m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

use vadd, vaddrelu, lstm and gru jitkernel

Browse Source
revert-15207-remove_op_handle_lock_and_fix_var
tensor-tang 6 years ago
parent 3713d08d40
commit 64a90b2f1c
4 changed files with 68 additions and 75 deletions
Show Stats Download Patch File Download Diff File

58
paddle/fluid/operators/fused/fusion_gru_op.cc
Unescape View File

@ -15,9 +15,9 @@ limitations under the License. */
#include "paddle/fluid/operators/fused/fusion_gru_op.h"
#include <cstring> // for memcpy
#include <string>
#include "paddle/fluid/operators/jit/kernels.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/fc_compute.h"
#include "paddle/fluid/operators/math/jit_kernel.h"
#include "paddle/fluid/operators/math/sequence2batch.h"
namespace paddle {
@ -183,27 +183,29 @@ class FusionGRUKernel : public framework::OpKernel<T> {
const int total_T = x_dims[0]; \
const int D3 = wh_dims[1]
#define INIT_OTHER_DEFINES \
auto* h0 = ctx.Input<Tensor>("H0"); \
auto* wx = ctx.Input<Tensor>("WeightX"); \
auto* bias = ctx.Input<Tensor>("Bias"); \
auto* hidden_out = ctx.Output<LoDTensor>("Hidden"); \
bool is_reverse = ctx.Attr<bool>("is_reverse"); \
const int M = x_dims[1]; \
const int D = wh_dims[0]; \
const int D2 = D * 2; \
const math::jitkernel::gru_attr_t attr( \
D, ctx.Attr<std::string>("gate_activation"), \
ctx.Attr<std::string>("activation")); \
math::jitkernel::gru_t one_step; \
const auto& ker = \
math::jitkernel::KernelPool::Instance() \
.template Get<math::jitkernel::GRUKernel<T>, \
const math::jitkernel::gru_attr_t&>(attr); \
const T* x_data = x->data<T>(); \
const T* wx_data = wx->data<T>(); \
const T* wh_data = wh->data<T>(); \
auto place = ctx.GetPlace(); \
#define INIT_OTHER_DEFINES \
auto* h0 = ctx.Input<Tensor>("H0"); \
auto* wx = ctx.Input<Tensor>("WeightX"); \
auto* bias = ctx.Input<Tensor>("Bias"); \
auto* hidden_out = ctx.Output<LoDTensor>("Hidden"); \
bool is_reverse = ctx.Attr<bool>("is_reverse"); \
const int M = x_dims[1]; \
const int D = wh_dims[0]; \
const int D2 = D * 2; \
const jit::gru_attr_t attr( \
D, jit::to_kerneltype(ctx.Attr<std::string>("gate_activation")), \
jit::to_kerneltype(ctx.Attr<std::string>("activation"))); \
jit::gru_t one_step; \
auto ComputeH1 = \
jit::Get<jit::gruh1, jit::GRUTuples, platform::CPUPlace>(attr); \
auto ComputeHtPart1 = \
jit::Get<jit::gruhtpart1, jit::GRUTuples, platform::CPUPlace>(attr); \
auto ComputeHtPart2 = \
jit::Get<jit::gruhtpart2, jit::GRUTuples, platform::CPUPlace>(attr); \
const T* x_data = x->data<T>(); \
const T* wx_data = wx->data<T>(); \
const T* wh_data = wh->data<T>(); \
auto place = ctx.GetPlace(); \
T* xx_data = xx->mutable_data<T>(place)
void SeqCompute(const framework::ExecutionContext& ctx) const {
@ -242,7 +244,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
} else {
one_step.gates = xx_data;
one_step.ht = hidden_out_data;
ker->ComputeH1(&one_step, &attr);
ComputeH1(&one_step, &attr);
prev_hidden_data = hidden_out_data;
tstart = 1;
move_step();
@ -255,12 +257,12 @@ class FusionGRUKernel : public framework::OpKernel<T> {
one_step.gates = xx_data;
one_step.ht_1 = prev_hidden_data;
one_step.ht = hidden_out_data;
ker->ComputeHtPart1(&one_step, &attr);
ComputeHtPart1(&one_step, &attr);
// gemm rt * Ws
blas.GEMM(CblasNoTrans, CblasNoTrans, 1, D, D, static_cast<T>(1),
hidden_out_data, D, wh_state_data, D, static_cast<T>(1),
xx_data + D2, D3);
ker->ComputeHtPart2(&one_step, &attr);
ComputeHtPart2(&one_step, &attr);
// save prev
prev_hidden_data = hidden_out_data;
move_step();
@ -324,7 +326,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
for (int i = 0; i < max_bs; ++i) {
one_step.gates = cur_in_data;
one_step.ht = cur_out_data;
ker->ComputeH1(&one_step, &attr);
ComputeH1(&one_step, &attr);
// add offset
cur_in_data += D3;
cur_out_data += D;
@ -352,7 +354,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
one_step.gates = cur_batched_data;
one_step.ht_1 = cur_prev_hidden_data;
one_step.ht = cur_out_data;
ker->ComputeHtPart1(&one_step, &attr);
ComputeHtPart1(&one_step, &attr);
cur_batched_data += D3;
cur_prev_hidden_data += D;
@ -370,7 +372,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
one_step.gates = cur_batched_data;
one_step.ht_1 = cur_prev_hidden_data;
one_step.ht = cur_out_data;
ker->ComputeHtPart2(&one_step, &attr);
ComputeHtPart2(&one_step, &attr);
cur_batched_data += D3;
cur_prev_hidden_data += D;
cur_out_data += D;

