[CVCUDA] Update CV-CUDA to v0.2.1, add vision processor C++ tutorial (#1678)

* update cvcuda 0.2.0 -> 0.2.1

* add cpp tutorials demo

* fix reviewed problem

FastDeploy.cmake.in

@@ -321,6 +321,8 @@ if(ENABLE_VISION)
     find_library(CVCUDA_LIB cvcuda ${CMAKE_CURRENT_LIST_DIR}/third_libs/install/cvcuda/lib NO_DEFAULT_PATH)
     find_library(NVCV_TYPES_LIB nvcv_types ${CMAKE_CURRENT_LIST_DIR}/third_libs/install/cvcuda/lib NO_DEFAULT_PATH)
     list(APPEND FASTDEPLOY_LIBS ${CVCUDA_LIB} ${NVCV_TYPES_LIB})
+    list(APPEND FASTDEPLOY_INCS ${CMAKE_CURRENT_LIST_DIR}/third_libs/install/cvcuda/include NO_DEFAULT_PATH)
+    add_definitions(-DENABLE_CVCUDA)
   endif()
 
 endif()

cmake/cvcuda.cmake

@@ -24,10 +24,10 @@ if(NOT (CMAKE_HOST_SYSTEM_PROCESSOR MATCHES "x86_64"))
   message(FATAL_ERROR "CV-CUDA only support x86_64.")
 endif()
 
-set(CVCUDA_LIB_URL https://github.com/CVCUDA/CV-CUDA/releases/download/v0.2.0-alpha/nvcv-lib-0.2.0_alpha-cuda11-x86_64-linux.tar.xz)
-set(CVCUDA_LIB_FILENAME nvcv-lib-0.2.0_alpha-cuda11-x86_64-linux.tar.xz)
-set(CVCUDA_DEV_URL https://github.com/CVCUDA/CV-CUDA/releases/download/v0.2.0-alpha/nvcv-dev-0.2.0_alpha-cuda11-x86_64-linux.tar.xz)
-set(CVCUDA_DEV_FILENAME nvcv-dev-0.2.0_alpha-cuda11-x86_64-linux.tar.xz)
+set(CVCUDA_LIB_URL https://github.com/CVCUDA/CV-CUDA/releases/download/v0.2.1-alpha/nvcv-lib-0.2.1_alpha-cuda11-x86_64-linux.tar.xz)
+set(CVCUDA_LIB_FILENAME nvcv-lib-0.2.1_alpha-cuda11-x86_64-linux.tar.xz)
+set(CVCUDA_DEV_URL https://github.com/CVCUDA/CV-CUDA/releases/download/v0.2.1-alpha/nvcv-dev-0.2.1_alpha-cuda11-x86_64-linux.tar.xz)
+set(CVCUDA_DEV_FILENAME nvcv-dev-0.2.1_alpha-cuda11-x86_64-linux.tar.xz)
 
 download_and_decompress(${CVCUDA_LIB_URL} ${CMAKE_CURRENT_BINARY_DIR}/${CVCUDA_LIB_FILENAME} ${THIRD_PARTY_PATH}/cvcuda)
 download_and_decompress(${CVCUDA_DEV_URL} ${CMAKE_CURRENT_BINARY_DIR}/${CVCUDA_DEV_FILENAME} ${THIRD_PARTY_PATH}/cvcuda)
@@ -40,4 +40,3 @@ execute_process(COMMAND cp -r ${THIRD_PARTY_PATH}/cvcuda/opt/nvidia/cvcuda0/incl
 link_directories(${THIRD_PARTY_PATH}/install/cvcuda/lib)
 include_directories(${THIRD_PARTY_PATH}/install/cvcuda/include)
 
-set(CMAKE_CXX_STANDARD 17)

docs/api_docs/python/index.rst

@@ -38,3 +38,10 @@ FastDeploy
 
     runtime.md
     runtime_option.md
+
+..  toctree::
+    :caption: Vision Processor (图像预处理库)
+    :maxdepth: 2
+    :titlesonly:
+
+    vision_processor.md

fastdeploy/vision/common/processors/cast.cc

@@ -96,7 +96,7 @@ bool Cast::ImplByCvCuda(FDMat* mat) {
   auto dst_tensor =
       CreateCvCudaTensorWrapData(*(mat->output_cache), mat->layout);
 
