Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

… dev/unify_gpu_context3

cmake/external/dgc.cmake

@@ -33,7 +33,7 @@ ExternalProject_Add(
   URL_MD5 "94e6fa1bc97169d0e1aad44570fe3251"
   PREFIX "${DGC_PREFIX_DIR}"
   CONFIGURE_COMMAND ""
-  BUILD_COMMAND make -j $(nproc)
+  BUILD_COMMAND make -j${NPROC}
   INSTALL_COMMAND
     mkdir -p ${DGC_INSTALL_DIR}/lib/ ${DGC_INCLUDE_DIR}/dgc && cp
     ${DGC_SOURCES_DIR}/build/lib/libdgc.a ${DGC_LIBRARIES} && cp

cmake/external/gflags.cmake

@@ -29,8 +29,8 @@ else()
   set(GFLAGS_LIBRARIES
       "${GFLAGS_INSTALL_DIR}/lib/libgflags.a"
       CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
-  set(BUILD_COMMAND $(MAKE) --silent)
-  set(INSTALL_COMMAND $(MAKE) install)
+  set(BUILD_COMMAND ${CMAKE_COMMAND} --build .)
+  set(INSTALL_COMMAND ${CMAKE_COMMAND} --build . --target install)
 endif()
 
 include_directories(${GFLAGS_INCLUDE_DIR})

cmake/external/mkldnn.cmake

@@ -61,6 +61,12 @@ else()
       CACHE FILEPATH "mkldnn library." FORCE)
 endif()
 
+if(LINUX)
+  set(BUILD_BYPRODUCTS_ARGS ${MKLDNN_LIB})
+else()
+  set(BUILD_BYPRODUCTS_ARGS "")
+endif()
+
 ExternalProject_Add(
   ${MKLDNN_PROJECT}
   ${EXTERNAL_PROJECT_LOG_ARGS} ${SHALLOW_CLONE}
@@ -83,7 +89,8 @@ ExternalProject_Add(
              -DCMAKE_POSITION_INDEPENDENT_CODE=ON
              -DDNNL_BUILD_TESTS=OFF
              -DDNNL_BUILD_EXAMPLES=OFF
-  CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR})
+  CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
+  BUILD_BYPRODUCTS ${BUILD_BYPRODUCTS_ARGS})
 
 message(STATUS "MKLDNN library: ${MKLDNN_LIB}")
 add_definitions(-DPADDLE_WITH_MKLDNN)

cmake/external/openblas.cmake

@@ -53,7 +53,7 @@ if(NOT WIN32)
     PREFIX ${CBLAS_PREFIX_DIR}
     INSTALL_DIR ${CBLAS_INSTALL_DIR}
     BUILD_IN_SOURCE 1
-    BUILD_COMMAND make -j$(nproc) ${COMMON_ARGS} ${OPTIONAL_ARGS}
+    BUILD_COMMAND make -j${NPROC} ${COMMON_ARGS} ${OPTIONAL_ARGS}
     INSTALL_COMMAND make install NO_SHARED=1 NO_LAPACK=1 PREFIX=<INSTALL_DIR>
     UPDATE_COMMAND ""
     CONFIGURE_COMMAND ""

cmake/external/rocksdb.cmake

@@ -44,7 +44,8 @@ ExternalProject_Add(
     ${ROCKSDB_PREFIX_DIR}/src/extern_rocksdb/librocksdb.a ${ROCKSDB_LIBRARIES}
     && cp -r ${ROCKSDB_PREFIX_DIR}/src/extern_rocksdb/include
     ${ROCKSDB_INSTALL_DIR}/
-  BUILD_IN_SOURCE 1)
+  BUILD_IN_SOURCE 1
+  BYPRODUCTS ${ROCKSDB_LIBRARIES})
 
 add_dependencies(extern_rocksdb snappy)
 

cmake/third_party.cmake

@@ -27,6 +27,9 @@ set(THIRD_PARTY_CACHE_PATH
 set(THIRD_PARTY_BUILD_TYPE Release)
 set(third_party_deps)
 
+include(ProcessorCount)
+ProcessorCount(NPROC)
+
 # cache funciton to avoid repeat download code of third_party.
 # This function has 4 parameters, URL / REPOSITOR / TAG / DIR:
 # 1. URL:           specify download url of 3rd party

paddle/fluid/operators/overlap_add_op.cc

@@ -12,7 +12,11 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/operators/overlap_add_op.h"
+#include "paddle/fluid/framework/infershape_utils.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/phi/core/infermeta_utils.h"
+#include "paddle/phi/infermeta/backward.h"
+#include "paddle/phi/infermeta/unary.h"
 
