diff --git a/real-esrgan/CMakeLists.txt b/real-esrgan/CMakeLists.txt new file mode 100644 index 0000000..f313a7a --- /dev/null +++ b/real-esrgan/CMakeLists.txt @@ -0,0 +1,44 @@ +cmake_minimum_required(VERSION 2.6) + +project(real-esrgan) + +add_definitions(-std=c++11) +add_definitions(-DAPI_EXPORTS) +option(CUDA_USE_STATIC_CUDA_RUNTIME OFF) +set(CMAKE_CXX_STANDARD 11) +set(CMAKE_BUILD_TYPE Debug) + +find_package(CUDA REQUIRED) + +if(WIN32) +enable_language(CUDA) +endif(WIN32) + +include_directories(${PROJECT_SOURCE_DIR}/include) +# include and link dirs of cuda and tensorrt, you need adapt them if yours are different +# cuda +include_directories(/usr/local/cuda/include) +link_directories(/usr/local/cuda/lib64) +# tensorrt +include_directories(/usr/include/x86_64-linux-gnu/) +link_directories(/usr/lib/x86_64-linux-gnu/) + +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -Wall -Ofast -g -Wfatal-errors -D_MWAITXINTRIN_H_INCLUDED") +cuda_add_library(myplugins SHARED preprocess.cu postprocess.cu) +target_link_libraries(myplugins nvinfer cudart) + +find_package(OpenCV) +include_directories(${OpenCV_INCLUDE_DIRS}) + +cuda_add_executable(real-esrgan real-esrgan.cpp) + +target_link_libraries(real-esrgan nvinfer) +target_link_libraries(real-esrgan cudart) +target_link_libraries(real-esrgan myplugins) +target_link_libraries(real-esrgan ${OpenCV_LIBS}) + +if(UNIX) +add_definitions(-O2 -pthread) +endif(UNIX) + + diff --git a/real-esrgan/README.md b/real-esrgan/README.md new file mode 100644 index 0000000..ac2de5a --- /dev/null +++ b/real-esrgan/README.md @@ -0,0 +1,59 @@ +# Real-ESRGAN +The Pytorch implementation is [real-esrgan](https://github.com/xinntao/Real-ESRGAN). + +## Config +- Input shape(**INPUT_H**, **INPUT_W**, **INPUT_C**) defined in real-esrgan.cpp +- GPU id(**DEVICE**) can be selected by the macro in real-esrgan.cpp +- **BATCH_SIZE** can be selected by the macro in real-esrgan.cpp +- FP16/FP32 can be selected by **PRECISION_MODE** in real-esrgan.cpp +- The example result can be visualized by **VISUALIZATION**. + +## How to Run, real-esrgan as example + +0. prepare test image +- download : [OST_009.png](https://drive.google.com/file/d/1KAyAiQ8qHc5jSBkk2Uft2LfIhzi9XSyH/view?usp=sharing) + +``` +cd {tensorrtx}/real-esrgan/ +mkdir sample +cp ~/Download/OST_009.png {tensorrtx}/real-esrgan/sample +``` + +1. generate .wts from pytorch with .pt, or download .wts from model zoo + +``` +git clone https://github.com/xinntao/Real-ESRGAN.git +cd Real-ESRGAN +pip install basicsr +pip install facexlib +pip install gfpgan +pip install -r requirements.txt +python setup.py develop + +// download https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth +cp ~/RealESRGAN_x4plus.pth {xinntao}/Real-ESRGAN/experiments/pretrained_models + +cp {tensorrtx}/Real-ESRGAN/gen_wts.py {xinntao}/Real-ESRGAN +cd {xinntao}/Real-ESRGAN +python gen_wts.py +// a file 'real-esrgan.wts' will be generated. +``` + +2. build tensorrtx/real-esrgan and run + +``` +cd {tensorrtx}/real-esrgan/ +mkdir build +cd build +cp {xinntao}/Real-ESRGAN/real-esrgan.wts {tensorrtx}/real-esrgan/build +cmake .. +make +sudo ./real-esrgan -s [.wts] [.engine] // serialize model to plan file +sudo ./real-esrgan -d [.engine] [image folder] // deserialize and run inference, the images in [image folder] will be processed. +// For example +// sudo ./real-esrgan -s ./real-esrgan.wts ./real-esrgan_f32.engine +// sudo ./real-esrgan -d ./real-esrgan_f32.engine ../samples + +``` + +3. check the images generated, as follows. _OST_009.png diff --git a/real-esrgan/common.hpp b/real-esrgan/common.hpp new file mode 100644 index 0000000..fd157b2 --- /dev/null +++ b/real-esrgan/common.hpp @@ -0,0 +1,137 @@ +#ifndef REAL_ESRGAN_COMMON_H_ +#define REAL_ESRGAN_COMMON_H_ + +#include +#include +#include +#include +#include +#include "NvInfer.h" + +using namespace nvinfer1; + +// TensorRT weight files have a simple space delimited format: +// [type] [size] +std::map loadWeights(const std::string file) { + std::cout << "Loading weights: " << file << std::endl; + std::map weightMap; + + // Open weights file + std::ifstream input(file); + assert(input.is_open() && "Unable to load weight file. please check if the .wts file path is right!!!!!!"); + + // Read number of weight blobs + int32_t count; + input >> count; + assert(count > 0 && "Invalid weight map file."); + + while (count--) + { + Weights wt{ DataType::kFLOAT, nullptr, 0 }; + uint32_t size; + + // Read name and type of blob + std::string name; + input >> name >> std::dec >> size; + wt.type = DataType::kFLOAT; + + // Load blob + uint32_t* val = reinterpret_cast(malloc(sizeof(val) * size)); + for (uint32_t x = 0, y = size; x < y; ++x) + { + input >> std::hex >> val[x]; + } + wt.values = val; + + wt.count = size; + weightMap[name] = wt; + } + + return weightMap; +} + +ITensor* residualDenseBlock(INetworkDefinition *network, std::map& weightMap, ITensor* x, std::string lname) +{ + IConvolutionLayer* conv_1 = network->addConvolutionNd(*x, 32, DimsHW{ 3, 3 }, weightMap[lname + ".conv1.weight"], weightMap[lname + ".conv1.bias"]); + conv_1->setStrideNd(DimsHW{ 1, 1 }); + conv_1->setPaddingNd(DimsHW{ 1, 1 }); + IActivationLayer* leaky_relu_1 = network->addActivation(*conv_1->getOutput(0), ActivationType::kLEAKY_RELU); + leaky_relu_1->setAlpha(0.2); + ITensor* x1 = leaky_relu_1->getOutput(0); + + ITensor* concat_input2[] = { x, x1 }; + IConcatenationLayer* concat2 = network->addConcatenation(concat_input2, 2); + concat2->setAxis(0); + IConvolutionLayer* conv_2 = network->addConvolutionNd(*concat2->getOutput(0), 32, DimsHW{ 3, 3 }, weightMap[lname + ".conv2.weight"], weightMap[lname + ".conv2.bias"]); + conv_2->setStrideNd(DimsHW{ 1, 1 }); + conv_2->setPaddingNd(DimsHW{ 1, 1 }); + IActivationLayer* leaky_relu_2 = network->addActivation(*conv_2->getOutput(0), ActivationType::kLEAKY_RELU); + leaky_relu_2->setAlpha(0.2); + ITensor* x2 = leaky_relu_2->getOutput(0); + + ITensor* concat_input3[] = { x, x1, x2 }; + IConcatenationLayer* concat3 = network->addConcatenation(concat_input3, 3); + concat3->setAxis(0); + IConvolutionLayer* conv_3 = network->addConvolutionNd(*concat3->getOutput(0), 32, DimsHW{ 3, 3 }, weightMap[lname + ".conv3.weight"], weightMap[lname + ".conv3.bias"]); + conv_3->setStrideNd(DimsHW{ 1, 1 }); + conv_3->setPaddingNd(DimsHW{ 1, 1 }); + IActivationLayer* leaky_relu_3 = network->addActivation(*conv_3->getOutput(0), ActivationType::kLEAKY_RELU); + leaky_relu_3->setAlpha(0.2); + ITensor* x3 = leaky_relu_3->getOutput(0); + + ITensor* concat_input4[] = { x, x1, x2, x3 }; + IConcatenationLayer* concat4 = network->addConcatenation(concat_input4, 4); + concat4->setAxis(0); + IConvolutionLayer* conv_4 = network->addConvolutionNd(*concat4->getOutput(0), 32, DimsHW{ 3, 3 }, weightMap[lname + ".conv4.weight"], weightMap[lname + ".conv4.bias"]); + conv_4->setStrideNd(DimsHW{ 1, 1 }); + conv_4->setPaddingNd(DimsHW{ 1, 1 }); + IActivationLayer* leaky_relu_4 = network->addActivation(*conv_4->getOutput(0), ActivationType::kLEAKY_RELU); + leaky_relu_4->setAlpha(0.2); + ITensor* x4 = leaky_relu_4->getOutput(0); + + ITensor* concat_input5[] = { x, x1, x2, x3, x4 }; + IConcatenationLayer* concat5 = network->addConcatenation(concat_input5, 5); + concat5->setAxis(0); + IConvolutionLayer* conv_5 = network->addConvolutionNd(*concat5->getOutput(0), 64, DimsHW{ 3, 3 }, weightMap[lname + ".conv5.weight"], weightMap[lname + ".conv5.bias"]); + conv_5->setStrideNd(DimsHW{ 1, 1 }); + conv_5->setPaddingNd(DimsHW{ 1, 1 }); + ITensor* x5 = conv_5->getOutput(0); + + float *scval = reinterpret_cast(malloc(sizeof(float))); + *scval = 0.2; + Weights scale{ DataType::kFLOAT, scval, 1 }; + float *shval = reinterpret_cast(malloc(sizeof(float))); + *shval = 0.0; + Weights shift{ DataType::kFLOAT, shval, 1 }; + float *pval = reinterpret_cast(malloc(sizeof(float))); + *pval = 1.0; + Weights power{ DataType::kFLOAT, pval, 1 }; + + IScaleLayer* scaled = network->addScale(*x5, ScaleMode::kUNIFORM, shift, scale, power); + IElementWiseLayer* ew1 = network->addElementWise(*scaled->getOutput(0), *x, ElementWiseOperation::kSUM); + return ew1->getOutput(0); +} + +ITensor* RRDB(INetworkDefinition *network, std::map& weightMap, ITensor* x, std::string lname) +{ + ITensor* out = residualDenseBlock(network, weightMap, x, lname + ".rdb1"); + out = residualDenseBlock(network, weightMap, out, lname + ".rdb2"); + out = residualDenseBlock(network, weightMap, out, lname + ".rdb3"); + + float *scval = reinterpret_cast(malloc(sizeof(float))); + *scval = 0.2; + Weights scale{ DataType::kFLOAT, scval, 1 }; + float *shval = reinterpret_cast(malloc(sizeof(float))); + *shval = 0.0; + Weights shift{ DataType::kFLOAT, shval, 1 }; + float *pval = reinterpret_cast(malloc(sizeof(float))); + *pval = 1.0; + Weights power{ DataType::kFLOAT, pval, 1 }; + + IScaleLayer* scaled = network->addScale(*out, ScaleMode::kUNIFORM, shift, scale, power); + IElementWiseLayer* ew1 = network->addElementWise(*scaled->getOutput(0), *x, ElementWiseOperation::kSUM); + return ew1->getOutput(0); +} + + +#endif \ No newline at end of file diff --git a/real-esrgan/cuda_utils.h b/real-esrgan/cuda_utils.h new file mode 100644 index 0000000..e7dc297 --- /dev/null +++ b/real-esrgan/cuda_utils.h @@ -0,0 +1,22 @@ +#ifndef TRTX_CUDA_UTILS_H_ +#define TRTX_CUDA_UTILS_H_ + +#include +#include +#include +#include +#include + +#ifndef CUDA_CHECK +#define CUDA_CHECK(callstr)\ + {\ + cudaError_t error_code = callstr;\ + if (error_code != cudaSuccess) {\ + std::cerr << "CUDA error " << error_code << " at " << __FILE__ << ":" << __LINE__;\ + assert(0);\ + }\ + } +#endif // CUDA_CHECK + +#endif // TRTX_CUDA_UTILS_H_ + diff --git a/real-esrgan/gen_wts.py b/real-esrgan/gen_wts.py new file mode 100644 index 0000000..db76fc9 --- /dev/null +++ b/real-esrgan/gen_wts.py @@ -0,0 +1,92 @@ +import argparse +import os +import struct +from basicsr.archs.rrdbnet_arch import RRDBNet +from realesrgan import RealESRGANer +from realesrgan.archs.srvgg_arch import SRVGGNetCompact + +def main(): + """Inference demo for Real-ESRGAN. + """ + parser = argparse.ArgumentParser() + #parser.add_argument('-i', '--input', type=str, default='../TestData3', help='Input image or folder') + parser.add_argument('-i', '--input', type=str, default='inputs', help='Input image or folder') + parser.add_argument( + '-n', + '--model_name', + type=str, + default='RealESRGAN_x4plus', + help=('Model names: RealESRGAN_x4plus | RealESRNet_x4plus | RealESRGAN_x4plus_anime_6B | RealESRGAN_x2plus | ' + 'realesr-animevideov3')) + parser.add_argument('-o', '--output', type=str, default='results', help='Output folder') + parser.add_argument('-s', '--outscale', type=float, default=4, help='The final upsampling scale of the image') + parser.add_argument('--suffix', type=str, default='out', help='Suffix of the restored image') + parser.add_argument('-t', '--tile', type=int, default=0, help='Tile size, 0 for no tile during testing') + parser.add_argument('--tile_pad', type=int, default=10, help='Tile padding') + parser.add_argument('--pre_pad', type=int, default=0, help='Pre padding size at each border') + parser.add_argument('--face_enhance', action='store_true', help='Use GFPGAN to enhance face') + parser.add_argument( + '--fp32', action='store_true', help='Use fp32 precision during inference. Default: fp16 (half precision).') + parser.add_argument( + '--alpha_upsampler', + type=str, + default='realesrgan', + help='The upsampler for the alpha channels. Options: realesrgan | bicubic') + parser.add_argument( + '--ext', + type=str, + default='auto', + help='Image extension. Options: auto | jpg | png, auto means using the same extension as inputs') + args = parser.parse_args() + + # determine models according to model names + args.model_name = args.model_name.split('.')[0] + if args.model_name in ['RealESRGAN_x4plus', 'RealESRNet_x4plus']: # x4 RRDBNet model + model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4) + netscale = 4 + elif args.model_name in ['RealESRGAN_x4plus_anime_6B']: # x4 RRDBNet model with 6 blocks + model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=6, num_grow_ch=32, scale=4) + netscale = 4 + elif args.model_name in ['RealESRGAN_x2plus']: # x2 RRDBNet model + model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2) + netscale = 2 + elif args.model_name in ['realesr-animevideov3']: # x4 VGG-style model (XS size) + model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, act_type='prelu') + netscale = 4 + + # determine model paths + model_path = os.path.join('experiments/pretrained_models', args.model_name + '.pth') + if not os.path.isfile(model_path): + model_path = os.path.join('realesrgan/weights', args.model_name + '.pth') + if not os.path.isfile(model_path): + raise ValueError(f'Model {args.model_name} does not exist.') + + # restorer + upsampler = RealESRGANer( + scale=netscale, + model_path=model_path, + model=model, + tile=args.tile, + tile_pad=args.tile_pad, + pre_pad=args.pre_pad, + half=args.fp32) + + if os.path.isfile('real-esrgan.wts'): + print('Already, real-esrgan.wts file exists.') + else: + print('making real-esrgan.wts file ...') + f = open("real-esrgan.wts", 'w') + f.write("{}\n".format(len(upsampler.model.state_dict().keys()))) + for k, v in upsampler.model.state_dict().items(): + print('key: ', k) + print('value: ', v.shape) + vr = v.reshape(-1).cpu().numpy() + f.write("{} {}".format(k, len(vr))) + for vv in vr: + f.write(" ") + f.write(struct.pack(">f", float(vv)).hex()) + f.write("\n") + print('Completed real-esrgan.wts file!') + +if __name__ == '__main__': + main() diff --git a/real-esrgan/logging.h b/real-esrgan/logging.h new file mode 100644 index 0000000..6b79a8b --- /dev/null +++ b/real-esrgan/logging.h @@ -0,0 +1,504 @@ +/* + * Copyright (c) 2019, NVIDIA CORPORATION. 