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

Merge remote-tracking branch 'upstream/develop' into windows/build

Browse Source
local_add_cudnn_lstm
peizhilin 6 years ago
parent 8580b7a130 7f17e561d7
commit 764f97deac
11 changed files with 632 additions and 505 deletions
Show Stats Download Patch File Download Diff File

18
paddle/fluid/operators/math/cpu_vec.h
Unescape View File

@ -33,11 +33,11 @@ namespace math {
#define SIGMOID_THRESHOLD_MIN -40.0
#define SIGMOID_THRESHOLD_MAX 13.0
#define AVX_FLOAT_BLOCK 8
#define YMM_FLOAT_BLOCK 8
#define AVX_DOUBLE_BLOCK 4
#define AVX2_FLOAT_BLOCK 8
#define YMM_FLOAT_BLOCK 8
#define AVX2_DOUBLE_BLOCK 4
#define AVX512_FLOAT_BLOCK 16
#define ZMM_FLOAT_BLOCK 16
#define AVX512_DOUBLE_BLOCK 8
template <typename T>
@ -88,7 +88,7 @@ template <>
inline void vec_scal<float, platform::jit::avx>(const int n, const float a,
const float* x, float* y) {
#ifdef __AVX__
constexpr int block = AVX_FLOAT_BLOCK;
constexpr int block = YMM_FLOAT_BLOCK;
if (n < block) {
vec_scal<float, platform::jit::isa_any>(n, a, x, y);
return;
@ -142,7 +142,7 @@ template <>
inline void vec_bias_sub<float, platform::jit::avx>(const int n, const float a,
const float* x, float* y) {
#ifdef __AVX__
constexpr int block = AVX_FLOAT_BLOCK;
constexpr int block = YMM_FLOAT_BLOCK;
if (n < block) {
vec_bias_sub<float, platform::jit::isa_any>(n, a, x, y);
return;
@ -200,7 +200,7 @@ inline void vec_cross<float, platform::jit::avx>(const int n, const float* x,
const float* y, const float* z,
float* out) {
#ifdef __AVX__
constexpr int block = AVX_FLOAT_BLOCK;
constexpr int block = YMM_FLOAT_BLOCK;
if (n < block) {
vec_cross<float, platform::jit::isa_any>(n, x, y, z, out);
return;
@ -257,7 +257,7 @@ template <>
inline void vec_add_bias<float, platform::jit::avx>(const int n, const float a,
const float* x, float* y) {
#ifdef __AVX__
constexpr int block = AVX_FLOAT_BLOCK;
constexpr int block = YMM_FLOAT_BLOCK;
if (n < block) {
vec_add_bias<float, platform::jit::isa_any>(n, a, x, y);
return;
@ -326,7 +326,7 @@ template <>
inline void vec_sigmoid<float, platform::jit::avx>(const int n, const float* x,
float* y) {
#ifdef __AVX__
constexpr int block = AVX_FLOAT_BLOCK;
constexpr int block = YMM_FLOAT_BLOCK;
if (n < block) {
vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
return;
@ -415,7 +415,7 @@ template <>
inline void vec_relu<float, platform::jit::avx>(const int n, const float* x,
float* y) {
#ifdef __AVX__
constexpr int block = AVX_FLOAT_BLOCK;
constexpr int block = YMM_FLOAT_BLOCK;
if (n < block * 4) {
vec_relu<float, platform::jit::isa_any>(n, x, y);
return;