62
paddle/fluid/operators/fused/fusion_lstm_op.cc
Unescape View File

@ -14,9 +14,9 @@ limitations under the License. */
#include "paddle/fluid/operators/fused/fusion_lstm_op.h"
#include <string>
#include "paddle/fluid/operators/jit/kernels.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/fc_compute.h"
#include "paddle/fluid/operators/math/jit_kernel.h"
#include "paddle/fluid/operators/math/sequence2batch.h"
namespace paddle {
@ -236,31 +236,33 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
const int D = wh_dims[0]; \
const int D4 = wh_dims[1]
#define INIT_OTHER_DEFINES \
const T* x_data = x->data<T>(); \
const T* wx_data = wx->data<T>(); \
const T* wh_data = wh->data<T>(); \
/* diagonal weight*/ \
const T* wp_data = bias->data<T>() + D4; \
/* for peephole only*/ \
T* checked_cell_data = nullptr; \
auto place = ctx.GetPlace(); \
if (use_peepholes) { \
/* w_ic * Ct-1, w_fc * Ct-1 ; w_oc * Ct => ih*/ \
auto* checked_cell = ctx.Output<Tensor>("CheckedCell"); \
checked_cell_data = checked_cell->mutable_data<T>(place); \
} \
const math::jitkernel::lstm_attr_t attr( \
D, ctx.Attr<std::string>("gate_activation"), \
ctx.Attr<std::string>("candidate_activation"), \
ctx.Attr<std::string>("cell_activation"), use_peepholes); \
math::jitkernel::lstm_t one_step; \
one_step.wp = wp_data; \
one_step.checked = checked_cell_data; \
const auto& ker = \
math::jitkernel::KernelPool::Instance() \
.template Get<math::jitkernel::LSTMKernel<T>, \
const math::jitkernel::lstm_attr_t&>(attr)
#define INIT_OTHER_DEFINES \
const T* x_data = x->data<T>(); \
const T* wx_data = wx->data<T>(); \
const T* wh_data = wh->data<T>(); \
/* diagonal weight*/ \
const T* wp_data = bias->data<T>() + D4; \
/* for peephole only*/ \
T* checked_cell_data = nullptr; \
auto place = ctx.GetPlace(); \
if (use_peepholes) { \
/* w_ic * Ct-1, w_fc * Ct-1 ; w_oc * Ct => ih*/ \
auto* checked_cell = ctx.Output<Tensor>("CheckedCell"); \
checked_cell_data = checked_cell->mutable_data<T>(place); \
} \
const jit \
: lstm_attr_t attr( \
D, jit::to_kerneltype(ctx.Attr<std::string>("gate_activation")), \
jit::to_kerneltype(ctx.Attr<std::string>("candidate_activation")), \
jit::to_kerneltype(ctx.Attr<std::string>("cell_activation")), \
use_peepholes); \
math::jitkernel::lstm_t one_step; \
one_step.wp = wp_data; \
one_step.checked = checked_cell_data; \
auto ComputeC1H1 = \
jit::Get<jit::lstmc1h1, jit::LSTMTuples, platform::CPUPlace>(attr); \
auto ComputeCtHt = \
jit::Get<jit::lstmctht, jit::LSTMTuples, platform::CPUPlace>(attr)
// Wh GEMM
#define GEMM_WH_ADDON(bs, prev, out) \
@ -306,7 +308,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
one_step.gates = xx_data;
one_step.ct = c_out_data;
one_step.ht = h_out_data;
ker->ComputeC1H1(&one_step, &attr);
ComputeC1H1(&one_step, &attr);
tstart = 1;
// move one step
prev_h_data = h_out_data;
@ -322,7 +324,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
one_step.ct_1 = prev_c_data;
one_step.ct = c_out_data;
one_step.ht = h_out_data;
ker->ComputeCtHt(&one_step, &attr);
ComputeCtHt(&one_step, &attr);
// move one step
prev_h_data = h_out_data;
prev_c_data = c_out_data;
@ -402,7 +404,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
one_step.gates = cur_in_data;
one_step.ct = cur_c_out_data;
one_step.ht = cur_h_out_data;
ker->ComputeC1H1(&one_step, &attr);
ComputeC1H1(&one_step, &attr);
cur_in_data += D4;
cur_c_out_data += D;
@ -432,7 +434,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
one_step.ct_1 = cur_prev_c_data;
one_step.ct = cur_c_out_data;
one_step.ht = cur_h_out_data;
ker->ComputeCtHt(&one_step, &attr);
ComputeC1H1(&one_step, &attr);
// move one batch
cur_in_data += D4;

