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
|
#include <stdlib.h>
#include <stdio.h>
#include "../../support/timer.h"
#ifndef T
#define T double
#endif
#if NUMA
#include <numaif.h>
#include <numa.h>
struct bitmask* bitmask_in;
struct bitmask* bitmask_out;
void* mp_pages[1];
int mp_status[1];
int mp_nodes[1];
int numa_node_in = -1;
int numa_node_out = -1;
int numa_node_cpu = -1;
#endif
#if NUMA_MEMCPY
struct bitmask* bitmask_cpu;
int numa_node_local = -1;
int numa_node_in_is_local = 0;
#endif
#define XSTR(x) STR(x)
#define STR(x) #x
#include "gemv_utils.h"
int main(int argc, char *argv[])
{
(void) argc;
/* // upstream config:
const size_t rows = 20480;
const size_t cols = 8192;
*/
// DPU config: 163840 -n 4096
const size_t rows = 163840;
const size_t cols = 4096;
T **A, *b, *x;
T **A_local, *x_local;
#if NUMA
bitmask_in = numa_parse_nodestring(argv[1]);
bitmask_out = numa_parse_nodestring(argv[2]);
numa_node_cpu = atoi(argv[3]);
#if NUMA_MEMCPY
bitmask_cpu = numa_parse_nodestring(argv[4]);
#endif // NUMA_MEMCPY
#else
(void) argv;
#endif // NUMA
#if NUMA
if (bitmask_out) {
numa_set_membind(bitmask_out);
numa_free_nodemask(bitmask_out);
}
b = (T*) numa_alloc(sizeof(T)*rows);
#else
b = (T*) malloc(sizeof(T)*rows);
#endif
#if NUMA
if (bitmask_in) {
numa_set_membind(bitmask_in);
// no free yet, re-used in allocate_dense
}
x = (T*) numa_alloc(sizeof(T)*cols);
#else
x = (T*) malloc(sizeof(T)*cols);
#endif
allocate_dense(rows, cols, &A);
#if NUMA
if (bitmask_in) {
numa_free_nodemask(bitmask_in);
}
#endif
make_hilbert_mat(rows,cols, &A);
#if NUMA
#if NUMA_MEMCPY
if (bitmask_cpu) {
numa_set_membind(bitmask_cpu);
numa_free_nodemask(bitmask_cpu);
}
#else
struct bitmask *bitmask_all = numa_allocate_nodemask();
numa_bitmask_setall(bitmask_all);
numa_set_membind(bitmask_all);
numa_free_nodemask(bitmask_all);
#endif // NUMA_MEMCPY
#endif // NUMA
A_local = A;
x_local = x;
#if NUMA
mp_pages[0] = A;
if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) {
perror("move_pages(A)");
}
else if (mp_status[0] < 0) {
printf("move_pages(A) error: %d", mp_status[0]);
}
else {
numa_node_in = mp_status[0];
}
mp_pages[0] = b;
if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) {
perror("move_pages(b)");
}
else if (mp_status[0] < 0) {
printf("move_pages(b) error: %d", mp_status[0]);
}
else {
numa_node_out = mp_status[0];
}
if (numa_node_cpu != -1) {
if (numa_run_on_node(numa_node_cpu) == -1) {
perror("numa_run_on_node");
numa_node_cpu = -1;
}
}
#endif
#if NUMA_MEMCPY
numa_node_in_is_local = ((numa_node_cpu == numa_node_in) || (numa_node_cpu + 8 == numa_node_in)) * 1;
#endif
Timer timer;
for (int i = 0; i < 40; i++) {
#pragma omp parallel
{
#pragma omp for
for (size_t i = 0; i < cols; i++) {
x[i] = (T) i+1 ;
}
#pragma omp for
for (size_t i = 0; i < rows; i++) {
b[i] = (T) 0;
}
}
#if NUMA_MEMCPY
start(&timer, 1, 0);
if (!numa_node_in_is_local) {
x_local = (T*) numa_alloc(sizeof(T)*cols);
allocate_dense(rows, cols, &A_local);
}
stop(&timer, 1);
if (x_local == NULL) {
return 1;
}
if (A_local == NULL) {
return 1;
}
start(&timer, 2, 0);
if (!numa_node_in_is_local) {
//for (size_t i=0; i < rows; i++ ) {