240
paddle/fluid/operators/math/jit_code.cc
View File

File diff suppressed because it is too large Load Diff

72
paddle/fluid/operators/math/jit_code.h
Unescape View File

@ -29,7 +29,16 @@ using ymm_t = const Xbyak::Ymm;
using zmm_t = const Xbyak::Zmm;
using Label = Xbyak::Label;
typedef enum { mul = 0, add } operand_type;
typedef enum {
mul = 0,
add,
sub,
relu,
exp,
sigmoid,
tanh,
identity
} operand_type;
// function: vec = Operand(vec(or scalar), vec(or scalar)) (maybe with relu)
class VXXJitCode : public JitCode {
@ -85,26 +94,65 @@ class VXXJitCode : public JitCode {
ymm_t ymm_zero = ymm_t(3);
};
class ReluJitCode : public JitCode {
class VActJitCode : public JitCode {
public:
DECLARE_JIT_CODE(ReluJitCode);
explicit ReluJitCode(int d, size_t code_size = 256 * 1024,
const char* name() const override {
std::string base = "VActJitCode";
switch (type_) {
case operand_type::relu:
base += "_Relu";
break;
case operand_type::exp:
base += "_Exp";
break;
case operand_type::sigmoid:
base += "_Sigmoid";
break;
case operand_type::tanh:
base += "_Tanh";
break;
case operand_type::identity:
base += "_Identity";
break;
default:
break;
}
return base.c_str();
}
explicit VActJitCode(int d, operand_type type, size_t code_size = 256 * 1024,
void* code_ptr = nullptr)
: JitCode(code_size, code_ptr), num_(d) {}
static bool init(int d);
: JitCode(code_size, code_ptr), num_(d), type_(type) {}
static bool init(int d, operand_type type);
void generate() override;
private:
protected:
// compute relu with ymm
void relu_ymm(const Xbyak::Ymm& dst, const Xbyak::Ymm& src,
const Xbyak::Ymm& zero);
// compute exp with ymm
void exp_ymm(const Xbyak::Ymm& dst, const Xbyak::Ymm& src, int fx_idx = 2,
int fy_idx = 3, int mask_idx = 4, int tmp_idx = 5);
// compute sigmoid with ymm
void sigmoid_ymm(const Xbyak::Ymm& dst, const Xbyak::Ymm& src, int fx_idx = 2,
int fy_idx = 3, int mask_idx = 4, int tmp_idx = 5);
// compute tanh with ymm
void tanh_ymm(const Xbyak::Ymm& dst, const Xbyak::Ymm& src, int fx_idx = 2,
int fy_idx = 3, int mask_idx = 4, int tmp_idx = 5);
protected:
int num_;
operand_type type_;
reg64_t param1{abi_param1};
reg64_t param2{abi_param2};
xmm_t xmm_zero = xmm_t(0);
xmm_t xmm_src = xmm_t(1);
xmm_t xmm_dst = xmm_t(1);
xmm_t xmm_src = xmm_t(0);
ymm_t ymm_src = ymm_t(0);
ymm_t ymm_zero = ymm_t(0);
ymm_t ymm_src = ymm_t(1);
xmm_t xmm_dst = xmm_t(1);
ymm_t ymm_dst = ymm_t(1);
};

34
paddle/fluid/operators/math/jit_kernel.h
Unescape View File

@ -29,9 +29,9 @@ namespace jitkernel {
#define SIGMOID_THRESHOLD_MIN -40.0
#define SIGMOID_THRESHOLD_MAX 13.0
#define EXP_MAX_INPUT 40.0
#define AVX_FLOAT_BLOCK 8
#define AVX2_FLOAT_BLOCK 8
#define AVX512_FLOAT_BLOCK 16
#define XMM_FLOAT_BLOCK 4
#define YMM_FLOAT_BLOCK 8
#define ZMM_FLOAT_BLOCK 16
typedef enum { kLT8, kEQ8, kGT8LT16, kEQ16, kGT16 } jit_block;
@ -97,39 +97,23 @@ class VAddBiasKernel : public Kernel {
template <typename T>
class VActKernel : public Kernel {
public:
virtual void ComputeDeprecated(const T *x, T *y) const = 0;
void (*Compute)(const T *, T *, int);
};
template <typename T>
class VReluKernel : public VActKernel<T> {
public:
virtual void ComputeDeprecated(const T *x, T *y) const = 0;
void (*Compute)(const T *, T *, int);
};
class VReluKernel : public VActKernel<T> {};
template <typename T>
class VIdentityKernel : public VActKernel<T> {
public:
virtual void ComputeDeprecated(const T *x, T *y) const = 0;
};
class VIdentityKernel : public VActKernel<T> {};
template <typename T>
class VExpKernel : public VActKernel<T> {
public:
virtual void ComputeDeprecated(const T *x, T *y) const = 0;
};
class VExpKernel : public VActKernel<T> {};
template <typename T>
class VSigmoidKernel : public VActKernel<T> {
public:
virtual void ComputeDeprecated(const T *x, T *y) const = 0;
};
class VSigmoidKernel : public VActKernel<T> {};
template <typename T>
class VTanhKernel : public VActKernel<T> {
public:
virtual void ComputeDeprecated(const T *x, T *y) const = 0;
};
class VTanhKernel : public VActKernel<T> {};
template <typename T>
class LSTMKernel : public Kernel {

