Merge C++ Standard Parallelism and SYCL2020 implementations

UoB-HPC · Jun 5, 2024 · b3786f6 · b3786f6
1 parent 9ff46ec
commit b3786f6
Show file tree

Hide file tree

Showing 22 changed files with 359 additions and 1,164 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -154,15 +154,12 @@ include(cmake/register_models.cmake)
 register_model(serial SERIAL SerialStream.cpp)
 register_model(omp OMP OMPStream.cpp)
 register_model(ocl OCL OCLStream.cpp)
-register_model(std-data STD_DATA STDDataStream.cpp)
-register_model(std-indices STD_INDICES STDIndicesStream.cpp)
-register_model(std-ranges STD_RANGES STDRangesStream.cpp)
+register_model(std STD STDStream.cpp)
 register_model(hip HIP HIPStream.cpp)
 register_model(cuda CUDA CUDAStream.cu)
 register_model(kokkos KOKKOS KokkosStream.cpp)
 register_model(sycl SYCL SYCLStream.cpp)
-register_model(sycl2020-acc SYCL2020 SYCLStream2020.cpp)
-register_model(sycl2020-usm SYCL2020 SYCLStream2020.cpp)
+register_model(sycl2020 SYCL2020 SYCLStream2020.cpp)
 register_model(acc ACC ACCStream.cpp)
 # defining RAJA collides with the RAJA namespace so USE_RAJA
 register_model(raja USE_RAJA RAJAStream.cpp)

diff --git a/src/Stream.h b/src/Stream.h
@@ -7,11 +7,14 @@
 
 #pragma once
 
+#include <cstdint>
 #include <array>
 #include <vector>
 #include <string>
 #include "benchmark.h"
 
+using std::intptr_t;
+
 template <class T>
 class Stream
 {

diff --git a/src/StreamModels.h b/src/StreamModels.h
@@ -3,12 +3,8 @@
 
 #if defined(CUDA)
 #include "CUDAStream.h"
-#elif defined(STD_DATA)
-#include "STDDataStream.h"
-#elif defined(STD_INDICES)
-#include "STDIndicesStream.h"
-#elif defined(STD_RANGES)
-#include "STDRangesStream.hpp"
+#elif defined(STD)
+#include "STDStream.h"
 #elif defined(TBB)
 #include "TBBStream.hpp"
 #elif defined(THRUST)
@@ -63,17 +59,9 @@ std::unique_ptr<Stream<T>> make_stream(Args... args) {
   // Use the Kokkos implementation
   return std::make_unique<KokkosStream<T>>(args...);
 
-#elif defined(STD_DATA)
+#elif defined(STD)
   // Use the C++ STD data-oriented implementation
-  return std::make_unique<STDDataStream<T>>(args...);
-
-#elif defined(STD_INDICES)
-  // Use the C++ STD index-oriented implementation
-  return std::make_unique<STDIndicesStream<T>>(args...);
-
-#elif defined(STD_RANGES)
-  // Use the C++ STD ranges implementation
-  return std::make_unique<STDRangesStream<T>>(args...);
+  return std::make_unique<STDStream<T>>(args...);
 
 #elif defined(TBB)
   // Use the C++20 implementation

diff --git a/src/ci-test-compile.sh b/src/ci-test-compile.sh
@@ -152,9 +152,10 @@ build_gcc() {
       *)   dpl_conditional_flags="-DFETCH_ONEDPL=ON -DFETCH_TBB=ON -DUSE_TBB=ON -DCXX_EXTRA_FLAGS=-D_GLIBCXX_USE_TBB_PAR_BACKEND=0" ;;
     esac
     # some distributions like Ubuntu bionic implements std par with TBB, so conditionally link it here
-    run_build $name "${GCC_CXX:?}" std-data    "$cxx $dpl_conditional_flags -DUSE_ONEDPL=$use_onedpl"
-    run_build $name "${GCC_CXX:?}" std-indices "$cxx $dpl_conditional_flags -DUSE_ONEDPL=$use_onedpl"
-    run_build $name "${GCC_CXX:?}" std-ranges  "$cxx $dpl_conditional_flags -DUSE_ONEDPL=$use_onedpl"
+    run_build $name "${GCC_CXX:?}" std    "$cxx $dpl_conditional_flags -DUSE_ONEDPL=$use_onedpl -DSTDIMPL=DATA17"
+    # Requires GCC 14 and newer CMake for C++23 support    
+    #run_build $name "${GCC_CXX:?}" std    "$cxx $dpl_conditional_flags -DUSE_ONEDPL=$use_onedpl -DSTDIMPL=DATA23"
+    run_build $name "${GCC_CXX:?}" std    "$cxx $dpl_conditional_flags -DUSE_ONEDPL=$use_onedpl -DSTDIMPL=INDICES"
   done
 
