter, since the code requires barriers for correctness. */ + h.parallel_for<KernelName>( + r, [aOut, aI, scratch, local, length](cl::sycl::nd_item<1> id) { + size_t globalid = id.get_global(0); + size_t localid = id.get_local(0); + /* All threads collectively read from global memory into local. + * The barrier ensures all threads' IO is resolved before + * execution continues (strictly speaking, all threads within + * a single work-group - there is no co-ordination between + * work-groups, only work-items). */ + if (globalid < length) { + scratch[localid] = aI[globalid]; + } + id.barrier(cl::sycl::access::fence_space::local_space); + + /* Apply the reduction operation between the current local + * id and the one on the other half of the vector. */ + if (globalid < length) { + int min = (length < local) ? length : local; + for (size_t offset = min / 2; offset > 0; offset /= 2) { + if (localid < offset) { + scratch[localid] += scratch[localid + offset]; + } + id.barrier(cl::sycl::access::fence_space::local_space); + } + /* The final result will be stored in local id 0. */ + if (localid == 0) { + aI[id.get_group(0)] = scratch[localid]; + if((length<=local) && globalid ==0){ + aOut[globalid]=scratch[localid]; + } + } + } + }); + }; + dev.m_queue.submit(f); + dev.m_queue.throw_asynchronous(); + + /* At this point, you could queue::wait_and_throw() to ensure that + * errors are caught quickly. However, this would likely impact + * performance negatively. */ + length = length / local; + + } while (length > 1); + + + +} + +}; + +/// For now let's start with a full reducer +/// Self is useless here because in expression construction we are going to treat reduction as a leafnode. +/// we want to take reduction child and then build a construction and apply the full reducer function on it. Fullreducre applies the +/// reduction operation on the child of the reduction. once it is done the reduction is an empty shell and can be thrown away and treated as +// a leafNode. +template <typename Self, typename Op, bool Vectorizable> +struct FullReducer<Self, Op, const Eigen::SyclDevice, Vectorizable> { + + typedef typename Self::CoeffReturnType CoeffReturnType; + static const bool HasOptimizedImplementation = false; + + static void run(const Self& self, Op& reducer, const Eigen::SyclDevice& dev, CoeffReturnType* output) { + typedef const typename Self::ChildType HostExpr; /// this is the child of reduction + typedef typename TensorSycl::internal::createPlaceHolderExpression<HostExpr>::Type PlaceHolderExpr; + auto functors = TensorSycl::internal::extractFunctors(self.impl()); + int red_factor =256; /// initial reduction. If the size is less than red_factor we only creates one thread. + size_t inputSize =self.impl().dimensions().TotalSize(); + size_t rng = inputSize/red_factor; // the total number of thread initially is half the size of the input + size_t remaining = inputSize% red_factor; + if(rng ==0) { + red_factor=1; + }; + size_t tileSize =dev.m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>()/2; + size_t GRange=std::max((size_t )1, rng); + + // convert global range to power of 2 for redecution + GRange--; + GRange |= GRange >> 1; + GRange |= GRange >> 2; + GRange |= GRange >> 4; + GRange |= GRange >> 8; + GRange |= GRange >> 16; +#if __x86_64__ || __ppc64__ || _WIN64 + GRange |= GRange >> 32; +#endif + GRange++; + size_t outTileSize = tileSize; + /// if the shared memory is less than the GRange, we set shared_mem size to the TotalSize and in this case one kernel would be created for recursion to reduce all to one. + if (GRange < outTileSize) outTileSize=GRange; + // getting final out buffer at the moment the created buffer is true because there is no need for assign + auto out_buffer =dev.template get_sycl_buffer<typename Eigen::internal::remove_all<CoeffReturnType>::type>(self.dimensions().TotalSize(), output); + /// creating the shared memory for calculating reduction. + /// This one is used to collect all the reduced value of shared memory as we dont have global barrier on GPU. Once it is saved we can + /// recursively apply reduction on it in order to reduce the whole. + auto temp_global_buffer =cl::sycl::buffer<CoeffReturnType, 1>(cl::sycl::range<1>(GRange)); + typedef typename Eigen::internal::remove_all<decltype(self.xprDims())>::type Dims; + Dims dims= self.xprDims(); + Op functor = reducer; + dev.m_queue.submit([&](cl::sycl::handler &cgh) { + // create a tuple of accessors from Evaluator + auto tuple_of_accessors = TensorSycl::internal::createTupleOfAccessors(cgh, self.impl()); + auto tmp_global_accessor = temp_global_buffer. template get_access<cl::sycl::access::mode::read_write, cl::sycl::access::target::global_buffer>(cgh); + + cgh.parallel_for<PlaceHolderExpr>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(outTileSize)), [=](cl::sycl::nd_item<1> itemID) { + typedef typename TensorSycl::internal::ConvertToDeviceExpression<const HostExpr>::Type DevExpr; + auto device_expr = TensorSycl::internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors); + /// reduction cannot be captured automatically through our device conversion recursion. The reason is that reduction has two behaviour + /// the first behaviour is when it is used as a root to lauch the sub-kernel. The second one is when it is treated as a leafnode to pass the + /// calculated result to its parent kernel. While the latter is automatically detected through our device expression generator. The former is created here. + const auto device_self_expr= TensorReductionOp<Op, Dims, decltype(device_expr.expr) ,MakeGlobalPointer>(device_expr.expr, dims, functor); + /// This is the evaluator for device_self_expr. This is exactly similar to the self which has been passed to run function. The difference is + /// the device_evaluator is detectable and recognisable on the device. + auto device_self_evaluator = Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice>(device_self_expr, Eigen::DefaultDevice()); + /// const cast added as a naive solution to solve the qualifier drop error + auto globalid=itemID.get_global_linear_id(); + + if(globalid<rng) + tmp_global_accessor.get_pointer()[globalid]=InnerMostDimReducer<decltype(device_self_evaluator), Op, false>::reduce(device_self_evaluator, red_factor*globalid, red_factor, const_cast<Op&>(functor)); + else + tmp_global_accessor.get_pointer()[globalid]=static_cast<CoeffReturnType>(0); + + if(remaining!