Merge pull request #10495 from luotao1/refine_relu_test

refine EngineIOConverter, and use io_convert in test_trt_activation_op

paddle/fluid/inference/analysis/dot.h

@@ -21,6 +21,7 @@
 
 #include <glog/logging.h>
 #include <sstream>
+#include <string>
 #include <unordered_map>
 #include <vector>
 

paddle/fluid/inference/engine.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <string>
 #include "paddle/fluid/framework/framework.pb.h"
 
 namespace paddle {
@@ -58,8 +59,8 @@ class EngineBase {
 
 struct Buffer {
   void* buffer{nullptr};               // buffer should be allocated only once.
-  int max_size;                        // buffer allocated space.
-  int size;                            // data size.
+  size_t max_size;                     // buffer allocated space.
+  size_t size;                         // data size.
   DeviceType device{DeviceType::UNK};  // tells which device this buffer is on.
 };
 

paddle/fluid/inference/tensorrt/CMakeLists.txt

 nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto)
 nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader)
 nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine)
-
 add_subdirectory(convert)

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

-nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc op_converter.h DEPS ${FLUID_CORE_MODULES})
-nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc
+nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES})
+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)
 nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor)

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

@@ -23,26 +23,42 @@ namespace tensorrt {
 using platform::is_gpu_place;
 using platform::is_cpu_place;
 
-class DefaultInputConverter : public EngineInputConverter {
+class DefaultIOConverter : public EngineIOConverter {
  public:
-  DefaultInputConverter() {}
+  DefaultIOConverter() {}
   // NOTE out is GPU memory.
   virtual void operator()(const LoDTensor& in, void* out,
                           size_t max_size) override {
     PADDLE_ENFORCE(out != nullptr);
-    PADDLE_ENFORCE_LE(in.memory_size(), max_size);
+    PADDLE_ENFORCE(stream_ != nullptr);
     const auto& place = in.place();
+    size_t size = in.memory_size();
+    PADDLE_ENFORCE_LE(size, max_size);
     if (is_cpu_place(place)) {
-      PADDLE_ENFORCE(stream_ != nullptr);
-      PADDLE_ENFORCE_EQ(0,
-                        cudaMemcpyAsync(out, in.data<float>(), in.memory_size(),
-                                        cudaMemcpyHostToDevice, *stream_));
-
+      PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out, in.data<float>(), size,
+                                           cudaMemcpyHostToDevice, *stream_));
     } else if (is_gpu_place(place)) {
-      PADDLE_ENFORCE_EQ(0,
-                        cudaMemcpyAsync(out, in.data<float>(), in.memory_size(),
-                                        cudaMemcpyHostToHost, *stream_));
-
+      PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out, in.data<float>(), size,
+                                           cudaMemcpyDeviceToDevice, *stream_));
+    } else {
+      PADDLE_THROW("Unknown device for converter");
+    }
+    cudaStreamSynchronize(*stream_);
+  }
+  // NOTE in is GPU memory.
+  virtual void operator()(const void* in, LoDTensor* out,
+                          size_t max_size) override {
+    PADDLE_ENFORCE(in != nullptr);
+    PADDLE_ENFORCE(stream_ != nullptr);
+    const auto& place = out->place();
+    size_t size = out->memory_size();
+    PADDLE_ENFORCE_LE(size, max_size);
+    if (is_cpu_place(place)) {
+      PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data<float>(), in, size,
+                                           cudaMemcpyDeviceToHost, *stream_));
+    } else if (is_gpu_place(place)) {
+      PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data<float>(), in, size,
+                                           cudaMemcpyDeviceToDevice, *stream_));
     } else {
       PADDLE_THROW("Unknown device for converter");
     }
@@ -50,7 +66,8 @@ class DefaultInputConverter : public EngineInputConverter {
   }
 };
 
-REGISTER_TENSORRT_INPUT_CONVERTER(default, DefaultInputConverter);
+// fluid LodTensor <-> tensorrt ITensor
+REGISTER_TENSORRT_IO_CONVERTER(default, DefaultIOConverter);
 
 }  // namespace tensorrt
 }  // namespace inference

paddle/fluid/inference/tensorrt/convert/io_converter.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <string>
 #include <unordered_map>
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/inference/utils/singleton.h"
@@ -25,43 +26,57 @@ namespace tensorrt {
 using framework::LoDTensor;
 
