├── image.bmp
├── kernel.cu
└── result.bmp


/image.bmp:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Sunagatov/MedianFilter/4421bd89a796a72a8699ddb122117cb7a0d4df62/image.bmp


--------------------------------------------------------------------------------
/kernel.cu:
--------------------------------------------------------------------------------
  1 | #include "cuda_runtime.h"
  2 | #include "device_launch_parameters.h"
  3 | 
  4 | #include <stdio.h>
  5 | #include <iostream>
  6 | #include <algorithm>
  7 | #include <ctime>
  8 | #include <opencv2/opencv.hpp>
  9 | 
 10 | #define BLOCK_SIZE 16
 11 | 
 12 | using namespace std;
 13 | using namespace cv;
 14 | 
 15 | texture<unsigned char, 2, cudaReadModeElementType> inTexture;
 16 | 
 17 | __device__ inline void sortArr(float arr[], int left, int right) {
 18 | 	int i = left, j = right;
 19 | 	int tmp;
 20 | 	int pivot = arr[(left + right) / 2];
 21 | 	while (i <= j) {
 22 | 		while (arr[i] < pivot)
 23 | 			i++;
 24 | 		while (arr[j] > pivot)
 25 | 			j--;
 26 | 		if (i <= j) {
 27 | 			tmp = arr[i];
 28 | 			arr[i] = arr[j];
 29 | 			arr[j] = tmp;
 30 | 			i++;
 31 | 			j--;
 32 | 		}
 33 | 	};
 34 | 	if (left < j)
 35 | 		sortArr(arr, left, j);
 36 | 	if (i < right)
 37 | 		sortArr(arr, i, right);
 38 | }
 39 | 
 40 | __global__ void gpuCalculation(unsigned char* input, unsigned char* output, int width, int height) {
 41 | 
 42 | 	int x = blockIdx.x * BLOCK_SIZE + threadIdx.x;
 43 | 	int y = blockIdx.y * BLOCK_SIZE + threadIdx.y;
 44 | 
 45 | 	float sortMas[9];
 46 | 
 47 | 	if ((x < width) && (y < height))
 48 | 	{
 49 | 		int vector_counter = 0;
 50 | 		for (int i = -1; i <= 1; i++) {
 51 | 			for (int j = -1; j <= 1; j++) {
 52 | 				sortMas[vector_counter] = tex2D(inTexture, x + j, y + i);
 53 | 				vector_counter++;
 54 | 			}
 55 | 		}
 56 | 		sortArr(sortMas, 0, 8);
 57 | 		output[y * width + x] = sortMas[4];
 58 | 	}
 59 | }
 60 | 
 61 | void medianFilter(const Mat& input, Mat& output) {
 62 | 
 63 | 	cudaEvent_t start, stop;
 64 | 	float time;
 65 | 	cudaEventCreate(&start);
 66 | 	cudaEventCreate(&stop);
 67 | 
 68 | 	int gray_size = input.step * input.rows;
 69 | 
 70 | 	size_t pitch;
 71 | 	unsigned char* d_input = NULL;
 72 | 	unsigned char* d_output;
 73 | 
 74 | 	cudaMallocPitch(&d_input, &pitch, sizeof(unsigned char) * input.step, input.rows);
 75 | 	cudaMemcpy2D(d_input, pitch, input.ptr(), sizeof(unsigned char) * input.step, sizeof(unsigned char) * input.step, input.rows, cudaMemcpyHostToDevice); 
 76 | 	cudaBindTexture2D(0, inTexture, d_input, input.step, input.rows, pitch);
 77 | 	cudaMalloc<unsigned char>(&d_output, gray_size);
 78 | 
 79 | 	dim3 block(BLOCK_SIZE, BLOCK_SIZE);
 80 | 	dim3 grid((input.cols + block.x - 1) / block.x, (input.rows + block.y - 1) / block.y);
 81 | 
 82 | 	cudaEventRecord(start, 0);
 83 | 
 84 | 	gpuCalculation <<<grid, block >>> (d_input, d_output, input.cols, input.rows);
 85 | 
 86 | 	cudaEventRecord(stop, 0);
 87 | 	cudaEventSynchronize(stop);
 88 | 
 89 | 	cudaMemcpy(output.ptr(), d_output, gray_size, cudaMemcpyDeviceToHost);
 90 | 
 91 | 	cudaFree(d_input);
 92 | 	cudaFree(d_output);
 93 | 
 94 | 	cudaEventElapsedTime(&time, start, stop);
 95 | 	cout << "Time on GPU: " << time << " milliseconds" << endl;
 96 | }
 97 | 
 98 | int main() {
 99 | 
100 | 	Mat input = imread("image.bmp", IMREAD_GRAYSCALE);
101 | 	//Mat output_own(input.rows, input.cols, CV_8UC1);
102 | 	//medianFilter(input, output_own);
103 | 	imwrite("result.bmp", input);
104 | 	getchar();
105 | 	return 0;
106 | }


--------------------------------------------------------------------------------
/result.bmp:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Sunagatov/MedianFilter/4421bd89a796a72a8699ddb122117cb7a0d4df62/result.bmp


--------------------------------------------------------------------------------