-  cvcuda_convert_op_(mat->Stream(), src_tensor, dst_tensor, 1.0f, 0.0f);
+  cvcuda_convert_op_(mat->Stream(), *src_tensor, *dst_tensor, 1.0f, 0.0f);
 
   mat->SetTensor(mat->output_cache);
   mat->mat_type = ProcLib::CVCUDA;

fastdeploy/vision/common/processors/center_crop.cc

@@ -70,7 +70,7 @@ bool CenterCrop::ImplByCvCuda(FDMat* mat) {
   int offset_x = static_cast<int>((mat->Width() - width_) / 2);
   int offset_y = static_cast<int>((mat->Height() - height_) / 2);
   NVCVRectI crop_roi = {offset_x, offset_y, width_, height_};
-  cvcuda_crop_op_(mat->Stream(), src_tensor, dst_tensor, crop_roi);
+  cvcuda_crop_op_(mat->Stream(), *src_tensor, *dst_tensor, crop_roi);
 
   mat->SetTensor(mat->output_cache);
   mat->SetWidth(width_);

fastdeploy/vision/common/processors/cvcuda_utils.cc

@@ -43,8 +43,8 @@ nvcv::ImageFormat CreateCvCudaImageFormat(FDDataType type, int channel,
   return nvcv::FMT_BGRf32;
 }
 
-nvcv::TensorWrapData CreateCvCudaTensorWrapData(const FDTensor& tensor,
-                                                Layout layout) {
+std::shared_ptr<nvcv::TensorWrapData> CreateCvCudaTensorWrapData(
+    const FDTensor& tensor, Layout layout) {
   FDASSERT(tensor.shape.size() == 3,
            "When create CVCUDA tensor from FD tensor,"
            "tensor shape should be 3-Dim,");
@@ -76,7 +76,7 @@ nvcv::TensorWrapData CreateCvCudaTensorWrapData(const FDTensor& tensor,
   nvcv::TensorDataStridedCuda tensor_data(
       nvcv::TensorShape{req.shape, req.rank, req.layout},
       nvcv::DataType{req.dtype}, buf);
-  return nvcv::TensorWrapData(tensor_data);
+  return std::make_shared<nvcv::TensorWrapData>(tensor_data, nullptr);
 }
 
 void* GetCvCudaTensorDataPtr(const nvcv::TensorWrapData& tensor) {

fastdeploy/vision/common/processors/cvcuda_utils.h

@@ -27,7 +27,7 @@ namespace vision {
 
 nvcv::ImageFormat CreateCvCudaImageFormat(FDDataType type, int channel,
                                           bool interleaved = true);
-nvcv::TensorWrapData CreateCvCudaTensorWrapData(const FDTensor& tensor,
+std::shared_ptr<nvcv::TensorWrapData> CreateCvCudaTensorWrapData(const FDTensor& tensor,
                                                 Layout layout = Layout::HWC);
 void* GetCvCudaTensorDataPtr(const nvcv::TensorWrapData& tensor);
 nvcv::ImageWrapData CreateImageWrapData(const FDTensor& tensor);

fastdeploy/vision/common/processors/hwc2chw.cc

@@ -75,7 +75,7 @@ bool HWC2CHW::ImplByCvCuda(FDMat* mat) {
   auto dst_tensor =
       CreateCvCudaTensorWrapData(*(mat->output_cache), Layout::CHW);
 
-  cvcuda_reformat_op_(mat->Stream(), src_tensor, dst_tensor);
+  cvcuda_reformat_op_(mat->Stream(), *src_tensor, *dst_tensor);
 
   mat->layout = Layout::CHW;
   mat->SetTensor(mat->output_cache);

fastdeploy/vision/common/processors/pad.cc

@@ -133,7 +133,7 @@ bool Pad::ImplByCvCuda(FDMat* mat) {
                             "output_cache", Device::GPU);
   auto dst_tensor = CreateCvCudaTensorWrapData(*(mat->output_cache));
 
-  cvcuda_pad_op_(mat->Stream(), src_tensor, dst_tensor, top_, left_,
+  cvcuda_pad_op_(mat->Stream(), *src_tensor, *dst_tensor, top_, left_,
                  NVCV_BORDER_CONSTANT, value);
 
   mat->SetTensor(mat->output_cache);

fastdeploy/vision/common/processors/pad_to_size.cc

@@ -206,7 +206,7 @@ static bool PadHWCByCvCuda(cvcuda::CopyMakeBorder& pad_op, FDMat* mat,
                             "output_cache", Device::GPU);
   auto dst_tensor = CreateCvCudaTensorWrapData(*(mat->output_cache));
 