=0 && globalid==0 ) + // this will add the rest of input buffer when the input size is not devidable to red_factor. + tmp_global_accessor.get_pointer()[globalid]+=InnerMostDimReducer<decltype(device_self_evaluator), Op, false>::reduce(device_self_evaluator, red_factor*(rng), remaining, const_cast<Op&>(functor)); + }); + }); + dev.m_queue.throw_asynchronous(); + +/// This is used to recursively reduce the tmp value to an element of 1; + syclGenericBufferReducer<CoeffReturnType,HostExpr>::run(out_buffer, temp_global_buffer,dev, GRange, outTileSize); + } + +}; + +template <typename Self, typename Op> +struct InnerReducer<Self, Op, const Eigen::SyclDevice> { + + typedef typename Self::CoeffReturnType CoeffReturnType; + static const bool HasOptimizedImplementation = false; + + static bool run(const Self& self, Op& reducer, const Eigen::SyclDevice& dev, CoeffReturnType* output, typename Self::Index , typename Self::Index num_coeffs_to_preserve) { + typedef const typename Self::ChildType HostExpr; /// this is the child of reduction + typedef typename TensorSycl::internal::createPlaceHolderExpression<HostExpr>::Type PlaceHolderExpr; + auto functors = TensorSycl::internal::extractFunctors(self.impl()); + + size_t tileSize =dev.m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>()/2; + + size_t GRange=num_coeffs_to_preserve; + if (tileSize>GRange) tileSize=GRange; + else if(GRange>tileSize){ + size_t xMode = GRange % tileSize; + if (xMode != 0) GRange += (tileSize - xMode); + } + // getting final out buffer at the moment the created buffer is true because there is no need for assign + /// creating the shared memory for calculating reduction. + /// This one is used to collect all the reduced value of shared memory as we dont have global barrier on GPU. Once it is saved we can + /// recursively apply reduction on it in order to reduce the whole. + typedef typename Eigen::internal::remove_all<decltype(self.xprDims())>::type Dims; + Dims dims= self.xprDims(); + Op functor = reducer; + + dev.m_queue.submit([&](cl::sycl::handler &cgh) { + // create a tuple of accessors from Evaluator + auto tuple_of_accessors = TensorSycl::internal::createTupleOfAccessors(cgh, self.impl()); + auto output_accessor = dev.template get_sycl_accessor<cl::sycl::access::mode::discard_write>(num_coeffs_to_preserve,cgh, output); + + cgh.parallel_for<Self>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), [=](cl::sycl::nd_item<1> itemID) { + typedef typename TensorSycl::internal::ConvertToDeviceExpression<const HostExpr>::Type DevExpr; + auto device_expr = TensorSycl::internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors); + /// reduction cannot be captured automatically through our device conversion recursion. The reason is that reduction has two behaviour + /// the first behaviour is when it is used as a root to lauch the sub-kernel. The second one is when it is treated as a leafnode to pass the + /// calculated result to its parent kernel. While the latter is automatically detected through our device expression generator. The former is created here. + const auto device_self_expr= TensorReductionOp<Op, Dims, decltype(device_expr.expr) ,MakeGlobalPointer>(device_expr.expr, dims, functor); + /// This is the evaluator for device_self_expr. This is exactly similar to the self which has been passed to run function. The difference is + /// the device_evaluator is detectable and recognisable on the device. + typedef Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice> DeiceSelf; + auto device_self_evaluator = Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice>(device_self_expr, Eigen::DefaultDevice()); + /// const cast added as a naive solution to solve the qualifier drop error + auto globalid=itemID.get_global_linear_id(); + if (globalid< static_cast<size_t>(num_coeffs_to_preserve)) { + typename DeiceSelf::CoeffReturnType accum = functor.initialize(); + GenericDimReducer<DeiceSelf::NumReducedDims-1, DeiceSelf, Op>::reduce(device_self_evaluator, device_self_evaluator.firstInput(globalid),const_cast<Op&>(functor), &accum); + functor.finalize(accum); + output_accessor.get_pointer()[globalid]= accum; + } + }); + }); + dev.m_queue.throw_asynchronous(); + return false; + } +}; + +} // end namespace internal +} // namespace Eigen + +#endif // UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_REDUCTION_SYCL_HPP
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSycl.h index da15f79..bb8800d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorSycl.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSycl.h
@@ -22,6 +22,13 @@ typedef typename cl::sycl::global_ptr<T>::pointer_t Type; }; +// global pointer to set different attribute state for a class +template <class T> +struct MakeLocalPointer { + typedef typename cl::sycl::local_ptr<T>::pointer_t Type; +}; + + namespace Eigen { namespace TensorSycl { namespace internal { @@ -43,9 +50,7 @@ // tuple construction #include "TensorSyclTuple.h" -// This file contains the PlaceHolder that replaces the actual data -#include "TensorSyclPlaceHolder.h" - +// counting number of leaf at compile time #include "TensorSyclLeafCount.h" // The index PlaceHolder takes the actual expression and replaces the actual @@ -57,9 +62,6 @@ // creation of an accessor tuple from a tuple of SYCL buffers #include "TensorSyclExtractAccessor.h" -// actual data extraction using accessors -//#include "GetDeviceData.h" - // this is used to change the address space type in tensor map for GPU #include "TensorSyclConvertToDeviceExpression.h" @@ -70,6 +72,9 @@ // this is used to construct the expression on the device #include "TensorSyclExprConstructor.h" +/// this is used for extracting tensor reduction +#include "TensorReductionSycl.h" + // kernel execution using fusion #include "TensorSyclRun.h"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h index a94c304..8729c86 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclConvertToDeviceExpression.h
@@ -102,6 +102,18 @@ KERNELBROKERCONVERT(const, true, TensorEvalToOp) KERNELBROKERCONVERT(, false, TensorEvalToOp) #undef KERNELBROKERCONVERT + +/// specialisation of the \ref ConvertToDeviceExpression struct when the node type is TensorReductionOp +#define KERNELBROKERCONVERTREDUCTION(CVQual)\ +template <typename OP, typename Dim, typename subExpr, template <class> class MakePointer_>\ +struct ConvertToDeviceExpression<CVQual TensorReductionOp<OP, Dim, subExpr, MakePointer_> > {\ + typedef CVQual TensorReductionOp<OP, Dim, typename ConvertToDeviceExpression<subExpr>::Type, MakeGlobalPointer> Type;\ +}; + +KERNELBROKERCONVERTREDUCTION(const) +KERNELBROKERCONVERTREDUCTION() +#undef KERNELBROKERCONVERTREDUCTION + } // namespace internal } // namespace TensorSycl } // namespace Eigen
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h index 833d5e2..7ed3a3a 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
@@ -33,8 +33,7 @@ EvalToLHSConstructor(const utility::tuple::Tuple<Params...> &t): expr((&(*(utility::tuple::get<N>(t).get_pointer())))) {} }; -/// \struct ExprConstructor is used to reconstruct the expression on the device -/// and +/// \struct ExprConstructor is used to reconstruct the expression on the device and /// recreate the expression with MakeGlobalPointer containing the device address /// space for the TensorMap pointers used in eval function. /// It receives the original expression type, the functor of the node, the tuple @@ -49,7 +48,7 @@ template <typename Scalar_, int Options_, int Options2_, int Options3_, int NumIndices_, typename IndexType_,\ template <class> class MakePointer_, size_t N, typename... Params>\ struct ExprConstructor< CVQual TensorMap<Tensor<Scalar_, NumIndices_, Options_, IndexType_>, Options2_, MakeGlobalPointer>,\ -CVQual Eigen::internal::PlaceHolder<CVQual TensorMap<Tensor<Scalar_, NumIndices_, Options_, IndexType_>, Options3_, MakePointer_>, N>, Params...>{\ +CVQual PlaceHolder<CVQual TensorMap<Tensor<Scalar_, NumIndices_, Options_, IndexType_>, Options3_, MakePointer_>, N>, Params...