Add XSIMD implementation

This commit is contained in:
mancini
2025-08-07 14:30:28 +02:00
parent b7c13be6c0
commit b55f456178
8 changed files with 255 additions and 0 deletions

View File

@@ -6,6 +6,7 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
option(USE_MKL "Enable Intel MKL backend" OFF) option(USE_MKL "Enable Intel MKL backend" OFF)
option(USE_GPU "Enable GPU backend" OFF) option(USE_GPU "Enable GPU backend" OFF)
option(USE_XSIMD "Enable XSIMD backend" OFF)
include_directories(${PROJECT_SOURCE_DIR}/include) include_directories(${PROJECT_SOURCE_DIR}/include)

View File

@@ -16,3 +16,8 @@ if(USE_GPU)
add_executable(benchmark_gpu benchmark_gpu.cpp) add_executable(benchmark_gpu benchmark_gpu.cpp)
target_link_libraries(benchmark_gpu PRIVATE trigdx gpu) target_link_libraries(benchmark_gpu PRIVATE trigdx gpu)
endif() endif()
if(USE_XSIMD)
add_executable(benchmark_lookup_xsimd benchmark_lookup_xsimd.cpp)
target_link_libraries(benchmark_lookup_xsimd PRIVATE trigdx)
endif()

View File

@@ -0,0 +1,13 @@
#include <trigdx/lookup_xsimd.hpp>
#include "benchmark_utils.hpp"
int main() {
benchmark_sinf<LookupXSIMDBackend<16384>>();
benchmark_cosf<LookupXSIMDBackend<16384>>();
benchmark_sincosf<LookupXSIMDBackend<16384>>();
benchmark_sinf<LookupXSIMDBackend<32768>>();
benchmark_cosf<LookupXSIMDBackend<32768>>();
benchmark_sincosf<LookupXSIMDBackend<32768>>();
}

View File

@@ -0,0 +1,22 @@
#pragma once
#include <cstddef>
#include <memory>
#include "interface.hpp"
template <std::size_t NR_SAMPLES> class LookupXSIMDBackend : public Backend {
public:
LookupXSIMDBackend();
~LookupXSIMDBackend() override;
void init(size_t n = 0) override;
void compute_sinf(std::size_t n, const float *x, float *s) const override;
void compute_cosf(std::size_t n, const float *x, float *c) const override;
void compute_sincosf(std::size_t n, const float *x, float *s,
float *c) const override;
private:
struct Impl;
std::unique_ptr<Impl> impl;
};

View File

@@ -18,3 +18,9 @@ if(USE_GPU)
target_link_libraries(trigdx PRIVATE CUDA::cudart) target_link_libraries(trigdx PRIVATE CUDA::cudart)
target_link_libraries(trigdx PRIVATE gpu) target_link_libraries(trigdx PRIVATE gpu)
endif() endif()
if(USE_XSIMD)
find_package(xsimd REQUIRED)
target_sources(trigdx PRIVATE lookup_xsimd.cpp)
target_link_libraries(trigdx PRIVATE xsimd)
endif()