-  pad_op(mat->Stream(), src_tensor, dst_tensor, 0, 0, NVCV_BORDER_CONSTANT,
+  pad_op(mat->Stream(), *src_tensor, *dst_tensor, 0, 0, NVCV_BORDER_CONSTANT,
          border_value);
 
   mat->SetTensor(mat->output_cache);
@@ -238,8 +238,8 @@ static bool PadCHWByCvCuda(cvcuda::CopyMakeBorder& pad_op, FDMat* mat,
                         input->device, input->device_id);
     auto dst_tensor = CreateCvCudaTensorWrapData(dst);
 
-    pad_op(mat->Stream(), src_tensor, dst_tensor, 0, 0, NVCV_BORDER_CONSTANT,
-           border_value);
+    pad_op(mat->Stream(), (*src_tensor), (*dst_tensor), 0, 0,
+           NVCV_BORDER_CONSTANT, border_value);
   }
   mat->SetTensor(mat->output_cache);
   mat->mat_type = ProcLib::CVCUDA;

fastdeploy/vision/common/processors/resize.cc

@@ -146,7 +146,7 @@ bool Resize::ImplByCvCuda(FDMat* mat) {
   auto dst_tensor = CreateCvCudaTensorWrapData(*(mat->output_cache));
 
   // CV-CUDA Interp value is compatible with OpenCV
-  cvcuda_resize_op_(mat->Stream(), src_tensor, dst_tensor,
+  cvcuda_resize_op_(mat->Stream(), *src_tensor, *dst_tensor,
                     CreateCvCudaInterp(interp_));
 
   mat->SetTensor(mat->output_cache);

fastdeploy/vision/common/processors/resize_by_short.cc

@@ -95,7 +95,7 @@ bool ResizeByShort::ImplByCvCuda(FDMat* mat) {
                             "output_cache", Device::GPU);
   auto dst_tensor = CreateCvCudaTensorWrapData(*(mat->output_cache));
 
-  cvcuda_resize_op_(mat->Stream(), src_tensor, dst_tensor,
+  cvcuda_resize_op_(mat->Stream(), *src_tensor, *dst_tensor,
                     CreateCvCudaInterp(interp_));
 
   mat->SetTensor(mat->output_cache);

fastdeploy/vision/common/processors/stride_pad.cc

@@ -167,7 +167,7 @@ bool StridePad::ImplByCvCuda(FDMat* mat) {
                             "output_cache", Device::GPU);
   auto dst_tensor = CreateCvCudaTensorWrapData(*(mat->output_cache));
 
-  cvcuda_pad_op_(mat->Stream(), src_tensor, dst_tensor, 0, 0,
+  cvcuda_pad_op_(mat->Stream(), *src_tensor, *dst_tensor, 0, 0,
                  NVCV_BORDER_CONSTANT, value);
 
   mat->SetTensor(mat->output_cache);

tutorials/vision_processor/cpp/CMakeLists.txt

@@ -0,0 +1,11 @@
+PROJECT(preprocessor_demo 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(preprocessor_demo ${PROJECT_SOURCE_DIR}/main.cc)
+target_link_libraries(preprocessor_demo ${FASTDEPLOY_LIBS})

tutorials/vision_processor/cpp/README.md

@@ -0,0 +1,27 @@
+English | [中文](README_CN.md)
+
+# Preprocessor Python Demo
+
+1. Compile FastDeploy and open CV-CUDA option
+    > [Compile FastDeploy](../../../docs/cn/build_and_install/gpu.md)  
+    > [Open CV-CUDA option](../../../docs/cn/faq/use_cv_cuda.md)
+
+2. Run the demo
+```bash
+# Download the test image
+wget https://gitee.com/paddlepaddle/PaddleClas/raw/release/2.4/deploy/images/ImageNet/ILSVRC2012_val_00000010.jpeg
+
+# Compile the Demo
+mkdir build
+cd build
+cmake .. -DFASTDEPLOY_INSTALL_DIR=${PWD}/../../../../build/compiled_fastdeploy_sdk/ # if build sdk in `FastDeploy/build/compiled_fastdeploy_sdk`
+make -j
+
+# Run the demo
+
+# Use OpenCV
+./preprocessor_demo ILSVRC2012_val_00000010.jpeg 0
+
+# Use CV-CUDA
+./preprocessor_demo ILSVRC2012_val_00000010.jpeg 1
+```

