Merge branch 'develop' of upstream into fix_seq_pad

Author: Yibing Liu

cmake/tensorrt.cmake

@@ -16,7 +16,9 @@ find_library(TENSORRT_LIBRARY NAMES libnvinfer.so libnvinfer.a
     DOC "Path to TensorRT library.")
 
 if(TENSORRT_INCLUDE_DIR AND TENSORRT_LIBRARY)
+  if(WITH_DSO)
     set(TENSORRT_FOUND ON)
+  endif(WITH DSO)
 else()
     set(TENSORRT_FOUND OFF)
 endif()

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -429,7 +429,7 @@ struct LSTM : public PatternBase {
 
 struct GRU : public PatternBase {
   GRU(PDPattern* pattern, const std::string& name_scope)
-      : PatternBase(pattern, name_scope, "lstm") {}
+      : PatternBase(pattern, name_scope, "gru") {}
 
   PDNode* operator()(PDNode* x);
 

paddle/fluid/inference/api/api.cc

@@ -9,8 +9,8 @@ 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 <glog/logging.h>
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
+#include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
 
@@ -64,13 +64,15 @@ PaddleBuf& PaddleBuf::operator=(PaddleBuf&& other) {
 
 void PaddleBuf::Resize(size_t length) {
   // Only the owned memory can be reset, the external memory can't be changed.
-  if (length_ == length) return;
+  if (length_ >= length) return;
   if (memory_owned_) {
     Free();
+    data_ = malloc(length);
+    length_ = length;
+    memory_owned_ = true;
+  } else {
+    PADDLE_THROW("The memory is allocated externally, can not Resized");
   }
-  data_ = new char[length];
-  length_ = length;
-  memory_owned_ = true;
 }
 
 void PaddleBuf::Reset(void* data, size_t length) {
@@ -82,8 +84,8 @@ void PaddleBuf::Reset(void* data, size_t length) {
 
 void PaddleBuf::Free() {
   if (memory_owned_ && data_) {
-    assert(length_ > 0);
-    delete[] static_cast<char*>(data_);
+    PADDLE_ENFORCE_GT(length_, 0);
+    free(static_cast<char*>(data_));
     data_ = nullptr;
     length_ = 0;
   }

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -53,7 +53,7 @@ set(TEXT_CLASSIFICATION_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/text_classifi
 download_model_and_data(${TEXT_CLASSIFICATION_INSTALL_DIR} "text-classification-Senta.tar.gz" "text_classification_data.txt.tar.gz")
 inference_analysis_test(test_analyzer_text_classification SRCS analyzer_text_classification_tester.cc
     EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
-    ARGS --infer_model=${TEXT_CLASSIFICATION_INSTALL_DIR}/text-classification-Senta
+    ARGS --infer_model=${TEXT_CLASSIFICATION_INSTALL_DIR}/model
          --infer_data=${TEXT_CLASSIFICATION_INSTALL_DIR}/data.txt)
 
 # ocr

paddle/fluid/operators/conv_mkldnn_op.cc

@@ -300,6 +300,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
     std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
     bool fuse_relu = ctx.Attr<bool>("fuse_relu");
+    bool fuse_eltwise = ctx.Attr<bool>("fuse_eltwise");
     int groups = ctx.Attr<int>("groups");
 
     // TODO: add support for dilation
@@ -366,12 +367,13 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
       bias_tz = paddle::framework::vectorize2int(bias->dims());
       auto bias_md = platform::MKLDNNMemDesc(
           bias_tz, platform::MKLDNNGetDataType<T>(), memory::format::x);
-      conv_pd =
-          ConvFwdPrimitiveDesc(src_md, weights_md, bias_md, dst_md, strides,
-                               paddings, mkldnn_engine, fuse_relu);
+      conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, bias_md, dst_md,
+                                     strides, paddings, mkldnn_engine,
+                                     fuse_relu, fuse_eltwise);
     } else {
-      conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides,
-                                     paddings, mkldnn_engine, fuse_relu);
+      conv_pd =
+          ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides, paddings,
+                               mkldnn_engine, fuse_relu, fuse_eltwise);
     }
     // Save conv_pd/src_memory/weights_memory for backward pass
     dev_ctx.SetBlob(key_conv_pd, conv_pd);
@@ -421,16 +423,26 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
   }
 