183
src/lookup_xsimd.cpp Normal file
View File

@@ -0,0 +1,183 @@
#include <algorithm>
#include <cmath>
#include <vector>
#include <xsimd/xsimd.hpp>
#include "trigdx/lookup_xsimd.hpp"
template <std::size_t NR_SAMPLES> struct lookup_table {
static constexpr std::size_t MASK = NR_SAMPLES - 1;
static constexpr float SCALE = NR_SAMPLES / (2.0f * float(M_PI));
lookup_table() : values{} {
for (uint_fast32_t i = 0; i < NR_SAMPLES; i++) {
values[i] = sinf(i * (2.0f * float(M_PI) / NR_SAMPLES));
}
}
std::array<float, NR_SAMPLES> values;
};
template <std::size_t NR_SAMPLES> struct cosf_dispatcher {
constexpr cosf_dispatcher() : lookup_table_(){};
template <class Tag, class Arch>
void operator()(Arch, const size_t n, const float *a, float *c, Tag) const {
using b_type = xsimd::batch<float, Arch>;
using m_type = xsimd::batch<int32_t, Arch>;
constexpr uint_fast32_t VL = b_type::size;
const uint_fast32_t VS = n - n % VL;
const uint_fast32_t Q_PI = NR_SAMPLES / 4U;
const b_type scale = b_type::broadcast(lookup_table_.SCALE);
const m_type mask = m_type::broadcast(lookup_table_.MASK);
const m_type quarter_pi = m_type::broadcast(Q_PI);
uint_fast32_t i;
for (i = 0; i < VS; i += VL) {
const b_type vx = b_type::load(a + i, Tag());
const b_type scaled = xsimd::mul(vx, scale);
m_type idx = xsimd::to_int(scaled);
m_type idx_cos = xsimd::add(idx, quarter_pi);
idx_cos = xsimd::bitwise_and(idx_cos, mask);
const b_type cosv = b_type::gather(lookup_table_.values.data(), idx_cos);
cosv.store(c + i, Tag());
}
for (; i < n; i++) {
std::size_t idx = static_cast<std::size_t>(a[i] * lookup_table_.SCALE) &
lookup_table_.MASK;
std::size_t idx_cos = (idx + Q_PI) & lookup_table_.MASK;
c[i] = lookup_table_.values[idx_cos];
}
}
lookup_table<NR_SAMPLES> lookup_table_;
};
template <std::size_t NR_SAMPLES> struct sinf_dispatcher {
constexpr sinf_dispatcher() : lookup_table_(){};
template <class Tag, class Arch>
void operator()(Arch, const size_t n, const float *a, float *s, Tag) const {
using b_type = xsimd::batch<float, Arch>;
using m_type = xsimd::batch<int32_t, Arch>;
constexpr uint_fast32_t VL = b_type::size;
const uint_fast32_t VS = n - n % VL;
const uint_fast32_t Q_PI = NR_SAMPLES / 4U;
const b_type scale = b_type::broadcast(lookup_table_.SCALE);
const m_type mask = m_type::broadcast(lookup_table_.MASK);
const m_type quarter_pi = m_type::broadcast(Q_PI);
uint_fast32_t i;
for (i = 0; i < VS; i += VL) {
const b_type vx = b_type::load(a + i, Tag());
const b_type scaled = xsimd::mul(vx, scale);
m_type idx = xsimd::to_int(scaled);
idx = xsimd::bitwise_and(idx, mask);
const b_type sinv = b_type::gather(lookup_table_.values.data(), idx);
sinv.store(s + i, Tag());
}
for (; i < n; i++) {
std::size_t idx = static_cast<std::size_t>(a[i] * lookup_table_.SCALE) &
lookup_table_.MASK;
s[i] = lookup_table_.values[idx];
}
}
lookup_table<NR_SAMPLES> lookup_table_;
};
template <std::size_t NR_SAMPLES> struct sin_cosf_dispatcher {
template <class Tag, class Arch>
void operator()(Arch, const size_t n, const float *a, float *s, float *c,
Tag) const {
using b_type = xsimd::batch<float, Arch>;
using m_type = xsimd::batch<int32_t, Arch>;
constexpr uint_fast32_t VL = b_type::size;
const uint_fast32_t VS = n - n % VL;
const uint_fast32_t Q_PI = NR_SAMPLES / 4U;
const b_type scale = b_type::broadcast(lookup_table_.SCALE);
const m_type mask = m_type::broadcast(lookup_table_.MASK);
const m_type quarter_pi = m_type::broadcast(Q_PI);
uint_fast32_t i;
for (i = 0; i < VS; i += VL) {
const b_type vx = b_type::load(a + i, Tag());
const b_type scaled = xsimd::mul(vx, scale);
m_type idx = xsimd::to_int(scaled);
m_type idx_cos = xsimd::add(idx, quarter_pi);
idx = xsimd::bitwise_and(idx, mask);
idx_cos = xsimd::bitwise_and(idx_cos, mask);
const b_type sinv = b_type::gather(lookup_table_.values.data(), idx);
const b_type cosv = b_type::gather(lookup_table_.values.data(), idx_cos);
sinv.store(s + i, Tag());
cosv.store(c + i, Tag());
}
for (; i < n; i++) {
std::size_t idx = static_cast<std::size_t>(a[i] * lookup_table_.SCALE) &
lookup_table_.MASK;
std::size_t idx_cos = (idx + Q_PI) & lookup_table_.MASK;
s[i] = lookup_table_.values[idx];
c[i] = lookup_table_.values[idx_cos];
}
}
lookup_table<NR_SAMPLES> lookup_table_;
};
template <std::size_t NR_SAMPLES> struct LookupXSIMDBackend<NR_SAMPLES>::Impl {
cosf_dispatcher<NR_SAMPLES> cosf_dispatcher_;
sinf_dispatcher<NR_SAMPLES> sinf_dispatcher_;
sin_cosf_dispatcher<NR_SAMPLES> sin_cosf_dispatcher_;
void init() {}
void compute_sincosf(std::size_t n, const float *x, float *s,
float *c) const {
xsimd::dispatch(sin_cosf_dispatcher_)(n, x, s, c, xsimd::unaligned_mode());
}
void compute_sinf(std::size_t n, const float *x, float *s) const {
xsimd::dispatch(sinf_dispatcher_)(n, x, s, xsimd::unaligned_mode());
}
void compute_cosf(std::size_t n, const float *x, float *c) const {
xsimd::dispatch(cosf_dispatcher_)(n, x, c, xsimd::unaligned_mode());
}
};
template <std::size_t NR_SAMPLES>
LookupXSIMDBackend<NR_SAMPLES>::LookupXSIMDBackend()
: impl(std::make_unique<Impl>()) {}
template <std::size_t NR_SAMPLES>
LookupXSIMDBackend<NR_SAMPLES>::~LookupXSIMDBackend() = default;
template <std::size_t NR_SAMPLES>
void LookupXSIMDBackend<NR_SAMPLES>::init(size_t) {
impl->init();
}
template <std::size_t NR_SAMPLES>
void LookupXSIMDBackend<NR_SAMPLES>::compute_sinf(const std::size_t n,
const float *x,
float *s) const {
impl->compute_sinf(n, x, s);
}
template <std::size_t NR_SAMPLES>
void LookupXSIMDBackend<NR_SAMPLES>::compute_cosf(const std::size_t n,
const float *x,
float *c) const {
impl->compute_cosf(n, x, c);
}
template <std::size_t NR_SAMPLES>
void LookupXSIMDBackend<NR_SAMPLES>::compute_sincosf(const std::size_t n,
const float *x, float *s,
float *c) const {
impl->compute_sincosf(n, x, s, c);
}
// Explicit instantiations
template class LookupXSIMDBackend<16384>;
template class LookupXSIMDBackend<32768>;

