-
Notifications
You must be signed in to change notification settings - Fork 0
/
attention.cuh
216 lines (192 loc) · 6.46 KB
/
attention.cuh
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
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
#include "kernel.cuh"
void GEMM_base(half *d_A, half *d_B, half *d_C, int M, int N, int K) {
const int block_size = 16;
dim3 threads(block_size, block_size);
dim3 grid((M + threads.x - 1) / threads.x, (N + threads.y - 1) / threads.y);
gemm_baseline<<<grid, threads>>>(d_A, d_B, d_C, M, N, K);
cudaError_t err = cudaGetLastError();
if (err != cudaSuccess) {
printf("CUDA Error: %s\n", cudaGetErrorString(err));
// Possibly: exit(-1) if program cannot continue....
}
}
void GEMM(half *d_A, half *d_B, half *d_C, int M, int N, int K) {
const int BM = 128, BN = 256, BK = 32;
dim3 blockDim(256);
int BX = (N + BN - 1) / BN;
int BY = (M + BM - 1) / BM;
dim3 gridDim(BX, BY);
cudaFuncSetAttribute(GEMM_sharedmem_wmma,
cudaFuncAttributeMaxDynamicSharedMemorySize, 98304);
unsigned int dsmem = 2 * (BM * (BK + 8) + BK * (BN + 8)) * sizeof(half);
GEMM_sharedmem_wmma<<<gridDim, blockDim, dsmem>>>(d_A, d_B, d_C, M, N, K);
cudaError_t err = cudaGetLastError();
if (err != cudaSuccess) {
printf("CUDA Error: %s\n", cudaGetErrorString(err));
// Possibly: exit(-1) if program cannot continue....
}
}
void Transpose(half *d_A, half *d_B, int M, int N) {
dim3 blockDim(256);
dim3 gridDim(128);
transpose<<<gridDim, blockDim>>>(d_A, d_B, M, N);
}
void SoftMax(half *p, half *s, int N) {
host_to_host(&s, &p, N * N);
half *h_mask;
host_malloc(&h_mask, N * N);
for(int i = 0; i < N; i++) {
for(int j = 0; j < N; j++) {
if(i <= j)
h_mask[i * N + j] = 1;
else
h_mask[i * N + j] = 0;
}
}
half *d_mask;
device_malloc(&d_mask, N * N);
copy_to_device(&d_mask, &h_mask, N * N);
int size_per_head = 1;
int batch_size = 1;
int head_num = 1;
half scaler = 1 / sqrtf((half)size_per_head * (half)1.0f);
// launch fusion kernel
dim3 grid(N);
dim3 block(N);
softmax_kernel<<<grid, block>>>(s, d_mask, batch_size, head_num, N, scaler);
}
void SoftMax_base(half *p, half *s, int N) {
host_to_host(&s, &p, N * N);
half *h_mask;
host_malloc(&h_mask, N * N);
for(int i = 0; i < N; i++) {
for(int j = 0; j < N; j++) {
if(i <= j)
h_mask[i * N + j] = 1;
else
h_mask[i * N + j] = 0;
}
}
half *d_mask;
device_malloc(&d_mask, N * N);
copy_to_device(&d_mask, &h_mask, N * N);
int size_per_head = 1;
int batch_size = 1;
int head_num = 1;
half scaler = 1 / sqrtf((half)size_per_head * (half)1.0f);
// launch fusion kernel
dim3 grid(N);
dim3 block(N);
addmask_kernel<<<grid, block>>>(s, d_mask, batch_size, head_num, N, scaler);
cudaDeviceSynchronize();
