Merge branch 'develop' into clean_inference_lib

paddle/fluid/API.spec

@@ -75,7 +75,8 @@ paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'outp
 paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
 paddle.fluid.layers.sequence_expand ArgSpec(args=['x', 'y', 'ref_level', 'name'], varargs=None, keywords=None, defaults=(-1, None))
 paddle.fluid.layers.sequence_expand_as ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen', 'name'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.layers.sequence_unpad ArgSpec(args=['x', 'length', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lstm_unit ArgSpec(args=['x_t', 'hidden_t_prev', 'cell_t_prev', 'forget_bias', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(0.0, None, None, None))
 paddle.fluid.layers.reduce_sum ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
 paddle.fluid.layers.reduce_mean ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
@@ -84,6 +85,7 @@ paddle.fluid.layers.reduce_min ArgSpec(args=['input', 'dim', 'keep_dim', 'name']
 paddle.fluid.layers.reduce_prod ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
 paddle.fluid.layers.sequence_first_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.sequence_last_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.layers.sequence_slice ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name'], varargs=None, keywords=None, defaults=(False, None, None))
 paddle.fluid.layers.split ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None))
 paddle.fluid.layers.ctc_greedy_decoder ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,))

paddle/fluid/framework/executor.cc

@@ -101,7 +101,7 @@ void InitializeVariable(Variable* var, proto::VarType::Type var_type) {
   } else if (var_type == proto::VarType::FETCH_LIST) {
     var->GetMutable<FeedFetchList>();
   } else if (var_type == proto::VarType::STEP_SCOPES) {
-    var->GetMutable<std::vector<framework::Scope>>();
+    var->GetMutable<std::vector<framework::Scope*>>();
   } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
     var->GetMutable<LoDRankTable>();
   } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {

paddle/fluid/framework/feed_fetch_method.cc

@@ -27,8 +27,7 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input,
   // be created.
   VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index;
   Variable* g_feed_value = scope->Var(var_name);
-  auto& feed_inputs =
-      *(g_feed_value->GetMutable<std::vector<paddle::framework::LoDTensor>>());
+  auto& feed_inputs = *(g_feed_value->GetMutable<FeedFetchList>());
   if (index >= feed_inputs.size()) {
     feed_inputs.resize(index + 1);
   }

paddle/fluid/framework/naive_executor.cc

@@ -37,7 +37,7 @@ static void InitializeVariable(Variable *var, proto::VarType::Type var_type) {
   } else if (var_type == proto::VarType::FETCH_LIST) {
     var->GetMutable<FeedFetchList>();
   } else if (var_type == proto::VarType::STEP_SCOPES) {
-    var->GetMutable<std::vector<framework::Scope>>();
+    var->GetMutable<std::vector<framework::Scope *>>();
   } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
     var->GetMutable<LoDRankTable>();
   } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {

paddle/fluid/framework/operator.cc

@@ -149,9 +149,17 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
     platform::SetDeviceId(dev_id);
 #endif
   }
-  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-  platform::RecordEvent record_event(Type(), pool.Get(place));
-  RunImpl(scope, place);
+
+  // The profile has a process-wide mutex, results in serious performance issue
+  // in concurrency scenerio. Here use an `if` to fix this issue.
+  // Please not remove the `if`, ask @Superjomn if there are any concern.
+  if (platform::IsProfileEnabled()) {
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    platform::RecordEvent record_event(Type(), pool.Get(place));
+    RunImpl(scope, place);
+  } else {
+    RunImpl(scope, place);
+  }
   VLOG(3) << place << " " << DebugStringEx(&scope);
 }
 

paddle/fluid/framework/tensor_util.cc

@@ -36,6 +36,11 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
   auto size = src.numel() * SizeOfType(src.type());
 
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
+    if (src_ptr == dst_ptr) {
+      VLOG(3) << "Skip copy the same data async from " << src_place << " to "
+              << dst_place;
+      return;
+    }
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                  boost::get<platform::CPUPlace>(src_place), src_ptr, size);
   }
@@ -71,6 +76,11 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
     auto stream =
         reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
     if (platform::is_same_place(src_place, dst_place)) {
+      if (src_ptr == dst_ptr) {
+        VLOG(3) << "Skip copy the same data async from " << src_place << " to "
+                << dst_place;
+        return;
+      }
       memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
                    stream);
     } else {
@@ -114,6 +124,11 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
   auto dst_ptr = dst->mutable_data(dst_place, src.type());
   auto size = src.numel() * SizeOfType(src.type());
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
+    if (src_ptr == dst_ptr) {
+      VLOG(3) << "Skip copy the same data from " << src_place << " to "
+              << dst_place;
+      return;
+    }
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                  boost::get<platform::CPUPlace>(src_place), src_ptr, size);
   }
@@ -130,6 +145,11 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
     memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, nullptr);
   } else if (platform::is_gpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
+    if (src_ptr == dst_ptr && platform::is_same_place(src_place, dst_place)) {
+      VLOG(3) << "Skip copy the same data from " << src_place << " to "
+              << dst_place;
+      return;
+    }
     auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
     auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
     memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);

paddle/fluid/framework/tensor_util_test.cc

@@ -41,6 +41,11 @@ TEST(TensorCopy, Tensor) {
     EXPECT_EQ(src_ptr[i], dst_ptr[i]);
   }
 
+  TensorCopy(dst_tensor, *cpu_place, &dst_tensor);
+  for (size_t i = 0; i < 9; ++i) {
+    EXPECT_EQ(src_ptr[i], dst_ptr[i]);
+  }
+
   EXPECT_TRUE(dst_tensor.layout() == src_tensor.layout());
 
   Tensor slice_tensor = src_tensor.Slice(1, 2);
@@ -82,6 +87,15 @@ TEST(TensorCopy, Tensor) {
       EXPECT_EQ(src_ptr[i], dst_ptr[i]);
     }
 
+    // Copy the same tensor
+    TensorCopy(gpu_tensor, *gpu_place, gpu_ctx, &gpu_tensor);
+    gpu_ctx.Wait();
+    const int* dst_ptr_tmp = dst_tensor.data<int>();
+    EXPECT_NE(src_ptr, dst_ptr_tmp);
+    for (size_t i = 0; i < 9; ++i) {
+      EXPECT_EQ(src_ptr[i], dst_ptr_tmp[i]);
+    }
+
     Tensor slice_tensor = src_tensor.Slice(1, 2);
 
     // CPU Slice Tensor to GPU Tensor

paddle/fluid/framework/var_desc.h

@@ -59,6 +59,7 @@ class VarDesc {
  public:
   explicit VarDesc(const std::string &name) {
     desc_.set_name(name);
+    // TODO(paddle-dev): Why default to lodtensor.
     desc_.mutable_type()->set_type(proto::VarType::LOD_TENSOR);
   }
 

paddle/fluid/framework/variable.h

@@ -38,8 +38,12 @@ class Variable {
 
   template <typename T>
   T* GetMutable() {
-    if (!IsType<T>()) {
+    if (!holder_) {
       holder_.reset(new PlaceholderImpl<T>(new T()));
+    } else {
+      PADDLE_ENFORCE(IsType<T>(),
+                     "Variable must be type %s, the holding type is %s",
+                     typeid(T).name(), holder_->Type().name());
     }
     return static_cast<T*>(holder_->Ptr());
   }

paddle/fluid/framework/variable_test.cc

@@ -33,9 +33,10 @@ TEST(Variable, GetMutable) {
   const Tensor& tt = v->Get<Tensor>();
   EXPECT_EQ(1234, tt.content_);
 
-  std::string* s = v->GetMutable<std::string>();
-  *s = "hello";
-
-  const std::string& ss = v->Get<std::string>();
-  EXPECT_EQ("hello", ss);
+  try {
+    v->GetMutable<std::string>();
+  } catch (std::exception& e) {
+    return;
+  }
+  EXPECT_TRUE(false);
 }

paddle/fluid/inference/api/analysis_predictor.cc

@@ -340,6 +340,19 @@ bool AnalysisPredictor::LoadProgramDesc() {
   }
   return true;
 }
+
+AnalysisPredictor::~AnalysisPredictor() {
+#if !defined(_WIN32)
+  if (FLAGS_profile) {
+    platform::DisableProfiler(platform::EventSortingKey::kTotal,
+                              "./profile.log");
+  }
+#endif
+  if (sub_scope_) {
+    scope_->DeleteScope(sub_scope_);
+  }
+}
+
 std::unique_ptr<PaddlePredictor> AnalysisPredictor::Clone() {
   auto *x = new AnalysisPredictor(config_);
   x->Init(scope_, inference_program_);

paddle/fluid/inference/api/analysis_predictor.h

@@ -72,6 +72,7 @@ class AnalysisPredictor : public PaddlePredictor {
   template <typename T>
   void GetFetchOne(const framework::LoDTensor &fetchs,
                    PaddleTensor *output_data);
+  ~AnalysisPredictor();
 
  private:
   contrib::AnalysisConfig config_;

paddle/fluid/operators/CMakeLists.txt

@@ -300,7 +300,7 @@ op_library(flatten_op DEPS reshape_op)
 op_library(sequence_pad_op DEPS sequence_padding)
 op_library(unstack_op DEPS stack_op)
 op_library(fake_quantize_op DEPS memory)
-op_library(fusion_lstm_op DEPS cpu_lstm_compute)
+op_library(fusion_lstm_op DEPS jit_kernel)
 if (WITH_GPU)
     op_library(conv_op DEPS vol2col depthwise_conv im2col)
     op_library(layer_norm_op DEPS cub)

paddle/fluid/operators/fill_constant_op.cc

@@ -70,6 +70,12 @@ class FillConstantOp : public framework::OperatorBase {
   }
 };
 
+class FillConstantOpVarTypeInference : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc &op_desc,
+                  framework::BlockDesc *block) const override {}
+};
+
 class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
@@ -102,4 +108,5 @@ Fill up a variable with specified constant value.
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(fill_constant, ops::FillConstantOp,
                   ops::FillConstantInferShape, ops::FillConstantOpMaker,
-                  paddle::framework::EmptyGradOpMaker);
+                  paddle::framework::EmptyGradOpMaker,
+                  ops::FillConstantOpVarTypeInference);

paddle/fluid/operators/fusion_lstm_op.cc

@@ -15,11 +15,9 @@ limitations under the License. */
 #include "paddle/fluid/operators/fusion_lstm_op.h"
 #include <string>
 #include "paddle/fluid/operators/math/blas.h"
-#include "paddle/fluid/operators/math/cpu_lstm_compute.h"
-#include "paddle/fluid/operators/math/cpu_vec.h"
 #include "paddle/fluid/operators/math/fc_compute.h"
+#include "paddle/fluid/operators/math/jit_kernel.h"
 #include "paddle/fluid/operators/math/sequence2batch.h"
-#include "paddle/fluid/platform/cpu_info.h"
 
 namespace paddle {
 namespace operators {
@@ -219,121 +217,55 @@ This operator fuse the X into LSTM, more details can refer to LSTM op.
 template <typename T>
 class FuisonLSTMKernel : public framework::OpKernel<T> {
  public:
-#define INIT_VEC_FUNC                                                          \
-  std::function<void(const int, const T *, T *)> act_gate, act_cell, act_cand; \
-  auto& act_gate_str = ctx.Attr<std::string>("gate_activation");               \
-  auto& act_cell_str = ctx.Attr<std::string>("cell_activation");               \
-  auto& act_cand_str = ctx.Attr<std::string>("candidate_activation");          \
-  if (platform::jit::MayIUse(platform::jit::avx)) {                            \
-    math::VecActivations<T, platform::jit::avx> act_functor;                   \
-    act_gate = act_functor(act_gate_str);                                      \
-    act_cell = act_functor(act_cell_str);                                      \
-    act_cand = act_functor(act_cand_str);                                      \
-  } else {                                                                     \
-    math::VecActivations<T, platform::jit::isa_any> act_functor;               \
-    act_gate = act_functor(act_gate_str);                                      \
-    act_cell = act_functor(act_cell_str);                                      \
-    act_cand = act_functor(act_cand_str);                                      \
-  }
-
-#define INIT_BASE_INPUT_OUTPUT                        \
-  auto* x = ctx.Input<LoDTensor>("X");                \
-  auto* h0 = ctx.Input<Tensor>("H0");                 \
-  auto* c0 = ctx.Input<Tensor>("C0");                 \
-  auto* wx = ctx.Input<Tensor>("WeightX");            \
-  auto* wh = ctx.Input<Tensor>("WeightH");            \
-  auto* bias = ctx.Input<Tensor>("Bias");             \
-  auto* xx = ctx.Output<LoDTensor>("XX");             \
-  auto* hidden_out = ctx.Output<LoDTensor>("Hidden"); \
-  auto* cell_out = ctx.Output<LoDTensor>("Cell");     \
-  bool is_reverse = ctx.Attr<bool>("is_reverse");     \
-  bool use_peepholes = ctx.Attr<bool>("use_peepholes");
-
-#define INIT_BASE_SIZES                  \
-  auto x_dims = x->dims();   /* T x M*/  \
-  auto wh_dims = wh->dims(); /* D x 4D*/ \
-  const int M = x_dims[1];               \
-  const int D = wh_dims[0];              \
-  const int D2 = D * 2;                  \
-  const int D3 = D * 3;                  \
-  const int D4 = wh_dims[1];
-
-#define INIT_BASE_INPUT_DATAS                                 \
-  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* wc_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); \
-  }
-
-/// Compute LSTM
+#define INIT_BASE_DEFINES                                   \
+  using DeviceContext = paddle::platform::CPUDeviceContext; \
+  auto* x = ctx.Input<LoDTensor>("X");                      \
+  auto* h0 = ctx.Input<Tensor>("H0");                       \
+  auto* c0 = ctx.Input<Tensor>("C0");                       \
+  auto* wx = ctx.Input<Tensor>("WeightX");                  \
+  auto* wh = ctx.Input<Tensor>("WeightH");                  \
+  auto* bias = ctx.Input<Tensor>("Bias");                   \
+  auto* xx = ctx.Output<LoDTensor>("XX");                   \
+  auto* hidden_out = ctx.Output<LoDTensor>("Hidden");       \
+  auto* cell_out = ctx.Output<LoDTensor>("Cell");           \
+  bool is_reverse = ctx.Attr<bool>("is_reverse");           \
+  bool use_peepholes = ctx.Attr<bool>("use_peepholes");     \
+  auto x_dims = x->dims();   /* T x M*/                     \
+  auto wh_dims = wh->dims(); /* D x 4D*/                    \
+  const int M = x_dims[1];                                  \
+  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 auto& ker =                                                         \
+      math::jitkernel::KernelPool::Instance()                               \
+          .template Get<math::jitkernel::LSTMKernel<T>, const std::string&, \
+                        const std::string&, const std::string&>(            \
+              ctx.Attr<std::string>("gate_activation"),                     \
+              ctx.Attr<std::string>("candidate_activation"),                \
+              ctx.Attr<std::string>("cell_activation"), D, use_peepholes)
+
+// Wh GEMM
 #define GEMM_WH_ADDON(bs, prev, out)                                           \
   blas.GEMM(CblasNoTrans, CblasNoTrans, bs, D4, D, static_cast<T>(1), prev, D, \
             wh_data, D4, static_cast<T>(1), out, D4)
 
-#define GET_Ct(ct_1, gates, ct)                   \
-  /* C_t = C_t-1 * fgated + cand_gated * igated*/ \
-  act_cand(D, gates, gates);                      \
-  blas.VMUL(D, gates, gates + D, gates + D);      \
-  blas.VMUL(D, ct_1, gates + D2, gates + D2);     \
-  blas.VADD(D, gates + D, gates + D2, ct)
-
-#define GET_Ht(ct, gates, ht)        \
-  /* H_t = act_cell(C_t) * ogated */ \
-  act_cell(D, ct, gates + D2);       \
-  blas.VMUL(D, gates + D2, gates + D3, ht)
-
-#define GET_Ct_NOH0C0(gates, ct)     \
-  /* C_t = igated * cgated*/         \
-  act_gate(D, gates + D, gates + D); \
-  act_cand(D, gates, gates);         \
-  blas.VMUL(D, gates, gates + D, ct)
-
-#define COMPUTE_CtHt_NOH0C0(gates, ct, ht) \
-  GET_Ct_NOH0C0(gates, ct);                \
-  act_gate(D, gates + D3, gates + D3);     \
-  GET_Ht(ct, gates, ht)
-
-#define COMPUTE_CtHt_PEEPHOLE_NOH0C0(gates, ct, ht) \
-  GET_Ct_NOH0C0(gates, ct);                         \
-  /* get outgated, put W_oc * C_t on igated */      \
-  blas.VMUL(D, wc_data + D2, ct, gates + D);        \
-  blas.VADD(D, gates + D, gates + D3, gates + D3);  \
-  act_gate(D, gates + D3, gates + D3);              \
-  GET_Ht(ct, gates, ht)
-
-#define COMPUTE_CtHt(gates, ct_1, ct, ht) \
-  act_gate(D3, gates + D, gates + D);     \
-  GET_Ct(ct_1, gates, ct);                \
-  GET_Ht(ct, gates, ht)
-
-#define COMPUTE_CtHt_PEEPHOLE(gates, ct_1, ct, ht)        \
-  /* get fgated and igated*/                              \
-  blas.VMUL(D, wc_data, ct_1, checked_cell_data);         \
-  blas.VMUL(D, wc_data + D, ct_1, checked_cell_data + D); \
-  blas.VADD(D2, checked_cell_data, gates + D, gates + D); \
-  act_gate(D2, gates + D, gates + D);                     \
-  GET_Ct(ct_1, gates, ct);                                \
-  /* get ogated*/                                         \
-  blas.VMUL(D, wc_data + D2, ct, gates + D);              \
-  blas.VADD(D, gates + D, gates + D3, gates + D3);        \
-  act_gate(D, gates + D3, gates + D3);                    \
-  GET_Ht(ct, gates, ht)
-
   void SeqCompute(const framework::ExecutionContext& ctx) const {
-    using DeviceContext = paddle::platform::CPUDeviceContext;
-    INIT_BASE_INPUT_OUTPUT
-    INIT_BASE_SIZES
-    INIT_VEC_FUNC
-    INIT_BASE_INPUT_DATAS
-
+    INIT_BASE_DEFINES;
+    INIT_OTHER_DEFINES;
     auto x_lod = x->lod();
     const int total_T = x_dims[0];
     const int N = x_lod[0].size() - 1;
@@ -357,89 +289,47 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
       gate_offset = -D;
     }
 
