[CVCUDA] Vision Processor Python API and Tutorial (#1394)

* bind success

* bind success fix

* FDMat pybind, ResizeByShort pybind

* FDMat pybind, ResizeByShort pybind, remove initialized_

* override BindProcessorManager::Run in python is available

* PyProcessorManager done

* vision_pybind fix

* manager.py fix

* add tutorials

* remove Apply() bind

* remove Apply() bind and fix

* fix reviewed problem

* fix reviewed problem

* fix reviewed problem readme

* fix reviewed problem readme etc

* apply return outputs

* nits

* update readme

* fix FDMatbatch

* add op pybind: CenterCrop, Pad

* add op overload for pass FDMatBatch

---------

Co-authored-by: Wang Xinyu <shaywxy@gmail.com>

tutorials/vision_processor/README.md

@@ -0,0 +1,43 @@
+English | [中文](README_CN.md)
+
+# Vision Processor
+
+Vision Processor is used to implement model preprocessing, postprocessing, etc. The following 3rd party vision libraries are integrated:
+- OpenCV, general CPU image processing
+- FlyCV, mainly optimized for ARM CPU
+- CV-CUDA, for NVIDIA GPU
+
+## C++
+
+TODO(guxukai)
+
+## Python
+
+Python API, Currently supported operators are as follows:
+
+- ResizeByShort
+- NormalizeAndPermute
+
+Users can implement a image processing modules by inheriting the `PyProcessorManager` class. The base class `PyProcessorManager` implements GPU memory management, CUDA stream management, etc. Users only need to implement the apply() function by calling vision processors in this library and implements processing logic. For specific implementation, please refer to the demo code.
+
+### Demo
+
+- [Python Demo](python)
+
+### Performance comparison between CV-CUDA and OpenCV:
+
+CPU: Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz
+
+GPU: T4
+
+CUDA: 11.6
+
+Processing logic: Resize -> NormalizeAndPermute
+
+Warmup 100 rounds，tested 1000 rounds and get avg. latency.
+
+| Input Image Shape | Target shape | Batch Size | OpenCV | CV-CUDA | Gain |
+| ----------- | -- | ---------- | ------- | ------ | ------ |
+| 1920x1080   | 640x360 | 1 | 1.1572ms | 0.9067ms | 16.44% |
+| 1280x720    | 640x360 | 1 | 2.7551ms | 0.5296ms | 80.78% |
+| 360x240     | 640x360 | 1 | 3.3450ms | 0.2421ms | 92.76% |

tutorials/vision_processor/README_CN.md

@@ -0,0 +1,42 @@
+中文 | [English](README.md)
+# 多硬件图像处理库
+
+多硬件图像处理库（Vision Processor）可用于实现模型的预处理、后处理等图像操作，底层封装了多个第三方图像处理库，包括：
+- OpenCV，用于通用CPU图像处理
+- FlyCV，主要针对ARM CPU加速
+- CV-CUDA，用于NVIDIA GPU
+
+## C++
+
+待编写
+
+## Python
+
+Python API目前支持的算子如下：
+
+- ResizeByShort
+- NormalizeAndPermute
+
+用户可通过继承PyProcessorManager类，实现自己的图像处理模块。基类PyProcessorManager实现了GPU内存管理、CUDA stream管理等，用户仅需要实现apply()函数，在其中调用多硬件图像处理库中的算子、实现处理逻辑即可，具体实现可参考示例代码。
+
+### 示例代码
+
+- [Python示例](python)
+
+### CV-CUDA与OpenCV性能对比
+
+CPU: Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz
+
+GPU: T4
+
+CUDA: 11.6
+
+Processing logic: Resize -> NormalizeAndPermute
+
+Warmup 100 rounds，tested 1000 rounds and get avg. latency.
+
+| Input Image Shape | Target shape | Batch Size | OpenCV | CV-CUDA | Gain |
+| ----------- | -- | ---------- | ------- | ------ | ------ |
+| 1920x1080   | 640x360 | 1 | 1.1572ms | 0.9067ms | 16.44% |
+| 1280x720    | 640x360 | 1 | 2.7551ms | 0.5296ms | 80.78% |
+| 360x240     | 640x360 | 1 | 3.3450ms | 0.2421ms | 92.76% |

