cory-master/BabelStream
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Split implementation for index/data oriented std variants

Browse Source 
Fix missing range iterator functions for std-indices
Rename std20 to std-ranges
This commit is contained in:
Tom Lin 2021-12-15 21:38:56 +00:00
parent 881c0cc706
commit f9bba3c0de
12 changed files with 251 additions and 57 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
CMakeLists.txt
View File

@ -34,9 +34,9 @@ endmacro()
#set(CUDA_ARCH sm_70)
#set(BLT_DIR /home/tom/Downloads/blt-0.3.6/)
#set(MODEL STD)
#set(ARCH cc70)
#set(CXX_EXTRA_FLAGS -v)
#set(MODEL std-data)
#set(CMAKE_CXX_COMPILER /home/tom/Downloads/nvhpc_2021_219_Linux_x86_64_cuda_multi/install_components/Linux_x86_64/21.9/compilers/bin/nvc++)
#set(NVHPC_OFFLOAD "cuda11.4,cc61")
#set(MODEL CUDA)
#set(ARCH sm_70)
@ -117,8 +117,9 @@ include(cmake/register_models.cmake)
# register out models <model_name> <preprocessor_def_name> <source files...>
register_model(omp OMP OMPStream.cpp)
register_model(ocl OCL OCLStream.cpp)
register_model(std STD STDStream.cpp)
register_model(std20 STD20 STD20Stream.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(hip HIP HIPStream.cpp)
register_model(cuda CUDA CUDAStream.cu)
register_model(kokkos KOKKOS KokkosStream.cpp)

16
src/ci-test-compile.sh
View File

@ -115,7 +115,7 @@ run_build() {
# GCC_STD_PAR_LIB="tbb"
# CLANG_STD_PAR_LIB="tbb"
# GCC_OMP_OFFLOAD_AMD=false
# GCC_OMP_OFFLOAD_NVIDIA=true
# GCC_OMP_OFFLOAD_NVIDIA=false
# CLANG_OMP_OFFLOAD_AMD=false
# CLANG_OMP_OFFLOAD_NVIDIA=false
###
@ -136,8 +136,9 @@ build_gcc() {
fi
# some distributions like Ubuntu bionic implements std par with TBB, so conditionally link it here
run_build $name "${GCC_CXX:?}" std "$cxx -DCXX_EXTRA_LIBRARIES=${GCC_STD_PAR_LIB:-}"
run_build $name "${GCC_CXX:?}" std20 "$cxx -DCXX_EXTRA_LIBRARIES=${GCC_STD_PAR_LIB:-}"
run_build $name "${GCC_CXX:?}" std-data "$cxx -DCXX_EXTRA_LIBRARIES=${GCC_STD_PAR_LIB:-}"
run_build $name "${GCC_CXX:?}" std-indices "$cxx -DCXX_EXTRA_LIBRARIES=${GCC_STD_PAR_LIB:-}"
run_build $name "${GCC_CXX:?}" std-ranges "$cxx -DCXX_EXTRA_LIBRARIES=${GCC_STD_PAR_LIB:-}"
run_build $name "${GCC_CXX:?}" tbb "$cxx -DONE_TBB_DIR=$TBB_LIB"
run_build $name "${GCC_CXX:?}" tbb "$cxx" # build TBB again with the system TBB
@ -211,7 +212,8 @@ build_clang() {
run_build $name "${CLANG_CXX:?}" cuda "$cxx -DCMAKE_CUDA_COMPILER=${NVHPC_NVCC:?} -DCUDA_ARCH=$NV_ARCH -DMEM=PAGEFAULT"
run_build $name "${CLANG_CXX:?}" kokkos "$cxx -DKOKKOS_IN_TREE=${KOKKOS_SRC:?} -DKokkos_ENABLE_OPENMP=ON"
run_build $name "${CLANG_CXX:?}" ocl "$cxx -DOpenCL_LIBRARY=${OCL_LIB:?}"
run_build $name "${CLANG_CXX:?}" std "$cxx -DCXX_EXTRA_LIBRARIES=${CLANG_STD_PAR_LIB:-}"
run_build $name "${CLANG_CXX:?}" std-data "$cxx -DCXX_EXTRA_LIBRARIES=${CLANG_STD_PAR_LIB:-}"
run_build $name "${CLANG_CXX:?}" std-indices "$cxx -DCXX_EXTRA_LIBRARIES=${CLANG_STD_PAR_LIB:-}"
# run_build $name "${LANG_CXX:?}" std20 "$cxx -DCXX_EXTRA_LIBRARIES=${CLANG_STD_PAR_LIB:-}" # not yet supported
run_build $name "${CLANG_CXX:?}" raja "$cxx -DRAJA_IN_TREE=${RAJA_SRC:?}"
run_build $name "${CLANG_CXX:?}" cuda "$cxx -DCMAKE_CUDA_COMPILER=${NVHPC_NVCC:?} -DCUDA_ARCH=$NV_ARCH"
@ -219,7 +221,8 @@ build_clang() {
run_build $name "${CLANG_CXX:?}" cuda "$cxx -DCMAKE_CUDA_COMPILER=${NVHPC_NVCC:?} -DCUDA_ARCH=$NV_ARCH -DMEM=PAGEFAULT"
run_build $name "${CLANG_CXX:?}" kokkos "$cxx -DKOKKOS_IN_TREE=${KOKKOS_SRC:?} -DKokkos_ENABLE_OPENMP=ON"
run_build $name "${CLANG_CXX:?}" ocl "$cxx -DOpenCL_LIBRARY=${OCL_LIB:?}"
run_build $name "${CLANG_CXX:?}" std "$cxx -DCXX_EXTRA_LIBRARIES=${CLANG_STD_PAR_LIB:-}"
run_build $name "${CLANG_CXX:?}" std-data "$cxx -DCXX_EXTRA_LIBRARIES=${CLANG_STD_PAR_LIB:-}"
run_build $name "${CLANG_CXX:?}" std-indices "$cxx -DCXX_EXTRA_LIBRARIES=${CLANG_STD_PAR_LIB:-}"
# run_build $name "${LANG_CXX:?}" std20 "$cxx -DCXX_EXTRA_LIBRARIES=${CLANG_STD_PAR_LIB:-}" # not yet supported
run_build $name "${CLANG_CXX:?}" tbb "$cxx -DONE_TBB_DIR=$TBB_LIB"
@ -232,7 +235,8 @@ build_clang() {
build_nvhpc() {
local name="nvhpc_build"
local cxx="-DCMAKE_CXX_COMPILER=${NVHPC_NVCXX:?}"
run_build $name "${NVHPC_NVCXX:?}" std "$cxx -DNVHPC_OFFLOAD=$NV_ARCH_CCXY"
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:?}" 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"
}

