13fe3c84
Laila Saadatifard
update the stimli...
|
1
2
3
|
#ifndef STIM_CUDA_LOCAL_MAX_H
#define STIM_CUDA_LOCAL_MAX_H
|
13fe3c84
Laila Saadatifard
update the stimli...
|
4
5
|
# include <iostream>
# include <cuda.h>
|
96f9b10f
Laila Saadatifard
change the header...
|
6
|
#include <stim/cuda/cudatools.h>
|
13fe3c84
Laila Saadatifard
update the stimli...
|
7
8
9
10
|
namespace stim{
namespace cuda{
|
13fe3c84
Laila Saadatifard
update the stimli...
|
11
12
|
// this kernel calculates the local maximum for finding the cell centers
template<typename T>
|
c01958ad
Pavel Govyadinov
merged the two st...
|
13
|
__global__ void cuda_local_max(T* gpuCenters, T* gpuVote, T final_t, int conn, int x, int y){
|
13fe3c84
Laila Saadatifard
update the stimli...
|
14
15
|
// calculate the 2D coordinates for this current thread.
|
c01958ad
Pavel Govyadinov
merged the two st...
|
16
17
|
int xi = blockIdx.x * blockDim.x + threadIdx.x;
int yi = blockIdx.y * blockDim.y + threadIdx.y;
|
84ca9bba
Laila Saadatifard
fix some bugs in ...
|
18
19
20
|
if(xi >= x || yi >= y)
return;
|
bf731970
Laila Saadatifard
fix some bugs in ...
|
21
|
|
13fe3c84
Laila Saadatifard
update the stimli...
|
22
|
// convert 2D coordinates to 1D
|
c01958ad
Pavel Govyadinov
merged the two st...
|
23
|
int i = yi * x + xi;
|
84ca9bba
Laila Saadatifard
fix some bugs in ...
|
24
|
|
84ca9bba
Laila Saadatifard
fix some bugs in ...
|
25
26
27
28
29
|
gpuCenters[i] = 0; //initialize the value at this location to zero
T val = gpuVote[i];
//compare to the threshold
|
c01958ad
Pavel Govyadinov
merged the two st...
|
30
|
if(val < final_t) return;
|
84ca9bba
Laila Saadatifard
fix some bugs in ...
|
31
|
|
11cd127f
Laila Saadatifard
Leila's ivote pro...
|
32
33
34
35
36
|
//define an array to store indices with same vote value
/*int * IdxEq;
IdxEq = new int [2*conn];
int n = 0;*/
|
84ca9bba
Laila Saadatifard
fix some bugs in ...
|
37
38
39
40
41
42
|
for(int xl = xi - conn; xl < xi + conn; xl++){
for(int yl = yi - conn; yl < yi + conn; yl++){
if(xl >= 0 && xl < x && yl >= 0 && yl < y){
int il = yl * x + xl;
if(gpuVote[il] > val){
return;
|
11cd127f
Laila Saadatifard
Leila's ivote pro...
|
43
44
45
46
47
48
49
50
51
|
}
if (gpuVote[il] == val){
/*IdxEq[n] = il;
n = n+1;*/
if( il > i){
return;
}
}
}
|
84ca9bba
Laila Saadatifard
fix some bugs in ...
|
52
53
|
}
}
|
11cd127f
Laila Saadatifard
Leila's ivote pro...
|
54
55
56
57
58
|
/*if (n!=0){
if(IdxEq[n/2] !=i){
return;
}
} */
|
c01958ad
Pavel Govyadinov
merged the two st...
|
59
|
gpuCenters[i] = 1;
|
13fe3c84
Laila Saadatifard
update the stimli...
|
60
|
}
|
bf731970
Laila Saadatifard
fix some bugs in ...
|
61
|
|
13fe3c84
Laila Saadatifard
update the stimli...
|
62
|
template<typename T>
|
c01958ad
Pavel Govyadinov
merged the two st...
|
63
|
void gpu_local_max(T* gpuCenters, T* gpuVote, T final_t, unsigned int conn, unsigned int x, unsigned int y){
|
13fe3c84
Laila Saadatifard
update the stimli...
|
64
|
|
13fe3c84
Laila Saadatifard
update the stimli...
|
65
|
unsigned int max_threads = stim::maxThreadsPerBlock();
|
03428452
Laila Saadatifard
update the ivote ...
|
66
67
|
/*dim3 threads(max_threads, 1);
dim3 blocks(x/threads.x + (x %threads.x == 0 ? 0:1) , y);*/
|
c01958ad
Pavel Govyadinov
merged the two st...
|
68
69
|
dim3 threads( sqrt(max_threads), sqrt(max_threads) );
dim3 blocks(x/threads.x + 1, y/threads.y + 1);
|
13fe3c84
Laila Saadatifard
update the stimli...
|
70
|
|
13fe3c84
Laila Saadatifard
update the stimli...
|
71
|
//call the kernel to find the local maximum.
|
c01958ad
Pavel Govyadinov
merged the two st...
|
72
|
cuda_local_max <<< blocks, threads >>>(gpuCenters, gpuVote, final_t, conn, x, y);
|
13fe3c84
Laila Saadatifard
update the stimli...
|
73
74
|
}
|
13fe3c84
Laila Saadatifard
update the stimli...
|
75
|
template<typename T>
|
c01958ad
Pavel Govyadinov
merged the two st...
|
76
|
void cpu_local_max(T* cpuCenters, T* cpuVote, T final_t, unsigned int conn, unsigned int x, unsigned int y){
|
13fe3c84
Laila Saadatifard
update the stimli...
|
77
|
|
13fe3c84
Laila Saadatifard
update the stimli...
|
78
79
80
81
82
83
84
|
//calculate the number of bytes in the array
unsigned int bytes = x * y * sizeof(T);
// allocate space on the GPU for the detected cell centes
T* gpuCenters;
cudaMalloc(&gpuCenters, bytes);
|
13fe3c84
Laila Saadatifard
update the stimli...
|
85
86
87
88
|
//allocate space on the GPU for the input Vote Image
T* gpuVote;
cudaMalloc(&gpuVote, bytes);
|
13fe3c84
Laila Saadatifard
update the stimli...
|
89
90
|
//copy the Vote image data to the GPU
HANDLE_ERROR(cudaMemcpy(gpuVote, cpuVote, bytes, cudaMemcpyHostToDevice));
|
bf731970
Laila Saadatifard
fix some bugs in ...
|
91
|
|
13fe3c84
Laila Saadatifard
update the stimli...
|
92
|
//call the GPU version of the local max function
|
c01958ad
Pavel Govyadinov
merged the two st...
|
93
|
gpu_local_max<T>(gpuCenters, gpuVote, final_t, conn, x, y);
|
bf731970
Laila Saadatifard
fix some bugs in ...
|
94
|
|
13fe3c84
Laila Saadatifard
update the stimli...
|
95
96
|
//copy the cell centers data to the CPU
cudaMemcpy(cpuCenters, gpuCenters, bytes, cudaMemcpyDeviceToHost) ;
|
bf731970
Laila Saadatifard
fix some bugs in ...
|
97
|
|
13fe3c84
Laila Saadatifard
update the stimli...
|
98
99
100
|
//free allocated memory
cudaFree(gpuCenters);
cudaFree(gpuVote);
|
13fe3c84
Laila Saadatifard
update the stimli...
|
101
102
103
104
105
106
107
108
|
}
}
}
#endif
|