   run_build $name "${GCC_CXX:?}" tbb "$cxx -DONE_TBB_DIR=$TBB_LIB"
@@ -251,9 +252,11 @@ build_clang() {
       OFF) dpl_conditional_flags="-DCXX_EXTRA_LIBRARIES=${CLANG_STD_PAR_LIB:-}" ;;
       *)   dpl_conditional_flags="-DFETCH_ONEDPL=ON -DFETCH_TBB=ON -DUSE_TBB=ON -DCXX_EXTRA_FLAGS=-D_GLIBCXX_USE_TBB_PAR_BACKEND=0"  ;;
     esac
-    run_build $name "${CLANG_CXX:?}" std-data     "$cxx $dpl_conditional_flags -DUSE_ONEDPL=$use_onedpl"
-    run_build $name "${CLANG_CXX:?}" std-indices  "$cxx $dpl_conditional_flags -DUSE_ONEDPL=$use_onedpl"
-    # run_build $name "${CLANG_CXX:?}" std-ranges "$cxx $dpl_conditional_flags -DUSE_ONEDPL=$use_onedpl" # not yet supported
+    run_build $name "${CLANG_CXX:?}" std  "$cxx $dpl_conditional_flags -DUSE_ONEDPL=$use_onedpl -DSTDIMPL=DATA17"
+    # Requires GCC 14 and newer CMake for C++23 support
+    # run_build $name "${CLANG_CXX:?}" std  "$cxx $dpl_conditional_flags -DUSE_ONEDPL=$use_onedpl -DSTDIMPL=DATA23"
+    # TODO: clang is too old
+    #run_build $name "${CLANG_CXX:?}" std  "$cxx $dpl_conditional_flags -DUSE_ONEDPL=$use_onedpl -DSTDIMPL=INDICES"    
   done
 
   run_build $name "${CLANG_CXX:?}" tbb "$cxx -DONE_TBB_DIR=$TBB_LIB"
@@ -270,8 +273,10 @@ build_clang() {
 build_nvhpc() {
   local name="nvhpc_build"
   local cxx="-DCMAKE_CXX_COMPILER=${NVHPC_NVCXX:?}"
-  run_build $name "${NVHPC_NVCXX:?}" std-data "$cxx -DNVHPC_OFFLOAD=$NV_ARCH_CCXY"
-  run_build $name "${NVHPC_NVCXX:?}" std-indices "$cxx -DNVHPC_OFFLOAD=$NV_ARCH_CCXY"
+  run_build $name "${NVHPC_NVCXX:?}" std "$cxx -DNVHPC_OFFLOAD=$NV_ARCH_CCXY -DSTDIMPL=DATA17"
+  # Requires GCC 14 and newer CMake for C++23 support
+  # run_build $name "${NVHPC_NVCXX:?}" std "$cxx -DNVHPC_OFFLOAD=$NV_ARCH_CCXY -DSTDIMPL=DATA23"
+  run_build $name "${NVHPC_NVCXX:?}" std "$cxx -DNVHPC_OFFLOAD=$NV_ARCH_CCXY -DSTDIMPL=INDICES"  
 
   run_build $name "${NVHPC_NVCXX:?}" acc "$cxx -DTARGET_DEVICE=gpu -DTARGET_PROCESSOR=px -DCUDA_ARCH=$NV_ARCH_CCXY"
   run_build $name "${NVHPC_NVCXX:?}" acc "$cxx -DTARGET_DEVICE=multicore -DTARGET_PROCESSOR=zen"

diff --git a/src/dpl_shim.h b/src/dpl_shim.h
@@ -29,12 +29,15 @@ T *alloc_raw(size_t size) { return sycl::malloc_shared<T>(size, exe_policy.queue
 template<typename T>
 void dealloc_raw(T *ptr) { sycl::free(ptr, exe_policy.queue()); }
 
+#define WORKAROUND
+
 #else
 
 // auto exe_policy = dpl::execution::seq;
 // auto exe_policy = dpl::execution::par;
 static constexpr auto exe_policy = dpl::execution::par_unseq;
 #define USE_STD_PTR_ALLOC_DEALLOC
+#define WORKAROUND
 
 #endif
 

diff --git a/src/main.cpp b/src/main.cpp
@@ -332,21 +332,22 @@ void check_solution(const size_t num_times, T const* a, T const* b, T const* c,
   size_t failed = 0;
   T max_rel = std::numeric_limits<T>::epsilon() * T(100.0);
   T max_rel_dot = std::numeric_limits<T>::epsilon() * T(10000000.0);
-  auto check = [&](const char* name, T is, T should, T max_rel, size_t i = size_t(-1)) {
+  auto check = [&](const char* name, T is, T should, T mrel, size_t i = size_t(-1)) {
     // Relative difference:
     T diff = std::abs(is - should);
     T abs_is = std::abs(is);
     T abs_sh = std::abs(should);
     T largest = std::max(abs_is, abs_sh);
-    T same = diff <= largest * max_rel;
+    T same = diff <= largest * mrel;
     if (!same || std::isnan(is)) {
       ++failed;
       if (failed > 10) return;
       std::cerr << "FAILED validation of " << name;
       if (i != size_t(-1)) std::cerr << "[" << i << "]";
       std::cerr << ": " << is << " (is) != " << should
 		<< " (should)" << ", diff=" << diff << " > "
-		<< largest * max_rel << std::endl;
+		<< largest * mrel << " (largest=" << largest
+		<< ", max_rel=" << mrel << ")" << std::endl;
     }
   };
 
@@ -360,9 +361,9 @@ void check_solution(const size_t num_times, T const* a, T const* b, T const* c,
 
   // Calculate the L^infty-norm relative error
   for (size_t i = 0; i < array_size; ++i) {
-    check("a", a[i], goldA, i, max_rel);
-    check("b", b[i], goldB, i, max_rel);
-    check("c", c[i], goldC, i, max_rel);
+    check("a", a[i], goldA, max_rel, i);
+    check("b", b[i], goldB, max_rel, i);
+    check("c", c[i], goldC, max_rel, i);
   }
 
   if (failed > 0 && !silence_errors)

diff --git a/src/std-data/STDDataStream.cpp b/src/std-data/STDDataStream.cpp