-#define MOVE_ONE_STEP                    \
-  prev_h_data = h_out_data;              \
-  prev_c_data = c_out_data;              \
-  xx_data = xx_data + xx_offset;         \
-  h_out_data = h_out_data + gate_offset; \
-  c_out_data = c_out_data + gate_offset
-
-#define PROCESS_H0C0_DEFINES                       \
-  int bid = is_reverse ? N - 1 - i : i;            \
-  int seq_len = x_lod[0][bid + 1] - x_lod[0][bid]; \
-  const T* prev_c_data = nullptr;                  \
-  const T* prev_h_data = nullptr;                  \
-  int tstart = 0
-
-#define PROCESS_H0C0_PEEPHOLE                                      \
-  PROCESS_H0C0_DEFINES;                                            \
-  if (h0_data) {                                                   \
-    prev_h_data = h0_data + bid * D;                               \
-    prev_c_data = c0_data + bid * D;                               \
-  } else {                                                         \
-    COMPUTE_CtHt_PEEPHOLE_NOH0C0(xx_data, c_out_data, h_out_data); \
-    MOVE_ONE_STEP;                                                 \
-    tstart = 1;                                                    \
-  }
-
-#define PROCESS_H0C0                                      \
-  PROCESS_H0C0_DEFINES;                                   \
-  if (h0_data) {                                          \
-    prev_h_data = h0_data + bid * D;                      \
-    prev_c_data = c0_data + bid * D;                      \
-  } else {                                                \
-    COMPUTE_CtHt_NOH0C0(xx_data, c_out_data, h_out_data); \
-    MOVE_ONE_STEP;                                        \
-    tstart = 1;                                           \
-  }
-
-    if (use_peepholes) {
-      for (int i = 0; i < N; ++i) {
-        PROCESS_H0C0_PEEPHOLE
-        for (int step = tstart; step < seq_len; ++step) {
-          GEMM_WH_ADDON(1, prev_h_data, xx_data);
-          COMPUTE_CtHt_PEEPHOLE(xx_data, prev_c_data, c_out_data, h_out_data);
-          MOVE_ONE_STEP;
-        }
-      }
-    } else {
-      // TODO(TJ): unly workaround, clean me
-      std::function<void(T*, const T*, T*, T*)> compute_ctht;
-      if (platform::jit::MayIUse(platform::jit::avx) &&
-          act_gate_str == "sigmoid" && act_cand_str == "tanh" &&
-          act_cell_str == "tanh" && D == 8) {
-        compute_ctht = math::lstm_compute_ctht<T>;
+    for (int i = 0; i < N; ++i) {
+      int bid = is_reverse ? N - 1 - i : i;
+      int seq_len = x_lod[0][bid + 1] - x_lod[0][bid];
+      const T* prev_c_data = nullptr;
+      const T* prev_h_data = nullptr;
+      int tstart = 0;
+      if (h0_data) {
+        prev_h_data = h0_data + bid * D;
+        prev_c_data = c0_data + bid * D;
       } else {
-        compute_ctht = [&](T* gates, const T* ct_1, T* ct, T* ht) {
-          COMPUTE_CtHt(gates, ct_1, ct, ht);
-        };
+        ker->ComputeC1H1(xx_data, c_out_data, h_out_data, wp_data);
+        tstart = 1;
+        // move one step
+        prev_h_data = h_out_data;
+        prev_c_data = c_out_data;
+        xx_data = xx_data + xx_offset;
+        h_out_data = h_out_data + gate_offset;
+        c_out_data = c_out_data + gate_offset;
       }
-      for (int i = 0; i < N; ++i) {
-        PROCESS_H0C0
-        for (int step = tstart; step < seq_len; ++step) {
-          GEMM_WH_ADDON(1, prev_h_data, xx_data);
-          compute_ctht(xx_data, prev_c_data, c_out_data, h_out_data);
-          MOVE_ONE_STEP;
-        }
+      for (int step = tstart; step < seq_len; ++step) {
+        GEMM_WH_ADDON(1, prev_h_data, xx_data);
+        ker->ComputeCtHt(xx_data, prev_c_data, c_out_data, h_out_data, wp_data,
+                         checked_cell_data);
+        // move one step
+        prev_h_data = h_out_data;
+        prev_c_data = c_out_data;
+        xx_data = xx_data + xx_offset;
+        h_out_data = h_out_data + gate_offset;
+        c_out_data = c_out_data + gate_offset;
       }
     }
-#undef PROCESS_H0C0_DEFINES
-#undef PROCESS_H0C0_PEEPHOLE
-#undef PROCESS_H0C0
-#undef MOVE_ONE_STEP
   }
 
   void BatchCompute(const framework::ExecutionContext& ctx) const {
-    using DeviceContext = platform::CPUDeviceContext;
-    INIT_BASE_INPUT_OUTPUT
-    INIT_BASE_SIZES
+    INIT_BASE_DEFINES;
     if (x->lod()[0].size() == 2) {
       xx->Resize({x_dims[0], D4});
       SeqCompute(ctx);
       return;
     }
-    INIT_VEC_FUNC
-    INIT_BASE_INPUT_DATAS
+    INIT_OTHER_DEFINES;
 
     auto* reordered_h0 = ctx.Output<Tensor>("ReorderedH0");
     auto* reordered_c0 = ctx.Output<Tensor>("ReorderedC0");
@@ -487,8 +377,8 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
       prev_c_data = reordered_c0_data;
       size_t sz = sizeof(T) * D;
       for (int i = 0; i < max_bs; ++i) {
-        std::memcpy(reordered_h0_data, h0_data + seq_order[i] * D, sz);
-        std::memcpy(reordered_c0_data, c0_data + seq_order[i] * D, sz);
+        blas.VCOPY(sz, h0_data + seq_order[i] * D, reordered_h0_data);
+        blas.VCOPY(sz, c0_data + seq_order[i] * D, reordered_c0_data);
         reordered_h0_data += D;
         reordered_c0_data += D;
       }
@@ -498,13 +388,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
       T* cur_h_out_data = batched_h_out_data;
       T* cur_c_out_data = batched_c_out_data;
       for (int i = 0; i < max_bs; ++i) {
-        GET_Ct_NOH0C0(cur_in_data, cur_c_out_data);
-        if (use_peepholes) {
-          blas.VMUL(D, wc_data + D2, cur_c_out_data, cur_in_data + D);
-          blas.VADD(D, cur_in_data + D, cur_in_data + D3, cur_in_data + D3);
-        }
-        act_gate(D, cur_in_data + D3, cur_in_data + D3);
-        GET_Ht(cur_c_out_data, cur_in_data, cur_h_out_data);
+        ker->ComputeC1H1(cur_in_data, cur_c_out_data, cur_h_out_data, wp_data);
         cur_in_data += D4;
         cur_c_out_data += D;
         cur_h_out_data += D;
@@ -513,71 +397,37 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
       prev_h_data = batched_h_out_data;
       prev_c_data = batched_c_out_data;
     }
+
+    // compute kernel part
     const auto& batch_starts = batched_lod[0];
     const int max_seq_len = batch_starts.size() - 1;
     const int offset = tstart * max_bs * D;
     batched_input_data = batched_input_data + offset * 4;
     batched_h_out_data = batched_h_out_data + offset;
     batched_c_out_data = batched_c_out_data + offset;
-
-#define DEFINE_CUR                        \
-  T* cur_in_data = batched_input_data;    \
-  T* cur_prev_c_data = prev_c_data;       \
-  T* cur_c_out_data = batched_c_out_data; \
-  T* cur_h_out_data = batched_h_out_data
-
-#define MOVE_ONE_BATCH  \
-  cur_in_data += D4;    \
-  cur_prev_c_data += D; \
-  cur_c_out_data += D;  \
-  cur_h_out_data += D
-
-#define MOVE_ONE_STEP                  \
-  prev_c_data = batched_c_out_data;    \
-  prev_h_data = batched_h_out_data;    \
-  batched_c_out_data = cur_c_out_data; \
-  batched_h_out_data = cur_h_out_data; \
-  batched_input_data = cur_in_data
-
-    if (use_peepholes) {
-      for (int step = tstart; step < max_seq_len; ++step) {
-        const int cur_bs = batch_starts[step + 1] - batch_starts[step];
-        GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data);
-        DEFINE_CUR;
-        for (int i = 0; i < cur_bs; ++i) {
-          COMPUTE_CtHt_PEEPHOLE(cur_in_data, cur_prev_c_data, cur_c_out_data,
-                                cur_h_out_data);
-          MOVE_ONE_BATCH;
-        }
-        MOVE_ONE_STEP;
-      }
-    } else {
-      // TODO(TJ): unly workaround, clean me
-      std::function<void(T*, const T*, T*, T*)> compute_ctht;
-      if (platform::jit::MayIUse(platform::jit::avx) &&
-          act_gate_str == "sigmoid" && act_cand_str == "tanh" &&
-          act_cell_str == "tanh" && D == 8) {
-        compute_ctht = math::lstm_compute_ctht<T>;
-      } else {
-        compute_ctht = [&](T* gates, const T* ct_1, T* ct, T* ht) {
-          COMPUTE_CtHt(gates, ct_1, ct, ht);
-        };
-      }
-      for (int step = tstart; step < max_seq_len; ++step) {
-        const int cur_bs = batch_starts[step + 1] - batch_starts[step];
-        GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data);
-        DEFINE_CUR;
-        for (int i = 0; i < cur_bs; ++i) {
-          compute_ctht(cur_in_data, cur_prev_c_data, cur_c_out_data,
-                       cur_h_out_data);
-          MOVE_ONE_BATCH;
-        }
-        MOVE_ONE_STEP;
+    for (int step = tstart; step < max_seq_len; ++step) {
+      const int cur_bs = batch_starts[step + 1] - batch_starts[step];
+      GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data);
+      T* cur_in_data = batched_input_data;
+      T* cur_prev_c_data = prev_c_data;
+      T* cur_c_out_data = batched_c_out_data;
+      T* cur_h_out_data = batched_h_out_data;
+      for (int i = 0; i < cur_bs; ++i) {
+        ker->ComputeCtHt(cur_in_data, cur_prev_c_data, cur_c_out_data,
+                         cur_h_out_data, wp_data, checked_cell_data);
+        // move one batch
+        cur_in_data += D4;
+        cur_prev_c_data += D;
+        cur_c_out_data += D;
+        cur_h_out_data += D;
       }
+      // move one step
+      prev_c_data = batched_c_out_data;
+      prev_h_data = batched_h_out_data;
+      batched_c_out_data = cur_c_out_data;
+      batched_h_out_data = cur_h_out_data;
+      batched_input_data = cur_in_data;
     }
-#undef MOVE_ONE_STEP
-#undef MOVE_ONE_BATCH
-#undef DEFINE_CUR
 
     math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
     batched_h_out->set_lod(batched_lod);
@@ -594,18 +444,9 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
     }
   }
 
-#undef COMPUTE_CtHt_PEEPHOLE
-#undef COMPUTE_CtHt
-#undef GET_Ct_NOH0C0
-#undef COMPUTE_CtHt_NOH0C0
-#undef COMPUTE_CtHt_PEEPHOLE_NOH0C0
-#undef GET_Ht
-#undef GET_Ct
 #undef GEMM_WH_ADDON
-#undef INIT_BASE_INPUT_DATAS
-#undef INIT_BASE_SIZES
-#undef INIT_BASE_INPUT_OUTPUT
-#undef INIT_VEC_FUNC
+#undef INIT_OTHER_DEFINES
+#undef INIT_BASE_DEFINES
 };
 
 }  // namespace operators

paddle/fluid/operators/isfinite_op.cc

@@ -60,7 +60,7 @@ class OverflowOpMaker : public framework::OpProtoAndCheckerMaker {
               "(Tensor) 1-dim tensor, contains a bool scalar. The output "
               "tensor of overflow operator.");
     AddComment(string::Sprintf(R"DOC(
-Overflow operator.
+Overflow %s operator.
 
 $$Out = any(X)$$
 
@@ -69,6 +69,8 @@ Out = Inf if any X contains Inf,
 Out = Nan if any X contains Nan,
 Out = 0 if no Inf/Nan detected.
 If X contains both Inf/Nan, it will return the first indicator it meeted.
+
+%s
 )DOC",
                                GetName(), GetComments()));
   }

paddle/fluid/operators/math/CMakeLists.txt

@@ -45,8 +45,6 @@ math_library(im2col)
 if (NOT WIN32) # windows do not support avx functions yet.
 math_library(gru_compute DEPS activation_functions math_function)
 math_library(lstm_compute DEPS activation_functions)
-# TODO(TJ): ugly workaround, clean me
-cc_library(cpu_lstm_compute SRCS cpu_lstm_compute.cc DEPS activation_functions cblas cpu_info)
 endif (NOT WIN32)
 
 cc_library(blas SRCS blas.cc DEPS cblas framework_proto device_context)
@@ -76,3 +74,7 @@ if(WITH_GPU)
 endif()
 cc_test(concat_test SRCS concat_test.cc DEPS concat)
 cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info)
+cc_library(jit_kernel 
+    SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_lstm.cc
+    DEPS cpu_info cblas activation_functions)
+cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel)

paddle/fluid/operators/math/cpu_lstm_compute.h

-/* 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 <string>
-#include "paddle/fluid/operators/math/cpu_vec.h"
-#include "paddle/fluid/platform/cpu_info.h"
-#ifdef __AVX__
-#include <immintrin.h>
-#endif
-
-namespace paddle {
-namespace operators {
-namespace math {
-
-// TODO(TJ): ugly workaround, clean me
-template <typename T>
-void lstm_compute_ctht(T* gates, const T* ct_1, T* ct, T* ht) {
-  // gates: W_ch, W_ih, W_fh, W_oh
-  vec_sigmoid<T, platform::jit::avx>(24, gates + 8, gates + 8);
-  vec_tanh<T, platform::jit::avx>(8, gates, gates);
-  const T *i = gates + 8, *f = gates + 16, *o = gates + 24;
-  const T min = SIGMOID_THRESHOLD_MIN;
-  const T max = SIGMOID_THRESHOLD_MAX;
-  for (int d = 0; d < 8; ++d) {
-    // C_t = C_t-1 * fgated + cand_gated * igated
-    ct[d] = ct_1[d] * f[d] + gates[d] * i[d];
-    // H_t = act_cell(C_t) * ogated
-    T tmp = ct[d] * 2;
-    tmp = static_cast<T>(0) - ((tmp < min) ? min : ((tmp > max) ? max : tmp));
-    vec_exp<T>(1, &tmp, &tmp);
-    tmp = static_cast<T>(2) / (static_cast<T>(1) + tmp) - static_cast<T>(1);
-    ht[d] = tmp * o[d];
-  }
-}
-
-#ifdef __AVX__
-namespace detail {
-namespace forward {
-namespace avx {
-__m256 Sigmoid(const __m256 a);
-__m256 Tanh(const __m256 a);
-
-}  // namespace avx
-}  // namespace forward
-}  // namespace detail
-
-template <>
-void lstm_compute_ctht<float>(float* gates, const float* ct_1, float* ct,
-                              float* ht);
-
-#endif
-
-}  // namespace math
-}  // namespace operators
-}  // namespace paddle

paddle/fluid/operators/math/cpu_vec.h

@@ -125,10 +125,8 @@ inline void vec_scal<float, platform::jit::avx2>(const int n, const float a,
 }
 
 template <>
-inline void vec_scal<float, platform::jit::avx512_common>(const int n,
-                                                          const float a,
-                                                          const float* x,
-                                                          float* y) {
+inline void vec_scal<float, platform::jit::avx512f>(const int n, const float a,
+                                                    const float* x, float* y) {
   // TODO(TJ): enable me
   vec_scal<float, platform::jit::avx2>(n, a, x, y);
 }
@@ -181,10 +179,10 @@ inline void vec_bias_sub<float, platform::jit::avx2>(const int n, const float a,
 }
 
 template <>
-inline void vec_bias_sub<float, platform::jit::avx512_common>(const int n,
-                                                              const float a,
-                                                              const float* x,
-                                                              float* y) {
+inline void vec_bias_sub<float, platform::jit::avx512f>(const int n,
+                                                        const float a,
+                                                        const float* x,
+                                                        float* y) {
   // TODO(TJ): enable me
   vec_bias_sub<float, platform::jit::avx2>(n, a, x, y);
 }
@@ -242,7 +240,7 @@ inline void vec_cross<float, platform::jit::avx2>(const int n, const float* x,
 }
 
 template <>
-inline void vec_cross<float, platform::jit::avx512_common>(
+inline void vec_cross<float, platform::jit::avx512f>(
     const int n, const float* x, const float* y, const float* z, float* out) {
   // TODO(TJ): enable me
   vec_cross<float, platform::jit::avx>(n, x, y, z, out);
@@ -296,10 +294,10 @@ inline void vec_add_bias<float, platform::jit::avx2>(const int n, const float a,
 }
 
 template <>
-inline void vec_add_bias<float, platform::jit::avx512_common>(const int n,
-                                                              const float a,
-                                                              const float* x,
-                                                              float* y) {
+inline void vec_add_bias<float, platform::jit::avx512f>(const int n,
+                                                        const float a,
+                                                        const float* x,
+                                                        float* y) {
   // TODO(TJ): enable me
   vec_add_bias<float, platform::jit::avx2>(n, a, x, y);
 }
@@ -390,9 +388,9 @@ inline void vec_sigmoid<float, platform::jit::avx2>(const int n, const float* x,
 }
 