 namespace paddle {
 namespace operators {
@@ -21,93 +25,6 @@ class OverlapAddOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
-  void InferShape(framework::InferShapeContext* ctx) const override {
-    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "overlap_add");
-    OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out", "overlap_add");
-
-    const int hop_length = ctx->Attrs().Get<int>("hop_length");
-    const int axis = ctx->Attrs().Get<int>("axis");
-
-    const auto x_dims = ctx->GetInputDim("X");
-    const int x_rank = x_dims.size();
-
-    PADDLE_ENFORCE_GE(
-        x_rank,
-        2,
-        platform::errors::InvalidArgument(
-            "Input(X) of OverlapAddOp should be a tensor which contains "
-            "at least 2 dimensions, but got rank %s.",
-            x_rank));
-
-    PADDLE_ENFORCE_GT(
-        hop_length,
-        0,
-        platform::errors::InvalidArgument(
-            "Attribute(hop_length) of OverlapAddOp should be greater "
-            "than 0, but got %s.",
-            hop_length));
-
-    PADDLE_ENFORCE_EQ(
-        (axis == 0 || axis == -1),
-        true,
-        platform::errors::InvalidArgument(
-            "Attribute(axis) of OverlapAddOp should 0 or -1, but got %s.",
-            axis));
-
-    std::vector<int64_t> output_shape;
-    int n_frames;
-    int frame_length;
-    int seq_length;
-
-    int start_axis;
-    int end_axis;
-    if (axis == 0) {
-      n_frames = x_dims[0];
-      frame_length = x_dims[1];
-      start_axis = 2;
-      end_axis = x_rank - 1;
-    } else {
-      n_frames = x_dims[x_rank - 1];
-      frame_length = x_dims[x_rank - 2];
-      start_axis = 0;
-      end_axis = x_rank - 3;
-    }
-
-    bool contain_unknown_dim = phi::contain_unknown_dim(x_dims);
-    bool check = ctx->IsRuntime() || !contain_unknown_dim;
-    if (check) {
-      PADDLE_ENFORCE_LE(
-          hop_length,
-          frame_length,
-          platform::errors::InvalidArgument(
-              "Attribute(hop_length) of OverlapAddOp should be less or equal "
-              "than frame_length, but got hop_length(%s) > frame_length(%s).",
-              hop_length,
-              frame_length));
-    }
-
-    if (n_frames == -1) {
-      seq_length = -1;
-    } else {
-      seq_length = (n_frames - 1) * hop_length + frame_length;
-    }
-
-    // It won't go into for loop when x_rank == 2U.
-    for (int i = start_axis; i <= end_axis; i++) {
-      output_shape.push_back(x_dims[i]);
-    }
-
-    if (axis == 0) {
-      // (seq_length, ...)
-      output_shape.insert(output_shape.begin(), seq_length);
-    } else {
-      // (..., seq_length)
-      output_shape.push_back(seq_length);
-    }
-
-    ctx->SetOutputDim("Out", phi::make_ddim(output_shape));
-  }
-
  protected:
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
@@ -137,17 +54,6 @@ class OverlapAddOpMaker : public framework::OpProtoAndCheckerMaker {
 class OverlapAddOpGrad : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
-  void InferShape(framework::InferShapeContext* ctx) const override {
-    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "overlap_add_grad");
-    OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")),
-                   "Input",
-                   "Out@GRAD",
-                   "overlap_add_grad");
-    const auto x_dims = ctx->GetInputDim("X");
-    if (ctx->HasOutput(framework::GradVarName("X"))) {
-      ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
-    }
-  }
 
  protected:
   framework::OpKernelType GetExpectedKernelType(
@@ -176,30 +82,21 @@ class OverlapAddOpGradMaker : public framework::SingleGradOpMaker<T> {
 
 namespace ops = paddle::operators;
 
+DECLARE_INFER_SHAPE_FUNCTOR(overlap_add,
+                            OverlapAddInferShapeFunctor,
+                            PD_INFER_META(phi::OverlapAddInferMeta));
+
+DECLARE_INFER_SHAPE_FUNCTOR(overlap_add_grad,
+                            OverlapAddGradInferShapeFunctor,
+                            PD_INFER_META(phi::OverlapAddGradInferMeta));
+
 REGISTER_OPERATOR(overlap_add,
                   ops::OverlapAddOp,
                   ops::OverlapAddOpMaker,
                   ops::OverlapAddOpGradMaker<paddle::framework::OpDesc>,
-                  ops::OverlapAddOpGradMaker<paddle::imperative::OpBase>);
-
-REGISTER_OPERATOR(overlap_add_grad, ops::OverlapAddOpGrad);
-
-REGISTER_OP_CPU_KERNEL(
-    overlap_add,
-    ops::OverlapAddKernel<phi::CPUContext, int>,
-    ops::OverlapAddKernel<phi::CPUContext, int64_t>,
-    ops::OverlapAddKernel<phi::CPUContext, float>,
-    ops::OverlapAddKernel<phi::CPUContext, double>,
-    ops::OverlapAddKernel<phi::CPUContext, paddle::platform::complex<float>>,
-    ops::OverlapAddKernel<phi::CPUContext, paddle::platform::complex<double>>);
+                  ops::OverlapAddOpGradMaker<paddle::imperative::OpBase>,
+                  OverlapAddInferShapeFunctor);
 
-REGISTER_OP_CPU_KERNEL(
-    overlap_add_grad,
-    ops::OverlapAddGradKernel<phi::CPUContext, int>,
-    ops::OverlapAddGradKernel<phi::CPUContext, int64_t>,
-    ops::OverlapAddGradKernel<phi::CPUContext, float>,
-    ops::OverlapAddGradKernel<phi::CPUContext, double>,
-    ops::OverlapAddGradKernel<phi::CPUContext,
-                              paddle::platform::complex<float>>,
-    ops::OverlapAddGradKernel<phi::CPUContext,
-                              paddle::platform::complex<double>>);
+REGISTER_OPERATOR(overlap_add_grad,
+                  ops::OverlapAddOpGrad,
+                  OverlapAddGradInferShapeFunctor);