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
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
|
//**************************************************************************
// Multi-threaded Matrix Multiply benchmark
//--------------------------------------------------------------------------
// TA : Christopher Celio
// Student:
//
//
// This benchmark multiplies two 2-D arrays together and writes the results to
// a third vector. The input data (and reference data) should be generated
// using the matmul_gendata.pl perl script and dumped to a file named
// dataset.h.
// print out arrays, etc.
//#define DEBUG
//--------------------------------------------------------------------------
// Includes
#include <string.h>
#include <stdlib.h>
#include <stdio.h>
//--------------------------------------------------------------------------
// Input/Reference Data
typedef float data_t;
#include "dataset.h"
//--------------------------------------------------------------------------
// Basic Utilities and Multi-thread Support
__thread unsigned long coreid;
unsigned long ncores;
#include "util.h"
#define stringify_1(s) #s
#define stringify(s) stringify_1(s)
#define stats(code) do { \
unsigned long _c = -rdcycle(), _i = -rdinstret(); \
code; \
_c += rdcycle(), _i += rdinstret(); \
if (coreid == 0) \
printf("%s: %ld cycles, %ld.%ld cycles/iter, %ld.%ld CPI\n", \
stringify(code), _c, _c/DIM_SIZE/DIM_SIZE/DIM_SIZE, 10*_c/DIM_SIZE/DIM_SIZE/DIM_SIZE%10, _c/_i, 10*_c/_i%10); \
} while(0)
//--------------------------------------------------------------------------
// Helper functions
void printArray( char name[], int n, data_t arr[] )
{
int i;
if (coreid != 0)
return;
printf( " %10s :", name );
for ( i = 0; i < n; i++ )
printf( " %3ld ", (long) arr[i] );
printf( "\n" );
}
void __attribute__((noinline)) verify(size_t n, const data_t* test, const data_t* correct)
{
if (coreid != 0)
return;
size_t i;
for (i = 0; i < n; i++)
{
if (test[i] != correct[i])
{
printf("FAILED test[%d]= %3ld, correct[%d]= %3ld\n",
i, (long)test[i], i, (long)correct[i]);
exit(-1);
}
}
return;
}
//--------------------------------------------------------------------------
// matmul function
// single-thread, naive version
void __attribute__((noinline)) matmul_naive(const int lda, const data_t A[], const data_t B[], data_t C[] )
{
int i, j, k;
if (coreid > 0)
return;
for ( i = 0; i < lda; i++ )
for ( j = 0; j < lda; j++ )
{
for ( k = 0; k < lda; k++ )
{
C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
}
}
}
void __attribute__((noinline)) matmul_MI_transpose(const int lda, const data_t A[], const data_t B[], data_t C[] )
{
int i, j, k;
data_t B_trans[32*32];
data_t acc_temp0, acc_temp1;
data_t *A_j, *B_i;
data_t *A_j_k, *B_i_k;
int z;
//for (i = 0; i < 32; i++) {
// for (j = 0; j < 32; j++) {
// B_trans[i*lda+j] = B[i+j*lda];
// }
//}
if (coreid == 0) {
for (i = 0; i < 32; i++) {
B_i = B_trans+i*32;
for (z = 0; z < 32; z++) {
*(B_i+z) = B[i+z*32];
}
for (j = 0; j < 16; j+=2) {
A_j = A+j*lda;
acc_temp0 = 0;
for (k = 0; k < 32; k+=8) {
A_j_k = A_j+k;
B_i_k = B_i+k;
acc_temp0 += *(A_j_k) * *(B_i_k);
acc_temp0 += *(A_j_k + 1) * *(B_i_k + 1);
acc_temp0 += *(A_j_k + 2) * *(B_i_k + 2);
acc_temp0 += *(A_j_k + 3) * *(B_i_k + 3);
acc_temp0 += *(A_j_k + 4) * *(B_i_k + 4);
acc_temp0 += *(A_j_k + 5) * *(B_i_k + 5);
acc_temp0 += *(A_j_k + 6) * *(B_i_k + 6);
acc_temp0 += *(A_j_k + 7) * *(B_i_k + 7);
}
A_j += 32;
acc_temp1 = 0;
for (k = 0; k < 32; k+=8) {
acc_temp1 += *(A_j+k) * *(B_i+k);
acc_temp1 += *(A_j+k + 1) * *(B_i+k + 1);
acc_temp1 += *(A_j+k + 2) * *(B_i+k + 2);
acc_temp1 += *(A_j+k + 3) * *(B_i+k + 3);
acc_temp1 += *(A_j+k + 4) * *(B_i+k + 4);
acc_temp1 += *(A_j+k + 5) * *(B_i+k + 5);
acc_temp1 += *(A_j+k + 6) * *(B_i+k + 6);
acc_temp1 += *(A_j+k + 7) * *(B_i+k + 7);
}
C[i + j*lda] = acc_temp0;
C[i + (j+1)*lda] = acc_temp1;
}
}
} else if (coreid == 1) {
for (i = 0; i < 32; i++) {
B_i = B_trans+i*32;
for (z = 0; z < 32; z++) {