 template <>
-inline void vec_sigmoid<float, platform::jit::avx512_common>(const int n,
-                                                             const float* x,
-                                                             float* y) {
+inline void vec_sigmoid<float, platform::jit::avx512f>(const int n,
+                                                       const float* x,
+                                                       float* y) {
   // TODO(TJ): enable me
   vec_sigmoid<float, platform::jit::avx2>(n, x, y);
 }
@@ -454,9 +452,8 @@ inline void vec_relu<float, platform::jit::avx2>(const int n, const float* x,
 }
 
 template <>
-inline void vec_relu<float, platform::jit::avx512_common>(const int n,
-                                                          const float* x,
-                                                          float* y) {
+inline void vec_relu<float, platform::jit::avx512f>(const int n, const float* x,
+                                                    float* y) {
   // TODO(TJ): enable me
   vec_relu<float, platform::jit::avx2>(n, x, y);
 }

paddle/fluid/operators/math/cpu_vec_test.cc

@@ -110,7 +110,7 @@ TEST(CpuVecTest, sigmoid) {
     TestAndBench<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
     TestAndBench<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
     TestAndBench<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
-    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx512_common>,
+    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx512f>,
                         ref_sigmoid<float>);
   }
   TestAndBench<double>(30, vec_sigmoid<double>, ref_sigmoid<double>);
@@ -123,8 +123,7 @@ TEST(CpuVecTest, tanh) {
     TestAndBench<float>(sz, vec_tanh<float>, ref_tanh<float>);
     TestAndBench<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
     TestAndBench<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
-    TestAndBench<float>(sz, vec_tanh<float, jit::avx512_common>,
-                        ref_tanh<float>);
+    TestAndBench<float>(sz, vec_tanh<float, jit::avx512f>, ref_tanh<float>);
   }
   TestAndBench<double>(30, vec_tanh<double>, ref_tanh<double>);
 }
@@ -136,8 +135,7 @@ TEST(CpuVecTest, relu) {
     TestAndBench<float>(sz, vec_relu<float>, ref_relu<float>);
     TestAndBench<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
     TestAndBench<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);
-    TestAndBench<float>(sz, vec_relu<float, jit::avx512_common>,
-                        ref_relu<float>);
+    TestAndBench<float>(sz, vec_relu<float, jit::avx512f>, ref_relu<float>);
   }
   TestAndBench<double>(30, vec_relu<double>, ref_relu<double>);
 }
@@ -170,7 +168,7 @@ TEST(CpuVecTest, inplace_sigmoid) {
     TestInplace<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
     TestInplace<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
     TestInplace<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
-    TestInplace<float>(sz, vec_sigmoid<float, jit::avx512_common>,
+    TestInplace<float>(sz, vec_sigmoid<float, jit::avx512f>,
                        ref_sigmoid<float>);
   }
   TestInplace<double>(30, vec_sigmoid<double>, ref_sigmoid<double>);
@@ -183,8 +181,7 @@ TEST(CpuVecTest, inplace_tanh) {
     TestInplace<float>(sz, vec_tanh<float>, ref_tanh<float>);
     TestInplace<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
     TestInplace<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
-    TestInplace<float>(sz, vec_tanh<float, jit::avx512_common>,
-                       ref_tanh<float>);
+    TestInplace<float>(sz, vec_tanh<float, jit::avx512f>, ref_tanh<float>);
   }
   TestInplace<double>(30, vec_tanh<double>, ref_tanh<double>);
 }
@@ -196,8 +193,7 @@ TEST(CpuVecTest, inplace_relu) {
     TestInplace<float>(sz, vec_relu<float>, ref_relu<float>);
     TestInplace<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
     TestInplace<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);
-    TestInplace<float>(sz, vec_relu<float, jit::avx512_common>,
-                       ref_relu<float>);
+    TestInplace<float>(sz, vec_relu<float, jit::avx512f>, ref_relu<float>);
   }
   TestInplace<double>(30, vec_relu<double>, ref_relu<double>);
 }

paddle/fluid/operators/math/cpu_lstm_compute.cc → paddle/fluid/operators/math/jit_kernel.cc

 /* 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/cpu_lstm_compute.h"
+#include "paddle/fluid/operators/math/jit_kernel.h"
+#include <iostream>
+#include <string>
 
 namespace paddle {
 namespace operators {
 namespace math {
-#ifdef __AVX__
-template <>
-void lstm_compute_ctht<float>(float* gates, const float* ct_1, float* ct,
-                              float* ht) {
-  namespace act = detail::forward::avx;
-  // gates: W_ch, W_ih, W_fh, W_oh
-  __m256 c, i, f, o;
-  c = _mm256_loadu_ps(gates);
-  i = _mm256_loadu_ps(gates + 8);
-  f = _mm256_loadu_ps(gates + 16);
-  o = _mm256_loadu_ps(gates + 24);
-
-  /* C_t = C_t-1 * fgated + cand_gated * igated*/
-  c = _mm256_mul_ps(act::Tanh(c), act::Sigmoid(i));
-  i = _mm256_loadu_ps(ct_1);
-  f = _mm256_mul_ps(i, act::Sigmoid(f));
-  f = _mm256_add_ps(c, f);
-  _mm256_storeu_ps(ct, f);
-
-  /* H_t = act_cell(C_t) * ogated */
-  o = _mm256_mul_ps(act::Tanh(f), act::Sigmoid(o));
-  _mm256_storeu_ps(ht, o);
+namespace jitkernel {
+
+namespace jit = platform::jit;
+
+KernelPool& KernelPool::Instance() {
+  static thread_local KernelPool g_jit_kernels;
+  return g_jit_kernels;
+}
+
+std::shared_ptr<const Kernel> KernelPool::Get(const std::string& key) const {
+  if (kers_.find(key) == kers_.end()) {
+    return nullptr;
+  }
+  return kers_.at(key);
 }
-#endif
+
+}  // namespace jitkernel
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/math/jit_kernel.h

+/* 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 <functional>
+#include <memory>  // for shared_ptr
+#include <string>
+#include <unordered_map>
+#include "paddle/fluid/platform/cpu_info.h"
+#include "paddle/fluid/platform/macros.h"
+
+// Note: Only support on CPU yet.
+namespace paddle {
+namespace operators {
+namespace math {
+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
+
+typedef enum { kLT8, kEQ8, kGT8LT16, kEQ16, kGT16 } jit_block;
+
+class Kernel {
+ public:
+  Kernel() = default;
+  virtual ~Kernel() = default;
+  int num_{0};
+  int end_{0};
+  int rest_{0};
+  DISABLE_COPY_AND_ASSIGN(Kernel);
+};
+
+class KernelPool {
+ public:
+  static KernelPool &Instance();
+
+  template <typename Ker, typename... ARGS>
+  std::shared_ptr<const Ker> Get(ARGS... args);
+
+  std::shared_ptr<const Kernel> Get(const std::string &key) const;
+
+ private:
+  KernelPool() = default;
+  std::unordered_map<std::string, std::shared_ptr<const Kernel>> kers_;
+
+  DISABLE_COPY_AND_ASSIGN(KernelPool);
+};
+
+template <typename T>
+class VMulKernel : public Kernel {
+ public:
+  virtual void Compute(const T *x, const T *y, T *z) const = 0;
+};
+
+template <typename T>
+class VAddKernel : public Kernel {
+ public:
+  virtual void Compute(const T *x, const T *y, T *z) const = 0;
+};
+
+template <typename T>
+class VScalKernel : public Kernel {
+ public:
+  virtual void Compute(const T a, const T *x, T *y) const = 0;
+  virtual void Compute(const T a, T *x) const = 0;
+};
+
+template <typename T>
+class VAddBiasKernel : public Kernel {
+ public:
+  virtual void Compute(const T a, const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class VActKernel : public Kernel {
+ public:
+  virtual void Compute(const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class VReluKernel : public VActKernel<T> {
+ public:
+  virtual void Compute(const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class VIdentityKernel : public VActKernel<T> {
+ public:
+  virtual void Compute(const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class VExpKernel : public VActKernel<T> {
+ public:
+  virtual void Compute(const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class VSigmoidKernel : public VActKernel<T> {
+ public:
+  virtual void Compute(const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class VTanhKernel : public VActKernel<T> {
+ public:
+  virtual void Compute(const T *x, T *y) const = 0;
+};
+
+template <typename T>
+class LSTMKernel : public Kernel {
+ public:
+  virtual void ComputeCtHt(T *gates, const T *ct_1, T *ct, T *ht,
+                           /* below only used in peephole*/
+                           const T *wp_data = nullptr,
+                           T *checked = nullptr) const = 0;
+
+  // compute c1 and h1 without c0 or h0
+  virtual void ComputeC1H1(T *gates, T *ct, T *ht,
+                           /* below only used in peephole*/
+                           const T *wp_data = nullptr) const = 0;
+};
+
+}  // namespace jitkernel
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_blas.cc

+/* 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_kernel.h"
+#include <string>
+#include "paddle/fluid/operators/math/jit_kernel_macro.h"
+#ifdef PADDLE_WITH_MKLML
+#include "paddle/fluid/platform/dynload/mklml.h"
+#endif
+
+#ifdef __AVX__
+#include <immintrin.h>
+#endif
+
+namespace paddle {
+namespace operators {
+namespace math {
+namespace jitkernel {
+
+namespace jit = platform::jit;
+
+/* VMUL JitKernel */
+template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+class VMulKernelImpl : public VMulKernel<T> {
+ public:
+  explicit VMulKernelImpl(int d) : VMulKernel<T>() { this->num_ = d; }
+  void Compute(const T* x, const T* y, T* z) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      z[i] = x[i] * y[i];
+    }
+  }
+};
+
+#ifdef PADDLE_WITH_MKLML
+#define MKL_FLOAT(isa, block)                           \
+  template <>                                           \
+  void VMulKernelImpl<float, isa, block>::Compute(      \
+      const float* x, const float* y, float* z) const { \
+    platform::dynload::vsMul(this->num_, x, y, z);      \
+  }
+
+#define MKL_DOUBLE(isa, block)                             \
+  template <>                                              \
+  void VMulKernelImpl<double, isa, block>::Compute(        \
+      const double* x, const double* y, double* z) const { \
+    platform::dynload::vdMul(this->num_, x, y, z);         \
+  }
+
+FOR_EACH_ISA(MKL_FLOAT, kGT16);
+FOR_EACH_ISA_BLOCK(MKL_DOUBLE);
+#endif
+
+#define INTRI8_FLOAT(isa)                               \
+  template <>                                           \
+  void VMulKernelImpl<float, isa, kEQ8>::Compute(       \
+      const float* x, const float* y, float* z) const { \
+    __m256 tmpx, tmpy;                                  \
+    tmpx = _mm256_loadu_ps(x);                          \
+    tmpy = _mm256_loadu_ps(y);                          \
+    tmpx = _mm256_mul_ps(tmpx, tmpy);                   \
+    _mm256_storeu_ps(z, tmpx);                          \
+  }
+
+// avx > for > mkl
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+#endif
+// TODO(TJ): eq16 test and complete avx512
+#undef INTRI8_FLOAT
+#undef MKL_FLOAT
+#undef MKL_DOUBLE
+
+/* VADD JitKernel */
+template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+class VAddKernelImpl : public VAddKernel<T> {
+ public:
+  explicit VAddKernelImpl(int d) : VAddKernel<T>() { this->num_ = d; }
+  void Compute(const T* x, const T* y, T* z) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      z[i] = x[i] + y[i];
+    }
+  }
+};
+
+#ifdef PADDLE_WITH_MKLML
+#define MKL_FLOAT(isa, block)                           \
+  template <>                                           \
+  void VAddKernelImpl<float, isa, block>::Compute(      \
+      const float* x, const float* y, float* z) const { \
+    platform::dynload::vsAdd(this->num_, x, y, z);      \
+  }
+
+#define MKL_DOUBLE(isa, block)                             \
+  template <>                                              \
+  void VAddKernelImpl<double, isa, block>::Compute(        \
+      const double* x, const double* y, double* z) const { \
+    platform::dynload::vdAdd(this->num_, x, y, z);         \
+  }
+
+FOR_EACH_ISA(MKL_FLOAT, kGT16);
+FOR_EACH_ISA_BLOCK(MKL_DOUBLE);
+#endif
+
+#define INTRI8_FLOAT(isa)                               \
+  template <>                                           \
+  void VAddKernelImpl<float, isa, kEQ8>::Compute(       \
+      const float* x, const float* y, float* z) const { \
+    __m256 tmpx, tmpy;                                  \
+    tmpx = _mm256_loadu_ps(x);                          \
+    tmpy = _mm256_loadu_ps(y);                          \
+    tmpx = _mm256_add_ps(tmpx, tmpy);                   \
+    _mm256_storeu_ps(z, tmpx);                          \
+  }
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+#endif
+// TODO(TJ): eq16 test and complete avx512
+
+#undef INTRI8_FLOAT
+#undef MKL_FLOAT
+#undef MKL_DOUBLE
+
+/* VSCAL JitKernel */
+template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+class VScalKernelImpl : public VScalKernel<T> {
+ public:
+  explicit VScalKernelImpl(int d) : VScalKernel<T>() { this->num_ = d; }
+  void Compute(const T a, const T* x, T* y) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      y[i] = a * x[i];
+    }
+  }
+  void Compute(const T a, T* x) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      x[i] = a * x[i];
+    }
+  }
+};
+
+#ifdef PADDLE_WITH_MKLML
+#define MKL_FLOAT(isa, block)                                               \
+  template <>                                                               \
+  void VScalKernelImpl<float, isa, block>::Compute(const float a, float* x) \
+      const {                                                               \
+    platform::dynload::cblas_sscal(this->num_, a, x, 1);                    \
+  }
+
+#define MKL_DOUBLE(isa, block)                                                 \
+  template <>                                                                  \
+  void VScalKernelImpl<double, isa, block>::Compute(const double a, double* x) \
+      const {                                                                  \
+    platform::dynload::cblas_dscal(this->num_, a, x, 1);                       \
+  }
+
+FOR_EACH_ISA(MKL_FLOAT, kGT16);
+FOR_EACH_ISA_BLOCK(MKL_DOUBLE);
+#endif
+
+#define INTRI8_FLOAT(isa)                              \
+  template <>                                          \
+  void VScalKernelImpl<float, isa, kEQ8>::Compute(     \
+      const float a, const float* x, float* y) const { \
+    __m256 tmp;                                        \
+    __m256 scalar = _mm256_set1_ps(a);                 \
+    tmp = _mm256_loadu_ps(x);                          \
+    tmp = _mm256_mul_ps(tmp, scalar);                  \
+    _mm256_storeu_ps(y, tmp);                          \
+  }
+#define INTRI8_INPLACE_FLOAT(isa)                                          \
+  template <>                                                              \
+  void VScalKernelImpl<float, isa, kEQ8>::Compute(const float a, float* x) \
+      const {                                                              \
+    __m256 tmp;                                                            \
+    __m256 scalar = _mm256_set1_ps(a);                                     \
+    tmp = _mm256_loadu_ps(x);                                              \
+    tmp = _mm256_mul_ps(tmp, scalar);                                      \
+    _mm256_storeu_ps(x, tmp);                                              \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI8_INPLACE_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI8_INPLACE_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+INTRI8_INPLACE_FLOAT(jit::avx512f);
+#endif
+// TODO(TJ): eq16 test and complete avx512
+
+#undef INTRI8_FLOAT
+#undef INTRI8_INPLACE_FLOAT
+#undef MKL_FLOAT
+#undef MKL_DOUBLE
+
+/* VAddBias JitKernel */
+template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+class VAddBiasKernelImpl : public VAddBiasKernel<T> {
+ public:
+  explicit VAddBiasKernelImpl(int d) : VAddBiasKernel<T>() { this->num_ = d; }
+  void Compute(const T a, const T* x, T* y) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      y[i] = x[i] + a;
+    }
+  }
+};
+
+#define INTRI8_FLOAT(isa)                              \
+  template <>                                          \
+  void VAddBiasKernelImpl<float, isa, kEQ8>::Compute(  \
+      const float a, const float* x, float* y) const { \
+    __m256 tmp = _mm256_loadu_ps(x);                   \
+    tmp = _mm256_add_ps(tmp, _mm256_set1_ps(a));       \
+    _mm256_storeu_ps(y, tmp);                          \
+  }
+
+#define INTRI16_FLOAT(isa)                             \
+  template <>                                          \
+  void VAddBiasKernelImpl<float, isa, kEQ16>::Compute( \
+      const float a, const float* x, float* y) const { \
+    __m256 tmp0 = _mm256_loadu_ps(x);                  \
+    __m256 tmp1 = _mm256_loadu_ps(x + 8);              \
+    tmp0 = _mm256_add_ps(tmp0, _mm256_set1_ps(a));     \
+    tmp1 = _mm256_add_ps(tmp1, _mm256_set1_ps(a));     \
+    _mm256_storeu_ps(y, tmp0);                         \
+    _mm256_storeu_ps(y + 8, tmp1);                     \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI16_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI16_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+INTRI16_FLOAT(jit::avx512f);
+#endif
+// TODO(TJ): eq16 test and complete avx512
+
+#undef INTRI8_FLOAT
+#undef INTRI16_FLOAT
+
+/* VRelu JitKernel */
+template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+class VReluKernelImpl : public VReluKernel<T> {
+ public:
+  explicit VReluKernelImpl(int d) : VReluKernel<T>() { this->num_ = d; }
+  void Compute(const T* x, T* y) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      y[i] = x[i] > 0 ? x[i] : 0;
+    }
+  }
+};
+
+#define INTRI8_FLOAT(isa)                                                   \
+  template <>                                                               \
+  void VReluKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
+      const {                                                               \
+    __m256 tmp = _mm256_loadu_ps(x);                                        \
+    tmp = _mm256_max_ps(tmp, _mm256_setzero_ps());                          \
+    _mm256_storeu_ps(y, tmp);                                               \
+  }
+
+#define INTRI16_FLOAT(isa)                                                   \
+  template <>                                                                \
+  void VReluKernelImpl<float, isa, kEQ16>::Compute(const float* x, float* y) \
+      const {                                                                \
+    __m256 zeros = _mm256_setzero_ps();                                      \
+    __m256 tmp0 = _mm256_loadu_ps(x);                                        \
+    __m256 tmp1 = _mm256_loadu_ps(x + 8);                                    \
+    tmp0 = _mm256_max_ps(tmp0, zeros);                                       \
+    tmp1 = _mm256_max_ps(tmp1, zeros);                                       \
+    _mm256_storeu_ps(y, tmp0);                                               \
+    _mm256_storeu_ps(y + 8, tmp1);                                           \
+  }
+
+#define INTRI_GT8LT16_FLOAT(isa)                                        \
+  template <>                                                           \
+  VReluKernelImpl<float, isa, kGT8LT16>::VReluKernelImpl(int d)         \
+      : VReluKernel<float>() {                                          \
+    this->num_ = d;                                                     \
+    this->end_ = AVX_FLOAT_BLOCK;                                       \
+    this->rest_ = d - AVX_FLOAT_BLOCK;                                  \
+  }                                                                     \
+  template <>                                                           \
+  void VReluKernelImpl<float, isa, kGT8LT16>::Compute(const float* x,   \
+                                                      float* y) const { \
+    __m256 zeros = _mm256_setzero_ps();                                 \
+    __m256 tmp0 = _mm256_loadu_ps(x);                                   \
+    __m256 tmp1 = _mm256_loadu_ps(x + this->rest_);                     \
+    tmp0 = _mm256_max_ps(tmp0, zeros);                                  \
+    tmp1 = _mm256_max_ps(tmp1, zeros);                                  \
+    _mm256_storeu_ps(y, tmp0);                                          \
+    _mm256_storeu_ps(y + this->rest_, tmp1);                            \
+  }
+
+#define INTRI_GT16_FLOAT(isa)                                                \
+  template <>                                                                \
+  VReluKernelImpl<float, isa, kGT16>::VReluKernelImpl(int d)                 \
+      : VReluKernel<float>() {                                               \
+    this->num_ = d;                                                          \
+    this->end_ = d - d % AVX_FLOAT_BLOCK;                                    \
+    this->rest_ = d - AVX_FLOAT_BLOCK;                                       \
+  }                                                                          \
+  template <>                                                                \
+  void VReluKernelImpl<float, isa, kGT16>::Compute(const float* x, float* y) \
+      const {                                                                \
+    __m256 zeros = _mm256_setzero_ps();                                      \
+    for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) {                  \
+      __m256 tmp = _mm256_loadu_ps(x + i);                                   \
+      tmp = _mm256_max_ps(tmp, zeros);                                       \
+      _mm256_storeu_ps(y + i, tmp);                                          \
+    }                                                                        \
+    __m256 tmp = _mm256_loadu_ps(x + this->rest_);                           \
+    tmp = _mm256_max_ps(tmp, zeros);                                         \
+    _mm256_storeu_ps(y + this->rest_, tmp);                                  \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI16_FLOAT(jit::avx);
+INTRI_GT8LT16_FLOAT(jit::avx);
+INTRI_GT16_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI16_FLOAT(jit::avx2);
+INTRI_GT8LT16_FLOAT(jit::avx2);
+INTRI_GT16_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+// TODO(TJ): refine avx512
+INTRI8_FLOAT(jit::avx512f);
+INTRI16_FLOAT(jit::avx512f);
+INTRI_GT8LT16_FLOAT(jit::avx512f);
+INTRI_GT16_FLOAT(jit::avx512f);
+#endif
+
+#undef INTRI8_FLOAT
+#undef INTRI16_FLOAT
+#undef INTRI_GT8LT16_FLOAT
+#undef INTRI_GT16_FLOAT
+
+/* 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 Compute(const T* x, T* y) const override {}
+};
+
+REGISTER_JITKERNEL(vmul, VMulKernel);
+REGISTER_JITKERNEL(vadd, VAddKernel);
+REGISTER_JITKERNEL(vscal, VScalKernel);
+REGISTER_JITKERNEL(vaddb, VAddBiasKernel);
+REGISTER_JITKERNEL(vrelu, VReluKernel);
+REGISTER_JITKERNEL(videntity, VIdentityKernel);
+
+}  // namespace jitkernel
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_exp.cc