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. + */ + +#ifndef TENSORRT_LOGGING_H +#define TENSORRT_LOGGING_H + +#include "NvInferRuntimeCommon.h" +#include +#include +#include +#include +#include +#include +#include +#include "macros.h" + +using Severity = nvinfer1::ILogger::Severity; + +class LogStreamConsumerBuffer : public std::stringbuf +{ +public: + LogStreamConsumerBuffer(std::ostream& stream, const std::string& prefix, bool shouldLog) + : mOutput(stream) + , mPrefix(prefix) + , mShouldLog(shouldLog) + { + } + + LogStreamConsumerBuffer(LogStreamConsumerBuffer&& other) + : mOutput(other.mOutput) + { + } + + ~LogStreamConsumerBuffer() + { + // std::streambuf::pbase() gives a pointer to the beginning of the buffered part of the output sequence + // std::streambuf::pptr() gives a pointer to the current position of the output sequence + // if the pointer to the beginning is not equal to the pointer to the current position, + // call putOutput() to log the output to the stream + if (pbase() != pptr()) + { + putOutput(); + } + } + + // synchronizes the stream buffer and returns 0 on success + // synchronizing the stream buffer consists of inserting the buffer contents into the stream, + // resetting the buffer and flushing the stream + virtual int sync() + { + putOutput(); + return 0; + } + + void putOutput() + { + if (mShouldLog) + { + // prepend timestamp + std::time_t timestamp = std::time(nullptr); + tm* tm_local = std::localtime(×tamp); + std::cout << "["; + std::cout << std::setw(2) << std::setfill('0') << 1 + tm_local->tm_mon << "/"; + std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_mday << "/"; + std::cout << std::setw(4) << std::setfill('0') << 1900 + tm_local->tm_year << "-"; + std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_hour << ":"; + std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_min << ":"; + std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_sec << "] "; + // std::stringbuf::str() gets the string contents of the buffer + // insert the buffer contents pre-appended by the appropriate prefix into the stream + mOutput << mPrefix << str(); + // set the buffer to empty + str(""); + // flush the stream + mOutput.flush(); + } + } + + void setShouldLog(bool shouldLog) + { + mShouldLog = shouldLog; + } + +private: + std::ostream& mOutput; + std::string mPrefix; + bool mShouldLog; +}; + +//! +//! \class LogStreamConsumerBase +//! \brief Convenience object used to initialize LogStreamConsumerBuffer before std::ostream in LogStreamConsumer +//! +class LogStreamConsumerBase +{ +public: + LogStreamConsumerBase(std::ostream& stream, const std::string& prefix, bool shouldLog) + : mBuffer(stream, prefix, shouldLog) + { + } + +protected: + LogStreamConsumerBuffer mBuffer; +}; + +//! +//! \class LogStreamConsumer +//! \brief Convenience object used to facilitate use of C++ stream syntax when logging messages. +//! Order of base classes is LogStreamConsumerBase and then std::ostream. +//! This is because the LogStreamConsumerBase class is used to initialize the LogStreamConsumerBuffer member field +//! in LogStreamConsumer and then the address of the buffer is passed to std::ostream. +//! This is necessary to prevent the address of an uninitialized buffer from being passed to std::ostream. +//! Please do not change the order of the parent classes. +//! +class LogStreamConsumer : protected LogStreamConsumerBase, public std::ostream +{ +public: + //! \brief Creates a LogStreamConsumer which logs messages with level severity. + //! Reportable severity determines if the messages are severe enough to be logged. + LogStreamConsumer(Severity reportableSeverity, Severity severity) + : LogStreamConsumerBase(severityOstream(severity), severityPrefix(severity), severity <= reportableSeverity) + , std::ostream(&mBuffer) // links the stream buffer with the stream + , mShouldLog(severity <= reportableSeverity) + , mSeverity(severity) + { + } + + LogStreamConsumer(LogStreamConsumer&& other) + : LogStreamConsumerBase(severityOstream(other.mSeverity), severityPrefix(other.mSeverity), other.mShouldLog) + , std::ostream(&mBuffer) // links the stream buffer with the stream + , mShouldLog(other.mShouldLog) + , mSeverity(other.mSeverity) + { + } + + void setReportableSeverity(Severity reportableSeverity) + { + mShouldLog = mSeverity <= reportableSeverity; + mBuffer.setShouldLog(mShouldLog); + } + +private: + static std::ostream& severityOstream(Severity severity) + { + return severity >= Severity::kINFO ? std::cout : std::cerr; + } + + static std::string severityPrefix(Severity severity) + { + switch (severity) + { + case Severity::kINTERNAL_ERROR: return "[F] "; + case Severity::kERROR: return "[E] "; + case Severity::kWARNING: return "[W] "; + case Severity::kINFO: return "[I] "; + case Severity::kVERBOSE: return "[V] "; + default: assert(0); return ""; + } + } + + bool mShouldLog; + Severity mSeverity; +}; + +//! \class Logger +//! +//! \brief Class which manages logging of TensorRT tools and samples +//! +//! \details This class provides a common interface for TensorRT tools and samples to log information to the console, +//! and supports logging two types of messages: +//! +//! - Debugging messages with an associated severity (info, warning, error, or internal error/fatal) +//! - Test pass/fail messages +//! +//! The advantage of having all samples use this class for logging as opposed to emitting directly to stdout/stderr is +//! that the logic for controlling the verbosity and formatting of sample output is centralized in one location. +//! +//! In the future, this class could be extended to support dumping test results to a file in some standard format +//! (for example, JUnit XML), and providing additional metadata (e.g. timing the duration of a test run). +//! +//! TODO: For backwards compatibility with existing samples, this class inherits directly from the nvinfer1::ILogger +//! interface, which is problematic since there isn't a clean separation between messages coming from the TensorRT +//! library and messages coming from the sample. +//! +//! In the future (once all samples are updated to use Logger::getTRTLogger() to access the ILogger) we can refactor the +//! class to eliminate the inheritance and instead make the nvinfer1::ILogger implementation a member of the Logger +//! object. + +class Logger : public nvinfer1::ILogger +{ +public: + Logger(Severity severity = Severity::kWARNING) + : mReportableSeverity(severity) + { + } + + //! + //! \enum TestResult + //! \brief Represents the state of a given test + //! + enum class TestResult + { + kRUNNING, //!