>{\ typedef CVQual TensorMap<Tensor<Scalar_, NumIndices_, Options_, IndexType_>, Options2_, MakeGlobalPointer> Type;\ Type expr;\ template <typename FuncDetector>\ @@ -187,7 +186,7 @@ #define FORCEDEVAL(CVQual)\ template <typename OrigExpr, typename DevExpr, size_t N, typename... Params>\ struct ExprConstructor<CVQual TensorForcedEvalOp<OrigExpr, MakeGlobalPointer>,\ -CVQual Eigen::internal::PlaceHolder<CVQual TensorForcedEvalOp<DevExpr>, N>, Params...> {\ +CVQual PlaceHolder<CVQual TensorForcedEvalOp<DevExpr>, N>, Params...> {\ typedef CVQual TensorMap<Tensor<typename TensorForcedEvalOp<DevExpr, MakeGlobalPointer>::Scalar,\ TensorForcedEvalOp<DevExpr, MakeGlobalPointer>::NumDimensions, 0, typename TensorForcedEvalOp<DevExpr>::Index>, 0, MakeGlobalPointer> Type;\ Type expr;\ @@ -200,14 +199,41 @@ FORCEDEVAL() #undef FORCEDEVAL +template <bool Conds, size_t X , size_t Y > struct ValueCondition { + static const size_t Res =X; +}; +template<size_t X, size_t Y> struct ValueCondition<false, X , Y> { + static const size_t Res =Y; +}; + +/// specialisation of the \ref ExprConstructor struct when the node type is TensorReductionOp +#define SYCLREDUCTIONEXPR(CVQual)\ +template <typename OP, typename Dim, typename OrigExpr, typename DevExpr, size_t N, typename... Params>\ +struct ExprConstructor<CVQual TensorReductionOp<OP, Dim, OrigExpr, MakeGlobalPointer>,\ +CVQual PlaceHolder<CVQual TensorReductionOp<OP, Dim, DevExpr>, N>, Params...> {\ + static const size_t NumIndices= ValueCondition< TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>::NumDimensions==0, 1, TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>::NumDimensions >::Res;\ + typedef CVQual TensorMap<Tensor<typename TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>::Scalar,\ + NumIndices, 0, typename TensorReductionOp<OP, Dim, DevExpr>::Index>, 0, MakeGlobalPointer> Type;\ + Type expr;\ + template <typename FuncDetector>\ + ExprConstructor(FuncDetector &fd, const utility::tuple::Tuple<Params...> &t)\ + : expr(Type((&(*(utility::tuple::get<N>(t).get_pointer()))), fd.dimensions())) {}\ +}; + +SYCLREDUCTIONEXPR(const) +SYCLREDUCTIONEXPR() +#undef SYCLREDUCTIONEXPR + /// template deduction for \ref ExprConstructor struct template <typename OrigExpr, typename IndexExpr, typename FuncD, typename... Params> auto createDeviceExpression(FuncD &funcD, const utility::tuple::Tuple<Params...> &t) -> decltype(ExprConstructor<OrigExpr, IndexExpr, Params...>(funcD, t)) { return ExprConstructor<OrigExpr, IndexExpr, Params...>(funcD, t); } -} -} -} // namespace Eigen + +} /// namespace TensorSycl +} /// namespace internal +} /// namespace Eigen + #endif // UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSORSYCL_EXPR_CONSTRUCTOR_HPP
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h index ceec528..b1da685 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
@@ -57,9 +57,9 @@ return utility::tuple::append(ExtractAccessor<Arg1>::getTuple(cgh, eval1),utility::tuple::append(ExtractAccessor<Arg2>::getTuple(cgh, eval2), ExtractAccessor<Arg3>::getTuple(cgh, eval3))); } template< cl::sycl::access::mode AcM, typename Arg> static inline auto getAccessor(cl::sycl::handler& cgh, Arg eval) - -> decltype(utility::tuple::make_tuple( eval.device().template get_sycl_accessor<AcM, true, + -> decltype(utility::tuple::make_tuple( eval.device().template get_sycl_accessor<AcM, typename Eigen::internal::remove_all<typename Arg::CoeffReturnType>::type>(eval.dimensions().TotalSize(), cgh,eval.data()))){ - return utility::tuple::make_tuple(eval.device().template get_sycl_accessor<AcM, true, typename Eigen::internal::remove_all<typename Arg::CoeffReturnType>::type>(eval.dimensions().TotalSize(), cgh,eval.data())); + return utility::tuple::make_tuple(eval.device().template get_sycl_accessor<AcM, typename Eigen::internal::remove_all<typename Arg::CoeffReturnType>::type>(eval.dimensions().TotalSize(), cgh,eval.data())); } }; @@ -73,14 +73,12 @@ } }; -/// specialisation of the \ref ExtractAccessor struct when the node type is -/// TensorCwiseNullaryOp, TensorCwiseUnaryOp and TensorBroadcastingOp +/// specialisation of the \ref ExtractAccessor struct when the node type is TensorCwiseNullaryOp, TensorCwiseUnaryOp and TensorBroadcastingOp template <template<class, class> class UnaryCategory, typename OP, typename RHSExpr, typename Dev> struct ExtractAccessor<TensorEvaluator<UnaryCategory<OP, RHSExpr>, Dev> > : ExtractAccessor<TensorEvaluator<const UnaryCategory<OP, RHSExpr>, Dev> > {}; -/// specialisation of the \ref ExtractAccessor struct when the node type is -/// const TensorCwiseBinaryOp +/// specialisation of the \ref ExtractAccessor struct when the node type is const TensorCwiseBinaryOp template <template<class, class, class> class BinaryCategory, typename OP, typename LHSExpr, typename RHSExpr, typename Dev> struct ExtractAccessor<TensorEvaluator<const BinaryCategory<OP, LHSExpr, RHSExpr>, Dev> > { static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<const BinaryCategory<OP, LHSExpr, RHSExpr>, Dev> eval) @@ -88,9 +86,7 @@ return AccessorConstructor::getTuple(cgh, eval.left_impl(), eval.right_impl()); } }; - -/// specialisation of the \ref ExtractAccessor struct when the node type is -/// TensorCwiseBinaryOp +/// specialisation of the \ref ExtractAccessor struct when the node type is TensorCwiseBinaryOp template <template<class, class, class> class BinaryCategory, typename OP, typename LHSExpr, typename RHSExpr, typename Dev> struct ExtractAccessor<TensorEvaluator<BinaryCategory<OP, LHSExpr, RHSExpr>, Dev> > : ExtractAccessor<TensorEvaluator<const BinaryCategory<OP, LHSExpr, RHSExpr>, Dev> >{}; @@ -105,8 +101,7 @@ } }; -/// specialisation of the \ref ExtractAccessor struct when the node type is -/// TensorCwiseTernaryOp +/// specialisation of the \ref ExtractAccessor struct when the node type is TensorCwiseTernaryOp template <template<class, class, class, class> class TernaryCategory, typename OP, typename Arg1Expr, typename Arg2Expr, typename Arg3Expr, typename Dev> struct ExtractAccessor<TensorEvaluator<TernaryCategory<OP, Arg1Expr, Arg2Expr, Arg3Expr>, Dev> > : ExtractAccessor<TensorEvaluator<const TernaryCategory<OP, Arg1Expr, Arg2Expr, Arg3Expr>, Dev> >{}; @@ -127,8 +122,7 @@ struct ExtractAccessor<TensorEvaluator<TensorSelectOp<IfExpr, ThenExpr, ElseExpr>, Dev> > : ExtractAccessor<TensorEvaluator<const TensorSelectOp<IfExpr, ThenExpr, ElseExpr>, Dev> >{}; -/// specialisation of the \ref ExtractAccessor struct when the node type is -/// const TensorAssignOp +/// specialisation of the \ref ExtractAccessor struct when the node type is const TensorAssignOp template <typename LHSExpr, typename RHSExpr, typename Dev> struct ExtractAccessor<TensorEvaluator<const TensorAssignOp<LHSExpr, RHSExpr>, Dev> > { static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<const TensorAssignOp<LHSExpr, RHSExpr>, Dev> eval) @@ -137,14 +131,12 @@ } }; -/// specialisation of the \ref ExtractAccessor struct when the node type is -/// TensorAssignOp +/// specialisation of the \ref ExtractAccessor struct when the node type is TensorAssignOp template <typename LHSExpr, typename RHSExpr, typename Dev> struct ExtractAccessor<TensorEvaluator<TensorAssignOp<LHSExpr, RHSExpr>, Dev> > : ExtractAccessor<TensorEvaluator<const