+/* 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_kernel.h"
+#include <cmath>  // for exp
+#include <string>
+#include "paddle/fluid/operators/math/jit_kernel_macro.h"
+#ifdef PADDLE_WITH_MKLML
+#include "paddle/fluid/platform/dynload/mklml.h"
+#endif
+
+#ifdef __AVX__
+#include <immintrin.h>
+#endif
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+#ifdef __AVX__
+namespace detail {
+__m256 Exp(__m256 a);
+}  // namespace detail
+#endif
+
+namespace jitkernel {
+namespace jit = platform::jit;
+
+/* VExp JitKernel */
+template <typename T, jit::cpu_isa_t isa, jit_block>
+class VExpKernelImpl : public VExpKernel<T> {
+ public:
+  explicit VExpKernelImpl(int d) : VExpKernel<T>() { this->num_ = d; }
+  void Compute(const T* x, T* y) const override {
+    for (int i = 0; i < this->num_; ++i) {
+      y[i] = std::exp(x[i]);
+    }
+  }
+};
+
+#ifdef PADDLE_WITH_MKLML
+#define MKL_FLOAT(isa, block)                                               \
+  template <>                                                               \
+  void VExpKernelImpl<float, isa, block>::Compute(const float* x, float* y) \
+      const {                                                               \
+    platform::dynload::vsExp(this->num_, x, y);                             \
+  }
+
+#define MKL_DOUBLE(isa, block)                                                 \
+  template <>                                                                  \
+  void VExpKernelImpl<double, isa, block>::Compute(const double* x, double* y) \
+      const {                                                                  \
+    platform::dynload::vdExp(this->num_, x, y);                                \
+  }
+FOR_EACH_ISA(MKL_FLOAT, kLT8);
+FOR_EACH_ISA(MKL_FLOAT, kGT8LT16);
+FOR_EACH_ISA(MKL_FLOAT, kGT16);
+FOR_EACH_ISA_BLOCK(MKL_DOUBLE);
+#endif
+
+#define INTRI8_FLOAT(isa)                                                  \
+  template <>                                                              \
+  void VExpKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
+      const {                                                              \
+    __m256 tmp = _mm256_loadu_ps(x);                                       \
+    _mm256_storeu_ps(y, detail::Exp(tmp));                                 \
+  }
+
+#define INTRI16_FLOAT(isa)                                                  \
+  template <>                                                               \
+  void VExpKernelImpl<float, isa, kEQ16>::Compute(const float* x, float* y) \
+      const {                                                               \
+    __m256 tmp0 = _mm256_loadu_ps(x);                                       \
+    __m256 tmp1 = _mm256_loadu_ps(x + 8);                                   \
+    tmp0 = detail::Exp(tmp0);                                               \
+    tmp1 = detail::Exp(tmp1);                                               \
+    _mm256_storeu_ps(y, tmp0);                                              \
+    _mm256_storeu_ps(y + 8, tmp1);                                          \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI16_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI16_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+INTRI16_FLOAT(jit::avx512f);
+#endif
+// TODO(TJ): eq16 test and complete avx512
+
+#undef INTRI8_FLOAT
+#undef INTRI16_FLOAT
+#undef MKL_FLOAT
+#undef MKL_DOUBLE
+
+REGISTER_JITKERNEL(vexp, VExpKernel);
+
+/* VSigmoid JitKernel */
+template <typename T, jit::cpu_isa_t isa, jit_block>
+class VSigmoidKernelImpl : public VSigmoidKernel<T> {
+ public:
+  explicit VSigmoidKernelImpl(int d) : VSigmoidKernel<T>() {
+    this->num_ = d;
+    vexp_ = KernelPool::Instance().template Get<VExpKernel<T>>(d);
+  }
+  void Compute(const T* x, T* y) const override {
+    const T min = SIGMOID_THRESHOLD_MIN;
+    const T max = SIGMOID_THRESHOLD_MAX;
+    for (int i = 0; i < this->num_; ++i) {
+      y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
+      y[i] = static_cast<T>(0) - y[i];
+    }
+    vexp_->Compute(y, y);
+    for (int i = 0; i < this->num_; ++i) {
+      y[i] = static_cast<T>(1) / (static_cast<T>(1) + y[i]);
+    }
+  }
+
+ private:
+  std::shared_ptr<const VExpKernel<T>> vexp_;
+};
+
+#define INTRI_SIGMOID(tmp, min, max)              \
+  tmp = _mm256_max_ps(tmp, min);                  \
+  tmp = _mm256_min_ps(tmp, max);                  \
+  tmp = _mm256_sub_ps(_mm256_set1_ps(0.0f), tmp); \
+  tmp = detail::Exp(tmp);                         \
+  tmp = _mm256_add_ps(_mm256_set1_ps(1.0f), tmp); \
+  tmp = _mm256_div_ps(_mm256_set1_ps(1.0f), tmp)
+
+#define INTRI8_FLOAT(isa)                                                      \
+  template <>                                                                  \
+  void VSigmoidKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
+      const {                                                                  \
+    __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);                        \
+    __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);                        \
+    __m256 tmp = _mm256_loadu_ps(x);                                           \
+    INTRI_SIGMOID(tmp, min, max);                                              \
+    _mm256_storeu_ps(y, tmp);                                                  \
+  }
+
+#define INTRI16_FLOAT(isa)                                              \
+  template <>                                                           \
+  void VSigmoidKernelImpl<float, isa, kEQ16>::Compute(const float* x,   \
+                                                      float* y) const { \
+    __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);                 \
+    __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);                 \
+    __m256 tmp0 = _mm256_loadu_ps(x);                                   \
+    __m256 tmp1 = _mm256_loadu_ps(x + 8);                               \
+    INTRI_SIGMOID(tmp0, min, max);                                      \
+    INTRI_SIGMOID(tmp1, min, max);                                      \
+    _mm256_storeu_ps(y, tmp0);                                          \
+    _mm256_storeu_ps(y + 8, tmp1);                                      \
+  }
+
+#define INTRI_GT8LT16_FLOAT(isa)                                             \
+  template <>                                                                \
+  VSigmoidKernelImpl<float, isa, kGT8LT16>::VSigmoidKernelImpl(int d)        \
+      : VSigmoidKernel<float>() {                                            \
+    this->num_ = d;                                                          \
+    this->end_ = AVX_FLOAT_BLOCK;                                            \
+    this->rest_ = d - this->end_;                                            \
+    vexp_ =                                                                  \
+        KernelPool::Instance().template Get<VExpKernel<float>>(this->rest_); \
+  }                                                                          \
+  template <>                                                                \
+  void VSigmoidKernelImpl<float, isa, kGT8LT16>::Compute(const float* x,     \
+                                                         float* y) const {   \
+    __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);                      \
+    __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);                      \
+    __m256 tmp = _mm256_loadu_ps(x);                                         \
+    INTRI_SIGMOID(tmp, min, max);                                            \
+    _mm256_storeu_ps(y, tmp);                                                \
+    const float min_ = SIGMOID_THRESHOLD_MIN;                                \
+    const float max_ = SIGMOID_THRESHOLD_MAX;                                \
+    for (int i = this->end_; i < this->num_; ++i) {                          \
+      y[i] = (x[i] < min_) ? min_ : ((x[i] > max_) ? max_ : x[i]);           \
+      y[i] = 0.f - y[i];                                                     \
+    }                                                                        \
+    vexp_->Compute(y + this->end_, y + this->end_);                          \
+    for (int i = this->end_; i < this->num_; ++i) {                          \
+      y[i] = 1.f / (1.f + y[i]);                                             \
+    }                                                                        \
+  }
+
+#define INTRI_GT16_FLOAT(isa)                                                \
+  template <>                                                                \
+  VSigmoidKernelImpl<float, isa, kGT16>::VSigmoidKernelImpl(int d)           \
+      : VSigmoidKernel<float>() {                                            \
+    this->num_ = d;                                                          \
+    this->rest_ = d % AVX_FLOAT_BLOCK;                                       \
+    this->end_ = d - this->rest_;                                            \
+    vexp_ =                                                                  \
+        KernelPool::Instance().template Get<VExpKernel<float>>(this->rest_); \
+  }                                                                          \
+  template <>                                                                \
+  void VSigmoidKernelImpl<float, isa, kGT16>::Compute(const float* x,        \
+                                                      float* y) const {      \
+    __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);                      \
+    __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);                      \
+    for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) {                  \
+      __m256 tmp = _mm256_loadu_ps(x + i);                                   \
+      INTRI_SIGMOID(tmp, min, max);                                          \
+      _mm256_storeu_ps(y + i, tmp);                                          \
+    }                                                                        \
+    const float min_ = SIGMOID_THRESHOLD_MIN;                                \
+    const float max_ = SIGMOID_THRESHOLD_MAX;                                \
+    for (int i = this->end_; i < this->num_; ++i) {                          \
+      y[i] = (x[i] < min_) ? min_ : ((x[i] > max_) ? max_ : x[i]);           \
+      y[i] = 0.f - y[i];                                                     \
+    }                                                                        \
+    vexp_->Compute(y + this->end_, y + this->end_);                          \
+    for (int i = this->end_; i < this->num_; ++i) {                          \
+      y[i] = 1.f / (1.f + y[i]);                                             \
+    }                                                                        \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI16_FLOAT(jit::avx);
+INTRI_GT8LT16_FLOAT(jit::avx);
+INTRI_GT16_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI16_FLOAT(jit::avx2);
+// INTRI_GT8LT16_FLOAT(jit::avx2);
+// INTRI_GT16_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+INTRI16_FLOAT(jit::avx512f);
+// INTRI_GT8LT16_FLOAT(jit::avx512f);
+// INTRI_GT16_FLOAT(jit::avx512f);
+#endif
+
+#undef INTRI8_FLOAT
+#undef INTRI16_FLOAT
+#undef INTRI_GT8LT16_FLOAT
+#undef INTRI_GT16_FLOAT
+#undef INTRI_VSIGMOID
+
+REGISTER_JITKERNEL(vsigmoid, VSigmoidKernel);
+
+/* VTanh JitKernel */
+template <typename T, jit::cpu_isa_t isa, jit_block>
+class VTanhKernelImpl : public VTanhKernel<T> {
+ public:
+  explicit VTanhKernelImpl(int d) : VTanhKernel<T>() {
+    this->num_ = d;
+    vscal_ = KernelPool::Instance().template Get<VScalKernel<T>>(d);
+    vsigmoid_ = KernelPool::Instance().template Get<VSigmoidKernel<T>>(d);
+    vaddbias_ = KernelPool::Instance().template Get<VAddBiasKernel<T>>(d);
+  }
+  void Compute(const T* x, T* y) const override {
+    vscal_->Compute(static_cast<T>(2), x, y);
+    vsigmoid_->Compute(y, y);
+    vscal_->Compute(static_cast<T>(2), y);
+    vaddbias_->Compute(static_cast<T>(-1), y, y);
+  }
+
+ private:
+  std::shared_ptr<const VScalKernel<T>> vscal_;
+  std::shared_ptr<const VSigmoidKernel<T>> vsigmoid_;
+  std::shared_ptr<const VAddBiasKernel<T>> vaddbias_;
+};
+
+#define INTRI_VTANH(tmp)                                   \
+  tmp = _mm256_mul_ps(_mm256_set1_ps(-2.0f), tmp);         \
+  tmp = _mm256_min_ps(tmp, _mm256_set1_ps(EXP_MAX_INPUT)); \
+  tmp = detail::Exp(tmp);                                  \
+  tmp = _mm256_add_ps(_mm256_set1_ps(1.0f), tmp);          \
+  tmp = _mm256_div_ps(_mm256_set1_ps(2.0f), tmp);          \
+  tmp = _mm256_sub_ps(tmp, _mm256_set1_ps(1.0f))
+
+#define INTRI8_FLOAT(isa)                                                   \
+  template <>                                                               \
+  void VTanhKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
+      const {                                                               \
+    __m256 tmp = _mm256_loadu_ps(x);                                        \
+    INTRI_VTANH(tmp);                                                       \
+    _mm256_storeu_ps(y, tmp);                                               \
+  }
+
+#define INTRI16_FLOAT(isa)                                                   \
+  template <>                                                                \
+  void VTanhKernelImpl<float, isa, kEQ16>::Compute(const float* x, float* y) \
+      const {                                                                \
+    __m256 tmp0 = _mm256_loadu_ps(x);                                        \
+    __m256 tmp1 = _mm256_loadu_ps(x + 8);                                    \
+    INTRI_VTANH(tmp0);                                                       \
+    INTRI_VTANH(tmp1);                                                       \
+    _mm256_storeu_ps(y, tmp0);                                               \
+    _mm256_storeu_ps(y + 8, tmp1);                                           \
+  }
+
+#define INTRI_GT8LT16_FLOAT(isa)                                              \
+  template <>                                                                 \
+  VTanhKernelImpl<float, isa, kGT8LT16>::VTanhKernelImpl(int d)               \
+      : VTanhKernel<float>() {                                                \
+    this->num_ = d;                                                           \
+    this->end_ = AVX_FLOAT_BLOCK;                                             \
+    this->rest_ = d - this->end_;                                             \
+    vscal_ =                                                                  \
+        KernelPool::Instance().template Get<VScalKernel<float>>(this->rest_); \
+    vsigmoid_ = KernelPool::Instance().template Get<VSigmoidKernel<float>>(   \
+        this->rest_);                                                         \
+    vaddbias_ = KernelPool::Instance().template Get<VAddBiasKernel<float>>(   \
+        this->rest_);                                                         \
+  }                                                                           \
+  template <>                                                                 \
+  void VTanhKernelImpl<float, isa, kGT8LT16>::Compute(const float* x,         \
+                                                      float* y) const {       \
+    __m256 tmp = _mm256_loadu_ps(x);                                          \
+    INTRI_VTANH(tmp);                                                         \
+    _mm256_storeu_ps(y, tmp);                                                 \
+    x += AVX_FLOAT_BLOCK;                                                     \
+    y += AVX_FLOAT_BLOCK;                                                     \
+    vscal_->Compute(2.f, x, y);                                               \
+    vsigmoid_->Compute(y, y);                                                 \
+    vscal_->Compute(2.f, y);                                                  \
+    vaddbias_->Compute(-1.f, y, y);                                           \
+  }
+
+#define INTRI_GT16_FLOAT(isa)                                                 \
+  template <>                                                                 \
+  VTanhKernelImpl<float, isa, kGT16>::VTanhKernelImpl(int d)                  \
+      : VTanhKernel<float>() {                                                \
+    this->num_ = d;                                                           \
+    this->rest_ = d % AVX_FLOAT_BLOCK;                                        \
+    this->end_ = d - this->rest_;                                             \
+    vscal_ =                                                                  \
+        KernelPool::Instance().template Get<VScalKernel<float>>(this->rest_); \
+    vsigmoid_ = KernelPool::Instance().template Get<VSigmoidKernel<float>>(   \
+        this->rest_);                                                         \
+    vaddbias_ = KernelPool::Instance().template Get<VAddBiasKernel<float>>(   \
+        this->rest_);                                                         \
+  }                                                                           \
+  template <>                                                                 \
+  void VTanhKernelImpl<float, isa, kGT16>::Compute(const float* x, float* y)  \
+      const {                                                                 \
+    for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) {                   \
+      __m256 tmp = _mm256_loadu_ps(x + i);                                    \
+      INTRI_VTANH(tmp);                                                       \
+      _mm256_storeu_ps(y + i, tmp);                                           \
+    }                                                                         \
+    x += this->end_;                                                          \
+    y += this->end_;                                                          \
+    vscal_->Compute(2.f, x, y);                                               \
+    vsigmoid_->Compute(y, y);                                                 \
+    vscal_->Compute(2.f, y);                                                  \
+    vaddbias_->Compute(-1.f, y, y);                                           \
+  }
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+INTRI16_FLOAT(jit::avx);
+INTRI_GT8LT16_FLOAT(jit::avx);
+INTRI_GT16_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+INTRI16_FLOAT(jit::avx2);
+// maybe use avx at gt8lt16 and gt16
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+INTRI16_FLOAT(jit::avx512f);
+// maybe use avx at gt8lt16 and gt16
+#endif
+
+#undef INTRI8_FLOAT
+#undef INTRI16_FLOAT
+#undef INTRI_GT8LT16_FLOAT
+#undef INTRI_GT16_FLOAT
+#undef INTRI_VTANH
+
+REGISTER_JITKERNEL(vtanh, VTanhKernel);
+
+#undef JITKERNEL_NEW_ACT_IMPL
+
+}  // namespace jitkernel
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_lstm.cc

