From 82527696e779dfa9dedd24b24c9d5f026ffb46f3 Mon Sep 17 00:00:00 2001 From: nhzlx Date: Mon, 23 Jul 2018 11:42:57 +0000 Subject: [PATCH] 1. we delelte mul op, 2.modify fc and action op 3. modify the test inferface --- .../inference/tensorrt/convert/CMakeLists.txt | 4 +- .../fluid/inference/tensorrt/convert/fc_op.cc | 17 +++--- .../inference/tensorrt/convert/mul_op.cc | 53 ------------------- .../tensorrt/convert/test_activation_op.cc | 4 +- .../inference/tensorrt/convert/test_fc_op.cc | 11 ++-- .../inference/tensorrt/convert/test_mul_op.cc | 49 ----------------- .../inference/tensorrt/convert/ut_helper.h | 8 +-- 7 files changed, 21 insertions(+), 125 deletions(-) delete mode 100644 paddle/fluid/inference/tensorrt/convert/mul_op.cc delete mode 100644 paddle/fluid/inference/tensorrt/convert/test_mul_op.cc diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt index 748f5a084e8..36bd3904e07 100644 --- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -1,14 +1,12 @@ # Add TRT tests nv_library(tensorrt_converter - SRCS mul_op.cc conv2d_op.cc fc_op.cc + SRCS conv2d_op.cc fc_op.cc DEPS tensorrt_engine mul_op) nv_test(test_op_converter SRCS test_op_converter.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine tensorrt_converter) 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 diff --git a/paddle/fluid/inference/tensorrt/convert/fc_op.cc b/paddle/fluid/inference/tensorrt/convert/fc_op.cc index bb603efaf30..c043c6aa3a4 100644 --- a/paddle/fluid/inference/tensorrt/convert/fc_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/fc_op.cc @@ -32,13 +32,13 @@ void Reorder2(nvinfer1::DimsHW shape, const T* idata, nvinfer1::DimsHW istrides, for (int h = 0; h < shape.h(); ++h) { for (int w = 0; w < shape.w(); ++w) { odata[h * ostrides.h() + w * ostrides.w()] = - idata[h * ostrides.h() + w * ostrides.w()]; + idata[h * istrides.h() + w * istrides.w()]; } } } - +// indata c * k // Reorder the data layout from CK to KC. -void ReorderCKtoKC(TensorRTEngine::Weight& iweights, +void ReorderCKtoKC(const TensorRTEngine::Weight& iweights, TensorRTEngine::Weight* oweights) { int c = iweights.dims[0]; int k = iweights.dims[1]; @@ -79,9 +79,8 @@ class FcOpConverter : public OpConverter { framework::LoDTensor tmp; tmp.Resize(Y_t->dims()); - memcpy(tmp.mutable_data(platform::CPUPlace()), Y_t->data(), - Y_t->dims()[0] * Y_t->dims()[1]); - + memcpy(tmp.mutable_data(platform::CPUPlace()), weight_data, + Y_t->dims()[0] * Y_t->dims()[1] * sizeof(float)); TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT, static_cast(weight_data), Y_t->memory_size() / sizeof(float)}; @@ -93,7 +92,7 @@ class FcOpConverter : public OpConverter { // The data layout of TRT FC layer's weight is different from fluid's FC, // need to reorder the elements. - ReorderCKtoKC(tmp_weight, &weight); + ReorderCKtoKC(weight, &tmp_weight); // Currently, the framework can only handle one fluid op -> one TRT layer, // but fc fuses `mul` and `bias` (2 fluid ops), so here is a trick, just @@ -103,7 +102,7 @@ class FcOpConverter : public OpConverter { auto* layer = TRT_ENGINE_ADD_LAYER(engine_, FullyConnected, *const_cast(X), - n_output, weight.get(), bias.get()); + n_output, tmp_weight.get(), bias.get()); auto output_name = op_desc.Output("Out").front(); engine_->SetITensor(output_name, layer->getOutput(0)); @@ -117,5 +116,5 @@ class FcOpConverter : public OpConverter { } // namespace inference } // namespace paddle -REGISTER_TRT_OP_CONVERTER(fc, FcOpConverter); +REGISTER_TRT_OP_CONVERTER(mul, FcOpConverter); USE_OP(mul); diff --git a/paddle/fluid/inference/tensorrt/convert/mul_op.cc b/paddle/fluid/inference/tensorrt/convert/mul_op.cc deleted file mode 100644 index 3c342957360..00000000000 --- a/paddle/fluid/inference/tensorrt/convert/mul_op.cc +++ /dev/null @@ -1,53 +0,0 @@ -/* 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/inference/tensorrt/convert/op_converter.h" - -namespace paddle { -namespace inference { -namespace tensorrt { - -/* - * MulOp, IMatrixMultiplyLayer in TRT. This Layer doesn't has weights. - */ -class MulOpConverter : public OpConverter { - public: - void operator()(const framework::proto::OpDesc& op, - const framework::Scope& scope, bool test_mode) override { - VLOG(4) << "convert a fluid mul op to tensorrt mul layer without bias"; - - framework::OpDesc op_desc(op, nullptr); - // Declare inputs - auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]); - auto* input2 = engine_->GetITensor(op_desc.Input("Y")[0]); - // Both the input1 and input2 do not need transpose. - auto* layer = TRT_ENGINE_ADD_LAYER( - engine_, MatrixMultiply, *const_cast(input1), false, - *const_cast(input2), false); - - auto output_name = op_desc.Output("Out")[0]; - engine_->SetITensor(output_name, layer->getOutput(0)); - if (test_mode) { // the