TensorAssignOp<LHSExpr, RHSExpr>, Dev> >{}; -/// specialisation of the \ref ExtractAccessor struct when the node type is -/// const TensorMap +/// specialisation of the \ref ExtractAccessor struct when the node type is const TensorMap #define TENSORMAPEXPR(CVQual, ACCType)\ template <typename PlainObjectType, int Options_, typename Dev>\ struct ExtractAccessor<TensorEvaluator<CVQual TensorMap<PlainObjectType, Options_>, Dev> > {\ @@ -157,8 +149,7 @@ TENSORMAPEXPR(, cl::sycl::access::mode::read_write) #undef TENSORMAPEXPR -/// specialisation of the \ref ExtractAccessor struct when the node type is -/// const TensorForcedEvalOp +/// specialisation of the \ref ExtractAccessor struct when the node type is const TensorForcedEvalOp template <typename Expr, typename Dev> struct ExtractAccessor<TensorEvaluator<const TensorForcedEvalOp<Expr>, Dev> > { static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<const TensorForcedEvalOp<Expr>, Dev> eval) @@ -167,14 +158,12 @@ } }; -/// specialisation of the \ref ExtractAccessor struct when the node type is -/// TensorForcedEvalOp +/// specialisation of the \ref ExtractAccessor struct when the node type is TensorForcedEvalOp template <typename Expr, typename Dev> struct ExtractAccessor<TensorEvaluator<TensorForcedEvalOp<Expr>, Dev> > : ExtractAccessor<TensorEvaluator<const TensorForcedEvalOp<Expr>, Dev> >{}; -/// specialisation of the \ref ExtractAccessor struct when the node type is -/// const TensorEvalToOp +/// specialisation of the \ref ExtractAccessor struct when the node type is const TensorEvalToOp template <typename Expr, typename Dev> struct ExtractAccessor<TensorEvaluator<const TensorEvalToOp<Expr>, Dev> > { static inline auto getTuple(cl::sycl::handler& cgh,const TensorEvaluator<const TensorEvalToOp<Expr>, Dev> eval) @@ -183,19 +172,33 @@ } }; -/// specialisation of the \ref ExtractAccessor struct when the node type is -/// TensorEvalToOp +/// specialisation of the \ref ExtractAccessor struct when the node type is TensorEvalToOp template <typename Expr, typename Dev> struct ExtractAccessor<TensorEvaluator<TensorEvalToOp<Expr>, Dev> > : ExtractAccessor<TensorEvaluator<const TensorEvalToOp<Expr>, Dev> >{}; +/// specialisation of the \ref ExtractAccessor struct when the node type is const TensorReductionOp +template <typename OP, typename Dim, typename Expr, typename Dev> +struct ExtractAccessor<TensorEvaluator<const TensorReductionOp<OP, Dim, Expr>, Dev> > { + static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<const TensorReductionOp<OP, Dim, Expr>, Dev> eval) + -> decltype(AccessorConstructor::template getAccessor<cl::sycl::access::mode::read>(cgh, eval)){ + return AccessorConstructor::template getAccessor<cl::sycl::access::mode::read>(cgh, eval); + } +}; + +/// specialisation of the \ref ExtractAccessor struct when the node type is TensorReductionOp +template <typename OP, typename Dim, typename Expr, typename Dev> +struct ExtractAccessor<TensorEvaluator<TensorReductionOp<OP, Dim, Expr>, Dev> > +: ExtractAccessor<TensorEvaluator<const TensorReductionOp<OP, Dim, Expr>, Dev> >{}; + /// template deduction for \ref ExtractAccessor template <typename Evaluator> auto createTupleOfAccessors(cl::sycl::handler& cgh, const Evaluator& expr) -> decltype(ExtractAccessor<Evaluator>::getTuple(cgh, expr)) { return ExtractAccessor<Evaluator>::getTuple(cgh, expr); } -} -} -} + +} /// namespace TensorSycl +} /// namespace internal +} /// namespace Eigen #endif // UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSORSYCL_EXTRACT_ACCESSOR_HPP
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h index 801b4f5..4271253 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
@@ -141,7 +141,30 @@ struct FunctorExtractor<TensorEvaluator<TensorEvalToOp<RHSExpr>, Dev> > : FunctorExtractor<TensorEvaluator<const TensorEvalToOp<RHSExpr>, Dev> > {}; +template<typename Dim, size_t NumOutputDim> struct DimConstr { +template<typename InDim> + static inline Dim getDim(InDim dims ) {return dims;} +}; +template<typename Dim> struct DimConstr<Dim, 0> { + template<typename InDim> + static inline Dim getDim(InDim dims ) {return Dim(dims.TotalSize());} +}; + +template<typename Op, typename Dims, typename ArgType, template <class> class MakePointer_, typename Device> +struct FunctorExtractor<TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device>>{ + typedef TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device> Evaluator; + typedef typename Eigen::internal::conditional<Evaluator::NumOutputDims==0, DSizes<typename Evaluator::Index, 1>, typename Evaluator::Dimensions >::type Dimensions; + const Dimensions m_dimensions; + const Dimensions& dimensions() const { return m_dimensions; } + FunctorExtractor(const TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device>& expr) + : m_dimensions(DimConstr<Dimensions, Evaluator::NumOutputDims>::getDim(expr.dimensions())) {} +}; + + +template<typename Op, typename Dims, typename ArgType, template <class> class MakePointer_, typename Device> +struct FunctorExtractor<TensorEvaluator<TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device>> +: FunctorExtractor<TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device>>{}; /// template deduction function for FunctorExtractor template <typename Evaluator> auto inline extractFunctors(const Evaluator& evaluator)-> FunctorExtractor<Evaluator> {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h index 8d520d2..25d1fac 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclLeafCount.h
@@ -43,8 +43,7 @@ static const size_t Count = LeafCount<Arg>::Count + CategoryCount<Args...>::Count; }; -/// specialisation of the \ref LeafCount struct when the node type is const -/// TensorMap +/// specialisation of the \ref LeafCount struct when the node type is const TensorMap template <typename PlainObjectType, int Options_, template <class> class MakePointer_> struct LeafCount<const TensorMap<PlainObjectType, Options_, MakePointer_> > { static const size_t Count =1; @@ -61,18 +60,15 @@ template <template <class, class...> class CategoryExpr, typename OP, typename... RHSExpr> struct LeafCount<CategoryExpr<OP, RHSExpr...> > :LeafCount<const CategoryExpr<OP, RHSExpr...