ca99f951
 David Mayerich
faster implementa...
|
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
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
|
#ifndef STIM_CUDA_UPDATE_DIR_GLOBALD_H
#define STIM_CUDA_UPDATE_DIR_GLOBAL_H
# include <iostream>
# include <cuda.h>
#include <stim/cuda/cudatools.h>
#include <stim/cuda/sharedmem.cuh>
#include <math.h>
#include "cpyToshare.cuh"
#define RMAX_TEST 8
namespace stim{
namespace cuda{
// this kernel calculates the voting direction for the next iteration based on the angle between the location of this voter and the maximum vote value in its voting area.
template<typename T>
__global__ void cuda_update_dir(T* gpuDir, T* gpuVote, T* gpuGrad, T* gpuTable, T phi, int rmax, int x, int y){
extern __shared__ T atan2_table[];
//calculate the start point for this block
//int bxi = blockIdx.x * blockDim.x;
stim::cuda::sharedMemcpy(atan2_table, gpuTable, (2 * rmax + 1) * (2 * rmax + 1), threadIdx.x, blockDim.x);
__syncthreads();
// calculate the 2D coordinates for this current thread.
//int xi = bxi + threadIdx.x;
int xi = blockIdx.x * blockDim.x + threadIdx.x;
int yi = blockIdx.y * blockDim.y + threadIdx.y;
if(xi >= x || yi >= y) return; //if the index is outside of the image, terminate the kernel
int i = yi * x + xi; // convert 2D coordinates to 1D
float theta = gpuGrad[2*i]; // calculate the voting direction based on the grtadient direction - global memory fetch
gpuDir[i] = 0; //initialize the vote direction to zero
float max = 0; // define a local variable to maximum value of the vote image in the voting area for this voter
int id_x = 0; // define two local variables for the x and y position of the maximum
int id_y = 0;
int x_table = 2*rmax +1; // compute the size of window which will be checked for finding the voting area for this voter
int rmax_sq = rmax * rmax;
int tx_rmax = threadIdx.x + rmax;
float atan_angle;
float vote_c;
unsigned int ind_t;
for(int yr = -rmax; yr <= rmax; yr++){ //for each counter in the y direction
if (yi+yr >= 0 && yi + yr < y){ //if the counter exists (we aren't looking outside of the image)
for(int xr = -rmax; xr <= rmax; xr++){ //for each counter in the x direction
if((xr * xr + yr *yr)< rmax_sq){ //if the counter is within range of the voter
ind_t = (rmax - yr) * x_table + rmax - xr; //calculate the index to the atan2 table
atan_angle = atan2_table[ind_t]; //retrieve the direction vector from the table
//atan_angle = atan2((float)yr, (float)xr);
if (abs(atan_angle - theta) <phi){ // check if the current pixel is located in the voting angle of this voter.
vote_c = gpuVote[(yi+yr)*x + (xi+xr)]; // find the vote value for the current counter
if(vote_c>max) { // compare the vote value of this pixel with the max value to find the maxima and its index.
max = vote_c;
id_x = xr;
id_y = yr;
}
}
}
}
}
}
unsigned int ind_m = (rmax - id_y) * x_table + (rmax - id_x);
float new_angle = gpuTable[ind_m];
if(xi < x && yi < y)
gpuDir[i] = new_angle;
} //end kernel
// this kernel updates the gradient direction by the calculated voting direction.
template<typename T>
__global__ void cuda_update_grad(T* gpuGrad, T* gpuDir, int x, int y){
// calculate the 2D coordinates for this current thread.
int xi = blockIdx.x * blockDim.x + threadIdx.x;
int yi = blockIdx.y * blockDim.y + threadIdx.y;
// convert 2D coordinates to 1D
int i = yi * x + xi;
//update the gradient image with the vote direction
gpuGrad[2*i] = gpuDir[i];
}
template<typename T>
void gpu_update_dir(T* gpuVote, T* gpuGrad, T* gpuTable, T phi, unsigned int rmax, unsigned int x, unsigned int y){
//calculate the number of bytes in the array
unsigned int bytes = x * y * sizeof(T);
unsigned int max_threads = stim::maxThreadsPerBlock();
dim3 threads(sqrt(max_threads), sqrt(max_threads));
dim3 blocks(x/threads.x + 1, y/threads.y + 1);
// allocate space on the GPU for the updated vote direction
T* gpuDir;
cudaMalloc(&gpuDir, bytes);
size_t shared_mem = sizeof(T) * std::pow((2 * rmax + 1), 2);
std::cout<<"Shared memory for atan2 table: "<<shared_mem<<std::endl;
//call the kernel to calculate the new voting direction
cuda_update_dir <<< blocks, threads, shared_mem>>>(gpuDir, gpuVote, gpuGrad, gpuTable, phi, rmax, x , y);
//call the kernel to update the gradient direction
cuda_update_grad <<< blocks, threads >>>(gpuGrad, gpuDir, x , y);
//free allocated memory
cudaFree(gpuDir);
}
template<typename T>
void cpu_update_dir(T* cpuVote, T* cpuGrad,T* cpuTable, T phi, unsigned int rmax, unsigned int x, unsigned int y){
//calculate the number of bytes in the array
unsigned int bytes = x * y * sizeof(T);
//calculate the number of bytes in the atan2 table
unsigned int bytes_table = (2*rmax+1) * (2*rmax+1) * sizeof(T);
//allocate space on the GPU for the Vote Image
T* gpuVote;
cudaMalloc(&gpuVote, bytes);
//copy the input vote image to the GPU
HANDLE_ERROR(cudaMemcpy(gpuVote, cpuVote, bytes, cudaMemcpyHostToDevice));
//allocate space on the GPU for the input Gradient image
T* gpuGrad;
HANDLE_ERROR(cudaMalloc(&gpuGrad, bytes*2));
//copy the Gradient data to the GPU
HANDLE_ERROR(cudaMemcpy(gpuGrad, cpuGrad, bytes*2, cudaMemcpyHostToDevice));
//allocate space on the GPU for the atan2 table
T* gpuTable;
HANDLE_ERROR(cudaMalloc(&gpuTable, bytes_table));
//copy the atan2 values to the GPU
HANDLE_ERROR(cudaMemcpy(gpuTable, cpuTable, bytes_table, cudaMemcpyHostToDevice));
//call the GPU version of the update direction function
gpu_update_dir<T>(gpuVote, gpuGrad, gpuTable, phi, rmax, x , y);
//copy the new gradient image back to the CPU
cudaMemcpy(cpuGrad, gpuGrad, bytes*2, cudaMemcpyDeviceToHost) ;
//free allocated memory
cudaFree(gpuTable);
cudaFree(gpuVote);
cudaFree(gpuGrad);
}
}
}
#endif
|