test framework can not determine which is the - // output, so place the declaration inside. - engine_->DeclareOutput(output_name); - } - } -}; - -} // namespace tensorrt -} // namespace inference -} // namespace paddle - -USE_OP(mul); -REGISTER_TRT_OP_CONVERTER(mul, MulOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc index 0a02a7bebf9..743c20adc79 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc @@ -23,7 +23,7 @@ namespace tensorrt { TEST(ReluOpConverter, main) { framework::Scope scope; std::unordered_set parameters; - TRTConvertValidation validator(10, parameters, scope, 1000); + TRTConvertValidation validator(1, parameters, scope, 1000); validator.DeclInputVar("relu-X", nvinfer1::Dims2(10, 6)); validator.DeclOutputVar("relu-Out", nvinfer1::Dims2(10, 6)); @@ -37,7 +37,7 @@ TEST(ReluOpConverter, main) { validator.SetOp(*desc.Proto()); LOG(INFO) << "execute"; - validator.Execute(10); + validator.Execute(1); } } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc b/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc index a30253072ac..c76f494f8cb 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc @@ -23,11 +23,12 @@ namespace tensorrt { TEST(fc_op, test) { std::unordered_set parameters({"mul-Y"}); framework::Scope scope; - TRTConvertValidation validator(20, parameters, scope, 1000); + TRTConvertValidation validator(1, parameters, scope, 1000); - validator.DeclInputVar("mul-X", nvinfer1::Dims4(8, 3, 1, 1)); - validator.DeclParamVar("mul-Y", nvinfer1::Dims2(3, 2)); - validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(8, 2)); + validator.DeclInputVar("mul-X", nvinfer1::Dims4(1, 10, 1, 1)); + validator.DeclParamVar("mul-Y", nvinfer1::Dims2(10, 2)); + // validator.DeclParamVar("mul-Y", nvinfer1::Dims2(8, 2)); + validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(1, 2)); // Prepare Op description framework::OpDesc desc; @@ -38,7 +39,7 @@ TEST(fc_op, test) { validator.SetOp(*desc.Proto()); - validator.Execute(10); + validator.Execute(1); } } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc b/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc deleted file mode 100644 index 1ce1130e5d6..00000000000 --- a/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc +++ /dev/null @@ -1,49 +0,0 @@ -/* 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 -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" - -namespace paddle { -namespace inference { -namespace tensorrt { - -TEST(MulOpConverter, main) { - framework::Scope scope; - std::unordered_set parameters; - TRTConvertValidation validator(10, parameters, scope, 1000); - validator.DeclInputVar("mul-X", nvinfer1::Dims2(10, 6)); - validator.DeclInputVar("mul-Y", nvinfer1::Dims2(6, 10)); - validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(10, 10)); - - // Prepare Op description - framework::OpDesc desc; - desc.SetType("mul"); - desc.SetInput("X", {"mul-X"}); - desc.SetInput("Y", {"mul-Y"}); - desc.SetOutput("Out", {"mul-Out"}); - - LOG(INFO) << "set OP"; - validator.SetOp(*desc.Proto()); - LOG(INFO) << "execute"; - - validator.Execute(10); -} - -} // namespace tensorrt -} // namespace inference -} // namespace paddle - -USE_OP(mul); diff --git a/paddle/fluid/inference/tensorrt/convert/ut_helper.h b/paddle/fluid/inference/tensorrt/convert/ut_helper.h index 3b1f531adc5..2e6c895b2e6 100644 --- a/paddle/fluid/inference/tensorrt/convert/ut_helper.h +++ b/paddle/fluid/inference/tensorrt/convert/ut_helper.h @@ -39,7 +39,7 @@ namespace tensorrt { float random(float low, float high) { static std::random_device rd; static std::mt19937 mt(rd()); - std::uniform_real_distribution dist(1.0, 10.0); + std::uniform_real_distribution dist(low, high); return dist(mt); } @@ -49,6 +49,7 @@ void RandomizeTensor(framework::LoDTensor* tensor, const platform::Place& place, size_t num_elements = analysis::AccuDims(dims, dims.size()); PADDLE_ENFORCE_GT(num_elements, 0); auto* data = tensor->mutable_data(place); + for (size_t i = 0; i < num_elements; i++) { *(data + i) = random(0., 1.); } @@ -68,7 +69,7 @@ class TRTConvertValidation { int workspace_size = 1 << 10) : parameters_(parameters), scope_(scope) { // create engine. - engine_.reset(new TensorRTEngine(10, 1 << 10, &stream_)); + engine_.reset(new TensorRTEngine(batch_size, workspace_size, &stream_)); engine_->InitNetwork(); PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0); @@ -142,8 +143,7 @@ class TRTConvertValidation { for (const auto& output : op_desc_->OutputArgumentNames()) { std::vector fluid_out; std::vector trt_out(output_space_size); - engine_->GetOutputInCPU(output, &trt_out[0], - output_space_size * sizeof(float)); + engine_->GetOutputInCPU(output, &trt_out[0]); cudaStreamSynchronize(*engine_->stream()); auto* var = scope_.FindVar(output); -- GitLab