-
Notifications
You must be signed in to change notification settings - Fork 0
/
asm_blocked.cu
154 lines (127 loc) · 5.25 KB
/
asm_blocked.cu
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
#include <stdio.h>
#include <stdlib.h>
#include <cuda.h>
#include <assert.h>
#define alphabet_size 26
#define num_block 128
#define num_thread 64
#define WARPSIZE 32
#define DEBUG 0
/*
Optimized from GPU to blocked version
*/
__device__ char POSSIBLE_CHAR[] = "ABCDEFGHIJKLMNOPQRSTUVWXYZ";
// TODO: Consider using openMP with vectorization to parallelize this loop.
__global__ void compute_X(int *X, int *T, int n) {
int row = threadIdx.x;
for (int j = 0; j <= n; j++) {
if (j == 0)
X[row*(n+1)+j] = 0;
else if (T[j-1] == int(POSSIBLE_CHAR[row]))
X[row*(n+1)+j] = j;
else
X[row*(n+1)+j] = X[row*(n+1)+(j-1)];
}
}
__global__ void compute_Dist(int *Dist, int *X, int *T, int *P, int n, int m, int rd) {
int num_tile = 1;
if (num_thread * num_block < (n+1))
num_tile = (n+1) / (num_block*num_thread) + 1;
int s_col = blockIdx.x * num_thread * num_tile + threadIdx.x, e_col = s_col + num_tile;
for (int col = s_col; col < e_col; col++) {
if (col > n)
return;
if (rd == 0)
Dist[rd*(n+1)+col] = 0;
else if (col == 0)
Dist[rd*(n+1)+col] = rd;
else if (T[col-1] == P[rd-1])
Dist[rd*(n+1)+col] = Dist[(rd-1)*(n+1)+(col-1)];
else if (X[(P[rd-1]-int('A'))*(n+1) + col] == 0)
Dist[rd*(n+1)+col] = 1 + min(Dist[(rd-1)*(n+1)+col], min(Dist[(rd-1)*(n+1)+(col-1)], rd + col - 1));
else
Dist[rd*(n+1)+col] = 1 + min(min(Dist[(rd-1)*(n+1)+col], Dist[(rd-1)*(n+1)+(col-1)]), Dist[(rd-1)*(n+1) + X[(P[rd-1]-int('A'))*(n+1) + col] - 1] + (col-1-X[(P[rd-1]-int('A'))*(n+1) + col]));
}
}
__global__ void compute_Dist_with_shuffle(int *Dist, int *X, int *T, int *P, int n, int m, int rd) {
// int col = blockIdx.x * num_thread + threadIdx.x;
// tile up
int num_tile = 1;
if (num_thread * num_block < (n+1))
num_tile = (n+1) / (num_block*num_thread) + 1;
int s_col = blockIdx.x * num_thread + threadIdx.x;
for (int i = 0; i < num_tile; i++) {
int col = s_col + num_block*num_thread*i;
if (col > n || rd == 0)
return;
int Dvar = Dist[(rd-1)*(n+1) + col], Avar, Bvar, Cvar;
if (col % WARPSIZE == 0) // edge between two warps, cannot use shuffle across warps
Avar = Dist[(rd-1)*(n+1) + col - 1];
else {
int test = __shfl_up_sync(0xffffffff, Dvar, 1);
// TODO: need explain why this is needed
if (col % WARPSIZE == 1) {
Avar = Dist[(rd-1)*(n+1) + col - 1];
}
else Avar = test;
}
Bvar = Dvar; // D[i-1][j]
Cvar = Dist[(rd-1)*(n+1) + X[(P[rd-1]-int('A'))*(n+1) + col] - 1]; // D[i-1][X[l][j]-1]
// compute D[i][j] in local memory
if (col == 0) Dvar = rd;
else if (T[col-1] == P[rd-1]) Dvar = Avar;
else if (X[(P[rd-1]-int('A'))*(n+1) + col] == 0) Dvar = 1 + min(Avar, min(Bvar, rd + col - 1));
else Dvar = 1 + min( min(Avar, Bvar), Cvar + (col-1-X[(P[rd-1]-int('A'))*(n+1) + col]));
Dist[rd*(n+1)+col] = Dvar; // write back to global memory
}
}
int main(int argc, char **argv) {
assert(argc == 4);
char *in_T = argv[1], *in_P = argv[2], *out = argv[3];
FILE *input_T, *input_P, *output;
input_T = fopen(in_T, "rb");
input_P = fopen(in_P, "rb");
int n, m;
int *T, *P, *device_T, *device_P;
fread(&n, 1, sizeof(int), input_T);
fread(&m, 1, sizeof(int), input_P);
T = (int*)malloc(sizeof(int)*n);
P = (int*)malloc(sizeof(int)*m);
// cudaMallocHost(&T, sizeof(int)*n);
// cudaMallocHost(&P, sizeof(int)*m);
fread(T, n, sizeof(int), input_T);
fread(P, m, sizeof(int), input_P);
fclose(input_T);
fclose(input_P);
int *host_Dist, *device_Dist, *host_X, *device_X;
// cudaMallocHost(&host_X, sizeof(int)*(n+1)*alphabet_size);
// cudaMallocHost(&host_Dist, sizeof(int)*(n+1)*(m+1));
host_X = (int*)malloc(sizeof(int)*(n+1)*alphabet_size);
host_Dist = (int*)malloc(sizeof(int)*(n+1)*(m+1));
cudaMalloc((void**)&device_Dist, sizeof(int)*(n+1)*(m+1));
cudaMalloc((void**)&device_X, sizeof(int)*(n+1)*alphabet_size);
cudaMalloc((void**)&device_T, sizeof(int)*n);
cudaMalloc((void**)&device_P, sizeof(int)*m);
cudaMemcpy(device_T, T, sizeof(int)*n, cudaMemcpyHostToDevice);
cudaMemcpy(device_P, P, sizeof(int)*m, cudaMemcpyHostToDevice);
// cudaMemcpy(device_Dist, host_Dist, sizeof(int)*n*m, cudaMemcpyHostToDevice);
compute_X <<< 1, alphabet_size >>> (device_X, device_T, n);
int nblocks = min((n+1)/num_thread+1, num_block);
for (int i = 0; i <= m; i++){
compute_Dist_with_shuffle <<< nblocks, num_thread >>> (device_Dist, device_X, device_T, device_P, n, m, i);
cudaDeviceSynchronize();
}
cudaMemcpy(host_Dist, device_Dist, sizeof(int)*(n+1)*(m+1), cudaMemcpyDeviceToHost);
cudaMemcpy(host_X, device_X, sizeof(int)*(n+1)*alphabet_size, cudaMemcpyDeviceToHost);
# if DEBUG
for (int i = 0; i <= m; i++) {
for (int j = 0; j <= n; j++)
printf("%d ", host_Dist[i*(n+1)+j]);
printf("\n");
}
# endif
output = fopen(out, "wb");
fwrite(host_Dist, (n+1)*(m+1), sizeof(int), output);
fclose(output);
return 0;
}