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
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
|
#include <stdio.h>
#include <curand.h>
// Define some error checking macros.
#define cudaErrCheck(stat) { cudaErrCheck_((stat), __FILE__, __LINE__); }
void cudaErrCheck_(cudaError_t stat, const char *file, int line) {
if (stat != cudaSuccess) {
fprintf(stderr, "CUDA Error: %s %s %d\n", cudaGetErrorString(stat), file, line);
}
}
#define curandErrCheck(stat) { curandErrCheck_((stat), __FILE__, __LINE__); }
void curandErrCheck_(curandStatus_t stat, const char *file, int line) {
if (stat != CURAND_STATUS_SUCCESS) {
fprintf(stderr, "cuRand Error: %d %s %d\n", stat, file, line);
}
}
#include <mma.h>
using namespace nvcuda;
// Must be multiples of 16 for wmma code to work
#define MATRIX_M (16)
#define MATRIX_N (16)
#define MATRIX_K (16)
// The only dimensions currently supported by WMMA
const int WMMA_M = 16;
const int WMMA_N = 16;
const int WMMA_K = 16;
__global__ void v4p_example(int *a_int32, int *b_int8, int *c,int *d_int32, int M, int N, int K) {
int registers_a[8];
int register_b; //contains 8 4bit b elements
int idx = blockDim.x * blockIdx.x + threadIdx.x;
asm("/*");
asm("CPTX_BEGIN");
asm("vp.load.b4.sync.row.m16n16k16.s32 {%0,%1,%2,%3,%4,%5,%6,%7},[%8],%9;" :
"=r"(registers_a[0]), "=r"(registers_a[1]),"=r"(registers_a[2]),"=r"(registers_a[3]),
"=r"(registers_a[4]),"=r"(registers_a[5]),"=r"(registers_a[6]),"=r"(registers_a[7]):
"l"(b_int8),"r"(M)
);
asm("CPTX_END");
asm("*/");
b_int8[0]=registers_a[7];
b_int8[1]=registers_a[7];
b_int8[2]=registers_a[7];
b_int8[3]=registers_a[7];
b_int8[4]=registers_a[7];
b_int8[5]=registers_a[7];
b_int8[6]=registers_a[7];
b_int8[7]=registers_a[7];
}
__global__ void convertFp32ToFp16 (half *out, float *in, int n) {
int idx = blockDim.x * blockIdx.x + threadIdx.x;
if (idx < n) {
out[idx] = in[idx];
}
}
__global__ void convertFp16ToFp32 (float *out, half *in, int n) {
int idx = blockDim.x * blockIdx.x + threadIdx.x;
if (idx < n) {
out[idx] = in[idx];
}
}
__global__ void convertInt32ToInt4 (int *out, int *in, int n) {
int idx = blockDim.x * blockIdx.x + threadIdx.x;
if (idx < n/8) {
out[idx] =(in[8*idx]&0xf)|(in[8*idx+1]&0xf)<<4|(in[8*idx+2]&0xf)<<8|(in[8*idx+3]&0xf)<<12|
(in[8*idx+4]&0xf)<<16|(in[8*idx+5]&0xf)<<20|(in[8*idx+6]&0xf)<<24|(in[8*idx+7]&0xf)<<28;
}
}
__global__ void convertInt32ToInt8 (int *out, int *in, int n) {
int idx = blockDim.x * blockIdx.x + threadIdx.x;
if (idx < n/4) {
out[idx] =(in[4*idx]&0xff)|(in[4*idx+1]&0xff)<<8|(in[4*idx+2]&0xff)<<16|(in[4*idx+3]&0xff)<<24;
}
}
__global__ void convertInt32ToInt16 (int *out, int *in, int n) {
int idx = blockDim.x * blockIdx.x + threadIdx.x;
if (idx < n/2) {
out[idx] =(in[2*idx]&0xffff)|(in[2*idx+1]&0xffff)<<16;
}
}
__global__ void convertInt4ToInt32 (int *out, int *in, int n) {
int idx = blockDim.x * blockIdx.x + threadIdx.x;
int shft_amt=4*(idx%8);
int shft_mask=0xf<<shft_amt;
if (idx < n) {
out[idx]= (in[idx/8]&shft_mask)>>shft_amt;
}
}
__global__ void convertInt8ToInt32 (int *out, int *in, int n) {
int idx = blockDim.x * blockIdx.x + threadIdx.x;