 /*
- * Convert Input from Fluid to an Engine.
- * TensorRT's ITensor follows row major, NCHW. Fluid is also row major, so in
- * most cases just need to copy the data.
+ * Convert Input from Fluid to TensorRT Engine.
+ * Convert Output from TensorRT Engine to Fluid.
+ *
+ * Note that TensorRT's ITensor follows row major, NCHW. Fluid is also row
+ * major,
+ * so in the default case just need to copy the data.
  */
-class EngineInputConverter {
+class EngineIOConverter {
  public:
-  EngineInputConverter() {}
+  EngineIOConverter() {}
 
   virtual void operator()(const LoDTensor& in, void* out, size_t max_size) {}
+  virtual void operator()(const void* in, LoDTensor* out, size_t max_size) {}
 
   void SetStream(cudaStream_t* stream) { stream_ = stream; }
 
-  static void Run(const std::string& in_op_type, const LoDTensor& in, void* out,
-                  size_t max_size, cudaStream_t* stream) {
+  static void ConvertInput(const std::string& op_type, const LoDTensor& in,
+                           void* out, size_t max_size, cudaStream_t* stream) {
     PADDLE_ENFORCE(stream != nullptr);
-    auto* converter = Registry<EngineInputConverter>::Lookup(
-        in_op_type, "default" /* default_type */);
+    auto* converter = Registry<EngineIOConverter>::Lookup(
+        op_type, "default" /* default_type */);
     PADDLE_ENFORCE_NOT_NULL(converter);
     converter->SetStream(stream);
     (*converter)(in, out, max_size);
   }
 
-  virtual ~EngineInputConverter() {}
+  static void ConvertOutput(const std::string& op_type, const void* in,
+                            LoDTensor* out, size_t max_size,
+                            cudaStream_t* stream) {
+    PADDLE_ENFORCE(stream != nullptr);
+    auto* converter = Registry<EngineIOConverter>::Lookup(
+        op_type, "default" /* default_type */);
+    PADDLE_ENFORCE_NOT_NULL(converter);
+    converter->SetStream(stream);
+    (*converter)(in, out, max_size);
+  }
+
+  virtual ~EngineIOConverter() {}
 
  protected:
   cudaStream_t* stream_{nullptr};
 };
 
+#define REGISTER_TENSORRT_IO_CONVERTER(op_type__, Converter__)        \
+  struct trt_io_##op_type__##_converter {                             \
+    trt_io_##op_type__##_converter() {                                \
+      Registry<EngineIOConverter>::Register<Converter__>(#op_type__); \
+    }                                                                 \
+  };                                                                  \
+  trt_io_##op_type__##_converter trt_io_##op_type__##_converter__;
+
 }  // namespace tensorrt
 }  // namespace inference
 }  // namespace paddle
-
-#define REGISTER_TENSORRT_INPUT_CONVERTER(in_op_type__, Converter__) \
-  struct trt_input_##in_op_type__##_converter {                      \
-    trt_input_##in_op_type__##_converter() {                         \
-      ::paddle::inference::Registry<EngineInputConverter>::Register< \
-          Converter__>(#in_op_type__);                               \
-    }                                                                \
-  };                                                                 \
-  trt_input_##in_op_type__##_converter trt_input_##in_op_type__##_converter__;

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

@@ -16,6 +16,7 @@ limitations under the License. */
 #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"
@@ -26,7 +27,7 @@ namespace paddle {
 namespace inference {
 namespace tensorrt {
 
-void Compare(float input, float expect) {
+void Compare(const std::string op_type, float input, float expect) {
   framework::Scope scope;
   platform::CUDAPlace place;
   platform::CUDADeviceContext ctx(place);
@@ -35,6 +36,7 @@ void Compare(float input, float expect) {
   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);
@@ -45,14 +47,15 @@ void Compare(float input, float expect) {
   out_tensor->mutable_data<float>(place);
 
   framework::OpDesc op_desc;
-  op_desc.SetType("relu");
+  op_desc.SetType(op_type);
   op_desc.SetInput("X", {"X"});
   op_desc.SetOutput("Out", {"Out"});
 
