Merge remote-tracking branch 'upstream/develop' into develop

Author: seemingwang

paddle/fluid/framework/CMakeLists.txt

@@ -360,46 +360,11 @@ set(FLUID_FRAMEWORK_MODULES proto_desc memory lod_tensor executor data_feed_prot
 
 cc_library(paddle_framework DEPS ${FLUID_FRAMEWORK_MODULES})
 
-# Old custom op extension mechanism related, will be removed in 2.1.0
-cc_library(paddle_framework_shared
-    SHARED SRCS executor.cc operator.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/c/c_api.cc
-    ${CMAKE_SOURCE_DIR}/paddle/fluid/imperative/layer.cc
-    DEPS ${FLUID_FRAMEWORK_MODULES})
-get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
-set_target_properties(paddle_framework_shared PROPERTIES OUTPUT_NAME paddle_framework)
-target_link_libraries(paddle_framework_shared ${os_dependency_modules})
-
-if (LINUX)
-  set(FLUID_FRAMEWORK_SHARED_LIB
-      ${PADDLE_BINARY_DIR}/paddle/fluid/framework/libpaddle_framework.so
-      CACHE INTERNAL "Fluid framework lib")
-endif()
-
-if (WIN32)
-  if("${CMAKE_GENERATOR}" STREQUAL "Ninja")
-    set(paddle_framework_lib_path ${CMAKE_CURRENT_BINARY_DIR})
-  else()
-    set(paddle_framework_lib_path ${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE})
-  endif()
-  set(FLUID_FRAMEWORK_IMPORT_LIB
-      ${paddle_framework_lib_path}/paddle_framework.lib
-      CACHE INTERNAL "Fluid framework lib")
-  set(FLUID_FRAMEWORK_SHARED_LIB
-      ${paddle_framework_lib_path}/paddle_framework.dll
-      CACHE INTERNAL "Fluid framework dll")
-endif()
-
-if(APPLE)
-  set(FLUID_FRAMEWORK_SHARED_LIB
-      ${PADDLE_BINARY_DIR}/paddle/fluid/framework/libpaddle_framework.dylib
-      CACHE INTERNAL "Fluid framework lib")
-endif()
 if(WITH_TESTING AND TEST selected_rows_test)
   set_tests_properties(selected_rows_test PROPERTIES TIMEOUT 120)
 endif()
 
-# New custom op extension mechanism related
+##### 2.0 New custom op extension mechanism related #####
 
 # if not deps `layer`, will cause: undefined symbol: _ZN6paddle10imperative7VarBase9name_set_
 set(PADDLE_CUSTOM_OP_MODULES custom_tensor op_meta_info custom_operator layer)

paddle/fluid/framework/c/c_api.cc

@@ -1192,6 +1192,7 @@ USE_TRT_CONVERTER(scale);
 USE_TRT_CONVERTER(stack);
 USE_TRT_CONVERTER(clip);
 USE_TRT_CONVERTER(gather);
+USE_TRT_CONVERTER(yolo_box);
 USE_TRT_CONVERTER(roi_align);
 USE_TRT_CONVERTER(affine_channel);
 USE_TRT_CONVERTER(multiclass_nms);

paddle/fluid/framework/c/c_api.h

@@ -6,6 +6,7 @@ nv_library(tensorrt_converter
                 shuffle_channel_op.cc swish_op.cc instance_norm_op.cc stack_op.cc transpose_op.cc flatten_op.cc
                 emb_eltwise_layernorm.cc skip_layernorm.cc scale_op.cc slice_op.cc hard_sigmoid_op.cc hard_swish_op.cc clip_op.cc
                 gather_op.cc
+                yolo_box_op.cc
                 roi_align_op.cc
                 affine_channel_op.cc
                 multiclass_nms_op.cc

paddle/fluid/framework/load_op_lib.h

@@ -158,17 +158,49 @@ class BatchNormOpConverter : public OpConverter {
     TensorRTEngine::Weight power_weights{nvinfer1::DataType::kFLOAT, nullptr,
                                          0};
 
-    nvinfer1::IScaleLayer* layer =
-        TRT_ENGINE_ADD_LAYER(engine_, Scale, *const_cast<nvinfer1::ITensor*>(X),
-                             nvinfer1::ScaleMode::kCHANNEL, shift_weights.get(),
-                             scale_weights.get(), power_weights.get());
+    int dynamic_shape_offset = engine_->with_dynamic_shape() ? 1 : 0;
+    nvinfer1::ILayer* layer = nullptr;
+    nvinfer1::IShuffleLayer* expand_layer = nullptr;
+    nvinfer1::IShuffleLayer* squeeze_layer = nullptr;
+
+    auto x_dim = X->getDimensions();
+    if (x_dim.nbDims < 3 + dynamic_shape_offset) {
+      nvinfer1::Dims expand_shape;
+      expand_shape.nbDims = 3 + dynamic_shape_offset;
+      for (int i = 0; i < 3 + dynamic_shape_offset; i++) {
+        if (i < x_dim.nbDims) {
+          expand_shape.d[i] = x_dim.d[i] < 0 ? 0 : x_dim.d[i];
+        } else {
+          expand_shape.d[i] = 1;
+        }
+      }
+      expand_layer = TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *X);
+      expand_layer->setReshapeDimensions(expand_shape);
+      X = expand_layer->getOutput(0);
+    }
+
+    layer = TRT_ENGINE_ADD_LAYER(
+        engine_, Scale, *X, nvinfer1::ScaleMode::kCHANNEL, shift_weights.get(),
+        scale_weights.get(), power_weights.get());
 
     auto output_name = op_desc.Output("Y").front();
     engine_->SetWeights(op_desc.Input("Bias").front(),
                         std::move(combile_bias_tensor));
     engine_->SetWeights(op_desc.Input("Scale").front(),
                         std::move(combile_scale_tensor));
-    RreplenishLayerAndOutput(layer, "pool2d", {output_name}, test_mode);
+    if (x_dim.nbDims < 3 + dynamic_shape_offset) {
+      nvinfer1::Dims squeeze_shape;
+      squeeze_shape.nbDims = x_dim.nbDims;
+      for (int i = 0; i < squeeze_shape.nbDims; i++) {
+        squeeze_shape.d[i] = x_dim.d[i] < 0 ? 0 : x_dim.d[i];
+      }
+      squeeze_layer =
+          TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *(layer->getOutput(0)));
+      squeeze_layer->setReshapeDimensions(squeeze_shape);
+      layer = static_cast<nvinfer1::ILayer*>(squeeze_layer);
+    }
+    RreplenishLayerAndOutput(layer, "batchnorm_add_scale", {output_name},
+                             test_mode);
   }
 };