+/* 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_kernel.h"
+#include <string>
+#include "paddle/fluid/operators/math/jit_kernel_macro.h"
+#include "paddle/fluid/platform/enforce.h"
+#include "paddle/fluid/platform/macros.h"
+
+#ifdef __AVX__
+#include <immintrin.h>
+#endif
+
+namespace paddle {
+namespace operators {
+namespace math {
+#ifdef __AVX__
+namespace detail {
+__m256 Exp(__m256 a);
+}  // namespace detail
+#endif
+
+namespace jitkernel {
+namespace jit = platform::jit;
+
+#ifdef __AVX__
+typedef enum { kSigmoid, kRelu, kTanh, kIdentity } act_type;
+
+class AVXAct {
+ public:
+  virtual ~AVXAct() = default;
+  virtual __m256 Compute(__m256 x) const = 0;
+};
+
+template <act_type type>
+class AVXActImpl : public AVXAct {
+ public:
+  __m256 Compute(__m256 x) const override { PADDLE_THROW("Unkown type!"); }
+};
+
+template <>
+__m256 AVXActImpl<kSigmoid>::Compute(__m256 x) const {
+  __m256 ones = _mm256_set1_ps(1.0f);
+  x = _mm256_max_ps(x, _mm256_set1_ps(SIGMOID_THRESHOLD_MIN));
+  x = _mm256_min_ps(x, _mm256_set1_ps(SIGMOID_THRESHOLD_MAX));
+  x = _mm256_sub_ps(_mm256_set1_ps(0.0f), x);
+  x = detail::Exp(x);
+  x = _mm256_add_ps(ones, x);
+  return _mm256_div_ps(ones, x);
+}
+
+template <>
+__m256 AVXActImpl<kTanh>::Compute(__m256 x) const {
+  __m256 ones = _mm256_set1_ps(1.0f);
+  x = _mm256_mul_ps(_mm256_set1_ps(-2.0f), x);
+  x = _mm256_min_ps(x, _mm256_set1_ps(EXP_MAX_INPUT));
+  x = detail::Exp(x);
+  x = _mm256_add_ps(ones, x);
+  x = _mm256_div_ps(_mm256_set1_ps(2.0f), x);
+  return _mm256_sub_ps(x, ones);
+}
+
+template <>
+__m256 AVXActImpl<kRelu>::Compute(__m256 x) const {
+  return _mm256_max_ps(x, _mm256_setzero_ps());
+}
+
+template <>
+__m256 AVXActImpl<kIdentity>::Compute(__m256 x) const {
+  return x;
+}
+#endif
+
+template <typename T>
+static std::shared_ptr<const VActKernel<T>> GetActKernel(
+    const std::string& type, int n) {
+  if (type == "sigmoid") {
+    return std::dynamic_pointer_cast<const VActKernel<T>>(
+        KernelPool::Instance().template Get<VSigmoidKernel<T>>(n));
+  } else if (type == "relu") {
+    return std::dynamic_pointer_cast<const VActKernel<T>>(
+        KernelPool::Instance().template Get<VReluKernel<T>>(n));
+  } else if (type == "tanh") {
+    return std::dynamic_pointer_cast<const VActKernel<T>>(
+        KernelPool::Instance().template Get<VTanhKernel<T>>(n));
+  } else if (type == "identity" || type == "") {
+    return std::dynamic_pointer_cast<const VActKernel<T>>(
+        KernelPool::Instance().template Get<VIdentityKernel<T>>(n));
+  }
+  PADDLE_THROW("Not support type: %s", type);
+  return nullptr;
+}
+
+/* LSTM JitKernel */
+template <typename T, jit::cpu_isa_t isa, jit_block>
+class LSTMKernelImpl : public LSTMKernel<T> {
+ public:
+  explicit LSTMKernelImpl(const std::string& act_gate,
+                          const std::string& act_cand,
+                          const std::string& act_cell, int d)
+      : LSTMKernel<T>() {
+    d_ = d;
+    d2_ = d * 2;
+    d3_ = d * 3;
+    act_gate_d3_ = GetActKernel<T>(act_gate, d3_);
+    act_gate_d_ = GetActKernel<T>(act_gate, d);
+    act_cand_d_ = GetActKernel<T>(act_cand, d);
+    act_cell_d_ = GetActKernel<T>(act_cell, d);
+    vmul_d_ = KernelPool::Instance().template Get<VMulKernel<T>>(d);
+    vadd_d_ = KernelPool::Instance().template Get<VAddKernel<T>>(d);
+#ifdef __AVX__
+    auto GetAVXAct = [&](const std::string& type) -> std::unique_ptr<AVXAct> {
+      if (type == "sigmoid") {
+        return std::unique_ptr<AVXAct>(new AVXActImpl<kSigmoid>());
+      } else if (type == "relu") {
+        return std::unique_ptr<AVXAct>(new AVXActImpl<kRelu>());
+      } else if (type == "tanh") {
+        return std::unique_ptr<AVXAct>(new AVXActImpl<kTanh>());
+      } else if (type == "identity" || type == "") {
+        return std::unique_ptr<AVXAct>(new AVXActImpl<kIdentity>());
+      }
+      PADDLE_THROW("Not support type: %s", type);
+    };
+    avx_act_gate_ = GetAVXAct(act_gate);
+    avx_act_cand_ = GetAVXAct(act_cand);
+    avx_act_cell_ = GetAVXAct(act_cell);
+#endif
+  }
+
+  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_->Compute(gates + d_, gates + d_);
+
+    /* C_t = C_t-1 * fgated + cand_gated * igated */
+    act_cand_d_->Compute(gates, gates);
+    vmul_d_->Compute(gates, gates + d_, gates + d_);
+    vmul_d_->Compute(ct_1, gates + d2_, gates + d2_);
+    vadd_d_->Compute(gates + d_, gates + d2_, ct);
+
+    /* H_t = act_cell(C_t) * ogated */
+    act_cell_d_->Compute(ct, gates + d2_);
+    vmul_d_->Compute(gates + d2_, gates + d3_, ht);
+  }
+  void ComputeC1H1(T* gates, T* ct, T* ht, const T* wp_data) const override {
+    /* C_t = igated * cgated*/
+    act_gate_d_->Compute(gates + d_, gates + d_);
+    act_cand_d_->Compute(gates, gates);
+    vmul_d_->Compute(gates, gates + d_, ct);
+    /* H_t = act_cell(C_t) * ogated */
+    act_gate_d_->Compute(gates + d3_, gates + d3_);
+    act_cell_d_->Compute(ct, gates + d2_);
+    vmul_d_->Compute(gates + d2_, gates + d3_, ht);
+  }
+
+ private:
+  int d_, d2_, d3_;
+  std::shared_ptr<const VActKernel<T>> act_gate_d3_, act_gate_d_, act_cand_d_,
+      act_cell_d_;
+  std::shared_ptr<const VMulKernel<T>> vmul_d_;
+  std::shared_ptr<const VAddKernel<T>> vadd_d_;
+#ifdef __AVX__
+  std::unique_ptr<const AVXAct> avx_act_gate_, avx_act_cand_, avx_act_cell_;
+#endif
+};
+
+#define INTRI8_FLOAT(isa)                                                    \
+  template <>                                                                \
+  void LSTMKernelImpl<float, isa, kEQ8>::ComputeCtHt(                        \
+      float* gates, const float* ct_1, float* ct, float* ht,                 \
+      const float* wp_data, float* checked) const {                          \
+    /* gates: W_ch, W_ih, W_fh, W_oh */                                      \
+    __m256 c, i, f, o;                                                       \
+    c = _mm256_loadu_ps(gates);                                              \
+    i = _mm256_loadu_ps(gates + 8);                                          \
+    f = _mm256_loadu_ps(gates + 16);                                         \
+    o = _mm256_loadu_ps(gates + 24);                                         \
+    /* C_t = C_t-1 * fgated + cand_gated * igated*/                          \
+    c = _mm256_mul_ps(avx_act_cand_->Compute(c), avx_act_gate_->Compute(i)); \
+    i = _mm256_loadu_ps(ct_1);                                               \
+    f = _mm256_mul_ps(i, avx_act_gate_->Compute(f));                         \
+    f = _mm256_add_ps(c, f);                                                 \
+    _mm256_storeu_ps(ct, f);                                                 \
+    /* H_t = act_cell(C_t) * ogated */                                       \
+    o = _mm256_mul_ps(avx_act_cell_->Compute(f), avx_act_gate_->Compute(o)); \
+    _mm256_storeu_ps(ht, o);                                                 \
+  }
+
+// TODO(TJ): optimize keq16
+
+#ifdef __AVX__
+INTRI8_FLOAT(jit::avx);
+#endif
+#ifdef __AVX2__
+INTRI8_FLOAT(jit::avx2);
+#endif
+#ifdef __AVX512F__
+INTRI8_FLOAT(jit::avx512f);
+#endif
+
+/* Peephole JitKernel */
+template <typename T, jit::cpu_isa_t isa, jit_block>
+class PeepholeKernelImpl : public LSTMKernel<T> {
+ public:
+  explicit PeepholeKernelImpl(const std::string& act_gate,
+                              const std::string& act_cand,
+                              const std::string& act_cell, int d)
+      : LSTMKernel<T>() {
+    d_ = d;
+    d2_ = d * 2;
+    d3_ = d * 3;
+    act_gate_d_ = GetActKernel<T>(act_gate, d);
+    act_cand_d_ = GetActKernel<T>(act_cand, d);
+    act_cell_d_ = GetActKernel<T>(act_cell, d);
+    vmul_d_ = KernelPool::Instance().template Get<VMulKernel<T>>(d);
+    vadd_d_ = KernelPool::Instance().template Get<VAddKernel<T>>(d);
+    vadd_d2_ = KernelPool::Instance().template Get<VAddKernel<T>>(d2_);
+    act_gate_d2_ = GetActKernel<T>(act_gate, d2_);
+  }
+
+  void ComputeCtHt(T* gates, const T* ct_1, T* ct, T* ht, const T* wp_data,
+                   T* checked) const override {
+    /* get fgated and igated*/
+    vmul_d_->Compute(wp_data, ct_1, checked);
+    vmul_d_->Compute(wp_data + d_, ct_1, checked + d_);
+    vadd_d2_->Compute(checked, gates + d_, gates + d_);
+    act_gate_d2_->Compute(gates + d_, gates + d_);
+    /* C_t = C_t-1 * fgated + cand_gated * igated*/
+    act_cand_d_->Compute(gates, gates);
+    vmul_d_->Compute(gates, gates + d_, gates + d_);
+    vmul_d_->Compute(ct_1, gates + d2_, gates + d2_);
+    vadd_d_->Compute(gates + d_, gates + d2_, ct);
+    /* get ogated*/
+    vmul_d_->Compute(wp_data + d2_, ct, gates + d_);
+    vadd_d_->Compute(gates + d_, gates + d3_, gates + d3_);
+    act_gate_d_->Compute(gates + d3_, gates + d3_);
+    /* H_t = act_cell(C_t) * ogated */
+    act_cell_d_->Compute(ct, gates + d2_);
+    vmul_d_->Compute(gates + d2_, gates + d3_, ht);
+  }
+
+  void ComputeC1H1(T* gates, T* ct, T* ht, const T* wp_data) const override {
+    /* C_t = igated * cgated*/
+    act_gate_d_->Compute(gates + d_, gates + d_);
+    act_cand_d_->Compute(gates, gates);
+    vmul_d_->Compute(gates, gates + d_, ct);
+    /* get outgated, put W_oc * C_t on igated */
+    vmul_d_->Compute(wp_data + d2_, ct, gates + d_);
+    vadd_d_->Compute(gates + d_, gates + d3_, gates + d3_);
+    /* H_t = act_cell(C_t) * ogated */
+    act_gate_d_->Compute(gates + d3_, gates + d3_);
+    act_cell_d_->Compute(ct, gates + d2_);
+    vmul_d_->Compute(gates + d2_, gates + d3_, ht);
+  }
+
+ private:
+  int d_, d2_, d3_;
+  std::shared_ptr<const VActKernel<T>> act_gate_d2_, act_gate_d_, act_cand_d_,
+      act_cell_d_;
+  std::shared_ptr<const VMulKernel<T>> vmul_d_;
+  std::shared_ptr<const VAddKernel<T>> vadd_d_, vadd_d2_;
+};
+
+#define JITKERNEL_DECLARE_LSTM(ker_class, ker_dtype)                  \
+  template <>                                                         \
+  std::shared_ptr<const LSTMKernel<ker_dtype>>                        \
+  KernelPool::Get<LSTMKernel<ker_dtype>, const std::string&,          \
+                  const std::string&, const std::string&, int, bool>( \
+      const std::string& act_gate, const std::string& act_cand,       \
+      const std::string& act_cell, int d, bool use_peephole)
+
+#define JITKERNEL_KEY_LSTM(ker_key, dtype_key)                               \
+  #ker_key #dtype_key + std::to_string(d) + act_gate + act_cand + act_cell + \
+                                       (use_peephole ? "p" : "n")
+
+#define JITKERNEL_NEW_LSTM_IMPL(ker, dtype, isa, k)                    \
+  if (use_peephole) {                                                  \
+    p = std::dynamic_pointer_cast<ker<dtype>>(                         \
+        std::make_shared<PeepholeKernelImpl<dtype, isa, k>>(           \
+            act_gate, act_cand, act_cell, d));                         \
+  } else {                                                             \
+    p = std::dynamic_pointer_cast<ker<dtype>>(                         \
+        std::make_shared<ker##Impl<dtype, isa, k>>(act_gate, act_cand, \
+                                                   act_cell, d));      \
+  }
+
+REGISTER_JITKERNEL_ARGS(lstm, LSTMKernel, JITKERNEL_DECLARE_LSTM,
+                        JITKERNEL_KEY_LSTM, JITKERNEL_NEW_LSTM_IMPL);
+
+#undef INTRI8_FLOAT
+#undef JITKERNEL_DECLARE_LSTM
+#undef JITKERNEL_KEY_LSTM
+#undef JITKERNEL_NEW_LSTM_IMPL
+}  // namespace jitkernel
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_macro.h