74
paddle/fluid/operators/math/jit_kernel_blas.cc
Unescape View File

@ -25,10 +25,6 @@ limitations under the License. */
#include "paddle/fluid/platform/dynload/mklml.h"
#endif
#ifdef __AVX__
#include <immintrin.h>
#endif
namespace paddle {
namespace operators {
namespace math {
@ -128,23 +124,16 @@ void VScalMKL<double>(const double* a, const double* x, double* y, int n) {
#endif
#define DECLARE_STATIC_FUNC \
static inline std::string name(int d) { \
PADDLE_THROW("DType should be either float or double"); \
} \
static inline bool useJIT(int d) { return false; } \
static inline bool useMKL(int d) { return false; }
/* VMUL JitKernel */
template <typename T>
class VMulKernelImpl : public VMulKernel<T> {
public:
DECLARE_STATIC_FUNC;
JITKERNEL_DECLARE_STATIC_FUNC;
explicit VMulKernelImpl(int d) : VMulKernel<T>() {
#ifdef PADDLE_WITH_XBYAK
if (useJIT(d)) {
// roughly estimate the size of code
size_t sz = 96 + d / AVX_FLOAT_BLOCK * 4 * 8;
size_t sz = 96 + d / YMM_FLOAT_BLOCK * 4 * 8;
jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::mul, 0, false,
sz > 4096 ? sz : 4096));
this->Compute =
@ -191,11 +180,11 @@ bool VMulKernelImpl<double>::useMKL(int d) {
template <typename T>
class VAddKernelImpl : public VAddKernel<T> {
public:
DECLARE_STATIC_FUNC;
JITKERNEL_DECLARE_STATIC_FUNC;
explicit VAddKernelImpl(int d) : VAddKernel<T>() {
#ifdef PADDLE_WITH_XBYAK
if (useJIT(d)) {
size_t sz = 96 + d / AVX_FLOAT_BLOCK * 4 * 8;
size_t sz = 96 + d / YMM_FLOAT_BLOCK * 4 * 8;
jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::add, 0, false,
sz > 4096 ? sz : 4096));
this->Compute =
@ -241,11 +230,11 @@ bool VAddKernelImpl<double>::useMKL(int d) {
template <typename T>
class VAddReluKernelImpl : public VAddReluKernel<T> {
public:
DECLARE_STATIC_FUNC;
JITKERNEL_DECLARE_STATIC_FUNC;
explicit VAddReluKernelImpl(int d) : VAddReluKernel<T>() {
#ifdef PADDLE_WITH_XBYAK
if (useJIT(d)) {
size_t sz = 96 + d / AVX_FLOAT_BLOCK * 4 * 8;
size_t sz = 96 + d / YMM_FLOAT_BLOCK * 4 * 8;
jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::add, 0, true,
sz > 4096 ? sz : 4096));
this->Compute =
@ -273,11 +262,11 @@ bool VAddReluKernelImpl<float>::useJIT(int d) {
template <typename T>
class VScalKernelImpl : public VScalKernel<T> {
public:
DECLARE_STATIC_FUNC;
JITKERNEL_DECLARE_STATIC_FUNC;
explicit VScalKernelImpl(int d) : VScalKernel<T>() {
#ifdef PADDLE_WITH_XBYAK
if (useJIT(d)) {
size_t sz = 96 + d / AVX_FLOAT_BLOCK * 4 * 8;
size_t sz = 96 + d / YMM_FLOAT_BLOCK * 4 * 8;
jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::mul, 1, false,
sz > 4096 ? sz : 4096));
this->Compute =
@ -322,11 +311,11 @@ bool VScalKernelImpl<double>::useMKL(int d) {
template <typename T>
class VAddBiasKernelImpl : public VAddBiasKernel<T> {
public:
DECLARE_STATIC_FUNC;
JITKERNEL_DECLARE_STATIC_FUNC;
explicit VAddBiasKernelImpl(int d) : VAddBiasKernel<T>() {
#ifdef PADDLE_WITH_XBYAK
if (useJIT(d)) {
size_t sz = 96 + d / AVX_FLOAT_BLOCK * 4 * 8;
size_t sz = 96 + d / YMM_FLOAT_BLOCK * 4 * 8;
jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::add, 1, false,
sz > 4096 ? sz : 4096));
this->Compute =
@ -355,15 +344,15 @@ bool VAddBiasKernelImpl<float>::useJIT(int d) {
template <typename T>
class VReluKernelImpl : public VReluKernel<T> {
public:
DECLARE_STATIC_FUNC;
JITKERNEL_DECLARE_STATIC_FUNC;
explicit VReluKernelImpl(int d) : VReluKernel<T>() {
this->num_ = d; // TODO(TJ): remove me when ComputeDeprecated done
#ifdef PADDLE_WITH_XBYAK
if (useJIT(d)) {
size_t sz = 96 /*init*/ +
d / AVX_FLOAT_BLOCK * 4 /* instructions*/ *
8 /*everage byte for each instruction*/;
jitcode_.reset(new gen::ReluJitCode(d, sz > 4096 ? sz : 4096));
size_t sz = 96 /* init size */ +
d / YMM_FLOAT_BLOCK * 4 /* instructions */ *
8 /* average bytes for each instruction */;
jitcode_.reset(new gen::VActJitCode(d, gen::operand_type::relu,
sz > 4096 ? sz : 4096));
this->Compute = jitcode_->getCode<void (*)(const T*, T*, int)>();
return;
}
@ -371,24 +360,32 @@ class VReluKernelImpl : public VReluKernel<T> {
this->Compute = VReluRefer<T>;
}
void ComputeDeprecated(const T* x, T* y) const override {
VReluRefer(x, y, this->num_);
}
#ifdef PADDLE_WITH_XBYAK
private:
std::unique_ptr<gen::ReluJitCode> jitcode_{nullptr};
std::unique_ptr<gen::VActJitCode> jitcode_{nullptr};
#endif
};
#ifdef PADDLE_WITH_XBYAK
template <>
bool VReluKernelImpl<float>::useJIT(int d) {
return gen::ReluJitCode::init(d);
return gen::VActJitCode::init(d, gen::operand_type::relu);
}
#endif
#undef DECLARE_STATIC_FUNC
template <typename T>
inline void VIdentityRefer(const T* x, T* y, int n) {}
/* An empty JitKernel */
template <typename T>
class VIdentityKernelImpl : public VIdentityKernel<T> {
public:
JITKERNEL_DECLARE_STATIC_FUNC;
explicit VIdentityKernelImpl(int d) : VIdentityKernel<T>() {
this->Compute = VIdentityRefer<T>;
}
};
REGISTER_JITKERNEL(vmul, VMulKernel);
REGISTER_JITKERNEL(vadd, VAddKernel);
@ -396,16 +393,7 @@ REGISTER_JITKERNEL(vaddrelu, VAddReluKernel);
REGISTER_JITKERNEL(vscal, VScalKernel);
REGISTER_JITKERNEL(vaddbias, VAddBiasKernel);
REGISTER_JITKERNEL(vrelu, VReluKernel);
/* An empty JitKernel */
template <typename T, platform::jit::cpu_isa_t isa, jit_block>
class VIdentityKernelImpl : public VIdentityKernel<T> {
public:
explicit VIdentityKernelImpl(int d) : VIdentityKernel<T>() { this->num_ = d; }
void ComputeDeprecated(const T* x, T* y) const override {}
};
REGISTER_JITKERNEL_DEPRECATED(videntity, VIdentityKernel);
REGISTER_JITKERNEL(videntity, VIdentityKernel);
} // namespace jitkernel
} // namespace math