tutorials/vision_processor/cpp/README_CN.md

@@ -0,0 +1,27 @@
+中文 | [English](README.md)
+
+# Preprocessor Python 示例代码
+
+1. 编译FastDeploy并开启CV-CUDA选项
+    > [编译FastDeploy](../../../docs/cn/build_and_install/gpu.md)  
+    > [开启CV-CUDA选项](../../../docs/cn/faq/use_cv_cuda.md)
+
+2. 运行示例代码
+```bash
+# 下载测试图片
+wget https://gitee.com/paddlepaddle/PaddleClas/raw/release/2.4/deploy/images/ImageNet/ILSVRC2012_val_00000010.jpeg
+
+# 编译示例代码
+mkdir build
+cd build
+cmake .. -DFASTDEPLOY_INSTALL_DIR=${PWD}/../../../../build/compiled_fastdeploy_sdk/ # 若编译FastDeploy在其他文件夹，请替换为相应的sdk路径
+make -j
+
+# 运行示例代码
+
+# 使用OpenCV处理图片
+./preprocessor_demo ILSVRC2012_val_00000010.jpeg 0
+
+# 使用CV-CUDA处理图片
+./preprocessor_demo ILSVRC2012_val_00000010.jpeg 1
+```

tutorials/vision_processor/cpp/main.cc

@@ -0,0 +1,78 @@
+#include "fastdeploy/vision.h"
+#include "fastdeploy/vision/common/processors/manager.h"
+#include "fastdeploy/vision/common/processors/transform.h"
+
+namespace fd = fastdeploy;
+
+// Define our custom processor
+class CustomPreprocessor : public fd::vision::ProcessorManager {
+ public:
+  explicit CustomPreprocessor(){};
+  ~CustomPreprocessor(){};
+
+  virtual bool Apply(fd::vision::FDMatBatch* image_batch,
+                     std::vector<fd::FDTensor>* outputs);
+
+ private:
+  // Create op
+  int width = 160;
+  int height = 160;
+  std::shared_ptr<fd::vision::Resize> resize_op =
+      std::make_shared<fd::vision::Resize>(width, height, -1.0, -1.0, 1, false);
+  std::shared_ptr<fd::vision::CenterCrop> crop =
+      std::make_shared<fd::vision::CenterCrop>(50, 50);
+  std::vector<float> mean = {0.485f, 0.456f, 0.406f};
+  std::vector<float> std = {0.229f, 0.224f, 0.225f};
+  std::shared_ptr<fd::vision::Normalize> normalize =
+      std::make_shared<fd::vision::Normalize>(mean, std);
+};
+
+// Implement our custom processor's Apply() method
+bool CustomPreprocessor::Apply(fd::vision::FDMatBatch* image_batch,
+                               std::vector<fd::FDTensor>* outputs) {
+  // Use op to transform the images
+  bool resize_ret = (*resize_op)(&(image_batch->mats->at(0)));
+  bool crop_ret = (*crop)(image_batch);
+  bool normalize_ret = (*normalize)(image_batch);
+
+  outputs->resize(1);
+  fd::FDTensor* tensor = image_batch->Tensor();
+  (*outputs)[0].SetExternalData(tensor->Shape(), tensor->Dtype(),
+                                tensor->Data(), tensor->device,
+                                tensor->device_id);
+  return true;
+}
+
+int main(int argc, char* argv[]) {
+  if (argc < 2) {
+    std::cout << "Usage: ./preprocessor_demo path/to/image run_option, "
+                 "e.g ././preprocessor_demo ./test.jpeg 0"
+              << std::endl;
+    std::cout << "Run_option 0: OpenCV; 1: CV-CUDA " << std::endl;
+    return -1;
+  }
+
+  // Prepare input images
+  auto im = cv::imread(argv[1]);
+  std::vector<cv::Mat> images = {im, im};
+  std::vector<fd::vision::FDMat> mats = fd::vision::WrapMat(images);
+  std::vector<fd::FDTensor> outputs;
+
+  // CustomPreprocessor processor;
+  CustomPreprocessor processor = CustomPreprocessor();
+
+  // Use CV-CUDA if parameter passed and detected
+  if (std::atoi(argv[2]) == 1) {
+    processor.UseCuda(true, 0);
+  }
+
+  // Run the processor
+  bool ret = processor.Run(&mats, &outputs);
+
+  // Print output
+  for (int i = 0; i < outputs.size(); i++) {
+    outputs[i].PrintInfo("out");
+  }
+
+  return 0;
+}