+/* 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 <string>
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+namespace jitkernel {
+
+namespace jit = platform::jit;
+
+#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_ISA_BLOCK(macro_, ker, dtype)        \
+  if (jit::MayIUse(jit::avx512f)) {                 \
+    SEARCH_BLOCK(macro_, ker, dtype, jit::avx512f); \
+  } else if (jit::MayIUse(jit::avx2)) {             \
+    SEARCH_BLOCK(macro_, ker, dtype, jit::avx2);    \
+  } else if (jit::MayIUse(jit::avx)) {              \
+    SEARCH_BLOCK(macro_, ker, dtype, jit::avx);     \
+  } else {                                          \
+    SEARCH_BLOCK(macro_, ker, dtype, jit::isa_any); \
+  }
+
+#define JITKERNEL_DECLARE(ker_class, ker_dtype) \
+  template <>                                   \
+  std::shared_ptr<const ker_class<ker_dtype>>   \
+  KernelPool::Get<ker_class<ker_dtype>, int>(int d)
+
+#define JITKERNEL_KEY(ker_key, dtype_key) \
+  #ker_key #dtype_key + std::to_string(d)
+
+#define JITKERNEL_NEW_IMPL(ker, dtype, isa, k) \
+  p = std::dynamic_pointer_cast<ker<dtype>>(   \
+      std::make_shared<ker##Impl<dtype, isa, k>>(d))
+
+#define JITKERNEL_WITH_DTYPE(ker_key, ker_class, ker_dtype, dtype_key, \
+                             marco_declare, macro_key, macro_impl)     \
+  marco_declare(ker_class, ker_dtype) {                                \
+    std::string key = macro_key(ker_key, dtype_key);                   \
+    if (kers_.find(key) == kers_.end()) {                              \
+      std::shared_ptr<ker_class<ker_dtype>> p;                         \
+      SEARCH_ISA_BLOCK(macro_impl, ker_class, ker_dtype);              \
+      kers_.insert({key, std::dynamic_pointer_cast<Kernel>(p)});       \
+      return p;                                                        \
+    }                                                                  \
+    return std::dynamic_pointer_cast<const ker_class<ker_dtype>>(      \
+        kers_.at(key));                                                \
+  }
+
+#define REGISTER_JITKERNEL(ker_key, ker_class)                           \
+  JITKERNEL_WITH_DTYPE(ker_key, ker_class, float, f, JITKERNEL_DECLARE,  \
+                       JITKERNEL_KEY, JITKERNEL_NEW_IMPL);               \
+  JITKERNEL_WITH_DTYPE(ker_key, ker_class, double, d, JITKERNEL_DECLARE, \
+                       JITKERNEL_KEY, JITKERNEL_NEW_IMPL)
+
+#define REGISTER_JITKERNEL_ARGS(ker_key, ker_class, marco_declare, macro_key,  \
+                                macro_impl)                                    \
+  JITKERNEL_WITH_DTYPE(ker_key, ker_class, float, f, marco_declare, macro_key, \
+                       macro_impl);                                            \
+  JITKERNEL_WITH_DTYPE(ker_key, ker_class, double, d, marco_declare,           \
+                       macro_key, macro_impl)
+
+#define FOR_EACH_ISA(macro_, block) \
+  macro_(jit::avx512f, block);      \
+  macro_(jit::avx2, block);         \
+  macro_(jit::avx, block);          \
+  macro_(jit::isa_any, block)
+
+#define FOR_EACH_BLOCK(macro_, isa) \
+  macro_(isa, kLT8);                \
+  macro_(isa, kEQ8);                \
+  macro_(isa, kGT8LT16);            \
+  macro_(isa, kEQ16);               \
+  macro_(isa, kGT16)
+
+#define FOR_EACH_ISA_BLOCK(macro_)      \
+  FOR_EACH_BLOCK(macro_, jit::avx512f); \
+  FOR_EACH_BLOCK(macro_, jit::avx2);    \
+  FOR_EACH_BLOCK(macro_, jit::avx);     \
+  FOR_EACH_BLOCK(macro_, jit::isa_any)
+
+}  // namespace jitkernel
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/jit_kernel_test.cc

+/* 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_kernel.h"
+#include <sys/time.h>
+#include <cmath>    // for exp
+#include <cstring>  // for memcpy
+#include <string>
+#include <vector>
+#include "gflags/gflags.h"
+#include "glog/logging.h"
+#include "gtest/gtest.h"
+
+#ifdef PADDLE_WITH_MKLML
+#include "paddle/fluid/platform/dynload/mklml.h"
+#endif
+
+#ifdef __AVX__
+#include <immintrin.h>
+#endif
+
+constexpr int repeat = 20000;
+
+inline double GetCurrentUS() {
+  struct timeval time;
+  gettimeofday(&time, NULL);
+  return 1e+6 * time.tv_sec + time.tv_usec;
+}
+
+template <typename T>
+void RandomVec(const int n, T* a, const T lower = static_cast<T>(-20.f),
+               const T upper = static_cast<T>(20.f)) {
+  static unsigned int seed = 100;
+  std::mt19937 rng(seed++);
+  std::uniform_real_distribution<double> uniform_dist(0, 1);
+  for (int i = 0; i < n; ++i) {
+    a[i] = static_cast<T>(uniform_dist(rng) * (upper - lower) + lower);
+  }
+}
+
+void vrelu_ref(const int n, const float* x, float* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] > 0.f ? x[i] : 0.f;
+  }
+}
+
+#if defined __AVX__ || defined __AVX2__
+void vrelu_intri8(const int n, const float* x, float* y) {
+  __m256 tmp = _mm256_loadu_ps(x);
+  tmp = _mm256_max_ps(tmp, _mm256_setzero_ps());
+  _mm256_storeu_ps(y, tmp);
+}
+#endif
+
+TEST(JitKernel, vrelu) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
+    std::vector<float> x(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data(), -10.f, 1.f);
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VReluKernel<float>>(d);
+    const float* x_data = x.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vrelu_ref(d, x_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+#if defined __AVX__ || defined __AVX2__
+    if (d == 8) {
+      auto si0 = GetCurrentUS();
+      for (int i = 0; i < repeat; ++i) {
+        vrelu_intri8(d, x_data, zref_data);
+      }
+      auto si1 = GetCurrentUS();
+      VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat;
+    }
+#endif
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+void vaddbias_ref(const int n, const float a, const float* x, float* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] + a;
+  }
+}
+
+TEST(JitKernel, vaddbias) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 64, 100, 128, 256}) {
+    std::vector<float> x(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data(), -2.f, 2.f);
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VAddBiasKernel<float>>(d);
+    const float a = 2.f;
+    const float* x_data = x.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vaddbias_ref(d, a, x_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(a, x_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+void vexp_ref(const int n, const float* x, float* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = std::exp(x[i]);
+  }
+}
+
+#ifdef PADDLE_WITH_MKLML
+void vexp_mkl(const int n, const float* x, float* y) {
+  paddle::platform::dynload::vsExp(n, x, y);
+}
+#endif
+
+TEST(JitKernel, vexp) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 128, 256}) {
+    std::vector<float> x(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data(), -2.f, 2.f);
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VExpKernel<float>>(d);
+    const float* x_data = x.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vexp_ref(d, x_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+
+#ifdef PADDLE_WITH_MKLML
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vexp_mkl(d, x_data, zref_data);
+    }
+    auto tmkle = GetCurrentUS();
+#endif
+
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+#ifdef PADDLE_WITH_MKLML
+            << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, "
+#else
+            << " us, "
+#endif
+            << "tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+inline float _sigmoid(float x) {
+  const float min = SIGMOID_THRESHOLD_MIN;
+  const float max = SIGMOID_THRESHOLD_MAX;
+  float tmp = (x < min) ? min : ((x > max) ? max : x);
+  return 1.f / (1.f + std::exp(-tmp));
+}
+
+void vsigmoid_ref(const int n, const float* x, float* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = _sigmoid(x[i]);
+  }
+}
+
+void vsigmoid_better(
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VExpKernel<float>>& vexp,
+    const int n, const float* x, float* y) {
+  const float min = SIGMOID_THRESHOLD_MIN;
+  const float max = SIGMOID_THRESHOLD_MAX;
+  for (int i = 0; i < n; ++i) {
+    y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
+    y[i] = 0.f - y[i];
+  }
+  vexp->Compute(y, y);
+  for (int i = 0; i < n; ++i) {
+    y[i] = 1.f / (1.f + y[i]);
+  }
+}
+
+TEST(JitKernel, vsigmoid) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) {
+    std::vector<float> x(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data(), -2.f, 2.f);
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VSigmoidKernel<float>>(d);
+    const auto& vexp =
+        jit::KernelPool::Instance().template Get<jit::VExpKernel<float>>(d);
+    const float* x_data = x.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vsigmoid_better(vexp, d, x_data, zref_data);
+    }
+    auto tmkle = GetCurrentUS();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vsigmoid_ref(d, x_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, better(jit exp) takes: " << (tmkle - tmkls) / repeat
+            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+inline float _tanh(float x) { return 2.f * _sigmoid(2.f * x) - 1.f; }
+
+void vtanh_ref(const int n, const float* x, float* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = _tanh(x[i]);
+  }
+}
+
+void vtanh_better(
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VScalKernel<float>>& vscal,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VSigmoidKernel<float>>&
+        vsigmoid,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VAddBiasKernel<float>>&
+        vaddbias,
+    const int n, const float* x, float* y) {
+  vscal->Compute(2.f, x, y);
+  vsigmoid->Compute(y, y);
+  vscal->Compute(2.f, y);
+  vaddbias->Compute(-1.f, y, y);
+}
+
+TEST(JitKernel, vtanh) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) {
+    std::vector<float> x(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data(), -2.f, 2.f);
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VTanhKernel<float>>(d);
+    const auto& vscal =
+        jit::KernelPool::Instance().template Get<jit::VScalKernel<float>>(d);
+    const auto& vsigmoid =
+        jit::KernelPool::Instance().template Get<jit::VSigmoidKernel<float>>(d);
+    const auto& vaddbias =
+        jit::KernelPool::Instance().template Get<jit::VAddBiasKernel<float>>(d);
+    const float* x_data = x.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vtanh_better(vscal, vsigmoid, vaddbias, d, x_data, zref_data);
+    }
+    auto tmkle = GetCurrentUS();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vtanh_ref(d, x_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, better(jit exp) takes: " << (tmkle - tmkls) / repeat
+            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+void lstm_ctht_ref(
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VSigmoidKernel<float>>&
+        vsigmoid_3d,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VTanhKernel<float>>& vtanh_d,
+    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->Compute(gates + d, gates + d);
+  vtanh_d->Compute(gates, gates);
+  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;
+  for (int k = 0; k < d; ++k) {
+    // C_t = C_t-1 * fgated + cand_gated * igated
+    ct[k] = ct_1[k] * f[k] + gates[k] * i[k];
+    // H_t = act_cell(C_t) * ogated
+    float tmp = ct[k] * 2;
+    tmp = 0.f - ((tmp < min) ? min : ((tmp > max) ? max : tmp));
+    vexp_1->Compute(&tmp, &tmp);
+    tmp = 2.f / (1.f + tmp) - 1.f;
+    ht[k] = tmp * o[k];
+  }
+}
+
+void lstm_ctht_better(
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VSigmoidKernel<float>>&
+        vsigmoid_3d,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VTanhKernel<float>>& vtanh_d,
+    const std::shared_ptr<
+        const paddle::operators::math::jitkernel::VMulKernel<float>>& vmul_d,
+    const std::shared_ptr<
+        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->Compute(gates + d, gates + d);
+  vtanh_d->Compute(gates, gates);
+  vmul_d->Compute(gates, gates + d, gates + d);
+  vmul_d->Compute(ct_1, gates + d2, gates + d2);
+  vadd_d->Compute(gates + d, gates + d2, ct);
+  /* H_t = act_cell(C_t) * ogated */
+  vtanh_d->Compute(ct, gates + d2);
+  vmul_d->Compute(gates + d2, gates + d * 3, ht);
+}
+
+TEST(JitKernel, lstm) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 32, 64, 100}) {
+    int d4 = d * 4;
+    int d3 = d * 3;
+    std::vector<float> x(d4), xref(d4);
+    std::vector<float> ct_1(d), ct_tgt(d), ht_tgt(d);
+    std::vector<float> ct_ref(d), ht_ref(d);
+    RandomVec<float>(d4, x.data(), -2.f, 2.f);
+    RandomVec<float>(d, ct_1.data(), -2.f, 2.f);
+    memcpy(xref.data(), x.data(), sizeof(float) * d4);
+    std::string act_gate = "sigmoid", act_cand = "tanh", act_cell = "tanh";
+    const auto& ker =
+        jit::KernelPool::Instance()
+            .template Get<jit::LSTMKernel<float>, const std::string&,
+                          const std::string&, const std::string&>(
+                act_gate, act_cand, act_cell, d, false);
+    // below kernels are used to compute refer
+    const auto& vsigmoid_3d =
+        jit::KernelPool::Instance().template Get<jit::VSigmoidKernel<float>>(
+            d3);
+    const auto& vtanh_d =
+        jit::KernelPool::Instance().template Get<jit::VTanhKernel<float>>(d);
+    const auto& vexp_1 =
+        jit::KernelPool::Instance().template Get<jit::VExpKernel<float>>(1);
+    const auto& vmul_d =
+        jit::KernelPool::Instance().template Get<jit::VMulKernel<float>>(d);
+    const auto& vadd_d =
+        jit::KernelPool::Instance().template Get<jit::VAddKernel<float>>(d);
+
+    float* x_data = x.data();
+    float* xref_data = xref.data();
+    const float* ct_1_data = ct_1.data();
+    float* ct_tgt_data = ct_tgt.data();
+    float* ht_tgt_data = ht_tgt.data();
+    float* ct_ref_data = ct_ref.data();
+    float* ht_ref_data = ht_ref.data();
+    // compute once to check correctness
+    lstm_ctht_ref(vsigmoid_3d, vtanh_d, vexp_1, d, xref_data, ct_1_data,
+                  ct_ref_data, ht_ref_data);
+    ker->ComputeCtHt(x_data, ct_1_data, ct_tgt_data, ht_tgt_data);
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ct_tgt_data[i], ct_ref_data[i], 1e-3);
+      EXPECT_NEAR(ht_tgt_data[i], ht_ref_data[i], 1e-3);
+    }
+
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      lstm_ctht_better(vsigmoid_3d, vtanh_d, vmul_d, vadd_d, d, xref_data,
+                       ct_1_data, ct_ref_data, ht_ref_data);
+    }
+    auto tmkle = GetCurrentUS();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      lstm_ctht_ref(vsigmoid_3d, vtanh_d, vexp_1, d, xref_data, ct_1_data,
+                    ct_ref_data, ht_ref_data);
+    }
+    auto trefe = GetCurrentUS();
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->ComputeCtHt(x_data, ct_1_data, ct_tgt_data, ht_tgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, better(jit) takes: " << (tmkle - tmkls) / repeat
+            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
+  }
+}
+
+void vscal_ref(const int n, const float a, const float* x, float* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = a * x[i];
+  }
+}
+void vscal_inp_ref(const int n, const float a, float* x) {
+  for (int i = 0; i < n; ++i) {
+    x[i] = a * x[i];
+  }
+}
+#if defined __AVX__ || defined __AVX2__
+void vscal_intri8(const int n, const float a, const float* x, float* y) {
+  __m256 tmp;
+  __m256 scalar = _mm256_set1_ps(a);
+  tmp = _mm256_loadu_ps(x);
+  tmp = _mm256_mul_ps(tmp, scalar);
+  _mm256_storeu_ps(y, tmp);
+}
+void vscal_inp_intri8(const int n, const float a, float* x) {
+  __m256 tmp;
+  __m256 scalar = _mm256_set1_ps(a);
+  tmp = _mm256_loadu_ps(x);
+  tmp = _mm256_mul_ps(tmp, scalar);
+  _mm256_storeu_ps(x, tmp);
+}
+#endif
+
+#ifdef PADDLE_WITH_MKLML
+void vscal_inp_mkl(const int n, const float a, float* x) {
+  paddle::platform::dynload::cblas_sscal(n, a, x, 1);
+}
+#endif
+
+TEST(JitKernel, vscal) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
+    std::vector<float> x(d), y(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data());
+    std::memcpy(y.data(), x.data(), sizeof(float) * d);
+    float a = 2.f;
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VScalKernel<float>>(d);
+    const float* x_data = x.data();
+    float* y_data = y.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vscal_ref(d, a, x_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+    auto trefs1 = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vscal_inp_ref(d, a, y_data);
+    }
+    auto trefe1 = GetCurrentUS();
+
+#ifdef PADDLE_WITH_MKLML
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vscal_inp_mkl(d, a, y_data);
+    }
+    auto tmkle = GetCurrentUS();
+#endif
+
+#if defined __AVX__ || defined __AVX2__
+    if (d == 8) {
+      auto si0 = GetCurrentUS();
+      for (int i = 0; i < repeat; ++i) {
+        vscal_intri8(d, a, x_data, zref_data);
+      }
+      auto si1 = GetCurrentUS();
+      auto si2 = GetCurrentUS();
+      for (int i = 0; i < repeat; ++i) {
+        vscal_inp_intri8(d, a, y_data);
+      }
+      auto si3 = GetCurrentUS();
+      VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat
+              << " us, inplace: " << (si3 - si2) / repeat;
+    }
+#endif
+
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(a, x_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+    auto ttgts1 = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(a, y_data);
+    }
+    auto ttgte1 = GetCurrentUS();
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+            << " us, inplace takes: " << (trefe1 - trefs1) / repeat
+#ifdef PADDLE_WITH_MKLML
+            << " us, mkl inplace takes: " << (tmkle - tmkls) / repeat << " us, "
+#else
+            << " us, "
+#endif
+            << "tgt takes: " << (ttgte - ttgts) / repeat
+            << "us, tgt inplace takes: " << (ttgte1 - ttgts1) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+void vmul_ref(const int n, const float* x, const float* y, float* z) {
+  for (int i = 0; i < n; ++i) {
+    z[i] = x[i] * y[i];
+  }
+}
+
+#if defined __AVX__ || defined __AVX2__
+void vmul_intri8(const int n, const float* x, const float* y, float* z) {
+  __m256 tmpx, tmpy;
+  tmpx = _mm256_loadu_ps(x);
+  tmpy = _mm256_loadu_ps(y);
+  tmpx = _mm256_mul_ps(tmpx, tmpy);
+  _mm256_storeu_ps(z, tmpx);
+}
+#endif
+
+#ifdef PADDLE_WITH_MKLML
+void vmul_mkl(const int n, const float* x, const float* y, float* z) {
+  paddle::platform::dynload::vsMul(n, x, y, z);
+}
+#endif
+
+TEST(JitKernel, vmul) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
+    std::vector<float> x(d), y(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data());
+    RandomVec<float>(d, y.data());
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VMulKernel<float>>(d);
+    const float* x_data = x.data();
+    const float* y_data = y.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vmul_ref(d, x_data, y_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+
+#ifdef PADDLE_WITH_MKLML
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vmul_mkl(d, x_data, y_data, zref_data);
+    }
+    auto tmkle = GetCurrentUS();
+#endif
+
+#if defined __AVX__ || defined __AVX2__
+    if (d == 8) {
+      auto si0 = GetCurrentUS();
+      for (int i = 0; i < repeat; ++i) {
+        vmul_intri8(d, x_data, y_data, zref_data);
+      }
+      auto si1 = GetCurrentUS();
+      VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat;
+    }
+#endif
+
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, y_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+#ifdef PADDLE_WITH_MKLML
+            << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, "
+#else
+            << " us, "
+#endif
+            << "tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+void vadd_ref(const int n, const float* x, const float* y, float* z) {
+  for (int i = 0; i < n; ++i) {
+    z[i] = x[i] + y[i];
+  }
+}
+
+#if defined __AVX__ || defined __AVX2__
+void vadd_intri8(const int n, const float* x, const float* y, float* z) {
+  __m256 tmpx, tmpy;
+  tmpx = _mm256_loadu_ps(x);
+  tmpy = _mm256_loadu_ps(y);
+  tmpx = _mm256_add_ps(tmpx, tmpy);
+  _mm256_storeu_ps(z, tmpx);
+}
+#endif
+
+#ifdef PADDLE_WITH_MKLML
+void vadd_mkl(const int n, const float* x, const float* y, float* z) {
+  paddle::platform::dynload::vsAdd(n, x, y, z);
+}
+#endif
+
+TEST(JitKernel, vadd) {
+  namespace jit = paddle::operators::math::jitkernel;
+  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
+    std::vector<float> x(d), y(d);
+    std::vector<float> zref(d), ztgt(d);
+    RandomVec<float>(d, x.data());
+    RandomVec<float>(d, y.data());
+    const auto& ker =
+        jit::KernelPool::Instance().template Get<jit::VAddKernel<float>>(d);
+    const float* x_data = x.data();
+    const float* y_data = y.data();
+    float* ztgt_data = ztgt.data();
+    float* zref_data = zref.data();
+    auto trefs = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vadd_ref(d, x_data, y_data, zref_data);
+    }
+    auto trefe = GetCurrentUS();
+
+#ifdef PADDLE_WITH_MKLML
+    auto tmkls = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      vadd_mkl(d, x_data, y_data, zref_data);
+    }
+    auto tmkle = GetCurrentUS();
+#endif
+
+#if defined __AVX__ || defined __AVX2__
+    if (d == 8) {
+      auto si0 = GetCurrentUS();
+      for (int i = 0; i < repeat; ++i) {
+        vadd_intri8(d, x_data, y_data, zref_data);
+      }
+      auto si1 = GetCurrentUS();
+      VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat;
+    }
+#endif
+
+    auto ttgts = GetCurrentUS();
+    for (int i = 0; i < repeat; ++i) {
+      ker->Compute(x_data, y_data, ztgt_data);
+    }
+    auto ttgte = GetCurrentUS();
+
+    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
+#ifdef PADDLE_WITH_MKLML
+            << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, "
+#else
+            << " us, "
+#endif
+            << "tgt takes: " << (ttgte - ttgts) / repeat;
+    for (int i = 0; i < d; ++i) {
+      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
+    }
+  }
+}
+
+TEST(JitKernel, pool) {
+  namespace jit = paddle::operators::math::jitkernel;
+  const int frame_size = 4;
+  std::string act_gate = "sigmoid", act_cand = "tanh", act_cell = "tanh";
+  const auto& plstm1 =
+      jit::KernelPool::Instance()
+          .template Get<jit::LSTMKernel<float>, const std::string&,
+                        const std::string&, const std::string&>(
+              act_gate, act_cand, act_cell, frame_size, false);
+  const auto& plstm2 =
+      jit::KernelPool::Instance()
+          .template Get<jit::LSTMKernel<float>, const std::string&,
+                        const std::string&, const std::string&>(
+              act_gate, act_cand, act_cell, frame_size, false);
+  const auto& peephole =
+      jit::KernelPool::Instance()
+          .template Get<jit::LSTMKernel<float>, const std::string&,
+                        const std::string&, const std::string&>(
+              act_gate, act_cand, act_cell, frame_size, true);
+  EXPECT_TRUE(plstm1 != peephole);
+
+  const auto& pvmul_f =
+      jit::KernelPool::Instance().template Get<jit::VMulKernel<float>>(4);
+  EXPECT_TRUE(std::dynamic_pointer_cast<const jit::Kernel>(plstm2) !=
+              std::dynamic_pointer_cast<const jit::Kernel>(pvmul_f));
+
+  const auto& pvmul_d =
+      jit::KernelPool::Instance().template Get<jit::VMulKernel<double>>(4);
+  EXPECT_TRUE(std::dynamic_pointer_cast<const jit::Kernel>(pvmul_f) !=
+              std::dynamic_pointer_cast<const jit::Kernel>(pvmul_d));
+
+  const auto& pvmul_from_key = jit::KernelPool::Instance().Get("vmulf4");
+  EXPECT_EQ(pvmul_f, pvmul_from_key);
+  const auto& pvmul_from_key2 = jit::KernelPool::Instance().Get("vmulf5");
+  EXPECT_TRUE(pvmul_from_key2 == nullptr);
+}