> >{}; -/// specialisation of the \ref LeafCount struct when the node type is -/// const TensorSelectOp is an exception +/// specialisation of the \ref LeafCount struct when the node type is const TensorSelectOp is an exception template <typename IfExpr, typename ThenExpr, typename ElseExpr> struct LeafCount<const TensorSelectOp<IfExpr, ThenExpr, ElseExpr> > : CategoryCount<IfExpr, ThenExpr, ElseExpr> {}; -/// specialisation of the \ref LeafCount struct when the node type is -/// TensorSelectOp +/// specialisation of the \ref LeafCount struct when the node type is TensorSelectOp template <typename IfExpr, typename ThenExpr, typename ElseExpr> struct LeafCount<TensorSelectOp<IfExpr, ThenExpr, ElseExpr> >: LeafCount<const TensorSelectOp<IfExpr, ThenExpr, ElseExpr> > {}; -/// specialisation of the \ref LeafCount struct when the node type is const -/// TensorAssignOp +/// specialisation of the \ref LeafCount struct when the node type is const TensorAssignOp template <typename LHSExpr, typename RHSExpr> struct LeafCount<const TensorAssignOp<LHSExpr, RHSExpr> >: CategoryCount<LHSExpr,RHSExpr> {}; @@ -81,31 +77,38 @@ template <typename LHSExpr, typename RHSExpr> struct LeafCount<TensorAssignOp<LHSExpr, RHSExpr> > :LeafCount<const TensorAssignOp<LHSExpr, RHSExpr> >{}; -/// specialisation of the \ref LeafCount struct when the node type is const -/// TensorForcedEvalOp +/// specialisation of the \ref LeafCount struct when the node type is const TensorForcedEvalOp template <typename Expr> struct LeafCount<const TensorForcedEvalOp<Expr> > { static const size_t Count =1; }; -/// specialisation of the \ref LeafCount struct when the node type is -/// TensorForcedEvalOp +/// specialisation of the \ref LeafCount struct when the node type is TensorForcedEvalOp template <typename Expr> struct LeafCount<TensorForcedEvalOp<Expr> >: LeafCount<const TensorForcedEvalOp<Expr> > {}; -/// specialisation of the \ref LeafCount struct when the node type is const -/// TensorEvalToOp +/// specialisation of the \ref LeafCount struct when the node type is const TensorEvalToOp template <typename Expr> struct LeafCount<const TensorEvalToOp<Expr> > { static const size_t Count = 1 + CategoryCount<Expr>::Count; }; -/// specialisation of the \ref LeafCount struct when the node type is -/// TensorEvalToOp +/// specialisation of the \ref LeafCount struct when the node type is const TensorReductionOp +template <typename OP, typename Dim, typename Expr> +struct LeafCount<const TensorReductionOp<OP, Dim, Expr> > { + static const size_t Count =1; +}; + +/// specialisation of the \ref LeafCount struct when the node type is TensorReductionOp +template <typename OP, typename Dim, typename Expr> +struct LeafCount<TensorReductionOp<OP, Dim, Expr> >: LeafCount<const TensorReductionOp<OP, Dim, Expr> >{}; + +/// specialisation of the \ref LeafCount struct when the node type is TensorEvalToOp template <typename Expr> struct LeafCount<TensorEvalToOp<Expr> >: LeafCount<const TensorEvalToOp<Expr> >{}; -} -} -} // namespace Eigen + +} /// namespace TensorSycl +} /// namespace internal +} /// namespace Eigen #endif // UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSORSYCL_LEAF_COUNT_HPP
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolder.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolder.h deleted file mode 100644 index 43a63c7..0000000 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolder.h +++ /dev/null
@@ -1,99 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Mehdi Goli Codeplay Software Ltd. -// Ralph Potter Codeplay Software Ltd. -// Luke Iwanski Codeplay Software Ltd. -// Contact: <eigen@codeplay.com> -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -/***************************************************************** - * TensorSyclPlaceHolder.h - * - * \brief: - * The PlaceHolder expression are nothing but a container preserving - * the order of actual data in the tuple of sycl buffer. - * -*****************************************************************/ - -#ifndef UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSORSYCL_PLACEHOLDER_HPP -#define UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSORSYCL_PLACEHOLDER_HPP - -namespace Eigen { -namespace internal { -/// \struct PlaceHolder -/// \brief PlaceHolder is used to replace the \ref TensorMap in the expression -/// tree. -/// PlaceHolder contains the order of the leaf node in the expression tree. -template <typename Scalar, size_t N> -struct PlaceHolder { - static constexpr size_t I = N; - typedef Scalar Type; -}; - -/// \brief specialisation of the PlaceHolder node for const TensorMap -#define TENSORMAPPLACEHOLDER(CVQual)\ -template <typename PlainObjectType, int Options_, template <class> class MakePointer_, size_t N>\ -struct PlaceHolder<CVQual TensorMap<PlainObjectType, Options_, MakePointer_>, N> {\ - static const size_t I = N;\ - typedef CVQual TensorMap<PlainObjectType, Options_, MakePointer_> Type;\ - typedef typename Type::Self Self;\ - typedef typename Type::Base Base;\ - typedef typename Type::Nested Nested;\ - typedef typename Type::StorageKind StorageKind;\ - typedef typename Type::Index Index;\ - typedef typename Type::Scalar Scalar;\ - typedef typename Type::RealScalar RealScalar;\ - typedef typename Type::CoeffReturnType CoeffReturnType;\ -}; - -TENSORMAPPLACEHOLDER(const) -TENSORMAPPLACEHOLDER() -#undef TENSORMAPPLACEHOLDER - -/// \brief specialisation of the PlaceHolder node for TensorForcedEvalOp. The -/// TensorForcedEvalOp acts as a leaf node for its parent node. -#define TENSORFORCEDEVALPLACEHOLDER(CVQual)\ -template <typename Expression, size_t N>\ -struct PlaceHolder<CVQual TensorForcedEvalOp<Expression>, N> {\ - static const size_t I = N;\ - typedef CVQual TensorForcedEvalOp<Expression> Type;\ - typedef typename Type::Nested Nested;\ - typedef typename Type::StorageKind StorageKind;\ - typedef typename Type::Index Index;\ - typedef typename Type::Scalar Scalar;\ - typedef typename Type::Packet Packet;\ - typedef typename Type::RealScalar RealScalar;\ - typedef typename Type::CoeffReturnType CoeffReturnType;\ - typedef typename Type::PacketReturnType PacketReturnType;\ -}; - -TENSORFORCEDEVALPLACEHOLDER(const) -TENSORFORCEDEVALPLACEHOLDER() -#undef TENSORFORCEDEVALPLACEHOLDER - -template <typename PlainObjectType, int Options_, template <class> class Makepointer_, size_t N> -struct traits<PlaceHolder<const TensorMap<PlainObjectType, Options_, Makepointer_>, N> >: public traits<PlainObjectType> { - typedef traits<PlainObjectType> BaseTraits; - typedef typename BaseTraits::Scalar Scalar; - typedef typename BaseTraits::StorageKind StorageKind; - typedef typename BaseTraits::Index Index; - static const int NumDimensions = BaseTraits::NumDimensions; - static const int Layout = BaseTraits::Layout; - enum { - Options = Options_, - Flags = BaseTraits::Flags, - }; -}; - -template <typename PlainObjectType, int Options_, template <class> class Makepointer_, size_t N> -struct traits<PlaceHolder<TensorMap<PlainObjectType, Options_, Makepointer_>, N> > -: traits<PlaceHolder<const TensorMap<PlainObjectType, Options_, Makepointer_>, N> > {}; - -} // end namespace internal -} // end namespoace Eigen - -#endif // UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSORSYCL_PLACEHOLDER_HPP
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h index f456c35..d4c250c 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclPlaceHolderExpr.h
