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

… while_loop_0d

CMakeLists.txt

@@ -157,8 +157,14 @@ if(WIN32)
     endforeach()
   endif()
 
-  # NOTE(zhouwei): msvc max/min macro conflict with std::min/max, define NOMINMAX globally
+  # msvc max/min macro conflict with std::min/max, define NOMINMAX globally
   add_definitions("-DNOMINMAX")
+
+  # 1. windows.h define 'small' cause CUDA11.6/11.7/11.8 's cub compile error,
+  # see https://github.com/microsoft/onnxruntime/issues/11227
+  # 2. WIN32_LEAN_AND_MEAN minimize the windows include files, avoid define 'small'
+  add_definitions(-DWIN32_LEAN_AND_MEAN)
+
   # windows build turn off warnings, use parallel compiling.
   foreach(
     flag_var

cmake/external/cinn.cmake

@@ -40,7 +40,10 @@ set(CINN_OPTIONAL_ARGS
     -DWITH_MKL_CBLAS=${WITH_MKL}
     -DWITH_MKLDNN=${WITH_MKL}
     -DPUBLISH_LIBS=ON
-    -DWITH_TESTING=ON)
+    -DWITH_TESTING=ON
+    -DPYTHON_EXECUTABLE=${PYTHON_EXECUTABLE}
+    -DPYTHON_INCLUDE_DIR=${PYTHON_INCLUDE_DIR}
+    -DPYTHON_LIBRARIES=${PYTHON_LIBRARIES})
 set(CINN_BUILD_COMMAND ${CMAKE_COMMAND} --build . --target cinnapi -j)
 set(CINN_BINARY_DIR ${CINN_PREFIX_DIR}/src/external_cinn-build)
 set(CINN_LIB_NAME "libcinnapi.so")

cmake/external/cub.cmake

@@ -14,7 +14,7 @@
 
 include(ExternalProject)
 
-# Note(zhouwei): extern_cub  has code __FILE_, If the path of extern_cub is changed,
+# extern_cub  has code __FILE_, If the path of extern_cub is changed,
 # it will effect about 30+ cu files sccache hit and slow compile speed  on windows.
 # Therefore, a fixed CUB_PATH will be input to increase the sccache hit rate.
 set(CUB_PATH
@@ -25,7 +25,7 @@ set(CUB_PREFIX_DIR ${CUB_PATH})
 set(CUB_REPOSITORY ${GIT_URL}/NVlabs/cub.git)
 
 if(${CMAKE_CUDA_COMPILER_VERSION} GREATER_EQUAL 11.6)
-  # cuda_11.6.2_511.65‘s own cub is 1.15.0, which will cause compiling error in windows.
+  # cuda_11.6/11.7/11.8‘s own cub is 1.15.0, which will cause compiling error in windows.
   set(CUB_TAG 1.16.0)
   # cub 1.16.0 is not compitable with current thrust version
   add_definitions(-DTHRUST_IGNORE_CUB_VERSION_CHECK)

cmake/operators.cmake

@@ -411,6 +411,17 @@ function(op_library TARGET)
       set(pybind_flag 1)
     endif()
 
+    # pybind USE_OP_ITSELF
+    set(op_name "")
+    # Add PHI Kernel Registry Message
+    find_register(${cc_src} "REGISTER_ACTIVATION_OP" op_name)
+    if(NOT ${op_name} EQUAL "")
+      file(APPEND ${pybind_file} "USE_OP_ITSELF(${op_name});\n")
+      # hack: for example, the target in conv_transpose_op.cc is conv2d_transpose, used in mkldnn
+      set(TARGET ${op_name})
+      set(pybind_flag 1)
+    endif()
+
     set(op_name "")
     find_register(${cc_src} "REGISTER_OP_WITHOUT_GRADIENT" op_name)
     if(NOT ${op_name} EQUAL "")

cmake/third_party.cmake

@@ -320,8 +320,7 @@ if(WITH_ONNXRUNTIME)
 endif()
 
 if(WITH_GPU)
-  if(${CMAKE_CUDA_COMPILER_VERSION} LESS 11.0
-     OR (WIN32 AND ${CMAKE_CUDA_COMPILER_VERSION} GREATER_EQUAL 11.6))
+  if(${CMAKE_CUDA_COMPILER_VERSION} LESS 11.0)
     include(external/cub) # download cub
     list(APPEND third_party_deps extern_cub)
   endif()

paddle/fluid/eager/auto_code_generator/generator/eager_gen.py

@@ -330,7 +330,7 @@ class {} : public egr::GradNodeBase {{
 #include "paddle/fluid/eager/nan_inf_utils.h"
 #include "paddle/phi/api/include/sparse_api.h"
 #include "paddle/fluid/eager/api/manual/eager_manual/nodes/nodes.h"
-#include "paddle/fluid/prim/api/manual/backward/composite_backward_api.h"
+#include "paddle/fluid/prim/api/composite_backward/composite_backward_api.h"
 #include "paddle/fluid/prim/api/all.h"
 #include "paddle/fluid/prim/utils/utils.h"
 DECLARE_bool(check_nan_inf);