24
paddle/fluid/operators/math/jit_kernel_crf_decode.cc
Unescape View File

@ -105,14 +105,14 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
int tag_num) \
: CRFDecodeKernel<float>() { \
this->num_ = tag_num; \
this->end_ = this->num_ / AVX_FLOAT_BLOCK; \
this->rest_ = this->num_ % AVX_FLOAT_BLOCK; \
this->end_ = this->num_ / YMM_FLOAT_BLOCK; \
this->rest_ = this->num_ % YMM_FLOAT_BLOCK; \
} \
template <> \
void CRFDecodeKernelImpl<float, jit::avx, block>::Compute( \
const int seq_len, const float* x, const float* w, float* alpha, \
int* track) const { \
INIT_ALPHA(AVX_FLOAT_BLOCK) \
INIT_ALPHA(YMM_FLOAT_BLOCK) \
/* Use the column-major strategy to get the location of maximum score.*/ \
int seq_offset = 0; \
constexpr int state_trans_base_idx = 2; \
@ -150,7 +150,7 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
max_score = _mm256_max_ps(max_score, score_v); \
trans_offset += this->num_; \
} \
UPDATE_ALPHA(AVX_FLOAT_BLOCK) \
UPDATE_ALPHA(YMM_FLOAT_BLOCK) \
} \
seq_offset += this->num_; \
} \
@ -161,14 +161,14 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
CRFDecodeKernelImpl<float, isa, block>::CRFDecodeKernelImpl(int tag_num) \
: CRFDecodeKernel<float>() { \
this->num_ = tag_num; \
this->end_ = this->num_ / AVX2_FLOAT_BLOCK; \
this->rest_ = this->num_ % AVX2_FLOAT_BLOCK; \
this->end_ = this->num_ / YMM_FLOAT_BLOCK; \
this->rest_ = this->num_ % YMM_FLOAT_BLOCK; \
} \
template <> \
void CRFDecodeKernelImpl<float, isa, block>::Compute( \
const int seq_len, const float* x, const float* w, float* alpha, \
int* track) const { \
INIT_ALPHA(AVX2_FLOAT_BLOCK) \
INIT_ALPHA(YMM_FLOAT_BLOCK) \
/* Use the column-major strategy to get the location of maximum score.*/ \
int seq_offset = 0; \
constexpr int state_trans_base_idx = 2; \
@ -196,7 +196,7 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
max_score = _mm256_max_ps(max_score, score_v); \
trans_offset += this->num_; \
} \
UPDATE_ALPHA(AVX2_FLOAT_BLOCK) \
UPDATE_ALPHA(YMM_FLOAT_BLOCK) \
} \
seq_offset += this->num_; \
} \
@ -208,14 +208,14 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
int tag_num) \
: CRFDecodeKernel<float>() { \
this->num_ = tag_num; \
this->end_ = this->num_ / AVX512_FLOAT_BLOCK; \
this->rest_ = this->num_ % AVX512_FLOAT_BLOCK; \
this->end_ = this->num_ / ZMM_FLOAT_BLOCK; \
this->rest_ = this->num_ % ZMM_FLOAT_BLOCK; \
} \
template <> \
void CRFDecodeKernelImpl<float, jit::avx512f, block>::Compute( \
const int seq_len, const float* x, const float* w, float* alpha, \
int* track) const { \
INIT_ALPHA(AVX512_FLOAT_BLOCK) \
INIT_ALPHA(ZMM_FLOAT_BLOCK) \
/* Use the column-major strategy to get the location of maximum score.*/ \
int seq_offset = 0; \
constexpr int state_trans_base_idx = 2; \
@ -250,7 +250,7 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
this->num_ + j_offset), \
max_j); \
/* Calculate the offset of next step*/ \
j_offset += AVX512_FLOAT_BLOCK; \
j_offset += ZMM_FLOAT_BLOCK; \
if (j == this->end_ - 1) { \
if (this->rest_ > 0) { \
j_offset += last_offset; \