// memcpy(A_local[i], A[i], cols * sizeof(T));
//}
memcpy(*A_local, *A, rows * cols * sizeof(T));
memcpy(x_local, x, cols * sizeof(T));
} else {
A_local = A;
x_local = x;
}
stop(&timer, 2);
mp_pages[0] = A_local;
if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) {
perror("move_pages(A_local)");
}
else if (mp_status[0] < 0) {
printf("move_pages error: %d", mp_status[0]);
}
else {
numa_node_local = mp_status[0];
}
#endif
unsigned int nr_threads = 0;
#pragma omp parallel
#pragma omp atomic
nr_threads++;
start(&timer, 0, 0);
gemv(A_local, x_local, rows, cols, &b);
stop(&timer, 0);
#if NUMA_MEMCPY
start(&timer, 3, 0);
if (!numa_node_in_is_local) {
numa_free(x_local, sizeof(T) * cols);
numa_free(*A_local, sizeof(T) * rows * cols);
numa_free(A_local, sizeof(void*) * rows);
}
stop(&timer, 3);
#endif
#if NUMA_MEMCPY
printf("[::] GEMV-CPU-MEMCPY | n_threads=%d e_type=%s n_elements=%ld"
" numa_node_in=%d numa_node_out=%d numa_node_cpu=%d numa_node_local=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
" | throughput_MBps=%f throughput_MOpps=%f",
nr_threads, XSTR(T), rows * cols,
numa_node_in, numa_node_out, numa_node_cpu, numa_node_local, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out),
rows * cols * sizeof(T) / timer.time[0],
rows * cols / timer.time[0]);
printf(" latency_kernel_us=%f latency_alloc_us=%f latency_memcpy_us=%f latency_free_us=%f latency_total_us=%f\n",
timer.time[0], timer.time[1], timer.time[2], timer.time[3],
timer.time[0] + timer.time[1] + timer.time[2] + timer.time[3]);
#else
printf("[::] GEMV-CPU | n_threads=%d e_type=%s n_elements=%ld"
#if NUMA
" numa_node_in=%d numa_node_out=%d numa_node_cpu=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
#endif
" | throughput_MBps=%f",
nr_threads, XSTR(T), rows * cols,
#if NUMA
numa_node_in, numa_node_out, numa_node_cpu, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out),
#endif
rows * cols * sizeof(T) / timer.time[0]);
printf(" throughput_MOpps=%f latency_us=%f\n",
rows * cols / timer.time[0], timer.time[0]);
#endif
}
#if 0
print_vec(x, rows);
print_mat(A, rows, cols);
print_vec(b, rows);
#endif
#if TYPE_double || TYPE_float
printf("sum(x) = %f, sum(Ax) = %f\n", sum_vec(x,cols), sum_vec(b,rows));
#else
printf("sum(x) = %d, sum(Ax) = %d\n", sum_vec(x,cols), sum_vec(b,rows));
#endif
#if NUMA
numa_free(b, sizeof(T)*rows);
numa_free(x, sizeof(T)*cols);
numa_free(*A, sizeof(T)*rows*cols);
numa_free(A, sizeof(void*)*rows);
#else
free(b);
free(x);
free(*A);
free(A);
#endif
return 0;
}
void gemv(T** A, T* x, size_t rows, size_t cols, T** b) {
#pragma omp parallel for
for (size_t i = 0; i < rows; i ++ )
for (size_t j = 0; j < cols; j ++ ) {
(*b)[i] = (*b)[i] + A[i][j]*x[j];
}
}
void make_hilbert_mat(size_t rows, size_t cols, T*** A) {
#pragma omp parallel for
for (size_t i = 0; i < rows; i++) {
for (size_t j = 0; j < cols; j++) {
#if TYPE_double || TYPE_float
(*A)[i][j] = 1.0/( (T) i + (T) j + 1.0);
#else
(*A)[i][j] = (T)(((i+j)%10));
#endif
}
}
}
T sum_vec(T* vec, size_t rows) {
T sum = 0;
#pragma omp parallel for reduction(+:sum)
for (int i = 0; i < rows; i++) sum = sum + vec[i];
return sum;
}
|
