Repository navigation
Expand file tree
/
Copy pathGPU_Implementation_1.cu
More file actions
116 lines (94 loc) · 2.72 KB
/
Copy pathGPU_Implementation_1.cu
File metadata and controls
116 lines (94 loc) · 2.72 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
#include<iostream>
#include<math.h>
using namespace std;
#define imgH 100
#define imgW 100
#define h 1
#define patchW 3
void printImg(float img[imgH][imgW])
{
for(int i=0; i<imgH; i++)
{
for(int j=0; j<imgW; j++)
{
cout<<img[i][j]<<" ";
}
cout<<endl;
}
return;
}
__global__ void NLM(float* img, float* imgTemp, float* C)
{
int tx = threadIdx.x;
int ty = threadIdx.y;
int bx = blockIdx.x;
int by = blockIdx.y;
int j = by*blockDim.y + ty;
int i = bx*blockDim.x + tx;
for(int k=i; k<imgH - patchW + 1; k++)
for(int l=0; l<imgW - patchW + 1; l++)
{
if(l != j)
{
float v = 0;
for(int p=k; p<k+patchW; p++)
for(int q=l; q<l+patchW; q++)
{
v += (img[(i+p-k)*imgW + j+q-l] - img[p*imgW + q]);
v = v*v;
}
float w = exp(-v/(h*h));
atomicAdd(imgTemp + i*imgW + j, w * img[k*imgW + l]);
atomicAdd(C + i*imgW + j, w);
atomicAdd(imgTemp + k*imgW + l, w * img[i*imgW + j]);
atomicAdd(C + k*imgW + l, w);
}
}
}
int main()
{
float img[imgH][imgW] = {0}, C[imgH][imgW] = {0};
float imgTemp[imgH][imgW] = {0};
cudaEvent_t start, stop;
cudaEventCreate(&start);
cudaEventCreate(&stop);
for(int i=0; i<imgH; i++)
for(int j=0; j<imgW; j++)
{
img[i][j] = 1.2;
}
float* deviceImg;
float* deviceImgTemp;
float* deviceC;
//Memory Allocation
cudaMalloc((void**)&deviceImg, imgH * imgW * sizeof(float));
cudaMalloc((void**)&deviceImgTemp, imgH * imgW * sizeof(float));
cudaMalloc((void**)&deviceC, imgH * imgW * sizeof(float));
cudaMemset(deviceImgTemp, 0, imgH * imgW * sizeof(float));
cudaMemset(deviceC, 0, imgH * imgW * sizeof(float));
//Memory Copy
cudaMemcpy(deviceImg, img, imgH * imgW * sizeof(float), cudaMemcpyHostToDevice);
dim3 block(1, 1, 1);
dim3 grid(imgH - patchW + 1, imgW - patchW + 1, 1);
cudaEventRecord(start);
//Invoke Kernel
NLM<<<grid, block>>>(deviceImg, deviceImgTemp, deviceC);
cudaEventRecord(stop);
//Memory
cudaMemcpy(imgTemp, deviceImgTemp, imgH * imgW * sizeof(float), cudaMemcpyDeviceToHost);
cudaMemcpy(C, deviceC, imgH * imgW * sizeof(float), cudaMemcpyDeviceToHost);
cudaEventSynchronize(stop);
float milliseconds = 0;
cudaEventElapsedTime(&milliseconds, start, stop);
for(int i=0; i<imgH - patchW + 1; i++)
for(int j=0; j<imgW - patchW + 1; j++)
{
img[i][j] = imgTemp[i][j]/C[i][j];
}
cout<<"Time elapsed for H = "<<imgH<<" and W = "<<imgW<<" is: "<<(float)milliseconds<<" milliseconds";
// printImg(img);
cudaFree(deviceImg);
cudaFree(deviceImgTemp);
cudaFree(deviceC);
return 0;
}