574
paddle/fluid/operators/math/jit_kernel_exp.cc
View File

File diff suppressed because it is too large Load Diff

30
paddle/fluid/operators/math/jit_kernel_macro.h
Unescape View File

@ -15,12 +15,20 @@ limitations under the License. */
#pragma once
#include <string>
#include "paddle/fluid/platform/cpu_info.h"
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace operators {
namespace math {
namespace jitkernel {
#define JITKERNEL_DECLARE_STATIC_FUNC \
static inline std::string name(int d) { \
PADDLE_THROW("DType should be either float or double"); \
} \
static inline bool useJIT(int d) { return false; } \
static inline bool useMKL(int d) { return false; }
#define JITKERNEL_DEFINE_NAME(ker_key, ker_class) \
template <> \
std::string ker_class##Impl<float>::name(int d) { \
@ -86,17 +94,17 @@ namespace jitkernel {
namespace jit = platform::jit;
// TODO(TJ): below defines are deprecated, would be remove recently
#define SEARCH_BLOCK(macro_, ker, dtype, isa) \
if (d < AVX_FLOAT_BLOCK) { \
macro_(ker, dtype, isa, kLT8); \
} else if (d == AVX_FLOAT_BLOCK) { \
macro_(ker, dtype, isa, kEQ8); \
} else if (d > AVX_FLOAT_BLOCK && d < AVX512_FLOAT_BLOCK) { \
macro_(ker, dtype, isa, kGT8LT16); \
} else if (d == AVX512_FLOAT_BLOCK) { \
macro_(ker, dtype, isa, kEQ16); \
} else { \
macro_(ker, dtype, isa, kGT16); \
#define SEARCH_BLOCK(macro_, ker, dtype, isa) \
if (d < YMM_FLOAT_BLOCK) { \
macro_(ker, dtype, isa, kLT8); \
} else if (d == YMM_FLOAT_BLOCK) { \
macro_(ker, dtype, isa, kEQ8); \
} else if (d > YMM_FLOAT_BLOCK && d < ZMM_FLOAT_BLOCK) { \
macro_(ker, dtype, isa, kGT8LT16); \
} else if (d == ZMM_FLOAT_BLOCK) { \
macro_(ker, dtype, isa, kEQ16); \
} else { \
macro_(ker, dtype, isa, kGT16); \
}
#define SEARCH_ISA_BLOCK(macro_, ker, dtype) \

