natas_hw
/
openvino
0
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge a0527fefad into 43de73cbd6

This commit is contained in:
Siddhant Chauhan 2024-06-12 03:28:00 +02:00 committed by GitHub
parent 43de73cbd6 a0527fefad
commit b4e89fdf3d
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 65 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

44
src/plugins/intel_cpu/src/nodes/common/cpu_convert.cpp
View File

@ -26,13 +26,13 @@ using namespace dnnl::impl::utils;
using namespace dnnl::impl::cpu::x64;
using namespace Xbyak;
template <typename src_t, typename dst_t>
template <typename src_t, typename dst_t, int isa>
void convert_vec(jit_generator & gen,
const RegExp & src,
const RegExp & dst);
template <>
void convert_vec<ov::float16, float>(jit_generator & gen,
void convert_vec<ov::float16, float, cpu_isa_t::avx2>(jit_generator & gen,
const RegExp & src,
const RegExp & dst) {
auto const & f16vec = gen.xmm3;
@ -44,7 +44,7 @@ void convert_vec<ov::float16, float>(jit_generator & gen,
}
template <>
void convert_vec<float, ov::float16>(jit_generator & gen,
void convert_vec<float, ov::float16, cpu_isa_t::avx2>(jit_generator & gen,
const RegExp & src,
const RegExp & dst) {
auto const & f16vec = gen.xmm3;
@ -55,6 +55,31 @@ void convert_vec<float, ov::float16>(jit_generator & gen,
gen.movdqu(gen.xword[dst], f16vec);
}
// AVX512 instructions for ov::float16 to float conversion
template <>
void convert_vec<ov::float16, float, cpu_isa_t::avx512_core>(jit_generator & gen,
const RegExp & src,
const RegExp & dst) {
auto const & f16vec = gen.ymm3;
auto const & f32vec = gen.zmm4;
gen.movdqu(f16vec, gen.yword[src]);
gen.vcvtph2ps(f32vec, f16vec);
gen.vmovups(gen.zword[dst], f32vec);
}
template <>
void convert_vec<float, ov::float16, cpu_isa_t::avx512_core>(jit_generator & gen,
const RegExp & src,
const RegExp & dst) {
auto const & f16vec = gen.ymm3;
auto const & f32vec = gen.zmm4;
gen.vmovups(f32vec, gen.zword[src]);
gen.vcvtps2ph(f16vec, f32vec, 0);
gen.movdqu(gen.yword[dst], f16vec);
}
class jit_convert_array : public jit_kernel {
DECLARE_CPU_JIT_AUX_FUNCTIONS(jit_convert_array)
@ -125,10 +150,15 @@ public:
template<typename src_t, typename dst_t>
static fn_t get() {
if (mayiuse(cpu_isa_t::avx2)
&& dnnl::impl::cpu::x64::cpu().has(Xbyak::util::Cpu::tF16C)) {
static jit_convert_array converter(convert_vec<src_t, dst_t>, sizeof(src_t), sizeof(dst_t));
auto & generator = static_cast<jit_generator&>(converter);
// fallback to AVX2 if AVX512 is not available
if (mayiuse(cpu_isa_t::avx512_core)) {
static jit_convert_array converter(convert_vec<src_t, dst_t, cpu_isa_t::avx512_core>, sizeof(src_t), sizeof(dst_t));
auto& generator = static_cast<jit_generator&>(converter);
generator.create_kernel();
return (fn_t)generator.jit_ker();
} else if (mayiuse(cpu_isa_t::avx2) && dnnl::impl::cpu::x64::cpu().has(Xbyak::util::Cpu::tF16C)) {
static jit_convert_array converter(convert_vec<src_t, dst_t, cpu_isa_t::avx2>, sizeof(src_t), sizeof(dst_t));
auto& generator = static_cast<jit_generator&>(converter);
generator.create_kernel();
return (fn_t)generator.jit_ker();
}

1
src/plugins/intel_cpu/tests/unit/CMakeLists.txt
View File

@ -25,6 +25,7 @@ if(NOT X86_64)
${CMAKE_CURRENT_SOURCE_DIR}/transformations/x64
${CMAKE_CURRENT_SOURCE_DIR}/snippets_transformations/x64
${CMAKE_CURRENT_SOURCE_DIR}/nodes/eltwise_node_test.cpp
${CMAKE_CURRENT_SOURCE_DIR}/nodes/avx_512_test.cpp)
${CMAKE_CURRENT_SOURCE_DIR}/brgemm_executor_test.cpp)
endif()

27
src/plugins/intel_cpu/tests/unit/nodes/avx_512_test.cpp Normal file
View File

@ -0,0 +1,27 @@
// Copyright (C) 2018-2024 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#include <chrono>
#include <iostream>
#include <vector>
#include "nodes/common/cpu_convert.h"
#include "gtest/gtest.h"
using namespace std::chrono;
using namespace ov::intel_cpu;
TEST(cpu_convert, AVX512_fp16_load_store) {
std::vector<size_t> sizes = {1000, 10000, 100000, 1000000};
for (size_t size : sizes) {
std::vector<float> input_data(size, 1.23f);
std::vector<short> output_data(size);
auto start = high_resolution_clock::now();
cpu_convert(input_data.data(), output_data.data(), ov::element::f32, ov::element::f16, size);
auto stop = high_resolution_clock::now();
auto duration = duration_cast<microseconds>(stop - start);
std::cout << "size " << size << ": " << duration.count() << " microseconds\n";
}
}