Skip to content

P
PaddleDetection

PaddlePaddle / PaddleDetection
大约 1 年 前同步成功

通知 695
Star 11112
Fork 2696
P
PaddleDetection
提交 6648995f 编写于 作者: T tensor-tang
浏览文件
操作

fix build

上级 74292f41
隐藏空白更改
内联 并排
Showing with 61 addition and 61 deletion
paddle/fluid/operators/crf_decoding_op.h
浏览文件 @ 6648995f
......@@ -82,8 +82,8 @@ class CRFDecodingOpKernel : public framework::OpKernel<T> {
Tensor track;
int* track_value =
track.mutable_data<int>(emission_dims, platform::CPUPlace());
auto ker = jit::Get<jit::crfdecoding, jit::CRFDecoding, platform::CPUPlace>(
tag_num);
auto ker = jit::Get<jit::crfdecoding, jit::CRFDecodingTuples<T>,
platform::CPUPlace>(tag_num);
ker(static_cast<int>(seq_len), x, w, alpha_value, track_value, tag_num);
T max_score = -std::numeric_limits<T>::max();
int max_i = 0;
......
paddle/fluid/operators/elementwise/elementwise_mul_mkldnn_op.cc
浏览文件 @ 6648995f
......@@ -108,7 +108,7 @@ class ElementwiseMulMKLDNNKernel : public framework::OpKernel<T> {
constexpr int simd_width = 16;
int C = c / simd_width;
auto multiply = jit::Get<jit::nchw16cmulnc, jit::NCHW16CMulNCTuples,
auto multiply = jit::Get<jit::nchw16cmulnc, jit::NCHW16CMulNCTuples<T>,
platform::CPUPlace>(0);
#pragma omp parallel for collapse(2)
for (int ni = 0; ni < n; ni++) {
......
paddle/fluid/operators/fused/fusion_gru_op.cc
浏览文件 @ 6648995f
......@@ -183,29 +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 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(); \
#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<T>, platform::CPUPlace>(attr); \
auto ComputeHtPart1 = \
jit::Get<jit::gruhtpart1, jit::GRUTuples<T>, platform::CPUPlace>(attr); \
auto ComputeHtPart2 = \
jit::Get<jit::gruhtpart2, jit::GRUTuples<T>, 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 {
......
paddle/fluid/operators/fused/fusion_lstm_op.cc
浏览文件 @ 6648995f
......@@ -236,33 +236,32 @@ 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 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)
#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); \
jit::lstm_t one_step; \
one_step.wp = wp_data; \
one_step.checked = checked_cell_data; \
auto ComputeC1H1 = \
jit::Get<jit::lstmc1h1, jit::LSTMTuples<T>, platform::CPUPlace>(attr); \
auto ComputeCtHt = \
jit::Get<jit::lstmctht, jit::LSTMTuples<T>, platform::CPUPlace>(attr)
// Wh GEMM
#define GEMM_WH_ADDON(bs, prev, out) \
......@@ -434,7 +433,7 @@ 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;
ComputeC1H1(&one_step, &attr);
ComputeCtHt(&one_step, &attr);
// move one batch
cur_in_data += D4;
......
paddle/fluid/operators/jit/helper.h
浏览文件 @ 6648995f
......@@ -32,7 +32,7 @@ inline typename std::enable_if<
std::is_same<typename KernelTuples::data_type, float>::value &&
std::is_same<PlaceType, platform::CPUPlace>::value,
typename KernelTuples::func_type>::type
GetJitCode(typename KernelTuples::attr_type attr) {
GetJitCode(const typename KernelTuples::attr_type& attr) {
using Func = typename KernelTuples::func_type;
using Attr = typename KernelTuples::attr_type;
size_t key = JitCodeKey<Attr>(attr);
......@@ -68,7 +68,7 @@ inline typename std::enable_if<
!std::is_same<typename KernelTuples::data_type, float>::value ||
!std::is_same<PlaceType, platform::CPUPlace>::value,
typename KernelTuples::func_type>::type
GetJitCode(typename KernelTuples::attr_type attr) {
GetJitCode(const typename KernelTuples::attr_type& attr) {
return nullptr;
}
......@@ -93,8 +93,8 @@ inline typename KernelTuples::func_type GetRefer() {
template <KernelType KT, typename KernelTuples,
typename PlaceType = platform::CPUPlace>
// TODO(TJ): const & attr
typename KernelTuples::func_type Get(typename KernelTuples::attr_type attr) {
typename KernelTuples::func_type Get(
const typename KernelTuples::attr_type& attr) {
auto jitfunc = GetJitCode<KT, KernelTuples, PlaceType>(attr);
if (jitfunc) {
return jitfunc;
......
paddle/fluid/operators/layer_norm_op.h
浏览文件 @ 6648995f
......@@ -230,7 +230,7 @@ class LayerNormKernel : public framework::OpKernel<T> {
PADDLE_ENFORCE_EQ(bias->numel(), right);
auto ker =
jit::Get<jit::layernorm, jit::LayerNormTuples, platform::CPUPlace>(
jit::Get<jit::layernorm, jit::LayerNormTuples<T>, platform::CPUPlace>(
right);
ker(x.data<T>(), out.data<T>(), mean->data<T>(), var->data<T>(),
scale->data<T>(), bias->data<T>(), static_cast<int>(left),
......
paddle/fluid/operators/math/fc_compute.h
浏览文件 @ 6648995f
......@@ -31,13 +31,14 @@ inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
}
if (relu) {
auto compute =
jit::Get<jit::vaddrelu, jit::XYZNTuples, platform::CPUPlcace>(N);
jit::Get<jit::vaddrelu, jit::XYZNTuples<T>, platform::CPUPlace>(N);
for (int i = 0; i < M; i++) {
T* dst = Y + i * N;
compute(B, dst, dst, N);
}
} else {
auto compute = jit::Get<jit::vadd, jit::XYZNTuples, platform::CPUPlcace>(N);
auto compute =
jit::Get<jit::vadd, jit::XYZNTuples<T>, platform::CPUPlace>(N);
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
#endif
......
Markdown is supported
0% 或 .
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 或