这个问题之前已经被问过，但是提问者没有提供足够的信息并且没有回答，我对这个程序很好奇。
我正在尝试使用opencv和cuda库进行sobel边缘检测，X方向的sobel内核是
-1 0 1
-2 0 2
-1 0 1
我的项目中有 3 个文件
main.cpp
CudaKernel.cu
CudaKernel.h
主文件
#include <stdlib.h>
#include <iostream>
#include <string.h>
#include <Windows.h>
#include <opencv2\core\core.hpp>
#include <opencv2\highgui\highgui.hpp>
#include <opencv2\gpu\gpu.hpp>
#include <cuda_runtime.h>
#include <cuda_gl_interop.h>
#include "CudaKernel.h"
using namespace cv;
using namespace std;
int main(int argc, char** argv)
{
IplImage* image;
try
{
image = cvLoadImage("4555472_460s.jpg", CV_LOAD_IMAGE_GRAYSCALE);
gpu::DeviceInfo info = gpu::getDevice();
cout << info.name() << endl;
cout << "Stream Processor : "<< info.multiProcessorCount() << endl;
cout << "Total Graphic Memory :" << info.totalMemory()/1048576 << " MB" << endl;
}
catch (const cv::Exception* ex)
{
cout << "Error: " << ex->what() << endl;
}
if(!image )
{
cout << "Could not open or find the image" << std::endl ;
return -1;
}
IplImage* image2=cvCreateImage(cvGetSize(image),IPL_DEPTH_32F,image->nChannels);
IplImage* image3=cvCreateImage(cvGetSize(image),IPL_DEPTH_32F,image->nChannels);
unsigned char * pseudo_input=(unsigned char *)image->imageData;
float *output=(float*)image2->imageData;
float *input=(float*)image3->imageData;
int s=image->widthStep/sizeof(float);
for(int w=0;w<=(image->height);w++)
for(int h=0;h<(image->width*image->nChannels);h++)
{
input[w*s+h]= pseudo_input[w*s+h];
}
Pixel *fagget = (unsigned char*) image->imageData;
kernelcall(input, output, image->width,image->height, image->widthStep);
// cv::namedWindow( "Display window", CV_WINDOW_AUTOSIZE );// Create a window for display.
cvShowImage( "Original Image", image ); // Show our image inside it.
cvShowImage("Sobeled Image", image2);
waitKey(0); // Wait for a keystroke in the window
return 0;
}
CudaKernel.cu
#include<cuda.h>
#include<iostream>
#include "CudaKernel.h"
using namespace std;
#define CudaSafeCall( err ) __cudaSafeCall( err, __FILE__, __LINE__ )
#define CudaCheckError() __cudaCheckError( __FILE__, __LINE__ )
#define checkCudaErrors(err) __checkCudaErrors (err, __FILE__, __LINE__)
texture <float,2,cudaReadModeElementType> tex1;
texture<unsigned char, 2> tex;
static cudaArray *array = NULL;
static cudaArray *cuArray = NULL;
//Kernel for x direction sobel
__global__ void implement_x_sobel(float* garbage,float* output,int width,int height,int widthStep)
{
int x=blockIdx.x*blockDim.x+threadIdx.x;
int y=blockIdx.y*blockDim.y+threadIdx.y;
float output_value=((0*tex2D(tex1,x,y))+(2*tex2D(tex1,x+1,y))+(-2*tex2D(tex1,x- 1,y))+(0*tex2D(tex1,x,y+1))+(1*tex2D(tex1,x+1,y+1))+(-1*tex2D(tex1,x-1,y+1))+ (1*tex2D(tex1,x+1,y-1))+(0*tex2D(tex1,x,y-1))+(-1*tex2D(tex1,x-1,y-1)));
output[y*widthStep+x]=output_value;
}
inline void __checkCudaErrors( cudaError err, const char *file, const int line )
{
if( cudaSuccess != err) {
fprintf(stderr, "%s(%i) : CUDA Runtime API error %d: %s.\n",
file, line, (int)err, cudaGetErrorString( err ) );
exit(-1);
}
}
//Host Code
inline void __cudaSafeCall( cudaError err, const char *file, const int line )
{
#ifdef CUDA_ERROR_CHECK
if ( cudaSuccess != err )
{
printf("cudaSafeCall() failed at %s:%i : %s\n",
file, line, cudaGetErrorString( err ) );
exit( -1 );
}
#endif
return;
}
inline void __cudaCheckError( const char *file, const int line )
{
#ifdef CUDA_ERROR_CHECK
cudaError err = cudaGetLastError();
if ( cudaSuccess != err )
{
printf("cudaCheckError() failed at %s:%i : %s\n",
file, line, cudaGetErrorString( err ) );
exit( -1 );
}
#endif
return;
}
void kernelcall(float* input,float* output,int width,int height,int widthStep){
//cudaChannelFormatDesc channelDesc=cudaCreateChannelDesc(32,32,0,0,cudaChannelFormatKindFloat);
cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc<float>();
//cudaArray *cuArray;
CudaSafeCall(cudaMallocArray(&cuArray,&channelDesc,width,height));
cudaMemcpyToArray(cuArray,0,0,input,widthStep*height,cudaMemcpyHostToDevice);
tex1.addressMode[0]=cudaAddressModeClamp;
tex1.addressMode[1]=cudaAddressModeClamp;
tex1.filterMode=cudaFilterModeLinear;
cudaBindTextureToArray(tex1,cuArray,channelDesc);
tex1.normalized=false;
float * D_output_x;
float * garbage=NULL;
CudaSafeCall(cudaMalloc(&D_output_x,widthStep*height));
dim3 blocksize(16,16);
dim3 gridsize;
gridsize.x=(width+blocksize.x-1)/blocksize.x;
gridsize.y=(height+blocksize.y-1)/blocksize.y;
implement_x_sobel<<<gridsize,blocksize>>>(garbage,D_output_x,width,height,widthStep/sizeof(float));
cudaThreadSynchronize();
CudaCheckError();
CudaSafeCall(cudaMemcpy(output,D_output_x,height*widthStep,cudaMemcpyDeviceToHost));
cudaFree(D_output_x);
cudaFree(garbage);
cudaFreeArray(cuArray);
}
结果真的一塌糊涂，根本不像原图
结果：
我将代码的某些行更改为
float *pseudo_input=(float *)image->imageData;
float *output=(float*)image2->imageData;
float *input=(float*)image3->imageData;
float *inputnormalized=(float *)image4->imageData;
int s=image->widthStep/sizeof(float);
for(int w=0;w<=(image->height);w++)
for(int h=0;h<(image->width*image->nChannels);h++)
{
input[w*s+h]= pseudo_input[w*s+h];
}
kernelcall(input, output, image->width,image->height, image->widthStep);
cvNormalize(input,inputnormalized,0,255,NORM_MINMAX, CV_8UC1);
cvShowImage( "Original Image", image ); // Show our image inside it.
cvShowImage("Sobeled Image", image2);
但现在我得到一个未处理的异常错误。