< The test is running + kPASSED, //!< The test passed + kFAILED, //!< The test failed + kWAIVED //!< The test was waived + }; + + //! + //! \brief Forward-compatible method for retrieving the nvinfer::ILogger associated with this Logger + //! \return The nvinfer1::ILogger associated with this Logger + //! + //! TODO Once all samples are updated to use this method to register the logger with TensorRT, + //! we can eliminate the inheritance of Logger from ILogger + //! + nvinfer1::ILogger& getTRTLogger() + { + return *this; + } + + //! + //! \brief Implementation of the nvinfer1::ILogger::log() virtual method + //! + //! Note samples should not be calling this function directly; it will eventually go away once we eliminate the + //! inheritance from nvinfer1::ILogger + //! + void log(Severity severity, const char* msg) TRT_NOEXCEPT override + { + LogStreamConsumer(mReportableSeverity, severity) << "[TRT] " << std::string(msg) << std::endl; + } + + //! + //! \brief Method for controlling the verbosity of logging output + //! + //! \param severity The logger will only emit messages that have severity of this level or higher. + //! + void setReportableSeverity(Severity severity) + { + mReportableSeverity = severity; + } + + //! + //! \brief Opaque handle that holds logging information for a particular test + //! + //! This object is an opaque handle to information used by the Logger to print test results. + //! The sample must call Logger::defineTest() in order to obtain a TestAtom that can be used + //! with Logger::reportTest{Start,End}(). + //! + class TestAtom + { + public: + TestAtom(TestAtom&&) = default; + + private: + friend class Logger; + + TestAtom(bool started, const std::string& name, const std::string& cmdline) + : mStarted(started) + , mName(name) + , mCmdline(cmdline) + { + } + + bool mStarted; + std::string mName; + std::string mCmdline; + }; + + //! + //! \brief Define a test for logging + //! + //! \param[in] name The name of the test. This should be a string starting with + //! "TensorRT" and containing dot-separated strings containing + //! the characters [A-Za-z0-9_]. + //! For example, "TensorRT.sample_googlenet" + //! \param[in] cmdline The command line used to reproduce the test + // + //! \return a TestAtom that can be used in Logger::reportTest{Start,End}(). + //! + static TestAtom defineTest(const std::string& name, const std::string& cmdline) + { + return TestAtom(false, name, cmdline); + } + + //! + //! \brief A convenience overloaded version of defineTest() that accepts an array of command-line arguments + //! as input + //! + //! \param[in] name The name of the test + //! \param[in] argc The number of command-line arguments + //! \param[in] argv The array of command-line arguments (given as C strings) + //! + //! \return a TestAtom that can be used in Logger::reportTest{Start,End}(). + static TestAtom defineTest(const std::string& name, int argc, char const* const* argv) + { + auto cmdline = genCmdlineString(argc, argv); + return defineTest(name, cmdline); + } + + //! + //! \brief Report that a test has started. + //! + //! \pre reportTestStart() has not been called yet for the given testAtom + //! + //! \param[in] testAtom The handle to the test that has started + //! + static void reportTestStart(TestAtom& testAtom) + { + reportTestResult(testAtom, TestResult::kRUNNING); + assert(!testAtom.mStarted); + testAtom.mStarted = true; + } + + //! + //! \brief Report that a test has ended. + //! + //! \pre reportTestStart() has been called for the given testAtom + //! + //! \param[in] testAtom The handle to the test that has ended + //! \param[in] result The result of the test. Should be one of TestResult::kPASSED, + //! TestResult::kFAILED, TestResult::kWAIVED + //! + static void reportTestEnd(const TestAtom& testAtom, TestResult result) + { + assert(result != TestResult::kRUNNING); + assert(testAtom.mStarted); + reportTestResult(testAtom, result); + } + + static int reportPass(const TestAtom& testAtom) + { + reportTestEnd(testAtom, TestResult::kPASSED); + return EXIT_SUCCESS; + } + + static int reportFail(const TestAtom& testAtom) + { + reportTestEnd(testAtom, TestResult::kFAILED); + return EXIT_FAILURE; + } + + static int reportWaive(const TestAtom& testAtom) + { + reportTestEnd(testAtom, TestResult::kWAIVED); + return EXIT_SUCCESS; + } + + static int reportTest(const TestAtom& testAtom, bool pass) + { + return pass ? reportPass(testAtom) : reportFail(testAtom); + } + + Severity getReportableSeverity() const + { + return mReportableSeverity; + } + +private: + //! + //! \brief returns an appropriate string for prefixing a log message with the given severity + //! + static const char* severityPrefix(Severity severity) + { + switch (severity) + { + case Severity::kINTERNAL_ERROR: return "[F] "; + case Severity::kERROR: return "[E] "; + case Severity::kWARNING: return "[W] "; + case Severity::kINFO: return "[I] "; + case Severity::kVERBOSE: return "[V] "; + default: assert(0); return ""; + } + } + + //! + //! \brief returns an appropriate string for prefixing a test result message with the given result + //! + static const char* testResultString(TestResult result) + { + switch (result) + { + case TestResult::kRUNNING: return "RUNNING"; + case TestResult::kPASSED: return "PASSED"; + case TestResult::kFAILED: return "FAILED"; + case TestResult::kWAIVED: return "WAIVED"; + default: assert(0); return ""; + } + } + + //! + //! \brief returns an appropriate output stream (cout or cerr) to use with the given severity + //! + static std::ostream& severityOstream(Severity severity) + { + return severity >= Severity::kINFO ? std::cout : std::cerr; + } + + //! + //! \brief method that implements logging test results + //! + static void reportTestResult(const TestAtom& testAtom, TestResult result) + { + severityOstream(Severity::kINFO) << "&&&& " << testResultString(result) << " " << testAtom.mName << " # " + << testAtom.mCmdline << std::endl; + } + + //! + //! \brief generate a command line string from the given (argc, argv) values + //! + static std::string genCmdlineString(int argc, char const* const* argv) + { + std::stringstream ss; + for (int i = 0; i < argc; i++) + { + if (i > 0) + ss << " "; + ss << argv[i]; + } + return ss.str(); + } + + Severity mReportableSeverity; +}; + +namespace +{ + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kVERBOSE +//! +//! Example usage: +//! +//! LOG_VERBOSE(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_VERBOSE(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kVERBOSE); +} + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kINFO +//! +//! Example usage: +//! +//! LOG_INFO(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_INFO(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kINFO); +} + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kWARNING +//! +//! Example usage: +//! +//! LOG_WARN(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_WARN(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kWARNING); +} + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kERROR +//! +//! Example usage: +//! +//! LOG_ERROR(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_ERROR(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kERROR); +} + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kINTERNAL_ERROR +// ("fatal" severity) +//! +//! Example usage: +//! +//! LOG_FATAL(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_FATAL(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kINTERNAL_ERROR); +} + +} // anonymous namespace + +#endif // TENSORRT_LOGGING_H diff --git a/real-esrgan/macros.h b/real-esrgan/macros.h new file mode 100644 index 0000000..0e5b71b --- /dev/null +++ b/real-esrgan/macros.h @@ -0,0 +1,27 @@ +#ifndef __MACROS_H +#define __MACROS_H + +#ifdef API_EXPORTS +#if defined(_MSC_VER) +#define API __declspec(dllexport) +#else +#define API __attribute__((visibility("default"))) +#endif +#else + +#if defined(_MSC_VER) +#define API __declspec(dllimport) +#else +#define API +#endif +#endif // API_EXPORTS + +#if NV_TENSORRT_MAJOR >= 8 +#define TRT_NOEXCEPT noexcept +#define TRT_CONST_ENQUEUE const +#else +#define TRT_NOEXCEPT +#define TRT_CONST_ENQUEUE +#endif + +#endif // __MACROS_H diff --git a/real-esrgan/postprocess.cu b/real-esrgan/postprocess.cu new file mode 100644 index 0000000..60ea083 --- /dev/null +++ b/real-esrgan/postprocess.cu @@ -0,0 +1,54 @@ +#include "cuda_utils.h" + +using namespace std; + +// postprocess (NCHW->NHWC, RGB->BGR, *255, ROUND, uint8) +__global__ void postprocess_kernel(uint8_t* output, float* input, + const int batchSize, const int height, const int width, const int channel, + const int thread_count) +{ + int index = threadIdx.x + blockIdx.x * blockDim.x; + if (index >= thread_count) return; + + const int c_idx = index % channel; + int idx = index / channel; + const int w_idx = idx % width; + idx /= width; + const int h_idx = idx % height; + const int b_idx = idx / height; + + int g_idx = b_idx * height * width * channel + (2 - c_idx)* height * width + h_idx * width + w_idx; + float tt = input[g_idx] * 255.f; + if (tt > 255) + tt = 255; + output[index] = tt; +} + +void postprocess(uint8_t* output, float*input, int batchSize, int height, int width, int channel, cudaStream_t stream) +{ + int thread_count = batchSize * height * width * channel; + int block = 512; + int grid = (thread_count - 1) / block + 1; + + postprocess_kernel << > > (output, input, batchSize, height, width, channel, thread_count); +} + + +#include "postprocess.hpp" + +namespace nvinfer1 +{ + int PostprocessPluginV2::enqueue(int batchSize, const void* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream) noexcept + { + float* input = (float*)inputs[0]; + uint8_t* output = (uint8_t*)outputs[0]; + + const int H = mPostprocess.H; + const int W = mPostprocess.W; + const int C = mPostprocess.C; + + postprocess(output, input, batchSize, H, W, C, stream); + + return 0; + } +} \ No newline at end of file diff --git a/real-esrgan/postprocess.hpp b/real-esrgan/postprocess.hpp new file mode 100644 index 0000000..6fd56bb --- /dev/null +++ b/real-esrgan/postprocess.hpp @@ -0,0 +1,206 @@ +#pragma once +#include +#include +#include "macros.h" +#include + +struct Postprocess { + int N; + int C; + int H; + int W; +}; + +namespace nvinfer1 +{ + class PostprocessPluginV2 : public IPluginV2IOExt + { + public: + PostprocessPluginV2(const Postprocess& arg) + { + mPostprocess = arg; + } + + PostprocessPluginV2(const void* data, size_t length) + { + const char* d = static_cast(data); + const char* const a = d; + mPostprocess = read(d); + assert(d == a + length); + } + PostprocessPluginV2() = delete; + + virtual ~PostprocessPluginV2() {} + + public: + int getNbOutputs() const noexcept override + { + return 1; + } + + Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) noexcept override + { + return Dims3(mPostprocess.H, mPostprocess.W, mPostprocess.C); + } + + int initialize() noexcept override + { + return 0; + } + + void terminate() noexcept override + { + } + + size_t getWorkspaceSize(int maxBatchSize) const noexcept override + { + return 0; + } + + int enqueue(int batchSize, const void* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream) noexcept override; + + size_t getSerializationSize() const noexcept override + { + size_t serializationSize = 0; + serializationSize += sizeof(mPostprocess); + return serializationSize; + } + + void serialize(void* buffer) const noexcept override + { + char* d = static_cast(buffer); + const char* const a = d; + write(d, mPostprocess); + assert(d == a + getSerializationSize()); + } + + void configurePlugin(const PluginTensorDesc* in, int nbInput, const PluginTensorDesc* out, int nbOutput) noexcept override + { + } + + //! The combination of kLINEAR + kINT8/kHALF/kFLOAT is supported. + bool supportsFormatCombination(int pos, const PluginTensorDesc* inOut, int nbInputs, int nbOutputs) const noexcept override + { + assert(nbInputs == 1 && nbOutputs == 1 && pos < nbInputs + nbOutputs); + bool condition = inOut[pos].format == TensorFormat::kLINEAR; + condition &= inOut[pos].type != DataType::kINT32; + condition &= inOut[pos].type == inOut[0].type; + return condition; + } + + DataType getOutputDataType(int index, const DataType* inputTypes, int nbInputs) const noexcept override + { + assert(inputTypes && nbInputs == 1); + return DataType::kFLOAT; // + } + + const char* getPluginType() const noexcept override + { + return "postprocess"; + } + + const char* getPluginVersion() const noexcept override + { + return "1"; + } + + void destroy() noexcept override + { + delete this; + } + + IPluginV2Ext* clone() const noexcept override + { + PostprocessPluginV2* plugin = new PostprocessPluginV2(*this); + return plugin; + } + + void setPluginNamespace(const char* libNamespace) noexcept override + { + mNamespace = libNamespace; + } + + const char* getPluginNamespace() const noexcept override + { + return mNamespace.data(); + } + + bool isOutputBroadcastAcrossBatch(int outputIndex, const bool* inputIsBroadcasted, int nbInputs) const noexcept override + { + return false; + } + + bool canBroadcastInputAcrossBatch(int inputIndex) const noexcept override + { + return false; + } + + private: + template + void write(char*& buffer, const T& val) const + { + *reinterpret_cast(buffer) = val; + buffer += sizeof(T); + } + + template + T read(const char*& buffer) const + { + T val = *reinterpret_cast(buffer); + buffer += sizeof(T); + return val; + } + + private: + Postprocess mPostprocess; + std::string mNamespace; + }; + + class PostprocessPluginV2Creator : public IPluginCreator + { + public: + const char* getPluginName() const noexcept override + { + return "postprocess"; + } + + const char* getPluginVersion() const noexcept override + { + return "1"; + } + + const PluginFieldCollection* getFieldNames() noexcept override + { + return nullptr; + } + + IPluginV2* createPlugin(const char* name, const PluginFieldCollection* fc) noexcept override + { + PostprocessPluginV2* plugin = new PostprocessPluginV2(*(Postprocess*)fc); + mPluginName = name; + return plugin; + } + + IPluginV2* deserializePlugin(const char* name, const void* serialData, size_t serialLength) noexcept override + { + auto plugin = new PostprocessPluginV2(serialData, serialLength); + mPluginName = name; + return plugin; + } + + void setPluginNamespace(const char* libNamespace) noexcept override + { + mNamespace = libNamespace; + } + + const char* getPluginNamespace() const noexcept override + { + return mNamespace.c_str(); + } + + private: + std::string mNamespace; + std::string mPluginName; + }; + REGISTER_TENSORRT_PLUGIN(PostprocessPluginV2Creator); +}; diff --git a/real-esrgan/preprocess.cu b/real-esrgan/preprocess.cu new file mode 100644 index 0000000..3789933 --- /dev/null +++ b/real-esrgan/preprocess.cu @@ -0,0 +1,51 @@ +#include "cuda_utils.h" + +using namespace std; + +// preprocess (NHWC->NCHW, BGR->RGB, [0, 255]->[0, 1](Normalize)) +__global__ void preprocess_kernel(float* output, uint8_t* input, + const int batchSize, const int height, const int width, const int channel, + const int thread_count) +{ + int index = threadIdx.x + blockIdx.x * blockDim.x; + if (index >= thread_count) return; + + const int w_idx = index % width; + int idx = index / width; + const int h_idx = idx % height; + idx /= height; + const int c_idx = idx % channel; + const int b_idx = idx / channel; + + int g_idx = b_idx * height * width * channel + h_idx * width * channel + w_idx * channel + 2 - c_idx; + + output[index] = input[g_idx] / 255.f; +} + +void preprocess(float* output, uint8_t*input, int batchSize, int height, int width, int channel, cudaStream_t stream) +{ + int thread_count = batchSize * height * width * channel; + int block = 512; + int grid = (thread_count - 1) / block + 1; + + preprocess_kernel << > > (output, input, batchSize, height, width, channel, thread_count); +} + +#include "preprocess.hpp" + +namespace nvinfer1 +{ + int PreprocessPluginV2::enqueue(int batchSize, const void* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream) noexcept + { + uint8_t* input = (uint8_t*)inputs[0]; + float* output = (float*)outputs[0]; + + const int H = mPreprocess.H; + const int W = mPreprocess.W; + const int C = mPreprocess.C; + + preprocess(output, input, batchSize, H, W, C, stream); + + return 0; + } +} \ No newline at end of file diff --git a/real-esrgan/preprocess.hpp b/real-esrgan/preprocess.hpp new file mode 100644 index 0000000..1b6c5b8 --- /dev/null +++ b/real-esrgan/preprocess.hpp @@ -0,0 +1,206 @@ +#pragma once +#include +#include +#include "macros.h" +#include + +struct Preprocess { + int N; + int C; + int H; + int W; +}; + +namespace nvinfer1 +{ + class PreprocessPluginV2 : public IPluginV2IOExt + { + public: + PreprocessPluginV2(const Preprocess& arg) + { + mPreprocess = arg; + } + + PreprocessPluginV2(const void* data, size_t length) + { + const char* d = static_cast(data); + const char* const a = d; + mPreprocess = read(d); + assert(d == a + length); + } + PreprocessPluginV2() = delete; + + virtual ~PreprocessPluginV2() {} + + public: + int getNbOutputs() const noexcept override + { + return 1; + } + + Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) noexcept override + { + return Dims3(mPreprocess.C, mPreprocess.H, mPreprocess.W); + } + + int initialize() noexcept override + { + return 0; + } + + void terminate() noexcept override + { + } + + size_t getWorkspaceSize(int maxBatchSize) const noexcept override + { + return 0; + } + + int enqueue(int batchSize, const void* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream) noexcept override; + + size_t getSerializationSize() const noexcept override + { + size_t serializationSize = 0; + serializationSize += sizeof(mPreprocess); + return serializationSize; + } + + void serialize(void* buffer) const noexcept override + { + char* d = static_cast(buffer); + const char* const a = d; + write(d, mPreprocess); + assert(d == a + getSerializationSize()); + } + + void configurePlugin(const PluginTensorDesc* in, int nbInput, const PluginTensorDesc* out, int nbOutput) noexcept override + { + } + + //! The combination of kLINEAR + kINT8/kHALF/kFLOAT is supported. + bool supportsFormatCombination(int pos, const PluginTensorDesc* inOut, int nbInputs, int nbOutputs) const noexcept override + { + assert(nbInputs == 1 && nbOutputs == 1 && pos < nbInputs + nbOutputs); + bool condition = inOut[pos].format == TensorFormat::kLINEAR; + condition &= inOut[pos].type != DataType::kINT32; + condition &= inOut[pos].type == inOut[0].type; + return condition; + } + + DataType getOutputDataType(int index, const DataType* inputTypes, int nbInputs) const noexcept override + { + assert(inputTypes && nbInputs == 1); + return DataType::kFLOAT; // + } + + const char* getPluginType() const noexcept override + { + return "preprocess"; + } + + const char* getPluginVersion() const noexcept override + { + return "1"; + } + + void destroy() noexcept override + { + delete this; + } + + IPluginV2Ext* clone() const noexcept override + { + PreprocessPluginV2* plugin = new PreprocessPluginV2(*this); + return plugin; + } + + void setPluginNamespace(const char* libNamespace) noexcept override + { + mNamespace = libNamespace; + } + + const char* getPluginNamespace() const noexcept override + { + return mNamespace.data(); + } + + bool isOutputBroadcastAcrossBatch(int outputIndex, const bool* inputIsBroadcasted, int nbInputs) const noexcept override + { + return false; + } + + bool canBroadcastInputAcrossBatch(int inputIndex) const noexcept override + { + return false; + } + + private: + template + void write(char*& buffer, const T& val) const + { + *reinterpret_cast(buffer) = val; + buffer += sizeof(T); + } + + template + T read(const char*& buffer) const + { + T val = *reinterpret_cast(buffer); + buffer += sizeof(T); + return val; + } + + private: + Preprocess mPreprocess; + std::string mNamespace; + }; + + class PreprocessPluginV2Creator : public IPluginCreator + { + public: + const char* getPluginName() const noexcept override + { + return "preprocess"; + } + + const char* getPluginVersion() const noexcept override + { + return "1"; + } + + const