rewrite unittest of trt_activation_op

cmake/configure.cmake

@@ -92,6 +92,9 @@ if(WITH_GPU)
         if(${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
             message(FATAL_ERROR "TensorRT needs CUDNN >= 7.0 to compile")
         endif()
+        if(${TENSORRT_MAJOR_VERSION} VERSION_LESS 4)
+            message(FATAL_ERROR "Paddle needs TensorRT >= 4.0 to compile")
+        endif()
         include_directories(${TENSORRT_INCLUDE_DIR})
     endif()
 elseif(WITH_AMD_GPU)

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

@@ -1,12 +1,9 @@
 # Add TRT tests
 nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine)
-# This test is not stable
-# See https://paddleci.ngrok.io/viewLog.html?tab=buildLog&buildTypeId=Paddle_PrCi2&buildId=36834&_focus=8828 
-#nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc io_converter.cc
-#    DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine
-#    SERIAL)
 nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor)
 nv_test(test_trt_mul_op SRCS test_mul_op.cc mul_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL)
 nv_test(test_trt_fc_op SRCS test_fc_op.cc fc_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL)
+nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc
+        DEPS ${FLUID_CORE_MODULES} tensorrt_engine activation_op SERIAL)

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

@@ -21,7 +21,8 @@ namespace tensorrt {
 class ReluOpConverter : public OpConverter {
  public:
   ReluOpConverter() {}
-  void operator()(const framework::proto::OpDesc& op) override {
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope) override {
     // Here the two nullptr looks strange, that's because the
     // framework::OpDesc's constructor is strange.
     framework::OpDesc op_desc(op, nullptr);
@@ -32,7 +33,7 @@ class ReluOpConverter : public OpConverter {
     nvinfer1::IActivationLayer* layer = TRT_ENGINE_ADD_LAYER(
         engine_, Activation, *const_cast<nvinfer1::ITensor*>(input_tensor),
         nvinfer1::ActivationType::kRELU);
-    engine_->SetITensor(op_desc.Output("Out")[0], layer->getOutput(0));
+    engine_->DeclareOutput(layer, 0, op_desc.Output("Out")[0]);
   }
 };
 

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

@@ -13,94 +13,35 @@ See the License for the specific language governing permissions and
    limitations under the License. */
 
 #include <gtest/gtest.h>
-#include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/framework/program_desc.h"
-#include "paddle/fluid/inference/tensorrt/convert/io_converter.h"
-#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
-#include "paddle/fluid/platform/device_context.h"
-#include "paddle/fluid/platform/place.h"
-
-USE_OP(relu);
+#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h"
 
 namespace paddle {
 namespace inference {
 namespace tensorrt {
 
-void Compare(const std::string op_type, float input, float expect) {
+TEST(ReluOpConverter, main) {
   framework::Scope scope;
-  platform::CUDAPlace place;
-  platform::CUDADeviceContext ctx(place);
-
-  // init fluid op and variable
-  auto x_var = scope.Var("X");
-  auto x_tensor = x_var->GetMutable<framework::LoDTensor>();
-  x_tensor->Resize({1, 1});
-  x_tensor->mutable_data<float>(place);
-  std::vector<float> init;
-  init.push_back(input);
-  framework::TensorFromVector(init, ctx, x_tensor);
-
-  auto out_var = scope.Var("Out");
-  auto out_tensor = out_var->GetMutable<framework::LoDTensor>();
-  out_tensor->Resize({1, 1});
-  out_tensor->mutable_data<float>(place);
-
-  framework::OpDesc op_desc;
-  op_desc.SetType(op_type);
-  op_desc.SetInput("X", {"X"});
-  op_desc.SetOutput("Out", {"Out"});
-
-  auto op = framework::OpRegistry::CreateOp(*op_desc.Proto());
-
-  // run fluid op
-  op->Run(scope, place);
-  // get fluid output
-  std::vector<float> out1;
-  framework::TensorToVector(*out_tensor, ctx, &out1);
-
-  // init tensorrt op
-  cudaStream_t stream;
-  ASSERT_EQ(0, cudaStreamCreate(&stream));
-  TensorRTEngine* engine = new TensorRTEngine(1, 1 << 10, &stream);
-  engine->InitNetwork();
-  engine->DeclareInput("X", nvinfer1::DataType::kFLOAT,
-                       nvinfer1::DimsCHW{1, 1, 1});
-  // convert op
-  OpConverter op_converter;
-  op_converter.ConvertOp(*op_desc.Proto(), engine);
-
-  engine->DeclareOutput("Out");
-  engine->FreezeNetwork();
-
-  // convert LoDTensor to ITensor
-  size_t size = x_tensor->memory_size();
-  EngineIOConverter::ConvertInput(op_type, *x_tensor,
-                                  engine->buffer("X").buffer, size, &stream);
-  // run tensorrt Outp
-  engine->Execute(1);
-  // convert ITensor to LoDTensor
-  EngineIOConverter::ConvertOutput(op_type, engine->buffer("Out").buffer,
-                                   out_tensor, size, &stream);
-  // get tensorrt output
-  std::vector<float> out2;
-  framework::TensorToVector(*out_tensor, ctx, &out2);
-
-  // compare
-  ASSERT_EQ(out1[0], out2[0]);
-  ASSERT_EQ(out1[0], expect);
-
-  delete engine;
-  cudaStreamDestroy(stream);
-}
-
-TEST(OpConverter, ConvertRelu) {
-  Compare("relu", 1, 1);   // relu(1) = 1
-  Compare("relu", -5, 0);  // relu(-5) = 0
+  std::unordered_set<std::string> parameters;
+  TRTConvertValidation validator(10, parameters, scope, 1000);
+  validator.DeclInputVar("relu-X", nvinfer1::Dims2(10, 6));
+  validator.DeclOutputVar("relu-Out", nvinfer1::Dims2(10, 6));
+
+  // Prepare Op description
+  framework::OpDesc desc;
+  desc.SetType("relu");
+  desc.SetInput("X", {"relu-X"});
+  desc.SetOutput("Out", {"relu-Out"});
+
+  LOG(INFO) << "set OP";
+  validator.SetOp(*desc.Proto());
+  LOG(INFO) << "execute";
+
+  validator.Execute(10);
 }
 
 }  // namespace tensorrt
 }  // namespace inference
 }  // namespace paddle
 
-USE_OP(activation);
+USE_OP(relu);