@@ -25,6 +25,17 @@ namespace Eigen { namespace TensorSycl { namespace internal { + +/// \struct PlaceHolder +/// \brief PlaceHolder is used to replace the \ref TensorMap in the expression +/// tree. +/// PlaceHolder contains the order of the leaf node in the expression tree. +template <typename Scalar, size_t N> +struct PlaceHolder { + static constexpr size_t I = N; + typedef Scalar Type; +}; + /// \sttruct PlaceHolderExpression /// \brief it is used to create the PlaceHolder expression. The PlaceHolder /// expression is a copy of expression type in which the TensorMap of the has @@ -113,7 +124,7 @@ #define TENSORMAPEXPR(CVQual)\ template <typename Scalar_, int Options_, int Options2_, int NumIndices_, typename IndexType_, template <class> class MakePointer_, size_t N>\ struct PlaceHolderExpression< CVQual TensorMap< Tensor<Scalar_, NumIndices_, Options_, IndexType_>, Options2_, MakePointer_>, N> {\ - typedef CVQual Eigen::internal::PlaceHolder<CVQual TensorMap<Tensor<Scalar_, NumIndices_, Options_, IndexType_>, Options2_, MakePointer_>, N> Type;\ + typedef CVQual PlaceHolder<CVQual TensorMap<Tensor<Scalar_, NumIndices_, Options_, IndexType_>, Options2_, MakePointer_>, N> Type;\ }; TENSORMAPEXPR(const) @@ -125,7 +136,7 @@ #define FORCEDEVAL(CVQual)\ template <typename Expr, size_t N>\ struct PlaceHolderExpression<CVQual TensorForcedEvalOp<Expr>, N> {\ - typedef CVQual Eigen::internal::PlaceHolder<CVQual TensorForcedEvalOp<Expr>, N> Type;\ + typedef CVQual PlaceHolder<CVQual TensorForcedEvalOp<Expr>, N> Type;\ }; FORCEDEVAL(const) @@ -144,6 +155,18 @@ EVALTO() #undef EVALTO + +/// specialisation of the \ref PlaceHolderExpression when the node is +/// TensorReductionOp +#define SYCLREDUCTION(CVQual)\ +template <typename OP, typename Dims, typename Expr, size_t N>\ +struct PlaceHolderExpression<CVQual TensorReductionOp<OP, Dims, Expr>, N>{\ + typedef CVQual PlaceHolder<CVQual TensorReductionOp<OP, Dims,Expr>, N> Type;\ +}; +SYCLREDUCTION(const) +SYCLREDUCTION() +#undef SYCLREDUCTION + /// template deduction for \ref PlaceHolderExpression struct template <typename Expr> struct createPlaceHolderExpression { @@ -151,8 +174,8 @@ typedef typename PlaceHolderExpression<Expr, TotalLeaves - 1>::Type Type; }; -} -} +} // internal +} // TensorSycl } // namespace Eigen #endif // UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSORSYCL_PLACEHOLDER_EXPR_HPP
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h index 57f2dda..7914b6f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h
@@ -37,20 +37,20 @@ typedef typename internal::createPlaceHolderExpression<Expr>::Type PlaceHolderExpr; auto functors = internal::extractFunctors(evaluator); + size_t tileSize =dev.m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>()/2; dev.m_queue.submit([&](cl::sycl::handler &cgh) { // create a tuple of accessors from Evaluator auto tuple_of_accessors = internal::createTupleOfAccessors<decltype(evaluator)>(cgh, evaluator); const auto range = utility::tuple::get<0>(tuple_of_accessors).get_range()[0]; - - size_t outTileSize = range; - if (range > 64) outTileSize = 64; - size_t yMode = range % outTileSize; - int yRange = static_cast<int>(range); - if (yMode != 0) yRange += (outTileSize - yMode); - + size_t GRange=range; + if (tileSize>GRange) tileSize=GRange; + else if(GRange>tileSize){ + size_t xMode = GRange % tileSize; + if (xMode != 0) GRange += (tileSize - xMode); + } // run the kernel - cgh.parallel_for<PlaceHolderExpr>( cl::sycl::nd_range<1>(cl::sycl::range<1>(yRange), cl::sycl::range<1>(outTileSize)), [=](cl::sycl::nd_item<1> itemID) { + cgh.parallel_for<PlaceHolderExpr>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), [=](cl::sycl::nd_item<1> itemID) { typedef typename internal::ConvertToDeviceExpression<Expr>::Type DevExpr; auto device_expr =internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors); auto device_evaluator = Eigen::TensorEvaluator<decltype(device_expr.expr), Eigen::DefaultDevice>(device_expr.expr, Eigen::DefaultDevice()); @@ -61,6 +61,7 @@ }); dev.m_queue.throw_asynchronous(); } + evaluator.cleanup(); } } // namespace TensorSycl
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index aeaea16..b5fa1c8 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt
@@ -1,6 +1,6 @@ # generate split test header file only if it does not yet exist # in order to prevent a rebuild everytime cmake is configured -if(NOT EXISTS ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h) +if(NOT EXISTS ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h) file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h "") foreach(i RANGE 1 999) file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h @@ -16,11 +16,11 @@ set_property(GLOBAL PROPERTY EIGEN_CURRENT_SUBPROJECT "Unsupported") add_custom_target(BuildUnsupported) -include_directories(../../test ../../unsupported ../../Eigen +include_directories(../../test ../../unsupported ../../Eigen ${CMAKE_CURRENT_BINARY_DIR}/../../test) find_package (Threads) - + find_package(GoogleHash) if(GOOGLEHASH_FOUND) add_definitions("-DEIGEN_GOOGLEHASH_SUPPORT") @@ -134,7 +134,7 @@ ei_add_test(cxx11_tensor_layout_swap) ei_add_test(cxx11_tensor_io) if("${CMAKE_SIZEOF_VOID_P}" EQUAL "8") - # This test requires __uint128_t which is only available on 64bit systems + # This test requires __uint128_t which is only available on 64bit systems ei_add_test(cxx11_tensor_uint128) endif() endif() @@ -145,6 +145,7 @@ ei_add_test_sycl(cxx11_tensor_forced_eval_sycl "-std=c++11") ei_add_test_sycl(cxx11_tensor_broadcast_sycl "-std=c++11") ei_add_test_sycl(cxx11_tensor_device_sycl "-std=c++11") + ei_add_test_sycl(cxx11_tensor_reduction_sycl "-std=c++11") endif(EIGEN_TEST_SYCL) # It should be safe to always run these tests as there is some fallback code for # older compiler that don't support cxx11.
diff --git a/unsupported/test/cxx11_tensor_broadcast_sycl.cpp b/unsupported/test/cxx11_tensor_broadcast_sycl.cpp index ecebf7d..7201bfe 100644 --- a/unsupported/test/cxx11_tensor_broadcast_sycl.cpp +++ b/unsupported/test/cxx11_tensor_broadcast_sycl.cpp
@@ -25,55 +25,50 @@ using Eigen::Tensor; using Eigen::TensorMap; -// Types used in tests: -using TestTensor = Tensor<float, 3>; -using TestTensorMap = TensorMap<Tensor<float, 3>>; -static void test_broadcast_sycl(){ +static void test_broadcast_sycl(const Eigen::SyclDevice &sycl_device){ - cl::sycl::gpu_selector s; - cl::sycl::queue q(s, [=](cl::sycl::exception_list l) { - for (const auto& e : l) { - try { - std::rethrow_exception(e); - } catch (cl::sycl::exception e) { - std::cout << e.what() << std::endl; + // BROADCAST test: + array<int, 4> in_range = {{2, 3, 5, 7}}; + array<int, 4> broadcasts = {{2, 3, 1, 4}}; + array<int, 4> out_range; // = in_range * broadcasts + for (size_t i = 0; i < out_range.size(); ++i) + out_range[i] = in_range[i] * broadcasts[i]; + + Tensor<float, 4> input(in_range); + Tensor<float, 4> out(out_range); + + for (size_t i = 0; i < in_range.size(); ++i) + VERIFY_IS_EQUAL(out.dimension(i), out_range[i]); + + + for (int i = 0; i < input.size(); ++i) + input(i) = static_cast<float>(i); + + float * gpu_in_data = static_cast<float*>(sycl_device.allocate(input.dimensions().TotalSize()*sizeof(float))); + float * gpu_out_data = static_cast<float*>(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(float))); + + TensorMap<Tensor<float, 4>> gpu_in(gpu_in_data, in_range); + TensorMap<Tensor<float, 4>> gpu_out(gpu_out_data, out_range); + sycl_device.memcpyHostToDevice(gpu_in_data, input.data(),(input.dimensions().TotalSize())*sizeof(float)); + gpu_out.device(sycl_device) = gpu_in.broadcast(broadcasts); + sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float)); + + for (int i = 0; i < 4; ++i) { + for (int j = 0; j < 9; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 28; ++l) { + VERIFY_IS_APPROX(input(i%2,j%3,k%5,l%7), out(i,j,k,l)); + } } } - }); - SyclDevice sycl_device(q); - // BROADCAST test: - array<int, 4> in_range = {{2, 3, 5, 7}}; - array<int, in_range.size()> broadcasts = {{2, 3, 1, 4}}; - array<int, in_range.size()> out_range; // = in_range * broadcasts - for (size_t i = 0; i < out_range.size(); ++i) - out_range[i] = in_range[i] * broadcasts[i]; - - Tensor<float, in_range.size()> input(in_range); - Tensor<float, out_range.size()> output(out_range); - - for (int i = 0; i < input.size(); ++i) - input(i) = static_cast<float>(i); - - TensorMap<decltype(input)> gpu_in(input.data(), in_range); - TensorMap<decltype(output)> gpu_out(output.data(), out_range); - gpu_out.device(sycl_device) = gpu_in.broadcast(broadcasts); - sycl_device.deallocate(output.data()); - - for (size_t i = 0; i < in_range.size(); ++i) - VERIFY_IS_EQUAL(output.dimension(i), out_range[i]); - - for (int i = 0; i < 4; ++i) { - for (int j = 0; j < 9; ++j) { - for (int k = 0; k < 5; ++k) { - for (int l = 0; l < 28; ++l) { - VERIFY_IS_APPROX(input(i%2,j%3,k%5,l%7), output(i,j,k,l)); - } - } - } - } - printf("Broadcast Test Passed\n"); + } + printf("Broadcast Test Passed\n"); + sycl_device.deallocate(gpu_in_data); + sycl_device.deallocate(gpu_out_data); } void test_cxx11_tensor_broadcast_sycl() { - CALL_SUBTEST(test_broadcast_sycl()); + cl::sycl::gpu_selector s; + Eigen::SyclDevice sycl_device(s); + CALL_SUBTEST(test_broadcast_sycl(sycl_device)); }
diff --git a/unsupported/test/cxx11_tensor_device_sycl.cpp b/unsupported/test/cxx11_tensor_device_sycl.cpp index f54fc87..7f79753 100644 --- a/unsupported/test/cxx11_tensor_device_sycl.cpp +++ b/unsupported/test/cxx11_tensor_device_sycl.cpp
@@ -20,20 +20,12 @@ #include "main.h" #include <unsupported/Eigen/CXX11/Tensor> -void test_device_sycl() { - cl::sycl::gpu_selector s; - cl::sycl::queue q(s, [=](cl::sycl::exception_list l) { - for (const auto& e : l) { - try { - std::rethrow_exception(e); - } catch (cl::sycl::exception e) { - std::cout << e.what() << std::endl; - } - } - }); - Eigen::SyclDevice sycl_device(q); - printf("Helo from ComputeCpp: Device Exists\n"); +void test_device_sycl(const Eigen::SyclDevice &sycl_device) { + std::cout <<"Helo from ComputeCpp: the requested device exists and the device name is : " + << sycl_device.m_queue.get_device(). template get_info<cl::sycl::info::device::name>() <<std::endl;; } void test_cxx11_tensor_device_sycl() { - CALL_SUBTEST(test_device_sycl()); + cl::sycl::gpu_selector s; + Eigen::SyclDevice sycl_device(s); + CALL_SUBTEST(test_device_sycl(sycl_device)); }
diff --git a/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp b/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp index 182ec7f..5690da7 100644 --- a/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp +++ b/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp
@@ -22,18 +22,7 @@ using Eigen::Tensor; -void test_forced_eval_sycl() { - cl::sycl::gpu_selector s; - cl::sycl::queue q(s, [=](cl::sycl::exception_list l) { - for (const auto& e : l) { - try { - std::rethrow_exception(e); - } catch (cl::sycl::exception e) { - std::cout << e.what() << std::endl; - } - } - }); - SyclDevice sycl_device(q); +void test_forced_eval_sycl(const Eigen::SyclDevice &sycl_device) { int sizeDim1 = 100; int sizeDim2 = 200; @@ -43,17 +32,22 @@ Eigen::Tensor<float, 3> in2(tensorRange); Eigen::Tensor<float, 3> out(tensorRange); + float * gpu_in1_data = static_cast<float*>(sycl_device.allocate(in1.dimensions().TotalSize()*sizeof(float))); + float * gpu_in2_data = static_cast<float*>(sycl_device.allocate(in2.dimensions().TotalSize()*sizeof(float))); + float * gpu_out_data = static_cast<float*>(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(float))); + in1 = in1.random() + in1.constant(10.0f); in2 = in2.random() + in2.constant(10.0f); - // creating TensorMap from tensor - Eigen::TensorMap<Eigen::Tensor<float, 3>> gpu_in1(in1.data(), tensorRange); - Eigen::TensorMap<Eigen::Tensor<float, 3>> gpu_in2(in2.data(), tensorRange); - Eigen::TensorMap<Eigen::Tensor<float, 3>> gpu_out(out.data(), tensorRange); - + // creating TensorMap from tensor + Eigen::TensorMap<Eigen::Tensor<float, 3>> gpu_in1(gpu_in1_data, tensorRange); + Eigen::TensorMap<Eigen::Tensor<float, 3>> gpu_in2(gpu_in2_data, tensorRange); + Eigen::TensorMap<Eigen::Tensor<float, 3>> gpu_out(gpu_out_data, tensorRange); + sycl_device.memcpyHostToDevice(gpu_in1_data, in1.data(),(in1.dimensions().TotalSize())*sizeof(float)); + sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in1.dimensions().TotalSize())*sizeof(float)); /// c=(a+b)*b - gpu_out.device(sycl_device) =(gpu_in1 + gpu_in2).eval() * gpu_in2; - sycl_device.deallocate(out.data()); + gpu_out.device(sycl_device) =(gpu_in1 + gpu_in2).eval() * gpu_in2; + sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -62,7 +56,15 @@ } } } - printf("(a+b)*b Test Passed\n"); + printf("(a+b)*b Test Passed\n"); + sycl_device.deallocate(gpu_in1_data); + sycl_device.deallocate(gpu_in2_data); + sycl_device.deallocate(gpu_out_data); + } -void test_cxx11_tensor_forced_eval_sycl() { CALL_SUBTEST(test_forced_eval_sycl()); } +void test_cxx11_tensor_forced_eval_sycl() { + cl::sycl::gpu_selector s; + Eigen::SyclDevice sycl_device(s); + CALL_SUBTEST(test_forced_eval_sycl(sycl_device)); +}
diff --git a/unsupported/test/cxx11_tensor_reduction_sycl.cpp b/unsupported/test/cxx11_tensor_reduction_sycl.cpp new file mode 100644 index 0000000..a9ef829 --- /dev/null +++ b/unsupported/test/cxx11_tensor_reduction_sycl.cpp
@@ -0,0 +1,138 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2015 +// Mehdi Goli Codeplay Software Ltd. +// Ralph Potter Codeplay Software Ltd. +// Luke Iwanski Codeplay Software Ltd. +// Contact: <eigen@codeplay.com> +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#define EIGEN_TEST_NO_LONGDOUBLE +#define EIGEN_TEST_NO_COMPLEX +#define EIGEN_TEST_FUNC cxx11_tensor_reduction_sycl +#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int +#define EIGEN_USE_SYCL + +#include "main.h" +#include <unsupported/Eigen/CXX11/Tensor> + + + +static void test_full_reductions_sycl(const Eigen::SyclDevice& sycl_device) { + + const int num_rows = 452; + const int num_cols = 765; + array<int, 2> tensorRange = {{num_rows, num_cols}}; + + Tensor<float, 2> in(tensorRange); + Tensor<float, 0> full_redux; + Tensor<float, 0> full_redux_gpu; + + in.setRandom(); + + full_redux = in.sum(); + + float* gpu_in_data = static_cast<float*>(sycl_device.allocate(in.dimensions().TotalSize()*sizeof(float))); + float* gpu_out_data =(float*)sycl_device.allocate(sizeof(float)); + + TensorMap<Tensor<float, 2> > in_gpu(gpu_in_data, tensorRange); + TensorMap<Tensor<float, 0> > out_gpu(gpu_out_data); + + sycl_device.memcpyHostToDevice(gpu_in_data, in.data(),(in.dimensions().TotalSize())*sizeof(float)); + out_gpu.device(sycl_device) = in_gpu.sum(); + sycl_device.memcpyDeviceToHost(full_redux_gpu.data(), gpu_out_data, sizeof(float)); + // Check that the CPU and GPU reductions return the same result. + VERIFY_IS_APPROX(full_redux_gpu(), full_redux()); + + sycl_device.deallocate(gpu_in_data); + sycl_device.deallocate(gpu_out_data); +} + +static void test_first_dim_reductions_sycl(const Eigen::SyclDevice& sycl_device) { + + int dim_x = 145; + int dim_y = 1; + int dim_z = 67; + + array<int, 3> tensorRange = {{dim_x, dim_y, dim_z}}; + Eigen::array<int, 1> red_axis; + red_axis[0] = 0; + array<int, 2> reduced_tensorRange = {{dim_y, dim_z}}; + + Tensor<float, 3> in(tensorRange); + Tensor<float, 2> redux(reduced_tensorRange); + Tensor<float, 