int shft_amt=8*(idx%4);
int shft_mask=0xff<<shft_amt;
if (idx < n) {
out[idx]= (in[idx/4]&shft_mask)>>shft_amt;
}
}
__global__ void convertInt16ToInt32 (int *out, int *in, int n) {
int idx = blockDim.x * blockIdx.x + threadIdx.x;
int shft_amt=16*(idx%2);
int shft_mask=0xffff<<shft_amt;
if (idx < n) {
out[idx]= (in[idx/2]&shft_mask)>>shft_amt;
}
}
int main(int argc, char* argv[]) {
int *a_int32;
int *b_int32;
int *c_int32;
int *d_int32;
int *a_int4;
int *b_int4;
int *a_int8;
int *b_int8;
int *a_int16;
int *b_int16;
int *a_host_wmma;
int *b_host_wmma;
int *c_host_wmma;
int *d_host_wmma;
int *d_cal_host_wmma;
cudaEvent_t startWMMA;
cudaEvent_t stopWMMA;
cudaErrCheck(cudaEventCreate(&startWMMA));
cudaErrCheck(cudaEventCreate(&stopWMMA));
// Use tensor cores
cudaErrCheck(cudaMalloc((void**)&a_int32, MATRIX_M * MATRIX_K * sizeof(int)));
cudaErrCheck(cudaMalloc((void**)&b_int32, MATRIX_K * MATRIX_N * sizeof(int)));
cudaErrCheck(cudaMalloc((void**)&c_int32, MATRIX_K * MATRIX_N * sizeof(int)));
cudaErrCheck(cudaMalloc((void**)&d_int32, MATRIX_K * MATRIX_N * sizeof(int)));
cudaErrCheck(cudaMalloc((void**)&a_int4, MATRIX_M * MATRIX_K * sizeof(int)/8));
cudaErrCheck(cudaMalloc((void**)&b_int4, MATRIX_K * MATRIX_N * sizeof(int)/8));
cudaErrCheck(cudaMalloc((void**)&a_int8, MATRIX_M * MATRIX_K * sizeof(int)/4));
cudaErrCheck(cudaMalloc((void**)&b_int8, MATRIX_K * MATRIX_N * sizeof(int)/4));
cudaErrCheck(cudaMalloc((void**)&a_int16, MATRIX_M * MATRIX_K * sizeof(int)/2));
cudaErrCheck(cudaMalloc((void**)&b_int16, MATRIX_K * MATRIX_N * sizeof(int)/2));
a_host_wmma = (int *)malloc(MATRIX_M * MATRIX_K * sizeof(int));
b_host_wmma = (int *)malloc(MATRIX_K * MATRIX_N * sizeof(int));
c_host_wmma = (int *)malloc(MATRIX_M * MATRIX_N * sizeof(int));
d_host_wmma = (int *)malloc(MATRIX_M * MATRIX_N * sizeof(int));
d_cal_host_wmma = (int *)malloc(MATRIX_M * MATRIX_N * sizeof(int));
printf("a_int32\n");
for(int m=0;m<MATRIX_M;m++){
for(int n=0;n<MATRIX_K;n++){
a_host_wmma[m*MATRIX_K+n]=(m*MATRIX_K+n)%4;
printf("%d ",a_host_wmma[m*MATRIX_K+n]);
}
printf(";\n");
}
printf("b_int32\n");
for(int m=0;m<MATRIX_K;m++){
for(int n=0;n<MATRIX_N;n++){
b_host_wmma[m*MATRIX_N+n]=(m*MATRIX_N+n)%4;
printf("%d ",b_host_wmma[m*MATRIX_N+n]);
}
printf(";\n");
}
printf("c_int32\n");
for(int m=0;m<MATRIX_M;m++){
for(int n=0;n<MATRIX_N;n++){
c_host_wmma[m*MATRIX_N+n]=(m*MATRIX_N+n)%4;
d_cal_host_wmma[m*MATRIX_N+n]=0;
printf("%d ",c_host_wmma[m*MATRIX_N+n]);
}
printf(";\n");
}
for(int m=0;m<MATRIX_M;m++){
for(int n=0;n<MATRIX_N;n++){
for(int k=0;k<MATRIX_K;k++){
d_cal_host_wmma[m*MATRIX_N+n]+= a_host_wmma[m*MATRIX_K+k]*b_host_wmma[k*MATRIX_K+n];
}
d_cal_host_wmma[m*MATRIX_N+n]+=c_host_wmma[m*MATRIX_N+n];
}
}
cudaErrCheck(cudaMemcpy(a_int32,a_host_wmma, MATRIX_M * MATRIX_K * sizeof(int), cudaMemcpyHostToDevice));
cudaErrCheck(cudaMemcpy(b_int32,b_host_wmma, MATRIX_K * MATRIX_N * sizeof(int), cudaMemcpyHostToDevice));
cudaErrCheck(cudaMemcpy(c_int32,c_host_wmma, MATRIX_M * MATRIX_N * sizeof(int), cudaMemcpyHostToDevice));
#define TEST4
#ifdef TEST16
convertInt32ToInt16 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (a_int16, a_int32, MATRIX_M * MATRIX_K);