  private:
-  mkldnn::primitive_attr AddRelu() const {
-    // Fusion with ReLU layer is executed through the PostOps feature. Create a
-    // PostOps object and configure it to execute an eltwise relu operation.
+  mkldnn::primitive_attr CreatePostOps(bool fuse_relu,
+                                       bool fuse_eltwise) const {
     mkldnn::primitive_attr conv_attr;
-    constexpr float scale = 1.0f;
-    constexpr float negative_slope = 0.0f;
-    constexpr float placeholder = 0.0f;
     mkldnn::post_ops post_operations;
-    post_operations.append_eltwise(scale, mkldnn::algorithm::eltwise_relu,
-                                   negative_slope, placeholder);
+    // Fusion with Elementwise layer relies on adding a sum post-operation with
+    // the scale parameter. It is assumed that when fuse_eltwise is true, the
+    // Output tensor contains the data coming from residual connection. The
+    // result of this post_op is: Output = scale * Output + Conv_Out.
+    if (fuse_eltwise) {
+      post_operations.append_sum(1.0f);
+    }
+    // Fusion with ReLU layer is executed through the PostOps feature. Create a
+    // PostOps object and configure it to execute an eltwise relu operation.
+    if (fuse_relu) {
+      constexpr float scale = 1.0f;
+      constexpr float negative_slope = 0.0f;
+      constexpr float placeholder = 0.0f;
+      post_operations.append_eltwise(scale, mkldnn::algorithm::eltwise_relu,
+                                     negative_slope, placeholder);
+    }
     conv_attr.set_post_ops(post_operations);
     return conv_attr;
   }
@@ -439,8 +451,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
   ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights,
                        const memory::desc& dst, const std::vector<int>& strides,
                        const std::vector<int>& paddings,
-                       const mkldnn::engine& engine,
-                       const bool fuse_relu) const {
+                       const mkldnn::engine& engine, const bool fuse_relu,
+                       const bool fuse_eltwise) const {
     memory::dims stride_dims = {strides[0], strides[1]};
     memory::dims padding_dims = {paddings[0], paddings[1]};
 
@@ -449,10 +461,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         dst, stride_dims, padding_dims, padding_dims,
         mkldnn::padding_kind::zero);
 
-    mkldnn::primitive_attr conv_attr;
-    if (fuse_relu) {
-      conv_attr = AddRelu();
-    }
+    mkldnn::primitive_attr conv_attr = CreatePostOps(fuse_relu, fuse_eltwise);
 
     auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
         conv_desc, conv_attr, engine);
@@ -466,8 +475,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
                        const memory::desc& bias, const memory::desc& dst,
                        const std::vector<int>& strides,
                        const std::vector<int>& paddings,
-                       const mkldnn::engine& engine,
-                       const bool fuse_relu) const {
+                       const mkldnn::engine& engine, const bool fuse_relu,
+                       const bool fuse_eltwise) const {
     memory::dims stride_dims = {strides[0], strides[1]};
     memory::dims padding_dims = {paddings[0], paddings[1]};
 
@@ -476,10 +485,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         bias, dst, stride_dims, padding_dims, padding_dims,
         mkldnn::padding_kind::zero);
 
-    mkldnn::primitive_attr conv_attr;
-    if (fuse_relu) {
-      conv_attr = AddRelu();
-    }
+    mkldnn::primitive_attr conv_attr = CreatePostOps(fuse_relu, fuse_eltwise);
 
     auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
         conv_desc, conv_attr, engine);

paddle/fluid/operators/conv_op.cc

@@ -164,6 +164,11 @@ void Conv2DOpMaker::Make() {
       .SetDefault(false);
   AddAttr<bool>("fuse_relu", "(bool, default false) Only used in mkldnn kernel")
       .SetDefault(false);
+  AddAttr<bool>("fuse_eltwise",
+                "(bool, default false) Only used in mkldnn kernel. Used "
+                "whenever convolution output is connected via skip connection "
+                "to a previous layer.")
+      .SetDefault(false);
   AddAttr<std::string>(
       "data_format",
       "(string, default NCHW) Only used in "

paddle/fluid/operators/distributed/grpc_client.cc

@@ -125,7 +125,7 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep,
   VarHandlePtr h(new VarHandle(ep, "Get", var_name_val, p_ctx, p_scope));
   s->Prepare(h, time_out);
 
-  framework::AsyncIO([var_name_val, p_scope, p_ctx, s, this] {
+  framework::AsyncIO([var_name_val, s, this] {
     // prepare input
     sendrecv::VariableMessage req;
     req.set_varname(var_name_val);
@@ -166,7 +166,7 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep,
   s->Prepare(h, time_out);
 
   framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx,
-                      time_out, s, this] {
+                      s, this] {
     auto* var = p_scope->FindVar(in_var_name_val);
 
     ::grpc::ByteBuffer req;

paddle/fluid/operators/distributed/proto_encoder_helper.h

@@ -82,8 +82,10 @@ class ProtoEncodeHelper {
       : base_(buf), p_(buf), limit_(base_ + max_size) {}
 
