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
|
/**
* app.c
* STREAM Host Application Source File
*
*/
#include <stdio.h>
#include <stdlib.h>
#include <stdbool.h>
#include <string.h>
#include <dpu.h>
#include <dpu_log.h>
#include <unistd.h>
#include <getopt.h>
#include <assert.h>
#include "../support/common.h"
#include "../support/timer.h"
#include "../support/params.h"
// Define the DPU Binary path as DPU_BINARY here
#ifndef DPU_BINARY
#define DPU_BINARY "./bin/dpu_code"
#endif
#define XSTR(x) STR(x)
#define STR(x) #x
// Pointer declaration
static T* A;
static T* B;
#if defined(add) || defined(triad)
static T* C;
#endif
static T* C2;
// Create input arrays
static void read_input(T* A, T* B, unsigned int nr_elements) {
srand(0);
printf("nr_elements\t%u\t", nr_elements);
for (unsigned int i = 0; i < nr_elements; i++) {
A[i] = (T) (rand());
B[i] = (T) (rand());
}
}
static char benchmark_name[] =
#ifdef scale
"SCALE"
#elif add
"ADD"
#elif triad
"TRIAD"
#elif copy
"COPY"
#elif copyw
"COPYW"
#endif
;
static char mem_name[] =
#ifdef WRAM
"WRAM"
#endif
#ifdef MRAM
"MRAM"
#endif
;
// Compute output in the host
#if defined(add) || defined(triad)
static void stream_host(T* C, T* B, T* A, unsigned int nr_elements) {
#else
static void stream_host(T* C, T* A, unsigned int nr_elements) {
#endif
for (unsigned int i = 0; i < nr_elements; i++) {
#ifdef scale
C[i] = (nr_elements / NR_DPUS) * A[i];
#elif add
C[i] = A[i] + B[i];
#elif triad
C[i] = A[i] + (nr_elements / NR_DPUS) * B[i];
#else // copy
C[i] = A[i];
#endif
}
}
// Main of the Host Application
int main(int argc, char **argv) {
struct Params p = input_params(argc, argv);
struct dpu_set_t dpu_set, dpu;
uint32_t nr_of_dpus;
// Allocate DPUs and load binary
DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set));
DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL));
DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus));
printf("Allocated %d DPU(s)\n", nr_of_dpus);
unsigned int i = 0;
double cc = 0;
double cc_min = 0;
const unsigned int input_size = p.exp == 0 ? p.input_size * nr_of_dpus : p.input_size;
#if defined(add) || defined(triad)
const unsigned int n_arrays = 3;
#else
const unsigned int n_arrays = 2;
#endif
// Input/output allocation
A = malloc(input_size * sizeof(T));
B = malloc(input_size * sizeof(T));
T *bufferA = A;
T *bufferB = B;
#if defined(add) || defined(triad)
C = malloc(input_size * sizeof(T));
T *bufferC = C;
#endif
C2 = malloc(input_size * sizeof(T));
// Create an input file with arbitrary data
read_input(A, B, input_size);
// Timer declaration
Timer timer;
printf("NR_TASKLETS\t%d\tBL\t%d\n", NR_TASKLETS, BL);
// Loop over main kernel
for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) {
// Compute output on CPU (performance comparison and verification purposes)
if(rep >= p.n_warmup)
start(&timer, 0, 0);
#if defined(add) || defined(triad)
stream_host(C2, B, A, input_size);
#else
stream_host(C2, A, input_size);
#endif
if(rep >= p.n_warmup)
stop(&timer, 0);
//printf("Load input data\n");
if(rep >= p.n_warmup)
start(&timer, 1, 0);
// Input arguments
const unsigned int input_size_dpu = input_size / nr_of_dpus;
unsigned int kernel = 0;
dpu_arguments_t input_arguments = {input_size_dpu * sizeof(T), kernel};
DPU_ASSERT(dpu_copy_to(dpu_set, "DPU_INPUT_ARGUMENTS", 0, (const void *)&input_arguments, sizeof(input_arguments)));
// Copy input arrays
i = 0;
DPU_FOREACH (dpu_set, dpu) {
DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, 0, bufferA + input_size_dpu * i, input_size_dpu * sizeof(T)));