convertInt16ToInt32 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (d_int32, a_int16, MATRIX_M * MATRIX_K);
cudaErrCheck(cudaMemcpy(d_host_wmma, d_int32, MATRIX_M * MATRIX_N * sizeof(int), cudaMemcpyDeviceToHost));
#endif
#ifdef TEST8
convertInt32ToInt8 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (a_int8, a_int32, MATRIX_M * MATRIX_K);
convertInt8ToInt32 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (d_int32, a_int8, MATRIX_M * MATRIX_K);
cudaErrCheck(cudaMemcpy(d_host_wmma, d_int32, MATRIX_M * MATRIX_N * sizeof(int), cudaMemcpyDeviceToHost));
#endif
#ifdef TEST4
convertInt32ToInt4 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (a_int4, a_int32, MATRIX_M * MATRIX_K);
convertInt4ToInt32 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (d_int32, a_int4, MATRIX_M * MATRIX_K);
cudaErrCheck(cudaMemcpy(d_host_wmma, d_int32, MATRIX_M * MATRIX_N * sizeof(int), cudaMemcpyDeviceToHost));
#endif
//convertFp32ToFp16 <<< (MATRIX_K * MATRIX_N + 255) / 256, 256 >>> (b_fp16, b_fp32, MATRIX_K * MATRIX_N);
//convertFp32ToFp16 <<< (MATRIX_M * MATRIX_N + 255) / 256, 256 >>> (c_fp16, c_fp32, MATRIX_K * MATRIX_N);
//AAMIR printf("\nM = %d, N = %d, K = %d. \n", MATRIX_M, MATRIX_N, MATRIX_K);
//AAMIR
//AAMIR printf("Running with wmma...\n");
//cudaErrCheck(cudaEventRecord(startWMMA));
//v4p_example <<< 1, 32>>> (a_int32, a_int8, a_int32, d_int32, MATRIX_M, MATRIX_N, MATRIX_K);
//cudaErrCheck(cudaEventRecord(stopWMMA));
//cudaErrCheck(cudaEventSynchronize(stopWMMA));
//AAMIR
//AAMIR // Error checking
//AAMIR printf("\nChecking results...\n");
//AAMIR cudaErrCheck(cudaMemcpy(d_host_wmma, d_fp32, MATRIX_M * MATRIX_N * sizeof(float), cudaMemcpyDeviceToHost));
//AAMIR
//AAMIR printf("Results verified: cublas and WMMA agree.\n\n");
//AAMIR float wmmaTime;
//AAMIR cudaErrCheck(cudaEventElapsedTime(&wmmaTime, startWMMA, stopWMMA));
//AAMIR printf("wmma took %fms\n", wmmaTime);
//AAMIR
//AAMIR cudaErrCheck(cudaEventDestroy(startWMMA));
//AAMIR cudaErrCheck(cudaEventDestroy(stopWMMA));
//AAMIR
//AAMIR int t=200000;
//AAMIR while(t-->0);
//AAMIR printf("D_CALCULATED\n");
//AAMIR
//AAMIR for(int m=0;m<MATRIX_M;m++){
//AAMIR for(int n=0;n<MATRIX_N;n++){
//AAMIR printf("%.2f,",d_cal_host_wmma[m*MATRIX_N+n]);
//AAMIR }
//AAMIR printf("\n");
//AAMIR }
printf("D_WMMA\n");
for(int m=0;m<MATRIX_M;m++){
for(int n=0;n<MATRIX_N;n++){
printf("%x,",d_host_wmma[m*MATRIX_N+n]);
}
printf("\n");
}
//AAMIR int suc=1;
//AAMIR for(int m=0;m<MATRIX_M;m++){
//AAMIR for(int n=0;n<MATRIX_N;n++){
//AAMIR if(abs(d_cal_host_wmma[m*MATRIX_N+n]-d_host_wmma[m*MATRIX_N+n])>1)
//AAMIR {
//AAMIR printf("ERROR:\n");
//AAMIR suc=0;
//AAMIR }
//AAMIR }
//AAMIR }
//AAMIR if(suc==1)
//AAMIR printf("COMPLETED_SUCCESSFULLY\n");
//AAMIR
cudaErrCheck(cudaFree(a_int32));
cudaErrCheck(cudaFree(b_int32));
cudaErrCheck(cudaFree(c_int32));
cudaErrCheck(cudaFree(d_int32));
cudaErrCheck(cudaFree(a_int8));
cudaErrCheck(cudaFree(b_int8));
free(a_host_wmma);
free(b_host_wmma);
free(c_host_wmma);
free(d_host_wmma);
cudaErrCheck(cudaDeviceReset());
return 0;
}
|