   ~ProtoEncodeHelper() {
+#define REPLACE_ENFORCE_GLOG 1
     // Make sure callers didn't do operations that went over max_size promised
-    PADDLE_ENFORCE_LE(p_, limit_);
+    paddle::platform::throw_on_error(p_ <= limit_);
+#undef REPLACE_ENFORCE_GLOG
   }
 
   const char* data() const { return base_; }

paddle/fluid/operators/listen_and_serv_op.cc

@@ -59,17 +59,16 @@ static void ParallelExecuteBlocks(
     framework::ProgramDesc *program, framework::Scope *scope) {
   std::vector<std::future<void>> fs;
   for (size_t idx : parallel_blkids) {
-    fs.push_back(
-        framework::Async([&executor, &prepared, &program, &scope, idx]() {
-          int run_block = idx;  // thread local
-          try {
-            VLOG(3) << "running server block: " << run_block
-                    << "pointer: " << prepared[run_block].get();
-            executor->RunPreparedContext(prepared[run_block].get(), scope);
-          } catch (const std::exception &e) {
-            LOG(ERROR) << "run sub program error " << e.what();
-          }
-        }));
+    fs.push_back(framework::Async([&executor, &prepared, &scope, idx]() {
+      int run_block = idx;  // thread local
+      try {
+        VLOG(3) << "running server block: " << run_block
+                << "pointer: " << prepared[run_block].get();
+        executor->RunPreparedContext(prepared[run_block].get(), scope);
+      } catch (const std::exception &e) {
+        LOG(ERROR) << "run sub program error " << e.what();
+      }
+    }));
   }
   for (size_t i = 0; i < fs.size(); ++i) fs[i].wait();
 }

paddle/fluid/operators/math/sequence_pooling.cc

@@ -103,6 +103,58 @@ class MaxSeqPoolGradFunctor {
   }
 };
 
+template <typename T>
+class LastSeqPoolFunctor {
+ public:
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::LoDTensor& input,
+                  framework::Tensor* output) {
+    // Create pointers to input and output data
+    auto* in_data = input.data<T>();
+    auto* out_data = output->data<T>();
+
+    // Calculate the size of each item in sequence
+    int64_t item_size = input.numel() / input.dims()[0];
+    auto lod = input.lod()[0];
+    int seq_num = static_cast<int>(lod.size()) - 1;
+    for (int i = 0; i < seq_num; ++i) {
+      // Calculate the length of each sequence
+      int64_t seq_len = static_cast<int64_t>(lod[i + 1] - lod[i]);
+      // Point to the begin of next sequence
+      in_data += seq_len * item_size;
+      // Copy the last item of sequence to output
+      std::memcpy(out_data, (in_data - item_size), item_size * sizeof(T));
+      out_data += item_size;
+    }
+  }
+};
+
+template <typename T>
+class FirstSeqPoolFunctor {
+ public:
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::LoDTensor& input,
+                  framework::Tensor* output) {
+    // Create pointers to input and output data
+    auto* in_data = input.data<T>();
+    auto* out_data = output->data<T>();
+
+    // Calculate the size of each item in sequence
+    int64_t item_size = input.numel() / input.dims()[0];
+    auto lod = input.lod()[0];
+    int seq_num = static_cast<int>(lod.size()) - 1;
+    for (int i = 0; i < seq_num; ++i) {
+      // Calculate the length of each sequence
+      int64_t seq_len = static_cast<int64_t>(lod[i + 1] - lod[i]);
+      // Copy the first item of sequence to output
+      std::memcpy(out_data, in_data, item_size * sizeof(T));
+      // Point to the next sequence
+      in_data += seq_len * item_size;
+      out_data += item_size;
+    }
+  }
+};
+
 template <typename T>
 class SequencePoolFunctor<platform::CPUDeviceContext, T> {
  public:
@@ -116,6 +168,16 @@ class SequencePoolFunctor<platform::CPUDeviceContext, T> {
       max_pool(context, input, output, index);
       return;
     }
+    if (pooltype == "LAST") {
+      math::LastSeqPoolFunctor<T> last_pool;
+      last_pool(context, input, output);
+      return;
+    }
+    if (pooltype == "FIRST") {
+      math::FirstSeqPoolFunctor<T> first_pool;
+      first_pool(context, input, output);
+      return;
+    }
     auto lod = input.lod()[0];
     auto& place = *context.eigen_device();
     for (int i = 0; i < static_cast<int>(lod.size()) - 1; ++i) {
@@ -133,10 +195,6 @@ class SequencePoolFunctor<platform::CPUDeviceContext, T> {
       } else if (pooltype == "SQRT") {
         out_e.device(place) = in_e.sum(Eigen::array<int, 1>({{0}})) /
                               std::sqrt(static_cast<T>(h));
-      } else if (pooltype == "LAST") {
-        out_e.device(place) = in_e.chip(h - 1, 0);
-      } else if (pooltype == "FIRST") {
-        out_e.device(place) = in_e.chip(0, 0);
       } else {
         PADDLE_THROW("unsupported pooling pooltype");
       }