9
paddle/fluid/operators/math/CMakeLists.txt
Unescape View File

@ -73,12 +73,3 @@ if(WITH_GPU)
endif()
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)
# set(JIT_KERNEL_SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_rnn.cc jit_kernel_crf_decode.cc jit_kernel_layer_norm.cc)
# set(JIT_KERNEL_DEPS cpu_info cblas gflags enforce)
# if(WITH_XBYAK)
# list(APPEND JIT_KERNEL_SRCS jit_gen.cc jit_code.cc)
# list(APPEND JIT_KERNEL_DEPS xbyak)
# endif()
# cc_library(jit_kernel SRCS ${JIT_KERNEL_SRCS} DEPS ${JIT_KERNEL_DEPS})
# cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel)

14
paddle/fluid/operators/math/fc_compute.h
Unescape View File

@ -14,8 +14,8 @@ limitations under the License. */
#pragma once
#include "paddle/fluid/operators/jit/kernels.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/jit_kernel.h"
namespace paddle {
namespace operators {
@ -30,22 +30,20 @@ inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
return;
}
if (relu) {
const auto& vaddrelu = jitkernel::KernelPool::Instance()
.template Get<jitkernel::VAddReluKernel<T>>(N);
auto compute =
jit::Get<jit::vaddrelu, jit::XYZNTuples, platform::CPUPlcace>(N);
for (int i = 0; i < M; i++) {
T* dst = Y + i * N;
vaddrelu->Compute(B, dst, dst, N);
compute(B, dst, dst, N);
}
} else {
const auto& vadd = jitkernel::KernelPool::Instance()
.template Get<jitkernel::VAddKernel<T>>(N);
auto compute = jit::Get<jit::vadd, jit::XYZNTuples, platform::CPUPlcace>(N);
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
#endif
for (int i = 0; i < M; i++) {
T* dst = Y + i * N;
vadd->Compute(B, dst, dst, N);
compute(B, dst, dst, N);
}
}
}

Write Preview
Loading…
Cancel
Save