softcal_kernel<<<grid, block>>>(s, d_mask, batch_size, head_num, N, scaler);
}
void attention(half *in,
half *Wq, half *Wk,half *Wv,
half *q, half *k, half *v,
half *kT,
half *p, half *s,
half *o, half *Wo,
half *out,
int N, int d,
int kd,
FILE* file)
{
struct timeval start, end;
struct timeval sub1, sub2;
float t1 = 0; // get q, k, v
float t2 = 0; // get kT
float t3 = 0; // q * k^T
float t4 = 0; // softmax
float t5 = 0; // s * v
float t6 = 0; // o * Wo
gettimeofday(&start, NULL);
printf("start attention...\n");
if(kd == 1) {
for(int i = 0; i < 1000; i++) {
// get q, k, v
gettimeofday(&sub1, NULL);
GEMM(in, Wq, q, N, d, d);
GEMM(in, Wk, k, N, d, d);
GEMM(in, Wv, v, N, d, d);
cudaDeviceSynchronize();
gettimeofday(&sub2, NULL);
t1 += time_diff(&sub1, &sub2);
// get kT
gettimeofday(&sub1, NULL);
Transpose(k, kT, N, d);
cudaDeviceSynchronize();
gettimeofday(&sub2, NULL);
t2 += time_diff(&sub1, &sub2);
// q * k^T
gettimeofday(&sub1, NULL);
GEMM(q, kT, p, N, N, d);
cudaDeviceSynchronize();
gettimeofday(&sub2, NULL);
t3 += time_diff(&sub1, &sub2);
// softmax
gettimeofday(&sub1, NULL);
SoftMax(p, s, N);
cudaDeviceSynchronize();
gettimeofday(&sub2, NULL);
t4 += time_diff(&sub1, &sub2);
// s * v
gettimeofday(&sub1, NULL);
GEMM(s, v, o, N, d, N);
cudaDeviceSynchronize();
gettimeofday(&sub2, NULL);
t5 += time_diff(&sub1, &sub2);
// o * Wo
gettimeofday(&sub1, NULL);
GEMM(o, Wo, out, N, d, d);
cudaDeviceSynchronize();
gettimeofday(&sub2, NULL);
t6 += time_diff(&sub1, &sub2);
}
}
else if(kd == 0) {
for(int i = 0; i < 1000; i++) {
// get q, k, v
gettimeofday(&sub1, NULL);
GEMM_base(in, Wq, q, N, d, d);
GEMM_base(in, Wk, k, N, d, d);
GEMM_base(in, Wv, v, N, d, d);
cudaDeviceSynchronize();
gettimeofday(&sub2, NULL);
t1 += time_diff(&sub1, &sub2);
// get kT
gettimeofday(&sub1, NULL);
Transpose(k, kT, N, d);
cudaDeviceSynchronize();
gettimeofday(&sub2, NULL);
t2 += time_diff(&sub1, &sub2);
// q * k^T
gettimeofday(&sub1, NULL);
GEMM_base(q, kT, p, N, N, d);
cudaDeviceSynchronize();
gettimeofday(&sub2, NULL);
t3 += time_diff(&sub1, &sub2);
// softmax
gettimeofday(&sub1, NULL);
SoftMax_base(p, s, N);
cudaDeviceSynchronize();
gettimeofday(&sub2, NULL);
t4 += time_diff(&sub1, &sub2);
// s * v
gettimeofday(&sub1, NULL);
GEMM_base(s, v, o, N, d, N);
cudaDeviceSynchronize();
gettimeofday(&sub2, NULL);
t5 += time_diff(&sub1, &sub2);
// o * Wo
gettimeofday(&sub1, NULL);
GEMM_base(o, Wo, out, N, d, d);
cudaDeviceSynchronize();
gettimeofday(&sub2, NULL);
t6 += time_diff(&sub1, &sub2);
}
}
gettimeofday(&end,NULL);
fprintf(file, "%d, %d, %d, %f, %f, %f, %f, %f, %f, %f\n", kd, N, d, time_diff(&start, &end)/1000, t1/1000, t2/1000, t3/1000, t4/1000, t5/1000, t6/1000);
printf("attention end...\n");
}