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
|
/**
* app.c
* UNI 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
#if ENERGY
#include <dpu_probe.h>
#endif
// Pointer declaration
static T* A;
static T* C;
static T* C2;
// Create input arrays
static void read_input(T* A, unsigned int nr_elements, unsigned int nr_elements_round) {
//srand(0);
printf("nr_elements\t%u\t", nr_elements);
for (unsigned int i = 0; i < nr_elements; i++) {
//A[i] = (T) (rand());
A[i] = i%2==0?i:i+1;
}
for (unsigned int i = nr_elements; i < nr_elements_round; i++) {
A[i] = A[nr_elements - 1];
}
}
// Compute output in the host
static unsigned int unique_host(T* C, T* A, unsigned int nr_elements) {
unsigned int pos = 0;
C[pos] = A[pos];
pos++;
for(unsigned int i = 1; i < nr_elements; i++) {
if(A[i] != A[i-1]) {
C[pos] = A[i];
pos++;
}
}
return pos;
}
// 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;
#if ENERGY
struct dpu_probe_t probe;
DPU_ASSERT(dpu_probe_init("energy_probe", &probe));
#endif
// 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;
uint32_t accum = 0;
uint32_t total_count = 0;
const unsigned int input_size = p.exp == 0 ? p.input_size * nr_of_dpus : p.input_size; // Total input size (weak or strong scaling)
const unsigned int input_size_dpu_ = divceil(input_size, nr_of_dpus); // Input size per DPU (max.)
const unsigned int input_size_dpu_round =
(input_size_dpu_ % (NR_TASKLETS * REGS) != 0) ? roundup(input_size_dpu_, (NR_TASKLETS * REGS)) : input_size_dpu_; // Input size per DPU (max.), 8-byte aligned
// Input/output allocation
A = malloc(input_size_dpu_round * nr_of_dpus * sizeof(T));
C = malloc(input_size_dpu_round * nr_of_dpus * sizeof(T));
C2 = malloc(input_size_dpu_round * nr_of_dpus * sizeof(T));
T *bufferA = A;
T *bufferC = C2;
// Create an input file with arbitrary data
read_input(A, input_size, input_size_dpu_round * nr_of_dpus);
// 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, rep - p.n_warmup);
total_count = unique_host(C, A, input_size);
if(rep >= p.n_warmup)
stop(&timer, 0);
printf("Load input data\n");
if(rep >= p.n_warmup)
start(&timer, 1, rep - p.n_warmup);
// Input arguments
const unsigned int input_size_dpu = input_size_dpu_round;
unsigned int kernel = 0;
dpu_arguments_t input_arguments = {input_size_dpu * sizeof(T), kernel};
// Copy input arrays
i = 0;
DPU_FOREACH(dpu_set, dpu, i) {
DPU_ASSERT(dpu_prepare_xfer(dpu, &input_arguments));
}
DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, "DPU_INPUT_ARGUMENTS", 0, sizeof(input_arguments), DPU_XFER_DEFAULT));
DPU_FOREACH(dpu_set, dpu, i) {
DPU_ASSERT(dpu_prepare_xfer(dpu, bufferA + input_size_dpu * i));
}
DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, DPU_MRAM_HEAP_POINTER_NAME, 0, input_size_dpu * sizeof(T), DPU_XFER_DEFAULT));
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, rep - p.n_warmup);
#if ENERGY
DPU_ASSERT(dpu_probe_start(&probe));
#endif
}
DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS));
if(rep >= p.n_warmup) {
stop(&timer, 2);
#if ENERGY
DPU_ASSERT(dpu_probe_stop(&probe));
#endif
}
#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");
dpu_results_t results[nr_of_dpus];
uint32_t* results_scan = malloc(nr_of_dpus * sizeof(uint32_t));
uint32_t* offset = calloc(nr_of_dpus, sizeof(uint32_t));
uint32_t* offset_scan = calloc(nr_of_dpus, sizeof(uint32_t));
i = 0;
accum = 0;
if(rep >= p.n_warmup)
start(&timer, 3, rep - p.n_warmup);
// PARALLEL RETRIEVE TRANSFER
dpu_results_t* results_retrieve[nr_of_dpus];
DPU_FOREACH(dpu_set, dpu, i) {
results_retrieve[i] = (dpu_results_t*)malloc(NR_TASKLETS * sizeof(dpu_results_t));
DPU_ASSERT(dpu_prepare_xfer(dpu, results_retrieve[i]));
}
DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_FROM_DPU, "DPU_RESULTS", 0, NR_TASKLETS * sizeof(dpu_results_t), DPU_XFER_DEFAULT));
DPU_FOREACH(dpu_set, dpu, i) {
// Retrieve tasklet timings
for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) {
// First output element of this DPU
if(each_tasklet == 0){
results[i].first = results_retrieve[i][each_tasklet].first;
}
// Last output element of this DPU and count
if(each_tasklet == NR_TASKLETS - 1){
results[i].t_count = results_retrieve[i][each_tasklet].t_count;
results[i].last = results_retrieve[i][each_tasklet].last;
}
}
// Check if first(i) == last(i-1) -- offset
if(i != 0){
if(results[i].first == results[i - 1].last)
offset[i] = 1;
// Sequential scan - offset
offset_scan[i] += offset[i];
}
// Sequential scan
uint32_t temp = results[i].t_count - offset[i];
results_scan[i] = accum;
accum += temp;
#if PRINT
printf("i=%d -- %u, %u, %u -- %u\n", i, results_scan[i], accum, temp, offset_scan[i]);
#endif
free(results_retrieve[i]);
}
if(rep >= p.n_warmup)
stop(&timer, 3);
i = 0;
if(rep >= p.n_warmup)
start(&timer, 4, rep - p.n_warmup);
DPU_FOREACH (dpu_set, dpu) {
// Copy output array
DPU_ASSERT(dpu_copy_from(dpu, DPU_MRAM_HEAP_POINTER_NAME, input_size_dpu * sizeof(T), bufferC + results_scan[i] - offset_scan[i], results[i].t_count * sizeof(T)));
i++;
}
if(rep >= p.n_warmup)
stop(&timer, 4);
// Free memory
free(results_scan);
free(offset);
free(offset_scan);
}
// Print timing results
printf("CPU ");
print(&timer, 0, p.n_reps);
printf("CPU-DPU ");
print(&timer, 1, p.n_reps);
printf("DPU Kernel ");
print(&timer, 2, p.n_reps);
printf("Inter-DPU ");
print(&timer, 3, p.n_reps);
printf("DPU-CPU ");
print(&timer, 4, p.n_reps);
#if ENERGY
double energy;
DPU_ASSERT(dpu_probe_get(&probe, DPU_ENERGY, DPU_AVERAGE, &energy));
printf("DPU Energy (J): %f\t", energy);
#endif
// Check output
bool status = true;
if(accum != total_count) status = false;
#if PRINT
printf("accum %u, total_count %u\n", accum, total_count);
#endif
for (i = 0; i < accum; i++) {
if(C[i] != bufferC[i]){
status = false;
#if PRINT
printf("%d: %lu -- %lu\n", i, C[i], bufferC[i]);
#endif
}
}
if (status) {
printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
} else {
printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
}
// Deallocation
free(A);
free(C);
free(C2);
DPU_ASSERT(dpu_free(dpu_set));
return status ? 0 : -1;
}
|