paddle/fluid/operators/parallel_do_op.cc

@@ -397,6 +397,24 @@ class ParallelDoGradOpShapeInference : public framework::InferShapeBase {
   }
 };
 
+class ParallelDoGradOpVarTypeInference : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc &op_desc,
+                  framework::BlockDesc *block) const override {
+    framework::BlockDesc *sub_block =
+        boost::get<framework::BlockDesc *>(op_desc.GetAttr(kParallelBlock));
+    for (auto &out_vars : op_desc.Outputs()) {
+      for (auto &out_var : out_vars.second) {
+        auto &var = block->FindRecursiveOrCreateVar(out_var);
+        auto sub_var = sub_block->FindRecursiveOrCreateVar(out_var);
+        if (sub_var.GetType() != var.GetType()) {
+          var.SetType(sub_var.GetType());
+        }
+      }
+    }
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -404,4 +422,5 @@ REGISTER_OPERATOR(parallel_do, paddle::operators::ParallelDoOp,
                   paddle::operators::ParallelDoOpProtoMaker,
                   paddle::operators::ParallelDoGradOpDescMaker);
 REGISTER_OPERATOR(parallel_do_grad, paddle::operators::ParallelDoGradOp,
-                  paddle::operators::ParallelDoGradOpShapeInference);
+                  paddle::operators::ParallelDoGradOpShapeInference,
+                  paddle::operators::ParallelDoGradOpVarTypeInference);

paddle/fluid/operators/reshape_op.cc

@@ -164,7 +164,7 @@ dimension value will be copied from Input(X) at runtime. Note that the index of
 [2, 3, 4], Attr(shape) = [2, 3, 2, 0] is an invalid input.
 
 3. Input(Shape) has a higher priority than Attr(shape) if it is provided, while
-Attr(shape) still should be set correctly to gurantee shape inference in 
+Attr(shape) still should be set correctly to gurantee shape inference in
 compile-time.
 
 )DOC");
@@ -259,7 +259,6 @@ class Reshape2Op : public ReshapeOp {
       : ReshapeOp(type, inputs, outputs, attrs) {}
 
   void InferShape(framework::InferShapeContext *ctx) const override {
-    ReshapeOp::InferShape(ctx);
     PADDLE_ENFORCE(ctx->HasOutput("XShape"),
                    "Output(XShape) of ReshapeOp should not be null.");
     const auto &x_dims = ctx->GetInputDim("X");
@@ -270,6 +269,8 @@ class Reshape2Op : public ReshapeOp {
     }
     ctx->SetOutputDim("XShape", framework::make_ddim(xshape_dims));
     ctx->ShareLoD("X", /*->*/ "XShape");
+
+    ReshapeOp::InferShape(ctx);
   }
 };
 

paddle/fluid/operators/sequence_concat_op.cc

@@ -90,11 +90,13 @@ REGISTER_OPERATOR(sequence_concat, paddle::framework::OperatorWithKernel,
                   paddle::framework::DefaultGradOpDescMaker<false>);
 template <typename T>
 using Kernel = op::SeqConcatKernel<paddle::platform::CPUDeviceContext, T>;
-REGISTER_OP_CPU_KERNEL(sequence_concat, Kernel<float>, Kernel<double>);
+REGISTER_OP_CPU_KERNEL(sequence_concat, Kernel<float>, Kernel<double>,
+                       Kernel<int64_t>);
+
 REGISTER_OPERATOR(sequence_concat_grad, paddle::framework::OperatorWithKernel,
                   op::SeqConcatGradShapeInferer);
 template <typename T>
 using GradKernel =
     op::SeqConcatGradKernel<paddle::platform::CPUDeviceContext, T>;
 REGISTER_OP_CPU_KERNEL(sequence_concat_grad, GradKernel<float>,
-                       GradKernel<double>);
+                       GradKernel<double>, GradKernel<int64_t>);

paddle/fluid/operators/sequence_unpad_op.cc

+/* 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/sequence_unpad_op.h"
+
+namespace paddle {
+namespace operators {
+
+class SequenceUnpadOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of SequenceUnpadOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Length"),
+                   "Input(Length) of SequenceUnpadOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of SequenceUnpadOp should not be null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    PADDLE_ENFORCE_GE(x_dims.size(), 2,
+                      "The rank of Input(X) can't be less than 2.");
+
+    auto len_dims = ctx->GetInputDim("Length");
+    PADDLE_ENFORCE(len_dims.size() == 2 && len_dims[1] == 1,
+                   "The shape of Input(Length) should be [batch_size, 1].");
+    PADDLE_ENFORCE(
+        len_dims[0] == x_dims[0],
+        "Input(X) and Input(Length) should have the same first dimension.");
+
+    int64_t out_dim_0 = -1;
+    if (ctx->IsRuntime()) {
+      out_dim_0 = x_dims[0] * x_dims[1];
+    }
+
+    std::vector<int64_t> out_dims_vec{out_dim_0};
+    if (x_dims.size() == 2) {
+      out_dims_vec.push_back(1);
+    } else {
+      for (size_t i = 2; i < x_dims.size(); ++i) {
+        out_dims_vec.push_back(x_dims[i]);
+      }
+    }
+    ctx->SetOutputDim("Out", framework::make_ddim(out_dims_vec));
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("X"));
+    return framework::OpKernelType(data_type, ctx.device_context());
+  }
+};
+
+class SequenceUnpadOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(LoDTensor, default LoDTensor<float>) Input tensor which "
+             "contains the padded sequences with equal length.");
+    AddInput("Length",
+             "(LoDTensor) The input tensor which specifies the actual ength of "
+             "sequences after unpadding.");
+    AddOutput(
+        "Out",
+        "(LoDTensor) The output tensor which contains unpadded sequences.");
+    AddComment(R"DOC(
+      Sequence Unpad Operator
+
+      This operator removes the padding data in the input sequences and convert 
+      them into sequences with actual length as output, identitied by lod 
+      information.
+
+      Example:
+
+      Given input tensor Input(X):
+          X.data = [[ 1.0,  2.0,  3.0,  4.0,  5.0],
+                    [ 6.0,  7.0,  8.0,  9.0, 10.0],
+                    [11.0, 12.0, 13.0, 14.0, 15.0]], 
+`     
+      in which there are 3 sequences padded to length 5, and the acutal length 
+      specified by Input(Length):
+
+          Length.data = [[2], [3], [4]],
+
+      after unpadding, Output(Out) will be:
+
+          Out.data = [[1.0, 2.0, 6.0, 7.0, 8.0, 11.0, 12.0, 13.0, 14.0]]
+          Out.lod = [[0, 2, 5, 9]]      
+
+    )DOC");
+  }
+};
+
+class SequenceUnpadGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of SequenceUnpadGradOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasInput(framework::GradVarName("Out")),
+        "Input(Out@GRAD) of SequenceUnpadGradOp should not be null.");
+
+    if (ctx->HasOutput(framework::GradVarName("X"))) {
+      ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
+      ctx->ShareLoD("X", /*->*/ framework::GradVarName("X"));
+    }
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("X"));
+    return framework::OpKernelType(data_type, ctx.device_context());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(sequence_unpad, ops::SequenceUnpadOp,
+                  ops::SequenceUnpadOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+REGISTER_OPERATOR(sequence_unpad_grad, ops::SequenceUnpadGradOp);
+REGISTER_OP_CPU_KERNEL(
+    sequence_unpad,
+    ops::SequenceUnpadOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SequenceUnpadOpKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::SequenceUnpadOpKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::SequenceUnpadOpKernel<paddle::platform::CPUDeviceContext, int64_t>);
+REGISTER_OP_CPU_KERNEL(
+    sequence_unpad_grad,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CPUDeviceContext,
+                                   int64_t>);

paddle/fluid/operators/sequence_unpad_op.cu

+/* 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/sequence_unpad_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    sequence_unpad,
+    ops::SequenceUnpadOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SequenceUnpadOpKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::SequenceUnpadOpKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::SequenceUnpadOpKernel<paddle::platform::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    sequence_unpad_grad,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::SequenceUnpadGradOpKernel<paddle::platform::CUDADeviceContext,
+                                   int64_t>);

paddle/fluid/operators/sequence_unpad_op.h

+/* 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 <vector>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/memory/memcpy.h"
+#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/operators/math/sequence_padding.h"
+
+namespace paddle {
+namespace operators {
+
+using LoDTensor = framework::LoDTensor;
+using LoD = framework::LoD;
+
+template <typename DeviceContext, typename T>
+class SequenceUnpadOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x_t = ctx.Input<LoDTensor>("X");
+    auto* len_t = ctx.Input<LoDTensor>("Length");
+    auto* out_t = ctx.Output<LoDTensor>("Out");
+    out_t->mutable_data<T>(ctx.GetPlace());
+
+    const int64_t* seq_len_ptr = nullptr;
+    if (platform::is_gpu_place(ctx.GetPlace())) {
+      LoDTensor seq_len_cpu;
+      seq_len_cpu.Resize(len_t->dims());
+      seq_len_ptr = seq_len_cpu.mutable_data<int64_t>(platform::CPUPlace());
+      framework::TensorCopy(*len_t, platform::CPUPlace(),
+                            ctx.template device_context<DeviceContext>(),
+                            &seq_len_cpu);
+    } else {
+      seq_len_ptr = len_t->data<int64_t>();
+    }
+
+    size_t batch_size = x_t->dims()[0];
+    std::vector<size_t> out_lod0(batch_size + 1, 0);
+    for (size_t i = 0; i < batch_size; ++i) {
+      out_lod0[i + 1] = out_lod0[i] + seq_len_ptr[i];
+    }
+
+    framework::LoD out_lod;
+    out_lod.push_back(out_lod0);
+    out_t->set_lod(out_lod);
+
+    std::vector<int64_t> out_dims_vec{static_cast<int64_t>(out_lod0.back())};
+    if (x_t->dims().size() == 2) {
+      out_dims_vec.push_back(1);
+    } else {
+      for (size_t i = 2; i < x_t->dims().size(); ++i) {
+        out_dims_vec.push_back(x_t->dims()[i]);
+      }
+    }
+    out_t->Resize(framework::make_ddim(out_dims_vec));
+
+    int64_t padded_length = x_t->dims()[1];
+    math::UnpaddingLoDTensorFunctor<DeviceContext, T>()(
+        ctx.template device_context<DeviceContext>(), *x_t, out_t,
+        padded_length, 0, false, math::kBatchLengthWidth);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class SequenceUnpadGradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* d_x = ctx.Output<LoDTensor>(framework::GradVarName("X"));
+    if (d_x) {
+      const auto* d_out = ctx.Input<LoDTensor>(framework::GradVarName("Out"));
+      const auto* x_t = ctx.Input<LoDTensor>("X");
+      d_x->mutable_data<T>(ctx.GetPlace());
+
+      int padded_length = x_t->dims()[1];
+
+      LoDTensor zero_pads;
+      zero_pads.Resize({1, 1});
+      zero_pads.mutable_data<T>(ctx.GetPlace());
+      math::SetConstant<DeviceContext, T> set_zero;
+      auto& dev_ctx = ctx.template device_context<DeviceContext>();
+      set_zero(dev_ctx, &zero_pads, static_cast<T>(0));
+
+      math::PaddingLoDTensorFunctor<DeviceContext, T>()(
+          ctx.template device_context<DeviceContext>(), *d_out, d_x, zero_pads,
+          padded_length, 0, false, math::kBatchLengthWidth);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/platform/cpu_info.cc

@@ -128,7 +128,7 @@ bool MayIUse(const cpu_isa_t cpu_isa) {
       return cpu.has(Cpu::tAVX);
     case avx2:
       return cpu.has(Cpu::tAVX2);
-    case avx512_common:
+    case avx512f:
       return cpu.has(Cpu::tAVX512F);
     case avx512_core:
       return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512BW) &&

paddle/fluid/platform/cpu_info.h

@@ -43,7 +43,7 @@ typedef enum {
   sse42,
   avx,
   avx2,
-  avx512_common,
+  avx512f,
   avx512_core,
   avx512_core_vnni,
   avx512_mic,

paddle/fluid/platform/init.cc

@@ -116,7 +116,7 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
   platform::SetNumThreads(FLAGS_paddle_num_threads);
 #endif
 
