resolve conflict

paddle/fluid/inference/api/analysis_predictor.cc

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

paddle/fluid/inference/tensorrt/convert/CMakeLists.txt

@@ -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
+                affine_channel_op.cc
                 multiclass_nms_op.cc
                 nearest_interp_op.cc
            DEPS tensorrt_engine tensorrt_plugin operator scope framework_proto op_registry)

paddle/fluid/inference/tensorrt/convert/affine_channel_op.cc

@@ -0,0 +1,94 @@
+/* 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/framework/data_layout.h"
+#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
+
+namespace paddle {
+namespace framework {
+class Scope;
+
+namespace proto {
+class OpDesc;
+}  // namespace proto
+}  // namespace framework
+}  // namespace paddle
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+/*
+ * Affine Channel Op
+ */
+class AffineChannelOpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope, bool test_mode) override {
+    VLOG(3) << "convert a fluid affine_channel op to tensorrt scale nd layer";
+
+    framework::OpDesc op_desc(op, nullptr);
+    std::string input_name = op_desc.Input("X").front();
+    std::string scale_name = op_desc.Input("Scale").front();
+    std::string bias_name = op_desc.Input("Bias").front();
+    std::string output_name = op_desc.Output("Out").front();
+
+    auto input_tensor = engine_->GetITensor(input_name);
+    auto idim = input_tensor->getDimensions();
+
+    auto* scale_v = scope.FindVar(scale_name);
+    auto* scale_t = scale_v->GetMutable<framework::LoDTensor>();
+    float* scale_ptr = engine_->GetWeightCPUData(scale_name, scale_t, false);
+
+    auto* bias_v = scope.FindVar(bias_name);
+    auto* bias_t = bias_v->GetMutable<framework::LoDTensor>();
+    float* bias_ptr = engine_->GetWeightCPUData(bias_name, bias_t, false);
+
+    auto data_layout = framework::StringToDataLayout(
+        BOOST_GET_CONST(std::string, op_desc.GetAttr("data_layout")));
+
+    PADDLE_ENFORCE_EQ(
+        data_layout, framework::DataLayout::kNCHW,
+        platform::errors::InvalidArgument(
+            "TensorRT affine channel converter can only convert NCHW format. "
+            "Other format should be run in fluid mode. Report a bug on github "
+            "issue if you see this line."));
+
+    // tensorrt scalend layer only support spatial dims >= 2,
+    // so nhwc is not availabe (spatial dims == 0)
+    const int channel_axis = engine_->with_dynamic_shape();
+
+    TensorRTEngine::Weight scale_weights{nvinfer1::DataType::kFLOAT,
+                                         static_cast<void*>(scale_ptr),
+                                         (size_t)idim.d[channel_axis]};
+    TensorRTEngine::Weight bias_weights{nvinfer1::DataType::kFLOAT,
+                                        static_cast<void*>(bias_ptr),
+                                        (size_t)idim.d[channel_axis]};
+    TensorRTEngine::Weight power_weights{nvinfer1::DataType::kFLOAT, nullptr,
+                                         0};
+
+    auto layer = TRT_ENGINE_ADD_LAYER(engine_, ScaleNd, *input_tensor,
+                                      nvinfer1::ScaleMode::kCHANNEL,
+                                      bias_weights.get(), scale_weights.get(),
+                                      power_weights.get(), channel_axis);
+
+    RreplenishLayerAndOutput(layer, "affine_channel", {output_name}, test_mode);
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(affine_channel, AffineChannelOpConverter);

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -111,6 +111,7 @@ struct SimpleOpTypeSetTeller : public Teller {
       "flatten2",
       "flatten",
       "gather",
+      "affine_channel",
       "multiclass_nms",
       "nearest_interp",
   };
@@ -196,6 +197,13 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
       if (!with_dynamic_shape || desc.Input("Axis").size() > 0) return false;
     }
 
+    if (op_type == "affine_channel") {
+      if (!desc.HasAttr("data_layout")) return false;
+      auto data_layout = framework::StringToDataLayout(
+          BOOST_GET_CONST(std::string, desc.GetAttr("data_layout")));
+      if (data_layout != framework::DataLayout::kNCHW) return false;
+    }
+
     if (op_type == "multiclass_nms") {
       if (with_dynamic_shape) return false;
       auto* block = desc.Block();
@@ -238,6 +246,7 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
         return false;
       }
     }
+
     if (op_type == "nearest_interp") {
       std::vector<std::string> attrs{"data_layout",   "interp_method",
                                      "align_corners", "scale",
@@ -254,7 +263,6 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
           BOOST_GET_CONST(std::string, desc.GetAttr("interp_method"));
       if (interp_method != "nearest") return false;
     }
-
     if ((*teller)(op_type, desc, use_no_calib_int8)) return true;
   }
   return false;