#if defined(add) || defined(triad)
DPU_ASSERT(dpu_copy_to(dpu, DPU_MRAM_HEAP_POINTER_NAME, input_size_dpu * sizeof(T), bufferB + input_size_dpu * i, input_size_dpu * sizeof(T)));
#endif
i++;
}
if(rep >= p.n_warmup)
stop(&timer, 1);
//printf("Run program on DPU(s) \n");
// Run DPU kernel
if(rep >= p.n_warmup)
start(&timer, 2, 0);
DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS));
if(rep >= p.n_warmup)
stop(&timer, 2);
#if PRINT
{
unsigned int each_dpu = 0;
printf("Display DPU Logs\n");
DPU_FOREACH (dpu_set, dpu) {
printf("DPU#%d:\n", each_dpu);
DPU_ASSERT(dpulog_read_for_dpu(dpu.dpu, stdout));
each_dpu++;
}
}
#endif
//printf("Retrieve results\n");
if(rep >= p.n_warmup)
start(&timer, 3, 0);
dpu_results_t results[nr_of_dpus];
i = 0;
DPU_FOREACH (dpu_set, dpu) {
// Copy output array
#if defined(add) || defined(triad)
DPU_ASSERT(dpu_copy_from(dpu, DPU_MRAM_HEAP_POINTER_NAME, 2 * input_size_dpu * sizeof(T), bufferC + input_size_dpu * i, input_size_dpu * sizeof(T)));
#else
DPU_ASSERT(dpu_copy_from(dpu, DPU_MRAM_HEAP_POINTER_NAME, input_size_dpu * sizeof(T), bufferB + input_size_dpu * i, input_size_dpu * sizeof(T)));
#endif
i++;
}
if(rep >= p.n_warmup)
stop(&timer, 3);
#if PERF
i = 0;
DPU_FOREACH (dpu_set, dpu) {
results[i].cycles = 0;
// Retrieve tasklet timings
for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) {
dpu_results_t result;
result.cycles = 0;
DPU_ASSERT(dpu_copy_from(dpu, "DPU_RESULTS", each_tasklet * sizeof(dpu_results_t), &result, sizeof(dpu_results_t)));
if (result.cycles > results[i].cycles)
results[i].cycles = result.cycles;
}
i++;
}
#endif
#if PERF
uint64_t max_cycles = 0;
uint64_t min_cycles = 0xFFFFFFFFFFFFFFFF;
// Print performance results
if(rep >= p.n_warmup){
i = 0;
DPU_FOREACH(dpu_set, dpu) {
if(results[i].cycles > max_cycles)
max_cycles = results[i].cycles;
if(results[i].cycles < min_cycles)
min_cycles = results[i].cycles;
i++;
}
cc += (double)max_cycles;
cc_min += (double)min_cycles;
}
#endif
// Check output
bool status = true;
for (i = 0; i < input_size; i++) {
#if defined(add) || defined(triad)
if(C2[i] != bufferC[i]){
#else
if(C2[i] != bufferB[i]){
#endif
status = false;
#if PRINT
#if defined(add) || defined(triad)
printf("%d: %u -- %u\n", i, C2[i], bufferC[i]);
#else
printf("%d: %u -- %u\n", i, C2[i], bufferB[i]);
#endif
#endif
}
}
if (status) {
printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
printf("[::] n_dpus=%d n_tasklets=%d e_benchmark=%-6s e_type=%s e_mem=%s b_unroll=%d block_size_B=%d | throughput_cpu_MBps=%f throughput_pim_MBps=%f throughput_MBps=%f \n", nr_of_dpus, NR_TASKLETS, benchmark_name, XSTR(T), mem_name, UNROLL, BLOCK_SIZE, input_size * n_arrays * sizeof(T) / timer.time[0], input_size * n_arrays * sizeof(T) / timer.time[2], input_size * n_arrays * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3]));
printf("[::] n_dpus=%d n_tasklets=%d e_benchmark=%-6s e_type=%s e_mem=%s b_unroll=%d block_size_B=%d | throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f \n", nr_of_dpus, NR_TASKLETS, benchmark_name, XSTR(T), mem_name, UNROLL, BLOCK_SIZE, input_size / timer.time[0], input_size / timer.time[2], input_size / (timer.time[1] + timer.time[2] + timer.time[3]));
} else {
printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
}
}
printf("DPU cycles = %g cc\n", cc / p.n_reps);
// Deallocation
free(A);
free(B);
#if defined(add) || defined(triad)
free(C);
#endif
free(C2);
DPU_ASSERT(dpu_free(dpu_set));
return 0;
}
|