-  if (platform::jit::MayIUse(platform::jit::avx512_common)) {
+  if (platform::jit::MayIUse(platform::jit::avx512f)) {
 #ifndef __AVX512F__
     LOG(WARNING) << "AVX512F is available, Please re-compile on local machine";
 #endif

paddle/fluid/platform/profiler.cc

@@ -370,8 +370,8 @@ void ParseEvents(const std::vector<std::vector<Event>>& events,
   std::vector<std::vector<Event>> merged_events_list;
   if (merge_thread) {
     std::vector<Event> merged_events;
-    for (int i = 0; i < events.size(); ++i) {
-      for (int j = 0; j < events[i].size(); ++j) {
+    for (size_t i = 0; i < events.size(); ++i) {
+      for (size_t j = 0; j < events[i].size(); ++j) {
         merged_events.push_back(events[i][j]);
       }
     }

python/paddle/fluid/layers/nn.py

@@ -56,6 +56,7 @@ __all__ = [
     'sequence_expand',
     'sequence_expand_as',
     'sequence_pad',
+    'sequence_unpad',
     'lstm_unit',
     'reduce_sum',
     'reduce_mean',
@@ -64,6 +65,7 @@ __all__ = [
     'reduce_prod',
     'sequence_first_step',
     'sequence_last_step',
+    'sequence_slice',
     'dropout',
     'split',
     'ctc_greedy_decoder',
@@ -1902,6 +1904,76 @@ def sequence_last_step(input):
     return sequence_pool(input=input, pool_type="last")
 
 
+def sequence_slice(input, offset, length, name=None):
+    """
+    **Sequence Slice Layer**
+
+    The layer crops a subsequence from given sequence with given start 
+    offset and subsequence length.
+
+    It only supports sequence data (LoDTensor with lod_level equal to 1).
+
+    .. code-block:: text
+    
+	- Case:
+
+            Given the input Variable **input**:
+                
+                input.data = [[a1, a2], [b1, b2], [c1, c2], [d1, d2], [e1, e2]],
+                input.lod = [[3, 2]],
+                input.dims = (5, 2),
+
+            with offset.data = [[0], [1]] and length.data = [[2], [1]],
+
+            the output Variable will be
+                
+                out.data = [[a1, a2], [b1, b2], [e1, e2]],
+                out.lod = [[2, 1]],
+                out.dims = (3, 2).
+	
+    NOTE: The first dimension size of **input**, **offset** and **length** 
+          should be equal. The **offset** should start from 0.
+    
+    Args:
+        input(Variable): The input Variable which consists of the complete 
+                         sequences.
+        offset(Variable): The offset to slice each sequence.
+        length(Variable): The length of each subsequence.
+        name(str|None): A name for this layer(optional). If set None, the
+                        layer will be named automatically.
+
+    Returns:
+        Variable: The output subsequences.
+
+    Examples:
+
+        .. code-block:: python
+
+             import numpy as np
+             seqs = fluid.layers.data(name='x', shape=[10, 5],
+                              dtype='float32', lod_level=1)
+             offset = fluid.layers.assign(input=np.array([[0, 1]]).astype("int32"))
+             length = fluid.layers.assign(input=np.array([[2, 1]]).astype("int32"))
+             subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset, 
+                                                   length=length)
+    """
+    helper = LayerHelper("sequence_slice", **locals())
+    dtype = helper.input_dtype()
+    out = helper.create_tmp_variable(dtype)
+
+    offset.stop_gradient = True
+    length.stop_gradient = True
+
+    helper.append_op(
+        type="sequence_slice",
+        inputs={"X": input,
+                "Offset": offset,
+                "Length": length},
+        outputs={"Out": out})
+
+    return out
+
+
 @templatedoc()
 def pool2d(input,
            pool_size=-1,
@@ -2793,7 +2865,7 @@ def sequence_expand_as(x, y, name=None):
 
 
 @templatedoc()
-def sequence_pad(x, pad_value, maxlen=None):
+def sequence_pad(x, pad_value, maxlen=None, name=None):
     """
     ${comment}
 
@@ -2807,7 +2879,9 @@ def sequence_pad(x, pad_value, maxlen=None):
             None or any positive int. When it is None, all sequences will be
             padded up to the length of the longest one among them; when it a
             certain positive value, it must be greater than the length of the
-            longest original sequence."
+            longest original sequence.
+        name(str|None): A name for this layer(optional). If set None, the layer
+            will be named automatically.
 
     Returns:
         Variable: The padded sequence batch and the original lengths before
@@ -2844,6 +2918,66 @@ def sequence_pad(x, pad_value, maxlen=None):
     return out, length
 
 
+def sequence_unpad(x, length, name=None):
+    """
+    **Sequence Unpad Layer**
+
+    This layer removes the padding data in the input sequences and convert 
+    them into sequences with actual length as output, identitied by lod 
+    information.
+
+    .. code-block:: text
+
+	Example:
+
+	Given input Variable **x**:
+	    x.data = [[ 1.0,  2.0,  3.0,  4.0,  5.0],
+		      [ 6.0,  7.0,  8.0,  9.0, 10.0],
+		      [11.0, 12.0, 13.0, 14.0, 15.0]], 
+     
+	in which there are 3 sequences padded to length 5, and the acutal length 
+	specified by input Variable **length**:
+
+	    length.data = [[2], [3], [4]],
+
+	after unpadding, the output Variable will be:
+
+	    out.data = [[1.0, 2.0, 6.0, 7.0, 8.0, 11.0, 12.0, 13.0, 14.0]]
+	    out.lod = [[2, 3, 4]]      
+
+    Args:
+        x(Variable): Input Variable which contains the padded sequences with
+            equal length.
+        length(Variable): The Variable that specifies the actual ength of
+            sequences after unpadding.
+        name(str|None): A name for this layer(optional). If set None, the layer
+            will be named automatically.
+
+    Returns:
+        Variable: The Variable contains the unpadded sequences.
+
+    Examples:
+        .. code-block:: python
+
+            x = fluid.layers.data(name='x', shape=[10, 5], dtype='float32')
+            len = fluid.layers.data(name='length', shape=[1], dtype='int64')
+            out = fluid.layers.sequence_unpad(x=x, length=len)
+    """
+
+    helper = LayerHelper('sequence_unpad', input=x, **locals())
+    dtype = helper.input_dtype()
+    out = helper.create_tmp_variable(dtype)
+
+    length.stop_gradient = True
+
+    helper.append_op(
+        type='sequence_unpad',
+        inputs={'X': x,
+                'Length': length},
+        outputs={'Out': out})
+    return out
+
+
 def beam_search(pre_ids,
                 pre_scores,
                 ids,

python/paddle/fluid/tests/unittests/dist_simnet_bow.py

@@ -81,7 +81,10 @@ def get_optimizer():
     return optimizer
 
 
-def train_network(batch_size, is_distributed=False, is_sparse=False):
+def train_network(batch_size,
+                  is_distributed=False,
+                  is_sparse=False,
+                  is_self_contained_lr=False):
     # query
     q = fluid.layers.data(
         name="query_ids", shape=[1], dtype="int64", lod_level=1)
@@ -93,7 +96,9 @@ def train_network(batch_size, is_distributed=False, is_sparse=False):
         param_attr=fluid.ParamAttr(
             initializer=fluid.initializer.Constant(value=0.01),
             name="__emb__",
-            learning_rate=emb_lr),
+            learning_rate=emb_lr) if is_self_contained_lr else fluid.ParamAttr(
+                initializer=fluid.initializer.Constant(value=0.01),
+                name="__emb__"),
         is_sparse=is_sparse)
     ## vsum
     q_sum = fluid.layers.sequence_pool(input=q_emb, pool_type='sum')
@@ -119,7 +124,9 @@ def train_network(batch_size, is_distributed=False, is_sparse=False):
         param_attr=fluid.ParamAttr(
             initializer=fluid.initializer.Constant(value=0.01),
             name="__emb__",
-            learning_rate=emb_lr),
+            learning_rate=emb_lr) if is_self_contained_lr else fluid.ParamAttr(
+                initializer=fluid.initializer.Constant(value=0.01),
+                name="__emb__"),
         is_sparse=is_sparse)
     ## vsum
     pt_sum = fluid.layers.sequence_pool(input=pt_emb, pool_type='sum')
@@ -144,7 +151,9 @@ def train_network(batch_size, is_distributed=False, is_sparse=False):
         param_attr=fluid.ParamAttr(
             initializer=fluid.initializer.Constant(value=0.01),
             name="__emb__",
-            learning_rate=emb_lr),
+            learning_rate=emb_lr) if is_self_contained_lr else fluid.ParamAttr(
+                initializer=fluid.initializer.Constant(value=0.01),
+                name="__emb__"),
         is_sparse=is_sparse)
     ## vsum
     nt_sum = fluid.layers.sequence_pool(input=nt_emb, pool_type='sum')
@@ -220,7 +229,10 @@ class TestDistSimnetBow2x2(TestDistRunnerBase):
     def get_model(self, batch_size=2):
         # Train program
         avg_cost, acc, predict = \
-            train_network(batch_size, bool(int(os.environ["IS_DISTRIBUTED"])), bool(int(os.environ["IS_SPARSE"])))
+            train_network(batch_size,
+                          bool(int(os.environ["IS_DISTRIBUTED"])),
+                          bool(int(os.environ["IS_SPARSE"])),
+                          bool(int(os.environ["IS_SELF_CONTAINED_LR"])))
 
         inference_program = fluid.default_main_program().clone()
 

python/paddle/fluid/tests/unittests/test_dist_simnet_bow.py

@@ -25,7 +25,11 @@ class TestDistSimnetBowDense2x2(TestDistBase):
         self._enforce_place = "CPU"
 
     def test_simnet_bow(self):
-        need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '0'}
+        need_envs = {
+            "IS_DISTRIBUTED": '0',
+            "IS_SPARSE": '0',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
         self.check_with_place(
             "dist_simnet_bow.py",
             delta=1e-5,
@@ -39,7 +43,11 @@ class TestDistSimnetBow2x2DenseAsync(TestDistBase):
         self._enforce_place = "CPU"
 
     def test_simnet_bow(self):
-        need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '0'}
+        need_envs = {
+            "IS_DISTRIBUTED": '0',
+            "IS_SPARSE": '0',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
         self.check_with_place(
             "dist_simnet_bow.py",
             delta=100,
@@ -53,7 +61,11 @@ class TestDistSimnetBowSparse2x2(TestDistBase):
         self._enforce_place = "CPU"
 
     def test_simnet_bow(self):
-        need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '1'}
+        need_envs = {
+            "IS_DISTRIBUTED": '0',
+            "IS_SPARSE": '1',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
         self.check_with_place(
             "dist_simnet_bow.py",
             delta=1e-5,
@@ -67,7 +79,11 @@ class TestDistSimnetBow2x2SparseAsync(TestDistBase):
         self._enforce_place = "CPU"
 
     def test_simnet_bow(self):
-        need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '1'}
+        need_envs = {
+            "IS_DISTRIBUTED": '0',
+            "IS_SPARSE": '1',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
         self.check_with_place(
             "dist_simnet_bow.py",
             delta=100,
@@ -75,5 +91,59 @@ class TestDistSimnetBow2x2SparseAsync(TestDistBase):
             need_envs=need_envs)
 
 
+class TestDistSimnetBow2x2LookupTableSync(TestDistBase):
+    def _setup_config(self):
+        self._sync_mode = True
+        self._enforce_place = "CPU"
+
+    def test_simnet_bow(self):
+        need_envs = {
+            "IS_DISTRIBUTED": '1',
+            "IS_SPARSE": '1',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
+        self.check_with_place(
+            "dist_simnet_bow.py",
+            delta=1e-5,
+            check_error_log=False,
+            need_envs=need_envs)
+
+
+class TestDistSimnetBow2x2LookupTableAsync(TestDistBase):
+    def _setup_config(self):
+        self._sync_mode = False
+        self._enforce_place = "CPU"
+
+    def test_simnet_bow(self):
+        need_envs = {
+            "IS_DISTRIBUTED": '1',
+            "IS_SPARSE": '1',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
+        self.check_with_place(
+            "dist_simnet_bow.py",
+            delta=100,
+            check_error_log=False,
+            need_envs=need_envs)
+
+
+class TestDistSimnetBow2x2LookupTableNotContainLRSync(TestDistBase):
+    def _setup_config(self):
+        self._sync_mode = True
+        self._enforce_place = "CPU"
+
+    def test_simnet_bow(self):
+        need_envs = {
+            "IS_DISTRIBUTED": '1',
+            "IS_SPARSE": '1',
+            'IS_SELF_CONTAINED_LR': '0'
+        }
+        self.check_with_place(
+            "dist_simnet_bow.py",
+            delta=1e-5,
+            check_error_log=False,
+            need_envs=need_envs)
+
+
 if __name__ == "__main__":
     unittest.main()

python/paddle/fluid/tests/unittests/test_layers.py

@@ -194,6 +194,14 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(layers.sequence_expand(x=x, y=y, ref_level=1))
         print(str(program))
 
+    def test_sequence_unpad(self):
+        program = Program()
+        with program_guard(program):
+            x = layers.data(name='x', shape=[10, 5], dtype='float32')
+            length = layers.data(name='length', shape=[1], dtype='int64')
+            self.assertIsNotNone(layers.sequence_unpad(x=x, length=length))
+        print(str(program))
+
     def test_lstm_unit(self):
         program = Program()
         with program_guard(program):
@@ -406,6 +414,19 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(out)
         print(str(program))
 
+    def test_sequence_slice(self):
+        program = Program()
+        with program_guard(program):
+            import numpy as np
+            seqs = layers.data(
+                name='x', shape=[10, 5], dtype='float32', lod_level=1)
+            offset = layers.assign(input=np.array([[0, 1]]).astype('int32'))
+            length = layers.assign(input=np.array([[2, 1]]).astype('int32'))
+            out = layers.sequence_slice(
+                input=seqs, offset=offset, length=length)
+            self.assertIsNotNone(out)
+        print(str(program))
+
     def test_lod_reset(self):
         program = Program()
         with program_guard(program):

python/paddle/fluid/tests/unittests/test_sequence_unpad_op.py

+# 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.
+
+import unittest
+import six
+import numpy as np
+from op_test import OpTest
+
+
+class TestSequenceUnpadOp(OpTest):
+    def init(self):
+        self.length = [2, 3, 4]
+        self.x_shape = (3, 5)
+        self.dtype = "float32"
+
+    def compute(self):
+        assert len(self.length) == self.x_shape[0]
+        x = np.random.random(self.x_shape).astype(self.dtype)
+        out_lod = [self.length]
+
+        out = x[0, 0:self.length[0]]
+        for i in six.moves.xrange(1, x.shape[0]):
+            out = np.append(out, x[i, 0:self.length[i]], axis=0)
+
+        out_shape = (sum(self.length), )
+        if len(self.x_shape) == 2:
+            out_shape = out_shape + (1, )
+        else:
+            out_shape = out_shape + self.x_shape[2:]
+
+        self.inputs = {
+            'X': x,
+            'Length': np.array(self.length).astype('int64').reshape(-1, 1)
+        }
+        self.outputs = {'Out': (out.reshape(out_shape), out_lod)}
+
+    def setUp(self):
+        self.op_type = 'sequence_unpad'
+        self.init()
+        self.compute()
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+
+class TestSequenceUnpadOp2(TestSequenceUnpadOp):
+    def init(self):
+        self.length = [2, 3, 4]
+        self.x_shape = (3, 5, 4, 3)
+        self.dtype = "float32"
+
+
+class TestSequenceUnpadOp3(TestSequenceUnpadOp):
+    def init(self):
+        self.length = [5, 2, 3, 4]
+        self.x_shape = (4, 5, 3, 3, 6)
+        self.dtype = "float64"
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -1119,6 +1119,7 @@ to transpile() call.")
 
     def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints):
         # 2. add split_ids_op and send_op to send gradient to pservers
+
         # there should only be one table_name
         all_ops = program.global_block().ops
         table_grad_name = grad_var_name(self.table_name)
@@ -1143,7 +1144,7 @@ to transpile() call.")
                         if self.sync_mode else []
                     },
                     attrs={
-                        "sync_mode": self.sync_mode,
+                        "sync_mode": not self.sync_mode,
                         "epmap": pserver_endpoints,
                         RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE,
                         OP_ROLE_VAR_ATTR_NAME: [
@@ -1189,7 +1190,15 @@ to transpile() call.")
     def _create_table_optimize_block(self, pserver_index, pserver_program,
                                      pre_block_idx, grad_to_block_id):
         # STEP: create table optimize block
+        table_opt_block = pserver_program._create_block(pre_block_idx)
         # create table param and grad var in pserver program
+        # create table optimize block in pserver program
+        table_opt_op = [
+            op for op in self.optimize_ops
+            if 'Param' in op.input_names and op.input("Param")[0] ==
+            self.table_name
+        ][0]
+
         origin_param_var = self.origin_program.global_block().vars[
             self.table_name]
 
@@ -1205,19 +1214,16 @@ to transpile() call.")
             dtype=origin_param_var.dtype,
             type=core.VarDesc.VarType.SELECTED_ROWS,
             persistable=True)
+
         # parameter must be selected rows
         param_var.desc.set_type(core.VarDesc.VarType.SELECTED_ROWS)
         grad_var = pserver_program.global_block()._clone_variable(
             self.origin_program.global_block().vars[grad_var_name(
                 self.table_name)])
 
-        # create table optimize block in pserver program
-        table_opt_op = [
-            op for op in self.optimize_ops
-            if 'Param' in op.input_names and op.input("Param")[0] ==
-            self.table_name
-        ][0]
-        table_opt_block = pserver_program._create_block(pre_block_idx)
+        lr_var = pserver_program.global_block()._clone_variable(
+            self.origin_program.global_block().vars[table_opt_op.input(
+                "LearningRate")[0]])
 
         if self.sync_mode:
             # create grad vars in pserver program
@@ -1249,8 +1255,6 @@ to transpile() call.")
             grad_var = pserver_program.global_block()._rename_var(
                 origin_grad_name, splited_grad_name)
 
-        lr_var = pserver_program.global_block().vars[table_opt_op.input(
-            "LearningRate")[0]]
         inputs = {
             "Param": [param_var],
             "Grad": [grad_var],