提交 a3377f7b 编写于 作者: T tensor-tang

refine jitcode and add vmul jitcode implementation

上级 f3badacd
...@@ -76,6 +76,6 @@ endif() ...@@ -76,6 +76,6 @@ endif()
cc_test(concat_test SRCS concat_test.cc DEPS concat_and_split) cc_test(concat_test SRCS concat_test.cc DEPS concat_and_split)
cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info) cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info)
cc_library(jit_kernel cc_library(jit_kernel
SRCS jit_kernel.cc jit_gen.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_rnn.cc jit_kernel_crf_decode.cc SRCS jit_kernel.cc jit_gen.cc jit_code.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_rnn.cc jit_kernel_crf_decode.cc
DEPS cpu_info cblas gflags enforce) DEPS cpu_info cblas gflags enforce)
cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel) cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel)
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file 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. */
#include "paddle/fluid/operators/math/jit_code.h"
#include "paddle/fluid/operators/math/jit_kernel.h"
#include "paddle/fluid/platform/cpu_info.h"
namespace paddle {
namespace operators {
namespace math {
namespace jitkernel {
namespace gen {
using namespace platform::jit; // NOLINT
bool VMulJitCode::init(int d) {
// TODO(TJ): maybe one AVX is enough, AVX above would slow down freq
// try more with avx2 or avx512
if (MayIUse(avx) || MayIUse(avx2)) {
return d % AVX_FLOAT_BLOCK == 0;
} else {
return false;
}
}
void VMulJitCode::generate() {
preCode();
int stride = sizeof(float) * AVX_FLOAT_BLOCK;
for (int i = 0; i < num_ / AVX_FLOAT_BLOCK; ++i) {
vmovups(ymm_src1, ptr[param1 + i * stride]);
vmovups(ymm_src2, ptr[param2 + i * stride]);
vmulps(ymm_dst, ymm_src1, ymm_src2);
vmovups(ptr[param3 + stride * i], ymm_dst);
}
postCode();
}
} // namespace gen
} // namespace jitkernel
} // namespace math
} // namespace operators
} // namespace paddle
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file 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. */
#pragma once
#include "paddle/fluid/operators/math/jit_gen.h"
namespace paddle {
namespace operators {
namespace math {
namespace jitkernel {
namespace gen {
using reg64_t = const Xbyak::Reg64;
using reg32_t = const Xbyak::Reg32;
using xmm_t = const Xbyak::Xmm;
using ymm_t = const Xbyak::Ymm;
using zmm_t = const Xbyak::Zmm;
using Label = Xbyak::Label;
class VMulJitCode : public JitCode {
public:
DECLARE_JIT_CODE(VMulJitCode);
explicit VMulJitCode(int d, size_t code_size = 256 * 1024,
void* code_ptr = nullptr)
: JitCode(code_size, code_ptr), num_(d) {}
static bool init(int d);
void generate() override;
private:
int num_;
reg64_t param1{abi_param1};
reg64_t param2{abi_param2};
reg64_t param3{abi_param3};
xmm_t xmm_src1 = xmm_t(0);
ymm_t ymm_src1 = ymm_t(0);
zmm_t zmm_src1 = zmm_t(0);
xmm_t xmm_src2 = xmm_t(1);
ymm_t ymm_src2 = ymm_t(1);
zmm_t zmm_src2 = zmm_t(1);
xmm_t xmm_dst = xmm_t(2);
ymm_t ymm_dst = ymm_t(2);
zmm_t zmm_dst = zmm_t(2);
};
} // namespace gen
} // namespace jitkernel
} // namespace math
} // namespace operators
} // namespace paddle
...@@ -14,7 +14,7 @@ limitations under the License. */ ...@@ -14,7 +14,7 @@ limitations under the License. */
#include "paddle/fluid/operators/math/jit_kernel.h" #include "paddle/fluid/operators/math/jit_kernel.h"
#include <string> #include <string>
#include "paddle/fluid/operators/math/jit_gen.h" #include "paddle/fluid/operators/math/jit_code.h"
#include "paddle/fluid/operators/math/jit_kernel_macro.h" #include "paddle/fluid/operators/math/jit_kernel_macro.h"
#include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/enforce.h"
...@@ -30,30 +30,7 @@ namespace paddle { ...@@ -30,30 +30,7 @@ namespace paddle {
namespace operators { namespace operators {
namespace math { namespace math {
namespace jitkernel { namespace jitkernel {
namespace jit = platform::jit;
namespace jit = platform::jit; // remove me
using namespace platform::jit; // NOLINT
/* VMUL JitKernel */
struct VMulJitCode : public gen::JitCode {
DECLARE_JIT_CODE(VMulJitCode);
explicit VMulJitCode(size_t code_size = 256 * 1024, void* code_ptr = nullptr)
: gen::JitCode(code_size, code_ptr) {}
static bool init(int d) {
if (MayIUse(avx) || MayIUse(avx2)) {
return d % AVX_FLOAT_BLOCK == 0;
} else if (MayIUse(avx512f)) {
return d % AVX512_FLOAT_BLOCK == 0;
} else {
return false;
}
}
void generate() override {
preCode();
postCode();
}
};
template <typename T> template <typename T>
void VMulRefer(const T* x, const T* y, T* z, int n) { void VMulRefer(const T* x, const T* y, T* z, int n) {
...@@ -76,6 +53,7 @@ void VMulMKL<double>(const double* x, const double* y, double* z, int n) { ...@@ -76,6 +53,7 @@ void VMulMKL<double>(const double* x, const double* y, double* z, int n) {
} }
#endif #endif
/* VMUL JitKernel */
template <typename T> template <typename T>
class VMulKernelImpl : public VMulKernel<T> { class VMulKernelImpl : public VMulKernel<T> {
public: public:
...@@ -88,7 +66,7 @@ class VMulKernelImpl : public VMulKernel<T> { ...@@ -88,7 +66,7 @@ class VMulKernelImpl : public VMulKernel<T> {
explicit VMulKernelImpl(int d) : VMulKernel<T>() { explicit VMulKernelImpl(int d) : VMulKernel<T>() {
if (useJIT(d)) { if (useJIT(d)) {
constexpr size_t sz = 256 * 1024; // TODO(TJ): should be related with d constexpr size_t sz = 256 * 1024; // TODO(TJ): should be related with d
jitcode_.reset(new VMulJitCode(sz)); jitcode_.reset(new gen::VMulJitCode(d, sz));
this->Compute = this->Compute =
jitcode_->getCode<void (*)(const T*, const T*, T*, int)>(); jitcode_->getCode<void (*)(const T*, const T*, T*, int)>();
return; return;
...@@ -103,12 +81,12 @@ class VMulKernelImpl : public VMulKernel<T> { ...@@ -103,12 +81,12 @@ class VMulKernelImpl : public VMulKernel<T> {
} }
private: private:
std::unique_ptr<VMulJitCode> jitcode_{nullptr}; std::unique_ptr<gen::VMulJitCode> jitcode_{nullptr};
}; };
template <> template <>
bool VMulKernelImpl<float>::useJIT(int d) { bool VMulKernelImpl<float>::useJIT(int d) {
return VMulJitCode::init(d); return gen::VMulJitCode::init(d);
} }
template <> template <>
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册