26
src/main.cpp
View File

@ -21,10 +21,12 @@
#if defined(CUDA)
#include "CUDAStream.h"
#elif defined(STD)
#include "STDStream.h"
#elif defined(STD20)
#include "STD20Stream.hpp"
#elif defined(STD_DATA)
#include "STDDataStream.h"
#elif defined(STD_INDICES)
#include "STDIndicesStream.h"
#elif defined(STD_RANGES)
#include "STDRangesStream.hpp"
#elif defined(TBB)
#include "TBBStream.hpp"
#elif defined(THRUST)
@ -264,13 +266,17 @@ void run()
// Use the Kokkos implementation
stream = new KokkosStream<T>(ARRAY_SIZE, deviceIndex);
#elif defined(STD)
// Use the STD implementation
stream = new STDStream<T>(ARRAY_SIZE, deviceIndex);
#elif defined(STD_DATA)
// Use the C++ STD data-oriented implementation
stream = new STDDataStream<T>(ARRAY_SIZE, deviceIndex);
#elif defined(STD20)
// Use the C++20 implementation
stream = new STD20Stream<T>(ARRAY_SIZE, deviceIndex);
#elif defined(STD_INDICES)
// Use the C++ STD index-oriented implementation
stream = new STDIndicesStream<T>(ARRAY_SIZE, deviceIndex);
#elif defined(STD_RANGES)
// Use the C++ STD ranges implementation
stream = new STDRangesStream<T>(ARRAY_SIZE, deviceIndex);
#elif defined(TBB)
// Use the C++20 implementation