PluginFieldCollection* getFieldNames() noexcept override + { + return nullptr; + } + + IPluginV2* createPlugin(const char* name, const PluginFieldCollection* fc) noexcept override + { + PreprocessPluginV2* plugin = new PreprocessPluginV2(*(Preprocess*)fc); + mPluginName = name; + return plugin; + } + + IPluginV2* deserializePlugin(const char* name, const void* serialData, size_t serialLength) noexcept override + { + auto plugin = new PreprocessPluginV2(serialData, serialLength); + mPluginName = name; + return plugin; + } + + void setPluginNamespace(const char* libNamespace) noexcept override + { + mNamespace = libNamespace; + } + + const char* getPluginNamespace() const noexcept override + { + return mNamespace.c_str(); + } + + private: + std::string mNamespace; + std::string mPluginName; + }; + REGISTER_TENSORRT_PLUGIN(PreprocessPluginV2Creator); +}; diff --git a/real-esrgan/real-esrgan.cpp b/real-esrgan/real-esrgan.cpp new file mode 100644 index 0000000..c930c8b --- /dev/null +++ b/real-esrgan/real-esrgan.cpp @@ -0,0 +1,285 @@ +#include "cuda_utils.h" +#include "common.hpp" +#include "preprocess.hpp"// preprocess plugin +#include "postprocess.hpp"// postprocess plugin +#include "logging.h" +#include "utils.h" +#include //access() + +#define DEVICE 0 // GPU id +#define BATCH_SIZE 1 + +// stuff we know about the network and the input/output blobs +static const int PRECISION_MODE = 32; // fp32 : 32, fp16 : 16 +static const bool VISUALIZATION = true; +static const int INPUT_H = 640; +static const int INPUT_W = 448; +static const int INPUT_C = 3; +static const int OUT_SCALE = 4; +static const int OUTPUT_SIZE = INPUT_C * INPUT_H * OUT_SCALE * INPUT_W * OUT_SCALE; +const char* INPUT_BLOB_NAME = "data"; +const char* OUTPUT_BLOB_NAME = "prob"; +static Logger gLogger; + +// Creat the engine using only the API and not any parser. +ICudaEngine* build_engine(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt, std::string& wts_name) { + INetworkDefinition* network = builder->createNetworkV2(0U); + + // Create input tensor of shape {INPUT_H, INPUT_W, INPUT_C} with name INPUT_BLOB_NAME + ITensor* data = network->addInput(INPUT_BLOB_NAME, dt, Dims3{ INPUT_H, INPUT_W, INPUT_C }); + assert(data); + std::map weightMap = loadWeights(wts_name); + + // Custom preprocess (NHWC->NCHW, BGR->RGB, [0, 255]->[0, 1](Normalize)) + Preprocess preprocess{ maxBatchSize, INPUT_C, INPUT_H, INPUT_W }; + IPluginCreator* preprocess_creator = getPluginRegistry()->getPluginCreator("preprocess", "1"); + IPluginV2 *preprocess_plugin = preprocess_creator->createPlugin("preprocess_plugin", (PluginFieldCollection*)&preprocess); + IPluginV2Layer* preprocess_layer = network->addPluginV2(&data, 1, *preprocess_plugin); + preprocess_layer->setName("preprocess_layer"); + ITensor* prep = preprocess_layer->getOutput(0); + + // conv_first + IConvolutionLayer* conv_first = network->addConvolutionNd(*prep, 64, DimsHW{ 3, 3 }, weightMap["conv_first.weight"], weightMap["conv_first.bias"]); + conv_first->setStrideNd(DimsHW{ 1, 1 }); + conv_first->setPaddingNd(DimsHW{ 1, 1 }); + conv_first->setName("conv_first"); + ITensor* feat = conv_first->getOutput(0); + + // conv_body + ITensor* body_feat = RRDB(network, weightMap, feat, "body.0"); + for (int idx = 1; idx < 23; idx++) { + body_feat = RRDB(network, weightMap, body_feat, "body." + std::to_string(idx)); + } + + IConvolutionLayer* conv_body = network->addConvolutionNd(*body_feat, 64, DimsHW{ 3, 3 }, weightMap["conv_body.weight"], weightMap["conv_body.bias"]); + conv_body->setStrideNd(DimsHW{ 1, 1 }); + conv_body->setPaddingNd(DimsHW{ 1, 1 }); + IElementWiseLayer* ew1 = network->addElementWise(*feat, *conv_body->getOutput(0), ElementWiseOperation::kSUM); + feat = ew1->getOutput(0); + + //upsample + IResizeLayer* interpolate_nearest = network->addResize(*feat); + float sclaes1[] = { 1, 2, 2 }; + interpolate_nearest->setScales(sclaes1, 3); + interpolate_nearest->setResizeMode(ResizeMode::kNEAREST); + + IConvolutionLayer* conv_up1 = network->addConvolutionNd(*interpolate_nearest->getOutput(0), 64, DimsHW{ 3, 3 }, weightMap["conv_up1.weight"], weightMap["conv_up1.bias"]); + conv_up1->setStrideNd(DimsHW{ 1, 1 }); + conv_up1->setPaddingNd(DimsHW{ 1, 1 }); + IActivationLayer* leaky_relu_1 = network->addActivation(*conv_up1->getOutput(0), ActivationType::kLEAKY_RELU); + leaky_relu_1->setAlpha(0.2); + + IResizeLayer* interpolate_nearest2 = network->addResize(*leaky_relu_1->getOutput(0)); + float sclaes2[] = { 1, 2, 2 }; + interpolate_nearest2->setScales(sclaes2, 3); + interpolate_nearest2->setResizeMode(ResizeMode::kNEAREST); + IConvolutionLayer* conv_up2 = network->addConvolutionNd(*interpolate_nearest2->getOutput(0), 64, DimsHW{ 3, 3 }, weightMap["conv_up2.weight"], weightMap["conv_up2.bias"]); + conv_up2->setStrideNd(DimsHW{ 1, 1 }); + conv_up2->setPaddingNd(DimsHW{ 1, 1 }); + IActivationLayer* leaky_relu_2 = network->addActivation(*conv_up2->getOutput(0), ActivationType::kLEAKY_RELU); + leaky_relu_2->setAlpha(0.2); + + IConvolutionLayer* conv_hr = network->addConvolutionNd(*leaky_relu_2->getOutput(0), 64, DimsHW{ 3, 3 }, weightMap["conv_hr.weight"], weightMap["conv_hr.bias"]); + conv_hr->setStrideNd(DimsHW{ 1, 1 }); + conv_hr->setPaddingNd(DimsHW{ 1, 1 }); + IActivationLayer* leaky_relu_hr = network->addActivation(*conv_hr->getOutput(0), ActivationType::kLEAKY_RELU); + leaky_relu_hr->setAlpha(0.2); + IConvolutionLayer* conv_last = network->addConvolutionNd(*leaky_relu_hr->getOutput(0), 3, DimsHW{ 3, 3 }, weightMap["conv_last.weight"], weightMap["conv_last.bias"]); + conv_last->setStrideNd(DimsHW{ 1, 1 }); + conv_last->setPaddingNd(DimsHW{ 1, 1 }); + ITensor* out = conv_last->getOutput(0); + + // Custom postprocess (RGB -> BGR, NCHW->NHWC, *255, ROUND, uint8) + Postprocess postprocess{ maxBatchSize, out->getDimensions().d[0], out->getDimensions().d[1], out->getDimensions().d[2] }; + IPluginCreator* postprocess_creator = getPluginRegistry()->getPluginCreator("postprocess", "1"); + IPluginV2 *postprocess_plugin = postprocess_creator->createPlugin("postprocess_plugin", (PluginFieldCollection*)&postprocess); + IPluginV2Layer* postprocess_layer = network->addPluginV2(&out, 1, *postprocess_plugin); + postprocess_layer->setName("postprocess_layer"); + + ITensor* final_tensor = postprocess_layer->getOutput(0); + final_tensor->setName(OUTPUT_BLOB_NAME); + network->markOutput(*final_tensor); + + // Build engine + builder->setMaxBatchSize(maxBatchSize); + config->setMaxWorkspaceSize(16 * (1 << 20)); // 16MB + + if (PRECISION_MODE == 16) { + std::cout << "==== precision f16 ====" << std::endl << std::endl; + config->setFlag(BuilderFlag::kFP16); + } + else { + std::cout << "==== precision f32 ====" << std::endl << std::endl; + } + + std::cout << "Building engine, please wait