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
|
/**
* app.c
* SCAN-SSA Host Application Source File
*
*/
#include <stdio.h>
#include <stdlib.h>
#include <stdbool.h>
#include <string.h>
#include <stdint.h>
#include <unistd.h>
#include <getopt.h>
#include <assert.h>
#include "../support/common.h"
#include "../support/timer.h"
#include "../support/params.h"
#define XSTR(x) STR(x)
#define STR(x) #x
#define NR_DPUS 1
extern int omp_get_num_threads();
// 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());
}
for (unsigned int i = nr_elements; i < nr_elements_round; i++) {
A[i] = 0;
}
}
// Compute output in the host (slow reference implementation)
static void scan_host(T* C, T* A, unsigned int nr_elements) {
C[0] = A[0];
for (unsigned int i = 1; i < nr_elements; i++) {
C[i] = C[i - 1] + A[i];
}
}
// Compute output in the host (OMP implementation)
static void scan_omp(T* C, T* A, unsigned int nr_elements, unsigned int nr_threads) {
unsigned int i;
unsigned int elem_per_block = nr_elements / nr_threads;
T accum = 0;
T tmp[nr_threads];
// parallel block-wise scan
#pragma omp parallel for
for (i=0; i < nr_threads; i++) {
unsigned int start_index = elem_per_block * i;
unsigned int stop_index = start_index + elem_per_block;
C[start_index] = A[start_index];
for (unsigned int j = start_index + 1; j < stop_index; j++) {
C[j] = C[j - 1] + A[j];
}
}
// sequential scan
tmp[0] = 0;
for (i=1; i < nr_threads; i++) {
unsigned int start_index = elem_per_block * i;
accum += C[start_index-1];
tmp[i] = accum;
}
// parallel block-wise add
#pragma omp parallel for
for (i=0; i < nr_threads; i++) {
unsigned int start_index = elem_per_block * i;
unsigned int stop_index = start_index + elem_per_block;
for (unsigned int j = start_index; j < stop_index; j++) {
C[j] += tmp[i];
}
}
}
/*
* "SCAN" := [a1, a1, ..., an] -> [a1, a1 + a2, ..., a1 + a2 + ... + an]
* From the paper:
* SCAN-SSA has three steps:
* * (copy data to the DPU) (timer 1)
* * compute the partial scan operation locally inside each DPU (timer 2)
* * retrieve last element of each local scan from the DPUs, do a local (complete) scan on the last elements, and push them to the next DPU (timer 3)
* * finish the scan operation in each DPU (add sum [a1 ... ax] to a(x+1), ..., a(x+m) for DPU size m) (timer 4)
* * (retrieve results from DPU) (timer 5)
*/
// Main of the Host Application
int main(int argc, char **argv) {
struct Params p = input_params(argc, argv);
unsigned int i = 0;
unsigned int nr_threads = 0;
#pragma omp parallel
#pragma omp atomic
nr_threads++;
const unsigned int input_size = p.input_size * nr_threads; // Total input size (weak or strong scaling)
// Input/output allocation
A = malloc(input_size * sizeof(T));
C = malloc(input_size * sizeof(T));
C2 = malloc(input_size * sizeof(T));
T *bufferC = C2;
// Create an input file with arbitrary data
read_input(A, input_size, input_size);
// Timer declaration
Timer timer;
// 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);
scan_host(C, A, input_size);
if(rep >= p.n_warmup)
stop(&timer, 0);
// Compute output on CPU (OMP)
if(rep >= p.n_warmup)
start(&timer, 1, 0);
scan_omp(C2, A, input_size, nr_threads);
if(rep >= p.n_warmup)
stop(&timer, 1);
// Check output
bool status = true;
for (i = 0; i < input_size; 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");
printf("[::] n_threads=%d e_type=%s n_elements=%d "
"| throughput_cpu_ref_MBps=%f throughput_cpu_omp_MBps=%f\n",
nr_threads, XSTR(T), input_size,
input_size * sizeof(T) / timer.time[0],
input_size * sizeof(T) / timer.time[1]);
printf("[::] n_threads=%d e_type=%s n_elements=%d "
"| throughput_cpu_ref_MOpps=%f throughput_cpu_omp_MOpps=%f\n",
nr_threads, XSTR(T), input_size,
input_size / timer.time[0],
input_size / timer.time[1]);
printf("[::] n_threads=%d e_type=%s n_elements=%d | ",
nr_threads, XSTR(T), input_size);
printall(&timer, 1);
} else {
printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
}
}
/*
// Print timing results
printf("CPU ");
print(&timer, 0, p.n_reps);
printf("CPU-DPU ");
print(&timer, 1, p.n_reps);
printf("DPU Kernel Scan ");
print(&timer, 2, p.n_reps);
printf("Inter-DPU (Scan) ");
print(&timer, 3, p.n_reps);
printf("DPU Kernel Add ");
print(&timer, 4, p.n_reps);
printf("DPU-CPU ");
print(&timer, 5, p.n_reps);
printf("\n");
*/
#if ENERGY
double energy;
DPU_ASSERT(dpu_probe_get(&probe, DPU_ENERGY, DPU_AVERAGE, &energy));
printf("DPU Energy (J): %f\t", energy);
#endif
// Deallocation
free(A);
free(C);
free(C2);
return 0;
}
|