View File

@@ -32,3 +32,9 @@ if(USE_GPU)
target_link_libraries(test_gpu PRIVATE trigdx Catch2::Catch2WithMain) target_link_libraries(test_gpu PRIVATE trigdx Catch2::Catch2WithMain)
add_test(NAME test_gpu COMMAND test_gpu) add_test(NAME test_gpu COMMAND test_gpu)
endif() endif()
if(USE_XSIMD)
add_executable(test_lookup_xsimd test_lookup_xsimd.cpp)
target_link_libraries(test_lookup_xsimd PRIVATE trigdx Catch2::Catch2WithMain)
add_test(NAME test_lookup_xsimd COMMAND test_lookup_xsimd)
endif()

View File

@@ -0,0 +1,19 @@
#include <catch2/catch_test_macros.hpp>
#include <trigdx/lookup_xsimd.hpp>
#include "test_utils.hpp"
TEST_CASE("sincosf") {
test_sincosf<LookupXSIMDBackend<16384>>(1e-2f);
test_sincosf<LookupXSIMDBackend<32768>>(1e-2f);
}
TEST_CASE("sinf") {
test_sinf<LookupXSIMDBackend<16384>>(1e-2f);
test_sinf<LookupXSIMDBackend<32768>>(1e-2f);
}
TEST_CASE("cosf") {
test_cosf<LookupXSIMDBackend<16384>>(1e-2f);
test_cosf<LookupXSIMDBackend<32768>>(1e-2f);
}