38
paddle/fluid/operators/math/jit_kernel_rnn.cc
Unescape View File

@ -175,26 +175,26 @@ class LSTMKernelImpl : public LSTMKernel<T> {
void ComputeCtHt(T* gates, const T* ct_1, T* ct, T* ht, const T* wp_data,
T* checked) const override {
// gates: W_ch, W_ih, W_fh, W_oh
act_gate_d3_->ComputeDeprecated(gates + d_, gates + d_);
act_gate_d3_->Compute(gates + d_, gates + d_, d3_);
/* C_t = C_t-1 * fgated + cand_gated * igated */
act_cand_d_->ComputeDeprecated(gates, gates);
act_cand_d_->Compute(gates, gates, d_);
vmul_d_->Compute(gates, gates + d_, gates + d_, d_);
vmul_d_->Compute(ct_1, gates + d2_, gates + d2_, d_);
vadd_d_->Compute(gates + d_, gates + d2_, ct, d_);
/* H_t = act_cell(C_t) * ogated */
act_cell_d_->ComputeDeprecated(ct, gates + d2_);
act_cell_d_->Compute(ct, gates + d2_, d_);
vmul_d_->Compute(gates + d2_, gates + d3_, ht, d_);
}
void ComputeC1H1(T* gates, T* ct, T* ht, const T* wp_data) const override {
/* C_t = igated * cgated*/
act_gate_d_->ComputeDeprecated(gates + d_, gates + d_);
act_cand_d_->ComputeDeprecated(gates, gates);
act_gate_d_->Compute(gates + d_, gates + d_, d_);
act_cand_d_->Compute(gates, gates, d_);
vmul_d_->Compute(gates, gates + d_, ct, d_);
/* H_t = act_cell(C_t) * ogated */
act_gate_d_->ComputeDeprecated(gates + d3_, gates + d3_);
act_cell_d_->ComputeDeprecated(ct, gates + d2_);
act_gate_d_->Compute(gates + d3_, gates + d3_, d_);
act_cell_d_->Compute(ct, gates + d2_, d_);
vmul_d_->Compute(gates + d2_, gates + d3_, ht, d_);
}
@ -292,32 +292,32 @@ class PeepholeKernelImpl : public LSTMKernel<T> {
vmul_d_->Compute(wp_data, ct_1, checked, d_);
vmul_d_->Compute(wp_data + d_, ct_1, checked + d_, d_);
vadd_d2_->Compute(checked, gates + d_, gates + d_, d2_);
act_gate_d2_->ComputeDeprecated(gates + d_, gates + d_);
act_gate_d2_->Compute(gates + d_, gates + d_, d2_);
/* C_t = C_t-1 * fgated + cand_gated * igated*/
act_cand_d_->ComputeDeprecated(gates, gates);
act_cand_d_->Compute(gates, gates, d_);
vmul_d_->Compute(gates, gates + d_, gates + d_, d_);
vmul_d_->Compute(ct_1, gates + d2_, gates + d2_, d_);
vadd_d_->Compute(gates + d_, gates + d2_, ct, d_);
/* get ogated*/
vmul_d_->Compute(wp_data + d2_, ct, gates + d_, d_);
vadd_d_->Compute(gates + d_, gates + d3_, gates + d3_, d_);
act_gate_d_->ComputeDeprecated(gates + d3_, gates + d3_);
act_gate_d_->Compute(gates + d3_, gates + d3_, d_);
/* H_t = act_cell(C_t) * ogated */
act_cell_d_->ComputeDeprecated(ct, gates + d2_);
act_cell_d_->Compute(ct, gates + d2_, d_);
vmul_d_->Compute(gates + d2_, gates + d3_, ht, d_);
}
void ComputeC1H1(T* gates, T* ct, T* ht, const T* wp_data) const override {
/* C_t = igated * cgated*/
act_gate_d_->ComputeDeprecated(gates + d_, gates + d_);
act_cand_d_->ComputeDeprecated(gates, gates);
act_gate_d_->Compute(gates + d_, gates + d_, d_);
act_cand_d_->Compute(gates, gates, d_);
vmul_d_->Compute(gates, gates + d_, ct, d_);
/* get outgated, put W_oc * C_t on igated */
vmul_d_->Compute(wp_data + d2_, ct, gates + d_, d_);
vadd_d_->Compute(gates + d_, gates + d3_, gates + d3_, d_);
/* H_t = act_cell(C_t) * ogated */
act_gate_d_->ComputeDeprecated(gates + d3_, gates + d3_);
act_cell_d_->ComputeDeprecated(ct, gates + d2_);
act_gate_d_->Compute(gates + d3_, gates + d3_, d_);
act_cell_d_->Compute(ct, gates + d2_, d_);
vmul_d_->Compute(gates + d2_, gates + d3_, ht, d_);
}
@ -376,20 +376,20 @@ class GRUKernelImpl : public GRUKernel<T> {
}
void ComputeH1(T* gates, T* ht) const override {
act_gate_d_->ComputeDeprecated(gates, gates);
act_state_d_->ComputeDeprecated(gates + d2_, gates + d2_);
act_gate_d_->Compute(gates, gates, d_);
act_state_d_->Compute(gates + d2_, gates + d2_, d_);
vmul_d_->Compute(gates, gates + d2_, ht, d_);
}
void ComputeHtPart1(T* gates, const T* ht_1, T* ht) const override {
// W: {W_update, W_reset; W_state}
act_gate_d2_->ComputeDeprecated(gates, gates);
act_gate_d2_->Compute(gates, gates, d2_);
vmul_d_->Compute(ht_1, gates + d_, ht, d_);
}
void ComputeHtPart2(T* gates, const T* ht_1, T* ht) const override {
T* y = gates + d2_;
act_state_d_->ComputeDeprecated(y, y);
act_state_d_->Compute(y, y, d_);
// out = zt*ht~ + (1-zt)*ht_1
for (int i = 0; i < d_; ++i) {
ht[i] = gates[i] * y[i] + (static_cast<T>(1) - gates[i]) * ht_1[i];