*(B_i+z) = B[i+z*32];
}
for (j = 16; j < 32; j+=2) {
A_j = A+j*lda;
acc_temp0 = 0;
for (k = 0; k < 32; k+=8) {
acc_temp0 += *(A_j+k) * *(B_i+k);
acc_temp0 += *(A_j+k + 1) * *(B_i+k + 1);
acc_temp0 += *(A_j+k + 2) * *(B_i+k + 2);
acc_temp0 += *(A_j+k + 3) * *(B_i+k + 3);
acc_temp0 += *(A_j+k + 4) * *(B_i+k + 4);
acc_temp0 += *(A_j+k + 5) * *(B_i+k + 5);
acc_temp0 += *(A_j+k + 6) * *(B_i+k + 6);
acc_temp0 += *(A_j+k + 7) * *(B_i+k + 7);
}
A_j += 32;
acc_temp1 = 0;
for (k = 0; k < 32; k+=8) {
acc_temp1 += *(A_j+k) * *(B_i+k);
acc_temp1 += *(A_j+k + 1) * *(B_i+k + 1);
acc_temp1 += *(A_j+k + 2) * *(B_i+k + 2);
acc_temp1 += *(A_j+k + 3) * *(B_i+k + 3);
acc_temp1 += *(A_j+k + 4) * *(B_i+k + 4);
acc_temp1 += *(A_j+k + 5) * *(B_i+k + 5);
acc_temp1 += *(A_j+k + 6) * *(B_i+k + 6);
acc_temp1 += *(A_j+k + 7) * *(B_i+k + 7);
}
C[i + j*lda] = acc_temp0;
C[i + (j+1)*lda] = acc_temp1;
}
}
}
}
void __attribute__((noinline)) matmul_MI(const int lda, const data_t A[], const data_t B[], data_t C[] )
{
int i, j, k;
data_t acc_temp;
data_t *A_j, *B_i;
int j_start = coreid*16;
int j_end = (coreid*16)+16;
if (coreid == 0) {
for ( i = 0; i < 32; i++ ) {
B_i = B + i;
for ( j = j_start; j < j_end; j++ )
{
acc_temp = 0;
A_j = A + j*32;
for ( k = 0; k < 32; k++ )
{
acc_temp += *(A_j + k) * *(B_i + k*32);
}
C[i + j*32] = acc_temp;
}
}
} else if (coreid == 1) {
for ( i = 16; i < 32; i++ ) {
B_i = B + i;
for ( j = j_start; j < j_end; j++ )
{
acc_temp = 0;
A_j = A + j*32;
for ( k = 0; k < 32; k+=4 )
{
acc_temp += *(A_j + k) * *(B_i + k*32);
acc_temp += *(A_j + k + 1) * *(B_i + (k+1)*32);
acc_temp += *(A_j + k + 2) * *(B_i + (k+2)*32);
acc_temp += *(A_j + k + 3) * *(B_i + (k+3)*32);
}
C[i + j*32] = acc_temp;
}
}
for ( i = 0; i < 16; i++ ) {
B_i = B + i;
for ( j = j_start; j < j_end; j++ )
{
acc_temp = 0;
A_j = A + j*32;
for ( k = 0; k < 32; k+=4 )
{
acc_temp += *(A_j + k) * *(B_i + k*32);
acc_temp += *(A_j + k + 1) * *(B_i + (k+1)*32);
acc_temp += *(A_j + k + 2) * *(B_i + (k+2)*32);
acc_temp += *(A_j + k + 3) * *(B_i + (k+3)*32);
}
C[i + j*32] = acc_temp;
}
}
}
}
void __attribute__((noinline)) matmul_MSI(const int lda, const data_t A[], const data_t B[], data_t C[] )
{
int i, j, k;
data_t acc_temp;
data_t *A_j, *B_i;
int j_start = coreid*16;
int j_end = (coreid*16)+16;
for ( i = 0; i < 32; i++ ) {
B_i = B + i;
for ( j = j_start; j < j_end; j++ )
{
acc_temp = 0;
A_j = A + j*32;
for ( k = 0; k < 32; k++ )
{
acc_temp += *(A_j + k) * *(B_i + k*32);
}
C[i + j*32] = acc_temp;
}
}
}
void __attribute__((noinline)) matmul(const int lda, const data_t A[], const data_t B[], data_t C[] )
{
// ***************************** //
// **** ADD YOUR CODE HERE ***** //
// ***************************** //
//
// feel free to make a separate function for MI and MSI versions.
// ENABLE_SHARING = false is MI
// ENABLE_SHARING = true is MSI
matmul_MI_transpose(lda, A, B, C);
//matmul_MSI(lda, A, B, C);
}
//--------------------------------------------------------------------------
// Main
//
// all threads start executing thread_entry(). Use their "coreid" to
// differentiate between threads (each thread is running on a separate core).
void thread_entry(int cid, int nc)
{
coreid = cid;
ncores = nc;
// static allocates data in the binary, which is visible to both threads
static data_t results_data[ARRAY_SIZE];
// // Execute the provided, naive matmul
// barrier();
// //stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
//
//
// // verify
// //verify(ARRAY_SIZE, results_data, verify_data);
//
// // clear results from the first trial
// size_t i;
// if (coreid == 0)
// for (i=0; i < ARRAY_SIZE; i++)
// results_data[i] = 0;
// barrier();
// Execute your faster matmul
barrier();
stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
#ifdef DEBUG
printArray("results:", ARRAY_SIZE, results_data);
printArray("verify :", ARRAY_SIZE, verify_data);
#endif
// verify
verify(ARRAY_SIZE, results_data, verify_data);
barrier();
exit(0);
}
|