2> redux_gpu(reduced_tensorRange); + + in.setRandom(); + + redux= in.sum(red_axis); + + float* gpu_in_data = static_cast<float*>(sycl_device.allocate(in.dimensions().TotalSize()*sizeof(float))); + float* gpu_out_data = static_cast<float*>(sycl_device.allocate(redux_gpu.dimensions().TotalSize()*sizeof(float))); + + TensorMap<Tensor<float, 3> > in_gpu(gpu_in_data, tensorRange); + TensorMap<Tensor<float, 2> > out_gpu(gpu_out_data, reduced_tensorRange); + + sycl_device.memcpyHostToDevice(gpu_in_data, in.data(),(in.dimensions().TotalSize())*sizeof(float)); + out_gpu.device(sycl_device) = in_gpu.sum(red_axis); + sycl_device.memcpyDeviceToHost(redux_gpu.data(), gpu_out_data, redux_gpu.dimensions().TotalSize()*sizeof(float)); + + // Check that the CPU and GPU reductions return the same result. + for(int j=0; j<reduced_tensorRange[0]; j++ ) + for(int k=0; k<reduced_tensorRange[1]; k++ ) + VERIFY_IS_APPROX(redux_gpu(j,k), redux(j,k)); + + sycl_device.deallocate(gpu_in_data); + sycl_device.deallocate(gpu_out_data); +} + +static void test_last_dim_reductions_sycl(const Eigen::SyclDevice &sycl_device) { + + int dim_x = 567; + int dim_y = 1; + int dim_z = 47; + + array<int, 3> tensorRange = {{dim_x, dim_y, dim_z}}; + Eigen::array<int, 1> red_axis; + red_axis[0] = 2; + array<int, 2> reduced_tensorRange = {{dim_x, dim_y}}; + + Tensor<float, 3> in(tensorRange); + Tensor<float, 2> redux(reduced_tensorRange); + Tensor<float, 2> redux_gpu(reduced_tensorRange); + + in.setRandom(); + + redux= in.sum(red_axis); + + float* gpu_in_data = static_cast<float*>(sycl_device.allocate(in.dimensions().TotalSize()*sizeof(float))); + float* gpu_out_data = static_cast<float*>(sycl_device.allocate(redux_gpu.dimensions().TotalSize()*sizeof(float))); + + TensorMap<Tensor<float, 3> > in_gpu(gpu_in_data, tensorRange); + TensorMap<Tensor<float, 2> > out_gpu(gpu_out_data, reduced_tensorRange); + + sycl_device.memcpyHostToDevice(gpu_in_data, in.data(),(in.dimensions().TotalSize())*sizeof(float)); + out_gpu.device(sycl_device) = in_gpu.sum(red_axis); + sycl_device.memcpyDeviceToHost(redux_gpu.data(), gpu_out_data, redux_gpu.dimensions().TotalSize()*sizeof(float)); + // Check that the CPU and GPU reductions return the same result. + for(int j=0; j<reduced_tensorRange[0]; j++ ) + for(int k=0; k<reduced_tensorRange[1]; k++ ) + VERIFY_IS_APPROX(redux_gpu(j,k), redux(j,k)); + + sycl_device.deallocate(gpu_in_data); + sycl_device.deallocate(gpu_out_data); + +} + +void test_cxx11_tensor_reduction_sycl() { + cl::sycl::gpu_selector s; + Eigen::SyclDevice sycl_device(s); + CALL_SUBTEST((test_full_reductions_sycl(sycl_device))); + CALL_SUBTEST((test_first_dim_reductions_sycl(sycl_device))); + CALL_SUBTEST((test_last_dim_reductions_sycl(sycl_device))); + +}
diff --git a/unsupported/test/cxx11_tensor_sycl.cpp b/unsupported/test/cxx11_tensor_sycl.cpp index 0f66cd8..6a9c334 100644 --- a/unsupported/test/cxx11_tensor_sycl.cpp +++ b/unsupported/test/cxx11_tensor_sycl.cpp
@@ -27,42 +27,33 @@ using Eigen::Tensor; using Eigen::TensorMap; -// Types used in tests: -using TestTensor = Tensor<float, 3>; -using TestTensorMap = TensorMap<Tensor<float, 3>>; - -void test_sycl_cpu() { - cl::sycl::gpu_selector s; - cl::sycl::queue q(s, [=](cl::sycl::exception_list l) { - for (const auto& e : l) { - try { - std::rethrow_exception(e); - } catch (cl::sycl::exception e) { - std::cout << e.what() << std::endl; - } - } - }); - SyclDevice sycl_device(q); +void test_sycl_cpu(const Eigen::SyclDevice &sycl_device) { int sizeDim1 = 100; int sizeDim2 = 100; int sizeDim3 = 100; array<int, 3> tensorRange = {{sizeDim1, sizeDim2, sizeDim3}}; - TestTensor in1(tensorRange); - TestTensor in2(tensorRange); - TestTensor in3(tensorRange); - TestTensor out(tensorRange); - in1 = in1.random(); + Tensor<float, 3> in1(tensorRange); + Tensor<float, 3> in2(tensorRange); + Tensor<float, 3> in3(tensorRange); + Tensor<float, 3> out(tensorRange); + in2 = in2.random(); in3 = in3.random(); - TestTensorMap gpu_in1(in1.data(), tensorRange); - TestTensorMap gpu_in2(in2.data(), tensorRange); - TestTensorMap gpu_in3(in3.data(), tensorRange); - TestTensorMap gpu_out(out.data(), tensorRange); + + float * gpu_in1_data = static_cast<float*>(sycl_device.allocate(in1.dimensions().TotalSize()*sizeof(float))); + float * gpu_in2_data = static_cast<float*>(sycl_device.allocate(in2.dimensions().TotalSize()*sizeof(float))); + float * gpu_in3_data = static_cast<float*>(sycl_device.allocate(in3.dimensions().TotalSize()*sizeof(float))); + float * gpu_out_data = static_cast<float*>(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(float))); + + TensorMap<Tensor<float, 3>> gpu_in1(gpu_in1_data, tensorRange); + TensorMap<Tensor<float, 3>> gpu_in2(gpu_in2_data, tensorRange); + TensorMap<Tensor<float, 3>> gpu_in3(gpu_in3_data, tensorRange); + TensorMap<Tensor<float, 3>> gpu_out(gpu_out_data, tensorRange); /// a=1.2f gpu_in1.device(sycl_device) = gpu_in1.constant(1.2f); - sycl_device.deallocate(in1.data()); + sycl_device.memcpyDeviceToHost(in1.data(), gpu_in1_data ,(in1.dimensions().TotalSize())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -74,7 +65,7 @@ /// a=b*1.2f gpu_out.device(sycl_device) = gpu_in1 * 1.2f; - sycl_device.deallocate(out.data()); + sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data ,(out.dimensions().TotalSize())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -86,8 +77,9 @@ printf("a=b*1.2f Test Passed\n"); /// c=a*b + sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in2.dimensions().TotalSize())*sizeof(float)); gpu_out.device(sycl_device) = gpu_in1 * gpu_in2; - sycl_device.deallocate(out.data()); + sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -101,7 +93,7 @@ /// c=a+b gpu_out.device(sycl_device) = gpu_in1 + gpu_in2; - sycl_device.deallocate(out.data()); + sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -115,7 +107,7 @@ /// c=a*a gpu_out.device(sycl_device) = gpu_in1 * gpu_in1; - sycl_device.deallocate(out.data()); + sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -125,12 +117,11 @@ } } } - printf("c= a*a Test Passed\n"); //a*3.14f + b*2.7f gpu_out.device(sycl_device) = gpu_in1 * gpu_in1.constant(3.14f) + gpu_in2 * gpu_in2.constant(2.7f); - sycl_device.deallocate(out.data()); + sycl_device.memcpyDeviceToHost(out.data(),gpu_out_data,(out.dimensions().TotalSize())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -143,8 +134,9 @@ printf("a*3.14f + b*2.7f Test Passed\n"); ///d= (a>0.5? b:c) + sycl_device.memcpyHostToDevice(gpu_in3_data, in3.data(),(in3.dimensions().TotalSize())*sizeof(float)); gpu_out.device(sycl_device) =(gpu_in1 > gpu_in1.constant(0.5f)).select(gpu_in2, gpu_in3); - sycl_device.deallocate(out.data()); + sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float)); for (int i = 0; i < sizeDim1; ++i) { for (int j = 0; j < sizeDim2; ++j) { for (int k = 0; k < sizeDim3; ++k) { @@ -155,8 +147,13 @@ } } printf("d= (a>0.5? b:c) Test Passed\n"); - + sycl_device.deallocate(gpu_in1_data); + sycl_device.deallocate(gpu_in2_data); + sycl_device.deallocate(gpu_in3_data); + sycl_device.deallocate(gpu_out_data); } void test_cxx11_tensor_sycl() { - CALL_SUBTEST(test_sycl_cpu()); + cl::sycl::gpu_selector s; + Eigen::SyclDevice sycl_device(s); + CALL_SUBTEST(test_sycl_cpu(sycl_device)); }