[nvJPEG] Integrate nvJPEG decoder (#1288)

* nvjpeg cmake

* add common decoder, nvjpeg decoder and add image name predict api

* ppclas support nvjpeg decoder

* remove useless comments

* image decoder support opencv

* nvjpeg decode fallback to opencv

* fdtensor add nbytes_allocated

* single image decode api

* fix bug

* add pybind

* ignore nvjpeg on jetson

* fix cmake in

* predict on fdmat

* remove image names predict api, add image decoder tutorial

* Update __init__.py

* fix pybind

tutorials/image_decoder/README.md

@@ -0,0 +1,16 @@
+English | [中文](README_CN.md)
+
+# Image Decoder
+
+Currently, we support below image decoder libs：
+- OpenCV
+- nvJPEG (Needs NVIDIA GPU, doesn't support Jetson)
+
+## Example
+
+- [C++ Example](cpp)
+- Python API(WIP)
+
+## nvJPEG vs. OpenCV performance benchmark
+
+Refer to: https://github.com/PaddlePaddle/FastDeploy/pull/1288#issuecomment-1427749772

tutorials/image_decoder/README_CN.md

@@ -0,0 +1,16 @@
+简体中文 | [English](README.md)
+
+# Image Decoder
+
+图片解码库，目前支持以下图片解码库：
+- OpenCV
+- nvJPEG (依赖NVIDIA GPU，不支持Jetson)
+
+## 示例代码
+
+- [C++示例](cpp)
+- Python API仍在开发中...
+
+## nvJPEG和OpenCV性能对比数据
+
+参见：https://github.com/PaddlePaddle/FastDeploy/pull/1288#issuecomment-1427749772

tutorials/image_decoder/cpp/CMakeLists.txt

@@ -0,0 +1,11 @@
+PROJECT(image_decoder C CXX)
+CMAKE_MINIMUM_REQUIRED (VERSION 3.10)
+
+option(FASTDEPLOY_INSTALL_DIR "Path of downloaded fastdeploy sdk.")
+
+include(${FASTDEPLOY_INSTALL_DIR}/FastDeploy.cmake)
+
+include_directories(${FASTDEPLOY_INCS})
+
+add_executable(image_decoder ${PROJECT_SOURCE_DIR}/main.cc)
+target_link_libraries(image_decoder ${FASTDEPLOY_LIBS})

tutorials/image_decoder/cpp/README.md

@@ -0,0 +1,22 @@
+English | [中文](README_CN.md)
+
+# Image Decoder C++ Example
+
+1. [Build FastDeploy](../docs/cn/build_and_install) or download [FastDeploy prebuilt library](../docs/cn/build_and_install/download_prebuilt_libraries.md)
+
+2. Build example
+```bash
+mkdir build
+cd build
+
+# [PATH-TO-FASTDEPLOY] is the install directory of FastDeploy
+cmake .. -DFASTDEPLOY_INSTALL_DIR=[PATH-TO-FASTDEPLOY]
+make -j
+
+# Download the test image
+wget https://gitee.com/paddlepaddle/PaddleClas/raw/release/2.4/deploy/images/ImageNet/ILSVRC2012_val_00000010.jpeg
+
+# OpenCV decoder
+./image_decoder ILSVRC2012_val_00000010.jpeg 0
+# nvJPEG
+./image_decoder ILSVRC2012_val_00000010.jpeg 1

tutorials/image_decoder/cpp/README_CN.md

@@ -0,0 +1,22 @@
+简体中文 | [English](README.md)
+
+# Image Decoder C++示例
+
+1. [编译FastDeploy](../docs/cn/build_and_install), 或直接下载[FastDeploy预编译库](../docs/cn/build_and_install/download_prebuilt_libraries.md)
+
+2. 编译示例
+```bash
+mkdir build
+cd build
+
+# [PATH-TO-FASTDEPLOY]需替换为FastDeploy的安装路径
+cmake .. -DFASTDEPLOY_INSTALL_DIR=[PATH-TO-FASTDEPLOY]
+make -j
+
+# 下载测试图片
+wget https://gitee.com/paddlepaddle/PaddleClas/raw/release/2.4/deploy/images/ImageNet/ILSVRC2012_val_00000010.jpeg
+
+# OpenCV解码
+./image_decoder ILSVRC2012_val_00000010.jpeg 0
+# nvJPEG
+./image_decoder ILSVRC2012_val_00000010.jpeg 1

tutorials/image_decoder/cpp/main.cc

@@ -0,0 +1,57 @@
+#include "fastdeploy/vision/common/image_decoder/image_decoder.h"
+
+namespace fdvis = fastdeploy::vision;
+namespace fd = fastdeploy;
+
+void OpenCVImageDecode(const std::string& img_name) {
+  fdvis::FDMat mat;
+  auto img_decoder = new fdvis::ImageDecoder();
+  img_decoder->Decode(img_name, &mat);
+  mat.PrintInfo("");
+  delete img_decoder;
+}
+
+void NvJpegImageDecode(const std::string& img_name) {
+  std::vector<fdvis::FDMat> mats(1);
+  std::vector<fastdeploy::FDTensor> caches(1);
+
+  cudaStream_t stream;
+  cudaStreamCreate(&stream);
+  // For nvJPEG decoder, we need set stream and output cache for the FDMat
+  for (size_t i = 0; i < mats.size(); i++) {
+    mats[i].output_cache = &caches[i];
+    mats[i].SetStream(stream);
+  }
+  auto img_decoder = new fdvis::ImageDecoder(fdvis::ImageDecoderLib::NVJPEG);
+
+  // This is batch decode API, for single image decode API,
+  // please refer to OpenCVImageDecode()
+  img_decoder->BatchDecode({img_name}, &mats);
+
+  for (size_t i = 0; i < mats.size(); i++) {
+    std::cout << "Mat type: " << mats[i].mat_type << ", "
+              << "DataType=" << mats[i].Type() << ", "
+              << "Channel=" << mats[i].Channels() << ", "
+              << "Height=" << mats[i].Height() << ", "
+              << "Width=" << mats[i].Width() << std::endl;
+  }
+
+  cudaStreamDestroy(stream);
+}
+
+int main(int argc, char* argv[]) {
+  if (argc < 3) {
+    std::cout << "Usage: image_decoder path/to/image run_option, "
+                 "e.g ./image_decoder ./test.jpeg 0"
+              << std::endl;
+    std::cout << "Run_option 0: OpenCV; 1: nvJPEG " << std::endl;
+    return -1;
+  }
+
+  if (std::atoi(argv[2]) == 0) {
+    OpenCVImageDecode(argv[1]);
+  } else if (std::atoi(argv[2]) == 1) {
+    NvJpegImageDecode(argv[1]);
+  }
+  return 0;
+}