tutorials/vision_processor/python/README.md

@@ -0,0 +1,19 @@
+English | [中文](README_CN.md)
+
+# Preprocessor Python Demo
+
+1. [build FastDeploy（Python）](../../../docs/cn/build_and_install), or download[FastDeploy prebuilt library（Python）](../../../docs/cn/build_and_install/download_prebuilt_libraries.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
+
+# Run the Demo
+
+# OpenCV
+python preprocess.py
+
+# CV-CUDA
+python preprocess.py --use_cvcuda True
+```

tutorials/vision_processor/python/README_CN.md

@@ -0,0 +1,19 @@
+中文 | [English](README.md)
+
+# Preprocessor Python 示例代码
+
+1. [编译FastDeploy（Python）](../docs/cn/build_and_install), 或直接下载[FastDeploy预编译库（Python）](../docs/cn/build_and_install/download_prebuilt_libraries.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
+
+# 运行示例代码
+
+# OpenCV
+python preprocess.py
+
+# CV-CUDA
+python preprocess.py --use_cvcuda True
+```

tutorials/vision_processor/python/preprocess.py

@@ -0,0 +1,89 @@
+import fastdeploy as fd
+import cv2
+
+from fastdeploy.vision.common.manager import PyProcessorManager
+
+
+def parse_arguments():
+    import argparse
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--use_cvcuda",
+        required=False,
+        type=bool,
+        help="Use CV-CUDA in preprocess")
+    return parser.parse_args()
+
+
+# define CustomProcessor
+class CustomProcessor(PyProcessorManager):
+    def __init__(self) -> None:
+        super().__init__()
+        # create op
+        hw = [500, 500]
+        self.resize_op = fd.C.vision.processors.ResizeByShort(100, 1, True, hw)
+
+        mean = [0.485, 0.456, 0.406]
+        std = [0.229, 0.224, 0.225]
+        is_scale = True
+        min = []
+        max = []
+        swap_rb = False
+        self.normalize_permute_op = fd.C.vision.processors.NormalizeAndPermute(
+            mean, std, is_scale, min, max, swap_rb)
+
+        width = 50
+        height = 50
+        self.centercrop_op = fd.C.vision.processors.CenterCrop(width, height)
+
+        top = 5
+        bottom = 5
+        left = 5
+        right = 5
+        pad_value = [225, 225, 225]
+        self.pad_op = fd.C.vision.processors.Pad(top, bottom, left, right,
+                                                 pad_value)
+
+    def apply(self, image_batch):
+        outputs = []
+        self.resize_op(image_batch)
+        self.centercrop_op(image_batch)
+        self.pad_op(image_batch)
+        self.normalize_permute_op(image_batch)
+
+        for i in range(len(image_batch.mats)):
+            outputs.append(image_batch.mats[i])
+
+        return outputs
+
+
+if __name__ == "__main__":
+
+    # read jpg
+    im1 = cv2.imread('ILSVRC2012_val_00000010.jpeg')
+    im2 = cv2.imread('ILSVRC2012_val_00000010.jpeg')
+
+    mat1 = fd.C.vision.FDMat()
+    mat1.from_numpy(im1)
+    mat2 = fd.C.vision.FDMat()
+    mat2.from_numpy(im2)
+    images = [mat1, mat2]
+
+    args = parse_arguments()
+    # creae processor
+    preprocessor = CustomProcessor()
+
+    # use CV-CUDA
+    if args.use_cvcuda:
+        preprocessor.use_cuda(True, -1)
+
+    # show input
+    for i in range(len(images)):
+        images[i].print_info('images' + str(i) + ': ')
+
+    # run the Processer with CVCUDA
+    outputs = preprocessor(images)
+
+    # show output
+    for i in range(len(outputs)):
+        outputs[i].print_info('outputs' + str(i) + ': ')