paddle/fluid/framework/dlpack_tensor.cc

@@ -134,6 +134,58 @@ struct DLDeviceVisitor
 };
 }  // namespace internal
 
+struct PaddleDLMTensor {
+  phi::DenseTensor handle;
+  DLManagedTensor tensor;
+};
+
+void deleter(DLManagedTensor *arg) {
+  delete[] arg->dl_tensor.shape;
+  delete[] arg->dl_tensor.strides;
+  delete static_cast<PaddleDLMTensor *>(arg->manager_ctx);
+}
+
+DLManagedTensor *toDLPack(const phi::DenseTensor &src) {
+  PaddleDLMTensor *pdDLMTensor(new PaddleDLMTensor);
+  pdDLMTensor->handle = const_cast<phi::DenseTensor &>(src);
+  pdDLMTensor->tensor.manager_ctx = pdDLMTensor;
+  pdDLMTensor->tensor.deleter = &deleter;
+  pdDLMTensor->tensor.dl_tensor.data = const_cast<void *>(src.data());
+
+  // init ndim
+  using DimType = decltype(pdDLMTensor->tensor.dl_tensor.ndim);  // int
+  pdDLMTensor->tensor.dl_tensor.ndim = static_cast<DimType>(src.dims().size());
+  DimType ndim = pdDLMTensor->tensor.dl_tensor.ndim;
+
+  // init shape
+  auto shape = new int64_t[ndim];
+  for (DimType i = 0; i < ndim; ++i) {
+    shape[i] = src.dims()[i];
+  }
+  pdDLMTensor->tensor.dl_tensor.shape = shape;
+
+  // init stride
+  auto strides = new int64_t[ndim];
+  for (DimType i = 0; i < ndim; ++i) {
+    strides[i] = 1;
+  }
+  for (DimType i = ndim - 2; i >= 0; --i) {
+    strides[i] = shape[i + 1] * strides[i + 1];
+  }
+  pdDLMTensor->tensor.dl_tensor.strides = strides;
+
+  // init device, DLDevice type with device_type and device_id
+  auto place = src.place();
+  pdDLMTensor->tensor.dl_tensor.device =
+      paddle::platform::VisitPlace(place, internal::DLDeviceVisitor());
+
+  pdDLMTensor->tensor.dl_tensor.dtype = internal::GetDLDataTypeFromTypeIndex(
+      framework::TransToProtoVarType(src.dtype()));
+
+  pdDLMTensor->tensor.dl_tensor.byte_offset = 0;
+  return &(pdDLMTensor->tensor);
+}
+
 DLPackTensor::DLPackTensor(const phi::DenseTensor &tensor, LaneType lanes) {
   // init data, data buffer
   t_.data = const_cast<void *>(tensor.data());

paddle/fluid/framework/dlpack_tensor.h

@@ -44,5 +44,7 @@ class DLPackTensor {
   ShapeType shape_[DDim::kMaxRank];
 };
 
+DLManagedTensor* toDLPack(const phi::DenseTensor& src);
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/feed_fetch_type.h

@@ -26,6 +26,11 @@ namespace framework {
 using FeedType =
     paddle::variant<phi::DenseTensor, Strings, phi::SparseCooTensor>;
 
+template <>
+struct PhiVectorType<FeedType> {
+  const char *type_name = "PhiVectorFeedType";
+};
+
 using FeedList = paddle::framework::PhiVector<FeedType>;
 
 using FetchType = paddle::variant<phi::DenseTensor,

paddle/fluid/framework/ir/CMakeLists.txt

@@ -387,22 +387,10 @@ if(WITH_MKLDNN)
     test_depthwise_conv_mkldnn_pass
     SRCS mkldnn/depthwise_conv_mkldnn_pass_tester.cc
     DEPS depthwise_conv_mkldnn_pass)
-  cc_test(
-    test_conv_bias_mkldnn_fuse_pass_cc
-    SRCS mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc
-    DEPS conv_bias_mkldnn_fuse_pass naive_executor)
   cc_test(
     test_conv_activation_mkldnn_fuse_pass
     SRCS mkldnn/conv_activation_mkldnn_fuse_pass_tester.cc
     DEPS conv_activation_mkldnn_fuse_pass)
-  cc_test(
-    test_conv_concat_relu_mkldnn_fuse_pass
-    SRCS mkldnn/conv_concat_relu_mkldnn_fuse_pass_tester.cc
-    DEPS conv_activation_mkldnn_fuse_pass)
-  cc_test_old(
-    test_conv_elementwise_add_mkldnn_fuse_pass SRCS
-    mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS
-    conv_elementwise_add_mkldnn_fuse_pass pass_test_util)
   cc_test_old(
     test_int8_scale_calculation_mkldnn_pass SRCS
     mkldnn/int8_scale_calculation_mkldnn_pass_tester.cc DEPS