103
src/std-data/STDDataStream.cpp Normal file
View File

@ -0,0 +1,103 @@
// Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
//
// For full license terms please see the LICENSE file distributed with this
// source code
#include "STDDataStream.h"
#include <algorithm>
#include <execution>
#include <numeric>
// There are three execution policies:
// auto exe_policy = std::execution::seq;
// auto exe_policy = std::execution::par;
auto exe_policy = std::execution::par_unseq;
template <class T>
STDDataStream<T>::STDDataStream(const int ARRAY_SIZE, int device)
noexcept : array_size{ARRAY_SIZE}, a(array_size), b(array_size), c(array_size)
{
}
template <class T>
void STDDataStream<T>::init_arrays(T initA, T initB, T initC)
{
std::fill(exe_policy, a.begin(), a.end(), initA);
std::fill(exe_policy, b.begin(), b.end(), initB);
std::fill(exe_policy, c.begin(), c.end(), initC);
}
template <class T>
void STDDataStream<T>::read_arrays(std::vector<T>& h_a, std::vector<T>& h_b, std::vector<T>& h_c)
{
h_a = a;
h_b = b;
h_c = c;
}
template <class T>
void STDDataStream<T>::copy()
{
// c[i] = a[i]
std::copy(exe_policy, a.begin(), a.end(), c.begin());
}
template <class T>
void STDDataStream<T>::mul()
{
// b[i] = scalar * c[i];
std::transform(exe_policy, c.begin(), c.end(), b.begin(), [scalar = startScalar](T ci){ return scalar*ci; });
}
template <class T>
void STDDataStream<T>::add()
{
// c[i] = a[i] + b[i];
std::transform(exe_policy, a.begin(), a.end(), b.begin(), c.begin(), std::plus<T>());
}
template <class T>
void STDDataStream<T>::triad()
{
// a[i] = b[i] + scalar * c[i];
std::transform(exe_policy, b.begin(), b.end(), c.begin(), a.begin(), [scalar = startScalar](T bi, T ci){ return bi+scalar*ci; });
}
template <class T>
void STDDataStream<T>::nstream()
{
// a[i] += b[i] + scalar * c[i];
// Need to do in two stages with C++11 STL.
// 1: a[i] += b[i]
// 2: a[i] += scalar * c[i];
std::transform(exe_policy, a.begin(), a.end(), b.begin(), a.begin(), [](T ai, T bi){ return ai + bi; });
std::transform(exe_policy, a.begin(), a.end(), c.begin(), a.begin(), [scalar = startScalar](T ai, T ci){ return ai + scalar*ci; });
}
template <class T>
T STDDataStream<T>::dot()
{
// sum = 0; sum += a[i]*b[i]; return sum;
return std::transform_reduce(exe_policy, a.begin(), a.end(), b.begin(), 0.0);
}
void listDevices(void)
{
std::cout << "Listing devices is not supported by the Parallel STL" << std::endl;
}
std::string getDeviceName(const int)
{
return std::string("Device name unavailable");
}
std::string getDeviceDriver(const int)
{
return std::string("Device driver unavailable");
}
template class STDDataStream<float>;
template class STDDataStream<double>;

42
src/std-data/STDDataStream.h Normal file
View File

@ -0,0 +1,42 @@
// Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
//
// For full license terms please see the LICENSE file distributed with this
// source code
#pragma once
#include <iostream>
#include <stdexcept>
#include "Stream.h"
#define IMPLEMENTATION_STRING "STD (data-oriented)"
template <class T>
class STDDataStream : public Stream<T>
{
protected:
// Size of arrays
int array_size;
// Device side pointers
std::vector<T> a;
std::vector<T> b;
std::vector<T> c;
public:
STDDataStream(const int, int) noexcept;
~STDDataStream() = default;
virtual void copy() override;
virtual void add() override;
virtual void mul() override;
virtual void triad() override;
virtual void nstream() override;
virtual T dot() override;
virtual void init_arrays(T initA, T initB, T initC) override;
virtual void read_arrays(std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
};

