conv2.h
3.87 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
#ifndef STIM_CUDA_CONV2_H
#define STIM_CUDA_CONV2_H
//#define __CUDACC__
#ifdef __CUDACC__
#include <stim/cuda/cudatools.h>
#endif
namespace stim {
//Kernel function that performs the 2D convolution.
template<typename T>
__global__ void kernel_conv2(T* out, T* in, T* kernel, size_t sx, size_t sy, size_t kx, size_t ky) {
size_t xi = blockIdx.x * blockDim.x + threadIdx.x; //threads correspond to indices into the output image
size_t yi = blockIdx.y * blockDim.y + threadIdx.y;
size_t X = sx - kx + 1; //calculate the size of the output image
size_t Y = sy - ky + 1;
if (xi >= X || yi >= Y) return; //returns if the thread is outside of the output image
//loop through the kernel
size_t kxi, kyi;
T v = 0;
for (kyi = 0; kyi < ky; kyi++) {
for (kxi = 0; kxi < kx; kxi++) {
v += in[(yi + kyi) * sx + xi + kxi] * kernel[kyi * kx + kxi];
}
}
out[yi * X + xi] = v; //write the result to global memory
}
//Performs a convolution of a 2D image using the GPU. All pointers are assumed to be to memory on the current device.
//@param out is a pointer to the output image
//@param in is a pointer to the input image
//@param sx is the size of the input image along X
//@param sy is the size of the input image along Y
//@param kx is the size of the kernel along X
//@param ky is the size of the kernel along Y
template<typename T>
void gpu_conv2(T* out, T* in, T* kernel, size_t sx, size_t sy, size_t kx, size_t ky) {
cudaDeviceProp p;
HANDLE_ERROR(cudaGetDeviceProperties(&p, 0));
size_t tmax = p.maxThreadsPerBlock;
dim3 tn(sqrt(tmax), sqrt(tmax)); //calculate the block dimensions
size_t X = sx - kx + 1; //calculate the size of the output image
size_t Y = sy - ky + 1;
dim3 bn(X / tn.x + 1, Y / tn.y + 1); //calculate the grid dimensions
kernel_conv2 <<<bn, tn >>> (out, in, kernel, sx, sy, kx, ky); //launch the kernel
}
//Performs a convolution of a 2D image. Only valid pixels based on the kernel are returned.
// As a result, the output image will be smaller than the input image by (kx-1, ky-1)
//@param out is a pointer to the output image
//@param in is a pointer to the input image
//@param sx is the size of the input image along X
//@param sy is the size of the input image along Y
//@param kx is the size of the kernel along X
//@param ky is the size of the kernel along Y
template<typename T>
void cpu_conv2(T* out, T* in, T* kernel, size_t sx, size_t sy, size_t kx, size_t ky) {
size_t X = sx - kx + 1; //x size of the output image
size_t Y = sy - ky + 1; //y size of the output image
#ifdef __CUDACC__
//allocate memory and copy everything to the GPU
T* gpu_in;
HANDLE_ERROR(cudaMalloc(&gpu_in, sx * sy * sizeof(T)));
HANDLE_ERROR(cudaMemcpy(gpu_in, in, sx * sy * sizeof(T), cudaMemcpyHostToDevice));
T* gpu_kernel;
HANDLE_ERROR(cudaMalloc(&gpu_kernel, kx * ky * sizeof(T)));
HANDLE_ERROR(cudaMemcpy(gpu_kernel, kernel, kx * ky * sizeof(T), cudaMemcpyHostToDevice));
T* gpu_out;
HANDLE_ERROR(cudaMalloc(&gpu_out, X * Y * sizeof(T)));
gpu_conv2(gpu_out, gpu_in, gpu_kernel, sx, sy, kx, ky); //execute the GPU kernel
HANDLE_ERROR(cudaMemcpy(out, gpu_out, X * Y * sizeof(T), cudaMemcpyDeviceToHost)); //copy the result to the host
HANDLE_ERROR(cudaFree(gpu_in));
HANDLE_ERROR(cudaFree(gpu_kernel));
HANDLE_ERROR(cudaFree(gpu_out));
#else
T v; //register stores the integral of the current pixel value
size_t yi, xi, kyi, kxi, yi_kyi_sx;
for (yi = 0; yi < Y; yi++) { //for each pixel in the output image
for (xi = 0; xi < X; xi++) {
v = 0;
for (kyi = 0; kyi < ky; kyi++) { //for each pixel in the kernel
yi_kyi_sx = (yi + kyi) * sx;
for (kxi = 0; kxi < kx; kxi++) {
v += in[yi_kyi_sx + xi + kxi] * kernel[kyi * kx + kxi];
}
}
out[yi * X + xi] = v; //save the result to the output array
}
}
#endif
}
}
#endif