diff --git a/paddle/fluid/inference/tensorrt/CMakeLists.txt b/paddle/fluid/inference/tensorrt/CMakeLists.txt index c8b656394b403c4965e01e96c9215d9406091907..288789d6e484100820c937e6081701f1e9245706 100644 --- a/paddle/fluid/inference/tensorrt/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/CMakeLists.txt @@ -1,5 +1,4 @@ nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader) nv_test(test_tensorrt_engine SRCS test_engine.cc engine.cc DEPS dynload_cuda) -nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor) set(ENGINE_FILE ${CMAKE_CURRENT_SOURCE_DIR}/engine.cc) add_subdirectory(convert) diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt index 572e29515f89685c528023cf6569ef8e559d6874..3c5909c0be1c690d5148ecfb32b1b6c2dd6f3211 100644 --- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -1,3 +1,4 @@ -nv_test(test_tensorrt_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES}) -nv_test(test_tensorrt_activation_op SRCS test_activation_op.cc ${ENGINE_FILE} 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 ${ENGINE_FILE} activation_op.cc DEPS ${FLUID_CORE_MODULES} activation_op) +nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor) diff --git a/paddle/fluid/inference/tensorrt/io_converter.cc b/paddle/fluid/inference/tensorrt/convert/io_converter.cc similarity index 93% rename from paddle/fluid/inference/tensorrt/io_converter.cc rename to paddle/fluid/inference/tensorrt/convert/io_converter.cc index 2baac96c26453af7e70e541d80b437df3d5c2657..32e8631fde3f748669d2008b4a060455a37e154e 100644 --- a/paddle/fluid/inference/tensorrt/io_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/io_converter.cc @@ -12,7 +12,7 @@ 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/inference/tensorrt/io_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/io_converter.h" #include #include "paddle/fluid/platform/enforce.h" @@ -50,7 +50,7 @@ class DefaultInputConverter : public EngineInputConverter { } }; -REGISTER_TENSORRT_INPUT_CONVERTER(mul, DefaultInputConverter); +REGISTER_TENSORRT_INPUT_CONVERTER(default, DefaultInputConverter); } // namespace tensorrt } // namespace inference diff --git a/paddle/fluid/inference/tensorrt/io_converter.h b/paddle/fluid/inference/tensorrt/convert/io_converter.h similarity index 95% rename from paddle/fluid/inference/tensorrt/io_converter.h rename to paddle/fluid/inference/tensorrt/convert/io_converter.h index 6ea61cbbac05f106f736b7d6a13912157c5ef48c..8972dae92be2c2d261a13c48d98e675f64e51d31 100644 --- a/paddle/fluid/inference/tensorrt/io_converter.h +++ b/paddle/fluid/inference/tensorrt/convert/io_converter.h @@ -40,7 +40,8 @@ class EngineInputConverter { static void Run(const std::string& in_op_type, const LoDTensor& in, void* out, size_t max_size, cudaStream_t* stream) { PADDLE_ENFORCE(stream != nullptr); - auto* converter = Registry::Lookup(in_op_type); + auto* converter = Registry::Lookup( + in_op_type, "default" /* default_type */); PADDLE_ENFORCE_NOT_NULL(converter); converter->SetStream(stream); (*converter)(in, out, max_size); diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h index f8ca219bb83c6db3ba640464233c6cbd337926a6..77c788550b2c7df1f483b926661789b2a54d8fff 100644 --- a/paddle/fluid/inference/tensorrt/convert/op_converter.h +++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h @@ -19,6 +19,7 @@ limitations under the License. */ #include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/inference/tensorrt/engine.h" +#include "paddle/fluid/inference/utils/singleton.h" namespace paddle { namespace inference { @@ -32,34 +33,23 @@ class OpConverter { OpConverter() {} virtual void operator()(const framework::OpDesc& op) {} - void Execute(const framework::OpDesc& op, TensorRTEngine* engine) { + void Run(const framework::OpDesc& op, TensorRTEngine* engine) { std::string type = op.Type(); - auto it = converters_.find(type); - PADDLE_ENFORCE(it != converters_.end(), "no OpConverter for optype [%s]", - type); - it->second->SetEngine(engine); - (*it->second)(op); - } - - static OpConverter& Global() { - static auto* x = new OpConverter; - return *x; - } - - template - void Register(const std::string& key) { - converters_[key] = new T; + auto* it = Registry::Lookup(type); + PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", type); + it->SetEngine(engine); + (*it)(op); } // convert fluid op to tensorrt layer void ConvertOp(const framework::OpDesc& op, TensorRTEngine* engine) { - OpConverter::Global().Execute(op, engine); + OpConverter::Run(op, engine); } // convert fluid block to tensorrt network void ConvertBlock(const framework::BlockDesc& block, TensorRTEngine* engine) { for (auto op : block.AllOps()) { - OpConverter::Global().Execute(*op, engine); + OpConverter::Run(*op, engine); } } @@ -78,12 +68,12 @@ class OpConverter { framework::Scope* scope_{nullptr}; }; -#define REGISTER_TRT_OP_CONVERTER(op_type__, Converter__) \ - struct trt_##op_type__##_converter { \ - trt_##op_type__##_converter() { \ - OpConverter::Global().Register(#op_type__); \ - } \ - }; \ +#define REGISTER_TRT_OP_CONVERTER(op_type__, Converter__) \ + struct trt_##op_type__##_converter { \ + trt_##op_type__##_converter() { \ + Registry::Register(#op_type__); \ + } \ + }; \ trt_##op_type__##_converter trt_##op_type__##_converter__; } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc index 0f390bee1f5967547ea21e6d42a1b4300de58d04..23e3435c21725328d3765fae0d158a83ac21478b 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc @@ -26,7 +26,7 @@ namespace paddle { namespace inference { namespace tensorrt { -void compare(float input, float expect) { +void Compare(float input, float expect) { framework::Scope scope; platform::CUDAPlace place; platform::CUDADeviceContext ctx(place); @@ -85,8 +85,8 @@ void compare(float input, float expect) { } TEST(OpConverter, ConvertRelu) { - compare(1, 1); // relu(1) = 1 - compare(-5, 0); // relu(-5) = 0 + Compare(1, 1); // relu(1) = 1 + Compare(-5, 0); // relu(-5) = 0 } } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/test_io_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc similarity index 87% rename from paddle/fluid/inference/tensorrt/test_io_converter.cc rename to paddle/fluid/inference/tensorrt/convert/test_io_converter.cc index 365e9366862bee25c70dba0cdd92f318ab3ee90f..afcc516e6b76d58e37ce0e60746704cf3933fac7 100644 --- a/paddle/fluid/inference/tensorrt/test_io_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/inference/tensorrt/io_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/io_converter.h" #include @@ -34,7 +34,7 @@ TEST_F(EngineInputConverterTester, DefaultCPU) { ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0); cudaStream_t stream; - EngineInputConverter::Run("mul", tensor, buffer, tensor.memory_size(), + EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(), &stream); } @@ -44,7 +44,7 @@ TEST_F(EngineInputConverterTester, DefaultGPU) { ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0); cudaStream_t stream; - EngineInputConverter::Run("mul", tensor, buffer, tensor.memory_size(), + EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(), &stream); } diff --git a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc index 5c5ac10394d3cb76ac4b8409c7fa79eb5e9a90b2..aa5fb726f1129eda65a6f39791330b795aad660d 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc @@ -20,7 +20,7 @@ namespace paddle { namespace inference { namespace tensorrt { -TEST(BlockConverter, ConvertBlock) { +TEST(OpConverter, ConvertBlock) { framework::ProgramDesc prog; auto* block = prog.MutableBlock(0); auto* mul_op = block->AppendOp(); diff --git a/paddle/fluid/inference/utils/singleton.h b/paddle/fluid/inference/utils/singleton.h index f05921067c45f156319375b613f51101cfda8e90..cfb89e704457a11a3cd6e89dba5efad5acae0bce 100644 --- a/paddle/fluid/inference/utils/singleton.h +++ b/paddle/fluid/inference/utils/singleton.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include #include "paddle/fluid/platform/enforce.h" @@ -49,9 +50,15 @@ struct Registry { items_[name] = new ItemChild; } - static ItemParent* Lookup(const std::string& name) { + static ItemParent* Lookup(const std::string& name, + const std::string& default_name = "") { auto it = items_.find(name); - if (it == items_.end()) return nullptr; + if (it == items_.end()) { + if (default_name == "") + return nullptr; + else + return items_.find(default_name)->second; + } return it->second; }