[CUDA] 與 OpenCV 共舞

取得 CUDA 版的 OpenCV 後, 使用 OpenCV 前, 先建立環境變數 OPENCV_DIR, 使其指向 opencv 目錄 ( C:\opencv\build\ ):

建立名稱為 OpenCvExample 的 CUDA 專案, CUDA精靈會自動建立一個 kernel.cu 檔 :

在專案中增加 main.cpp  與 OpenCvExample.h檔:

專案預設將編譯為 x64 版本, 因此在設定專案的 VC++ 的 include 與 lib 目錄屬性, 將 include 設定為 $(OPENCV_DIR)\include, lib目錄設定為: $(OPENCV_DIR)\x64\vc15\lib :

設定 VC 的連結程式庫名稱, 加入  opencv_world343d.lib (Debug版) 或 opencv_world343.lib (Release版):

本例只有一個很簡單的複製功能, 主要是展示如何在 kernel 存取 Mat 物件的影像資料 data,  這樣便能利用 OpenCV 的檔案讀寫功能來進行檔案處理, 配合改變 kernel 函數即可做出不同的影像處理結果. 專案程式的函數說明: 

函 數	說  明
main()	1.      建立 srcMat 並讀取影像檔 2.      建立 destMat 準備放置處理結果 3.      呼叫在 kernel.cu的處理函數 CopyImage() 4.      顯示處理結果
CopyImage()	1.      準備 CUDA 的 device memory 2.      呼叫 kernel 函數 copy() 3.      複製處理結果至輸出 dest
__global__ void copy()	平行處理複製來源影像 src至輸出影像 dest

3個程式檔案的內容如下 :

OpenCvExample.h :

#pragma once #include "cuda_runtime.h" #include "device_launch_parameters.h" #include <opencv2/core.hpp> #include <opencv2/imgcodecs.hpp> #include <opencv2/highgui.hpp> #include "opencv2/imgproc.hpp" #include <opencv2/core/cuda.hpp> #include <Windows.h> using namespace cv; using namespace cuda; __global__ void copy(byte *dest, byte *src, int channels); void CopyImage(Mat *dest, Mat *src);

kernel.cu :

#include "OpenCvExample.h" #define WindowRadius   1 __global__ void copy(byte *dest, byte *src, int channels) {      int p = (blockIdx.x*blockDim.x + threadIdx.x)*channels;      for (int c = 0; c < channels; c++)      {          dest[p + c] = src[p + c];      } } void CopyImage(Mat *dest, Mat *src) {      byte *dev_src = 0;      byte *dev_dest = 0;      size_t pixelBytes = src->total() *src->elemSize();      cudaError_t cudaStatus;      // Choose which GPU to run on      cudaStatus = cudaSetDevice(0);      if (cudaStatus != cudaSuccess) {          fprintf(stderr, "cudaSetDevice failed!");          return;      }      // Allocate GPU buffers      cudaStatus = cudaMalloc((void**)&dev_src, pixelBytes);      if (cudaStatus != cudaSuccess) {          fprintf(stderr, "cudaMalloc failed!");          return;      }      cudaStatus = cudaMalloc((void**)&dev_dest, pixelBytes);      if (cudaStatus != cudaSuccess) {          fprintf(stderr, "cudaMalloc failed!");          return;      }      // Copy from host memory to GPU buffers.      cudaStatus = cudaMemcpy(dev_src, src->data, pixelBytes, cudaMemcpyHostToDevice);      if (cudaStatus != cudaSuccess) {          fprintf(stderr, "cudaMemcpy failed!");          return;      }      int Channel = src->channels();      copy<< <src->rows, src->cols >> > (dev_dest, dev_src, Channel);      // Check for any errors launching the kernel      cudaStatus = cudaGetLastError();      if (cudaStatus != cudaSuccess) {          fprintf(stderr, "addKernel launch failed: %s\n", cudaGetErrorString(cudaStatus));          return;      }      // cudaDeviceSynchronize waits for the kernel to finish      cudaStatus = cudaDeviceSynchronize();      if (cudaStatus != cudaSuccess) {          fprintf(stderr, "cudaDeviceSynchronize returned error code %d after launching addKernel!\n", cudaStatus);          return;      }      // Copy output from GPU buffer to host memory.      cudaStatus = cudaMemcpy(dest->data, dev_dest, pixelBytes, cudaMemcpyDeviceToHost);      if (cudaStatus != cudaSuccess) {          fprintf(stderr, "cudaMemcpy failed!");          return;      }      //check result      /*      bool ok = true;      uint32_t Width = src->cols;      uint32_t Height = src->rows;      for (uint32_t x = 0; x < Width; x++)      {          for (uint32_t y = 0; y < Height; y++)          {               for (int c = 0; c < Channel; c++) {                   if (dest->data[(y*Width + x)*Channel + c] != src->data[(y*Width + x)*Channel + c])                   {                        ok = false;                        break;                   }               }          }      }      fprintf(stderr, "ok=%s\n",ok?"ok":"fail");      */      // All done, reset the device      cudaStatus = cudaDeviceReset(); if (cudaStatus != cudaSuccess) {          fprintf(stderr, "cudaDeviceReset failed!");          return;      } }

main.cpp :

#include <stdio.h> #include <iostream> #include "OpenCvExample.h" using namespace std; int main(int argc, char** argv) {      cudaError_t cudaStatus;      char *filename = "D:\\LENA.JPG";      // Load source image      Mat srcMat = imread(filename, IMREAD_COLOR); // GRAYSCALE); // Read the file and convert to grayscale      if (srcMat.empty()) // Check for invalid input      {          cout << "Could not open or find the image" << std::endl;          return 1;      }      //create output Mat of the same size as srcMt      Mat destMat;      destMat.create(srcMat.rows, srcMat.cols, srcMat.type());      // Run the parallel algorithm which should be in the .cu file      CopyImage(&destMat, &srcMat);      namedWindow("Source window", WINDOW_AUTOSIZE); // Create a window for display.      imshow("Source window", srcMat); // Show our image inside it.      namedWindow("Dest window", WINDOW_AUTOSIZE); // Create a window for display.      imshow("Dest window", destMat); // Show our image inside it.      waitKey(0); // Wait for a keystroke in the window      return 0; }

 執行結果畫面 :