0
src/std/model.cmake → src/std-data/model.cmake
View File

24
src/std/STDStream.cpp → src/std-indices/STDIndicesStream.cpp
View File

@ -3,7 +3,7 @@
// For full license terms please see the LICENSE file distributed with this
// source code
#include "STDStream.h"
#include "STDIndicesStream.h"
#include <algorithm>
#include <execution>
@ -16,13 +16,13 @@ auto exe_policy = std::execution::par_unseq;
template <class T>
STDStream<T>::STDStream(const int ARRAY_SIZE, int device)
STDIndicesStream<T>::STDIndicesStream(const int ARRAY_SIZE, int device)
noexcept : array_size{ARRAY_SIZE}, range(0, array_size), a(array_size), b(array_size), c(array_size)
{
}
template <class T>
void STDStream<T>::init_arrays(T initA, T initB, T initC)
void STDIndicesStream<T>::init_arrays(T initA, T initB, T initC)
{
std::fill(exe_policy, a.begin(), a.end(), initA);
std::fill(exe_policy, b.begin(), b.end(), initB);
@ -30,7 +30,7 @@ void STDStream<T>::init_arrays(T initA, T initB, T initC)
}
template <class T>
void STDStream<T>::read_arrays(std::vector<T>& h_a, std::vector<T>& h_b, std::vector<T>& h_c)
void STDIndicesStream<T>::read_arrays(std::vector<T>& h_a, std::vector<T>& h_b, std::vector<T>& h_c)
{
h_a = a;
h_b = b;
@ -38,14 +38,14 @@ void STDStream<T>::read_arrays(std::vector<T>& h_a, std::vector<T>& h_b, std::ve
}
template <class T>
void STDStream<T>::copy()
void STDIndicesStream<T>::copy()
{
// c[i] = a[i]
std::copy(exe_policy, a.begin(), a.end(), c.begin());
}
template <class T>
void STDStream<T>::mul()
void STDIndicesStream<T>::mul()
{
// b[i] = scalar * c[i];
std::transform(exe_policy, range.begin(), range.end(), b.begin(), [&, scalar = startScalar](int i) {
@ -54,7 +54,7 @@ void STDStream<T>::mul()
}
template <class T>
void STDStream<T>::add()
void STDIndicesStream<T>::add()
{
// c[i] = a[i] + b[i];
std::transform(exe_policy, range.begin(), range.end(), c.begin(), [&](int i) {
@ -63,7 +63,7 @@ void STDStream<T>::add()
}
template <class T>
void STDStream<T>::triad()
void STDIndicesStream<T>::triad()
{
// a[i] = b[i] + scalar * c[i];
std::transform(exe_policy, range.begin(), range.end(), a.begin(), [&, scalar = startScalar](int i) {
@ -72,7 +72,7 @@ void STDStream<T>::triad()
}
template <class T>
void STDStream<T>::nstream()
void STDIndicesStream<T>::nstream()
{
// a[i] += b[i] + scalar * c[i];
// Need to do in two stages with C++11 STL.
@ -85,7 +85,7 @@ void STDStream<T>::nstream()
template <class T>
T STDStream<T>::dot()
T STDIndicesStream<T>::dot()
{
// sum = 0; sum += a[i]*b[i]; return sum;
return std::transform_reduce(exe_policy, a.begin(), a.end(), b.begin(), 0.0);
@ -105,6 +105,6 @@ std::string getDeviceDriver(const int)
{
return std::string("Device driver unavailable");
}
template class STDStream<float>;
template class STDStream<double>;
template class STDIndicesStream<float>;
template class STDIndicesStream<double>;

17
src/std/STDStream.h → src/std-indices/STDIndicesStream.h
View File

@ -9,9 +9,9 @@
#include <stdexcept>
#include "Stream.h"
#define IMPLEMENTATION_STRING "STD"
#define IMPLEMENTATION_STRING "STD (index-oriented)"
template <typename N = size_t>
template <typename N>
class ranged {
N from, to;
public:
@ -22,13 +22,18 @@ public:
using difference_type = N;
using value_type = N;
using pointer = const N*;
using reference = N&;
using reference = const N&;
using iterator_category = std::random_access_iterator_tag;
explicit iterator(N _num = 0) : num(_num) {}
iterator& operator++() { num++; return *this; }
iterator operator++(int) { iterator retval = *this; ++(*this); return retval; }
iterator operator+(const value_type v) const { return iterator(num + v); }
bool operator==(iterator other) const { return num == other.num; }
bool operator!=(iterator other) const { return *this != other; }
bool operator<(iterator other) const { return num < other.num; }
reference operator*() const { return num;}
difference_type operator-(const iterator &it) const { return num - it.num; }
value_type operator[](const difference_type &i) const { return num + i; }
@ -39,7 +44,7 @@ public:
};
template <class T>
class STDStream : public Stream<T>
class STDIndicesStream : public Stream<T>
{
protected:
// Size of arrays
@ -55,8 +60,8 @@ class STDStream : public Stream<T>
public:
STDStream(const int, int) noexcept;
~STDStream() = default;
STDIndicesStream(const int, int) noexcept;
~STDIndicesStream() = default;
virtual void copy() override;
virtual void add() override;