25
paddle/fluid/operators/math/jit_kernel_test.cc
Unescape View File

@ -181,7 +181,8 @@ TEST(JitKernel, vexp) {
auto ttgts = GetCurrentUS();
for (int i = 0; i < repeat; ++i) {
ker->ComputeDeprecated(x_data, ztgt_data);
// ker->Compute(x_data, ztgt_data);
ker->Compute(x_data, ztgt_data, d);
}
auto ttgte = GetCurrentUS();
@ -222,7 +223,7 @@ void vsigmoid_better(
y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
y[i] = 0.f - y[i];
}
vexp->ComputeDeprecated(y, y);
vexp->Compute(y, y, n);
for (int i = 0; i < n; ++i) {
y[i] = 1.f / (1.f + y[i]);
}
@ -253,7 +254,7 @@ TEST(JitKernel, vsigmoid) {
auto trefe = GetCurrentUS();
auto ttgts = GetCurrentUS();
for (int i = 0; i < repeat; ++i) {
ker->ComputeDeprecated(x_data, ztgt_data);
ker->Compute(x_data, ztgt_data, d);
}
auto ttgte = GetCurrentUS();
@ -287,7 +288,7 @@ void vtanh_better(
const int n, const float* x, float* y) {
const float a = 2.f, b = -1.f;
vscal->Compute(&a, x, y, n);
vsigmoid->ComputeDeprecated(y, y);
vsigmoid->Compute(y, y, n);
vscal->Compute(&a, y, y, n);
vaddbias->Compute(&b, y, y, n);
}
@ -321,7 +322,7 @@ TEST(JitKernel, vtanh) {
auto trefe = GetCurrentUS();
auto ttgts = GetCurrentUS();
for (int i = 0; i < repeat; ++i) {
ker->ComputeDeprecated(x_data, ztgt_data);
ker->Compute(x_data, ztgt_data, d);
}
auto ttgte = GetCurrentUS();
@ -344,8 +345,8 @@ void lstm_ctht_ref(
const std::shared_ptr<
const paddle::operators::math::jitkernel::VExpKernel<float>>& vexp_1,
const int d, float* gates, const float* ct_1, float* ct, float* ht) {
vsigmoid_3d->ComputeDeprecated(gates + d, gates + d);
vtanh_d->ComputeDeprecated(gates, gates);
vsigmoid_3d->Compute(gates + d, gates + d, 3 * d);
vtanh_d->Compute(gates, gates, d);
const float *i = gates + d, *f = gates + d * 2, *o = gates + d * 3;
const float min = SIGMOID_THRESHOLD_MIN;
const float max = SIGMOID_THRESHOLD_MAX;
@ -355,7 +356,7 @@ void lstm_ctht_ref(
// H_t = act_cell(C_t) * ogated
float tmp = ct[k] * 2;
tmp = 0.f - ((tmp < min) ? min : ((tmp > max) ? max : tmp));
vexp_1->ComputeDeprecated(&tmp, &tmp);
vexp_1->Compute(&tmp, &tmp, 1);
tmp = 2.f / (1.f + tmp) - 1.f;
ht[k] = tmp * o[k];
}
@ -373,13 +374,13 @@ void lstm_ctht_better(
const paddle::operators::math::jitkernel::VAddKernel<float>>& vadd_d,
const int d, float* gates, const float* ct_1, float* ct, float* ht) {
int d2 = d * 2;
vsigmoid_3d->ComputeDeprecated(gates + d, gates + d);
vtanh_d->ComputeDeprecated(gates, gates);
vsigmoid_3d->Compute(gates + d, gates + d, 3 * d);
vtanh_d->Compute(gates, gates, d);
vmul_d->Compute(gates, gates + d, gates + d, d);
vmul_d->Compute(ct_1, gates + d2, gates + d2, d);
vadd_d->Compute(gates + d, gates + d2, ct, d);
/* H_t = act_cell(C_t) * ogated */
vtanh_d->ComputeDeprecated(ct, gates + d2);
vtanh_d->Compute(ct, gates + d2, d);
vmul_d->Compute(gates + d2, gates + d * 3, ht, d);
}
@ -736,7 +737,7 @@ void vaddrelu_better(
const paddle::operators::math::jitkernel::VReluKernel<float>>& vrelu,
const float* x, const float* y, float* z, int d) {
vadd->Compute(x, y, z, d);
vrelu->ComputeDeprecated(z, z);
vrelu->Compute(z, z, d);
}
TEST(JitKernel, vaddrelu) {

8
python/paddle/fluid/tests/book/test_label_semantic_roles.py
Unescape View File

@ -38,7 +38,7 @@ depth = 8
mix_hidden_lr = 1e-3
IS_SPARSE = True
PASS_NUM = 1
PASS_NUM = 2
BATCH_SIZE = 10
embedding_name = 'emb'
@ -196,7 +196,7 @@ def train(use_cuda, save_dirname=None, is_local=True):
print("second per batch: " + str((time.time(
) - start_time) / batch_id))
# Set the threshold low to speed up the CI test
if float(cost) < 60.0:
if float(cost) < 80.0:
if save_dirname is not None:
# TODO(liuyiqun): Change the target to crf_decode
fluid.io.save_inference_model(save_dirname, [
@ -208,6 +208,10 @@ def train(use_cuda, save_dirname=None, is_local=True):
batch_id = batch_id + 1
raise RuntimeError(
"This model should save_inference_model and return, but not reach here, please check!"
)
if is_local:
train_loop(fluid.default_main_program())
else:

Write Preview
Loading…
Cancel
Save