-  auto relu_op = framework::OpRegistry::CreateOp(*op_desc.Proto());
+  auto op = framework::OpRegistry::CreateOp(*op_desc.Proto());
 
   // run fluid op
-  relu_op->Run(scope, place);
+  op->Run(scope, place);
+  // get fluid output
   std::vector<float> out1;
   framework::TensorToVector(*out_tensor, ctx, &out1);
 
@@ -63,21 +66,28 @@ void Compare(float input, float expect) {
   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();
-  engine->SetInputFromCPU("X", &input, 1 * sizeof(float));
 
-  // run tensorrt op
+  // 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);
-
-  float out2;
-  engine->GetOutputInCPU("Out", &out2, 1 * sizeof(float));
-
-  ASSERT_EQ(out1[0], out2);
+  // 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;
@@ -85,8 +95,8 @@ void Compare(float input, float expect) {
 }
 
 TEST(OpConverter, ConvertRelu) {
-  Compare(1, 1);   // relu(1) = 1
-  Compare(-5, 0);  // relu(-5) = 0
+  Compare("relu", 1, 1);   // relu(1) = 1
+  Compare("relu", -5, 0);  // relu(-5) = 0
 }
 
 }  // namespace tensorrt

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

@@ -12,40 +12,63 @@ 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 <gtest/gtest.h>
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/inference/tensorrt/convert/io_converter.h"
 
-#include <gtest/gtest.h>
-
 namespace paddle {
 namespace inference {
 namespace tensorrt {
 
-class EngineInputConverterTester : public ::testing::Test {
- public:
-  void SetUp() override { tensor.Resize({10, 10}); }
+void IOConverterTester(const platform::DeviceContext& ctx) {
+  cudaStream_t stream;
+  ASSERT_EQ(0, cudaStreamCreate(&stream));
 
-  framework::LoDTensor tensor;
-};
+  // init fluid in_tensor
+  framework::LoDTensor in_tensor;
+  in_tensor.Resize({10, 10});
+  auto place = ctx.GetPlace();
+  in_tensor.mutable_data<float>(place);
+  std::vector<float> init;
+  for (int64_t i = 0; i < 10 * 10; ++i) {
+    init.push_back(i);
+  }
+  framework::TensorFromVector(init, ctx, &in_tensor);
 
-TEST_F(EngineInputConverterTester, DefaultCPU) {
+  // init tensorrt buffer
   void* buffer;
-  tensor.mutable_data<float>(platform::CPUPlace());
-  ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0);
+  size_t size = in_tensor.memory_size();
+  ASSERT_EQ(cudaMalloc(&buffer, size), 0);
 
-  cudaStream_t stream;
-  EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(),
-                            &stream);
+  // convert fluid in_tensor to tensorrt buffer
+  EngineIOConverter::ConvertInput("test", in_tensor, buffer, size, &stream);
+
+  // convert tensorrt buffer to fluid out_tensor
+  framework::LoDTensor out_tensor;
+  out_tensor.Resize({10, 10});
+  out_tensor.mutable_data<float>(place);
+  EngineIOConverter::ConvertOutput("test", buffer, &out_tensor, size, &stream);
+
+  // compare in_tensor and out_tensor
+  std::vector<float> result;
+  framework::TensorToVector(out_tensor, ctx, &result);
+  EXPECT_EQ(init.size(), result.size());
+  for (size_t i = 0; i < init.size(); i++) {
+    EXPECT_EQ(init[i], result[i]);
+  }
+  cudaStreamDestroy(stream);
 }
 
-TEST_F(EngineInputConverterTester, DefaultGPU) {
-  void* buffer;
-  tensor.mutable_data<float>(platform::CUDAPlace());
-  ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0);
+TEST(EngineIOConverterTester, DefaultCPU) {
+  platform::CPUPlace place;
+  platform::CPUDeviceContext ctx(place);
+  IOConverterTester(ctx);
+}
 
-  cudaStream_t stream;
-  EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(),
-                            &stream);
+TEST(EngineIOConverterTester, DefaultGPU) {
+  platform::CUDAPlace place;
+  platform::CUDADeviceContext ctx(place);
+  IOConverterTester(ctx);
 }
 
 }  // namespace tensorrt