33
src/std-indices/model.cmake Normal file
View File

@ -0,0 +1,33 @@
register_flag_optional(CMAKE_CXX_COMPILER
"Any CXX compiler that is supported by CMake detection"
"c++")
register_flag_optional(NVHPC_OFFLOAD
"Enable offloading support (via the non-standard `-stdpar`) for the new NVHPC SDK.
The values are Nvidia architectures in ccXY format will be passed in via `-gpu=` (e.g `cc70`)
Possible values are:
cc35 - Compile for compute capability 3.5
cc50 - Compile for compute capability 5.0
cc60 - Compile for compute capability 6.0
cc62 - Compile for compute capability 6.2
cc70 - Compile for compute capability 7.0
cc72 - Compile for compute capability 7.2
cc75 - Compile for compute capability 7.5
cc80 - Compile for compute capability 8.0
ccall - Compile for all supported compute capabilities"
"")
macro(setup)
set(CMAKE_CXX_STANDARD 17)
if (NVHPC_OFFLOAD)
set(NVHPC_FLAGS -stdpar -gpu=${NVHPC_OFFLOAD})
# propagate flags to linker so that it links with the gpu stuff as well
register_append_cxx_flags(ANY ${NVHPC_FLAGS})
register_append_link_flags(${NVHPC_FLAGS})
endif ()
endmacro()

24
src/std20/STD20Stream.cpp → src/std-ranges/STDRangesStream.cpp
View File

@ -4,14 +4,14 @@
// For full license terms please see the LICENSE file distributed with this
// source code
#include "STD20Stream.hpp"
#include "STDRangesStream.hpp"
#include <algorithm>
#include <execution>
#include <ranges>
template <class T>
STD20Stream<T>::STD20Stream(const int ARRAY_SIZE, int device)
STDRangesStream<T>::STDRangesStream(const int ARRAY_SIZE, int device)
: array_size{ARRAY_SIZE}
{
a = std::vector<T>(array_size);
@ -20,7 +20,7 @@ STD20Stream<T>::STD20Stream(const int ARRAY_SIZE, int device)
}
template <class T>
void STD20Stream<T>::init_arrays(T initA, T initB, T initC)
void STDRangesStream<T>::init_arrays(T initA, T initB, T initC)
{
std::for_each_n(
std::execution::par_unseq,
@ -34,7 +34,7 @@ void STD20Stream<T>::init_arrays(T initA, T initB, T initC)
}
template <class T>
void STD20Stream<T>::read_arrays(std::vector<T>& h_a, std::vector<T>& h_b, std::vector<T>& h_c)
void STDRangesStream<T>::read_arrays(std::vector<T>& h_a, std::vector<T>& h_b, std::vector<T>& h_c)
{
// Element-wise copy.
h_a = a;
@ -43,7 +43,7 @@ void STD20Stream<T>::read_arrays(std::vector<T>& h_a, std::vector<T>& h_b, std::
}
template <class T>
void STD20Stream<T>::copy()
void STDRangesStream<T>::copy()
{
std::for_each_n(
std::execution::par_unseq,
@ -55,7 +55,7 @@ void STD20Stream<T>::copy()
}
template <class T>
void STD20Stream<T>::mul()
void STDRangesStream<T>::mul()
{
const T scalar = startScalar;
@ -69,7 +69,7 @@ void STD20Stream<T>::mul()
}
template <class T>
void STD20Stream<T>::add()
void STDRangesStream<T>::add()
{
std::for_each_n(
std::execution::par_unseq,
@ -81,7 +81,7 @@ void STD20Stream<T>::add()
}
template <class T>
void STD20Stream<T>::triad()
void STDRangesStream<T>::triad()
{
const T scalar = startScalar;
@ -95,7 +95,7 @@ void STD20Stream<T>::triad()
}
template <class T>
void STD20Stream<T>::nstream()
void STDRangesStream<T>::nstream()
{
const T scalar = startScalar;
@ -109,7 +109,7 @@ void STD20Stream<T>::nstream()
}
template <class T>
T STD20Stream<T>::dot()
T STDRangesStream<T>::dot()
{
// sum += a[i] * b[i];
return
@ -133,6 +133,6 @@ std::string getDeviceDriver(const int)
return std::string("Device driver unavailable");
}
template class STD20Stream<float>;
template class STD20Stream<double>;
template class STDRangesStream<float>;
template class STDRangesStream<double>;

8
src/std20/STD20Stream.hpp → src/std-ranges/STDRangesStream.hpp
View File

@ -11,10 +11,10 @@
#include "Stream.h"
#define IMPLEMENTATION_STRING "C++20"
#define IMPLEMENTATION_STRING "STD C++ ranges"
template <class T>
class STD20Stream : public Stream<T>
class STDRangesStream : public Stream<T>
{
protected:
// Size of arrays
@ -26,8 +26,8 @@ class STD20Stream : public Stream<T>
std::vector<T> c;
public:
STD20Stream(const int, int);
~STD20Stream() = default;
STDRangesStream(const int, int);
~STDRangesStream() = default;
virtual void copy() override;
virtual void add() override;

0
src/std20/model.cmake → src/std-ranges/model.cmake
View File