for a while..." << std::endl; + ICudaEngine* engine = builder->buildEngineWithConfig(*network, *config); + std::cout << "Build engine successfully!" << std::endl; + + // Don't need the network any more + delete network; + + // Release host memory + for (auto& mem : weightMap) + { + free((void*)(mem.second.values)); + } + + return engine; +} + +void APIToModel(unsigned int maxBatchSize, IHostMemory** modelStream, std::string& wts_name) { + // Create builder + IBuilder* builder = createInferBuilder(gLogger); + IBuilderConfig* config = builder->createBuilderConfig(); + + // Create model to populate the network, then set the outputs and create an engine + ICudaEngine *engine = build_engine(maxBatchSize, builder, config, DataType::kFLOAT, wts_name); + + assert(engine != nullptr); + + // Serialize the engine + (*modelStream) = engine->serialize(); + + // Close everything down + delete engine; + delete builder; + delete config; +} + +void doInference(IExecutionContext& context, cudaStream_t& stream, void **buffers, uint8_t* output, int batchSize) { + // infer on the batch asynchronously, and DMA output back to host + context.enqueue(batchSize, buffers, stream, nullptr); + CUDA_CHECK(cudaMemcpyAsync(output, buffers[1], batchSize * OUTPUT_SIZE * sizeof(uint8_t), cudaMemcpyDeviceToHost, stream)); + cudaStreamSynchronize(stream); +} + +bool parse_args(int argc, char** argv, std::string& wts, std::string& engine, std::string& img_dir) { + if (argc < 4) return false; + if (std::string(argv[1]) == "-s" && argc == 4) { + wts = std::string(argv[2]); + engine = std::string(argv[3]); + } + else if (std::string(argv[1]) == "-d" && argc == 4) { + engine = std::string(argv[2]); + img_dir = std::string(argv[3]); + } + else { + return false; + } + return true; +} + +// ./real-esrgan -s ./real-esrgan.wts ./real-esrgan_f32.engine +// ./real-esrgan -d ./real-esrgan_f32.engine ../samples + +int main(int argc, char** argv) { + std::string wts_name = ""; + std::string engine_name = ""; + std::string img_dir; + if (!parse_args(argc, argv, wts_name, engine_name, img_dir)) { + std::cerr << "arguments not right!" << std::endl; + std::cerr << "./real-esrgan -s [.wts] [.engine] // serialize model to plan file" << std::endl; + std::cerr << "./real-esrgan -d [.engine] ../samples // deserialize plan file and run inference" << std::endl; + return -1; + } + + // create a model using the API directly and serialize it to a stream + if (!wts_name.empty()) { + IHostMemory* modelStream{ nullptr }; + APIToModel(BATCH_SIZE, &modelStream, wts_name); + assert(modelStream != nullptr); + std::ofstream p(engine_name, std::ios::binary); + if (!p) { + std::cerr << "could not open plan output file" << std::endl; + return -1; + } + p.write(reinterpret_cast(modelStream->data()), modelStream->size()); + delete modelStream; + return 0; + } + + // deserialize the .engine and run inference + std::ifstream file(engine_name, std::ios::binary); + if (!file.good()) { + std::cerr << "read " << engine_name << " error!" << std::endl; + return -1; + } + char *trtModelStream = nullptr; + size_t size = 0; + file.seekg(0, file.end); + size = file.tellg(); + file.seekg(0, file.beg); + trtModelStream = new char[size]; + assert(trtModelStream); + file.read(trtModelStream, size); + file.close(); + + std::vector file_names; + if (read_files_in_dir(img_dir.c_str(), file_names) < 0) { + std::cerr << "read_files_in_dir failed." << std::endl; + return -1; + } + + IRuntime* runtime = createInferRuntime(gLogger); + assert(runtime != nullptr); + ICudaEngine* engine = runtime->deserializeCudaEngine(trtModelStream, size); + assert(engine != nullptr); + IExecutionContext* context = engine->createExecutionContext(); + assert(context != nullptr); + delete[] trtModelStream; + assert(engine->getNbBindings() == 2); + void* buffers[2]; + // In order to bind the buffers, we need to know the names of the input and output tensors. + // Note that indices are guaranteed to be less than IEngine::getNbBindings() + const int inputIndex = engine->getBindingIndex(INPUT_BLOB_NAME); + const int outputIndex = engine->getBindingIndex(OUTPUT_BLOB_NAME); + assert(inputIndex == 0); + assert(outputIndex == 1); + + // Create GPU buffers on device + CUDA_CHECK(cudaMalloc(&buffers[inputIndex], BATCH_SIZE * INPUT_C * INPUT_H * INPUT_W * sizeof(uint8_t))); + CUDA_CHECK(cudaMalloc(&buffers[outputIndex], BATCH_SIZE * OUTPUT_SIZE * sizeof(uint8_t))); + + std::vector input(BATCH_SIZE * INPUT_H * INPUT_W * INPUT_C); + std::vector outputs(BATCH_SIZE * OUTPUT_SIZE); + + // Create stream + cudaStream_t stream; + CUDA_CHECK(cudaStreamCreate(&stream)); + + std::vector imgs_buffer(BATCH_SIZE); + for (int f = 0; f < (int)file_names.size(); f++) { + + for (int b = 0; b < BATCH_SIZE; b++) { + cv::Mat img = cv::imread(img_dir + "/" + file_names[f]); + if (img.empty()) continue; + memcpy(input.data() + b * INPUT_H * INPUT_W * INPUT_C, img.data, INPUT_H * INPUT_W * INPUT_C); + } + + CUDA_CHECK(cudaMemcpyAsync(buffers[inputIndex], input.data(), BATCH_SIZE * INPUT_C * INPUT_H * INPUT_W * sizeof(uint8_t), cudaMemcpyHostToDevice, stream)); + + // Run inference + auto start = std::chrono::system_clock::now(); + doInference(*context, stream, (void**)buffers, outputs.data(), BATCH_SIZE); + auto end = std::chrono::system_clock::now(); + std::cout << "inference time: " << std::chrono::duration_cast(end - start).count() << "ms" << std::endl; + } + + cv::Mat frame = cv::Mat(INPUT_H * OUT_SCALE, INPUT_W * OUT_SCALE, CV_8UC3, outputs.data()); + cv::imwrite("../_" + file_names[0] + ".png", frame); + + if (VISUALIZATION) { + cv::imshow("result : " + file_names[0], frame); + cv::waitKey(0); + } + + // Release stream and buffers + cudaStreamDestroy(stream); + CUDA_CHECK(cudaFree(buffers[inputIndex])); + CUDA_CHECK(cudaFree(buffers[outputIndex])); + // Destroy the engine + delete context; + delete engine; + delete runtime; +} \ No newline at end of file diff --git a/real-esrgan/utils.h b/real-esrgan/utils.h new file mode 100644 index 0000000..b4468d8 --- /dev/null +++ b/real-esrgan/utils.h @@ -0,0 +1,30 @@ +#ifndef TRTX_REAL_ESRGAN_UTILS_H_ +#define TRTX_REAL_ESRGAN_UTILS_H_ + +#include +#include + +static inline int read_files_in_dir(const char *p_dir_name, std::vector &file_names) { + DIR *p_dir = opendir(p_dir_name); + if (p_dir == nullptr) { + return -1; + } + + struct dirent* p_file = nullptr; + while ((p_file = readdir(p_dir)) != nullptr) { + if (strcmp(p_file->d_name, ".") != 0 && + strcmp(p_file->d_name, "..") != 0) { + //std::string cur_file_name(p_dir_name); + //cur_file_name += "/"; + //cur_file_name += p_file->d_name; + std::string cur_file_name(p_file->d_name); + file_names.push_back(cur_file_name); + } + } + + closedir(p_dir); + return 0; +} + +#endif // TRTX_YOLOV5_UTILS_H_ +