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
|
#include <cstdint>
#include <cstdlib>
#include <cstring>
#include <iostream>
#include "utest_helper.hpp"
#include <sys/time.h>
#include <iomanip>
#include <algorithm>
using namespace std;
/* NDRANGE */
#define WG_GLOBAL_SIZE (512 * 256)
#define WG_LOCAL_SIZE 128
#define WG_LOOP_COUNT 10000
enum WG_FUNCTION
{
WG_REDUCE_ADD,
WG_REDUCE_MIN,
WG_REDUCE_MAX,
WG_SCAN_EXCLUSIVE_ADD,
WG_SCAN_EXCLUSIVE_MAX,
WG_SCAN_EXCLUSIVE_MIN,
WG_SCAN_INCLUSIVE_ADD,
WG_SCAN_INCLUSIVE_MAX,
WG_SCAN_INCLUSIVE_MIN
};
/*
* Generic compute-expected on CPU function for any workgroup type
* and any variable type
*/
template<class T>
static void benchmark_expected(WG_FUNCTION wg_func,
T* input,
T* expected)
{
if(wg_func == WG_REDUCE_ADD)
{
T wg_sum = input[0];
for(uint32_t i = 1; i < WG_LOCAL_SIZE; i++)
wg_sum += input[i];
for(uint32_t i = 0; i < WG_LOCAL_SIZE; i++)
expected[i] = wg_sum;
}
else if(wg_func == WG_REDUCE_MAX)
{
T wg_max = input[0];
for(uint32_t i = 1; i < WG_LOCAL_SIZE; i++)
wg_max = max(input[i], wg_max);
for(uint32_t i = 0; i < WG_LOCAL_SIZE; i++)
expected[i] = wg_max;
}
else if(wg_func == WG_REDUCE_MIN)
{
T wg_min = input[0];
for(uint32_t i = 1; i < WG_LOCAL_SIZE; i++)
wg_min = min(input[i], wg_min);
for(uint32_t i = 0; i < WG_LOCAL_SIZE; i++)
expected[i] = wg_min;
}
else if(wg_func == WG_SCAN_INCLUSIVE_ADD)
{
expected[0] = input[0];
for(uint32_t i = 1; i < WG_LOCAL_SIZE; i++)
expected[i] = input[i] + expected[i - 1];
}
else if(wg_func == WG_SCAN_INCLUSIVE_MAX)
{
expected[0] = input[0];
for(uint32_t i = 1; i < WG_LOCAL_SIZE; i++)
expected[i] = max(input[i], expected[i - 1]);
}
else if(wg_func == WG_SCAN_INCLUSIVE_MIN)
{
expected[0] = input[0];
for(uint32_t i = 1; i < WG_LOCAL_SIZE; i++)
expected[i] = min(input[i], expected[i - 1]);
}
}
/*
* Generic input-expected generate function for any workgroup type
* and any variable type
*/
template<class T>
static void benchmark_data(WG_FUNCTION wg_func,
T* &input,
T* &expected)
{
input = new T[WG_GLOBAL_SIZE];
expected = new T[WG_GLOBAL_SIZE];
/* seed for random inputs */
srand (time(NULL));
/* generate inputs and expected values */
for(uint32_t gid = 0; gid < WG_GLOBAL_SIZE; gid += WG_LOCAL_SIZE)
{
/* input values */
for(uint32_t lid = 0; lid < WG_LOCAL_SIZE; lid++)
input[gid + lid] = (rand() % 112) / 3.1415f;
/* expected values */
benchmark_expected(wg_func, input + gid, expected + gid);
}
}
/*
* Generic benchmark function for any workgroup type
* and any variable type
*/
template<class T>
static double benchmark_generic(WG_FUNCTION wg_func,
T* input,
T* expected)
{
double elapsed = 0;
const uint32_t reduce_loop = 10000;
struct timeval start,stop;
/* input and expected data */
benchmark_data(wg_func, input, expected);
/* prepare input for datatype */
OCL_CREATE_BUFFER(buf[0], 0, WG_GLOBAL_SIZE * sizeof(T), NULL);
OCL_CREATE_BUFFER(buf[1], 0, WG_GLOBAL_SIZE * sizeof(T), NULL);
OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]);
OCL_SET_ARG(1, sizeof(cl_mem), &buf[1]);
OCL_SET_ARG(2, sizeof(cl_uint), &reduce_loop);
/* set input data for GPU */
OCL_MAP_BUFFER(0);
memcpy(buf_data[0], input, WG_GLOBAL_SIZE * sizeof(T));
OCL_UNMAP_BUFFER(0);
/* run the kernel on GPU */
globals[0] = WG_GLOBAL_SIZE;
locals[0] = WG_LOCAL_SIZE;
gettimeofday(&start,0);
OCL_NDRANGE(1);
clFinish(queue);
gettimeofday(&stop,0);
elapsed = time_subtract(&stop, &start, 0);
/* check if mistmatch, display execution time */
OCL_MAP_BUFFER(1);
uint32_t mistmatches = 0;
for (uint32_t i = 0; i < WG_GLOBAL_SIZE; i++)
if(((T *)buf_data[1])[i] != *(expected + i)){
cout << "Err at " << i << ", " <<
((T *)buf_data[1])[i] << " != " << *(expected + i) << endl;
mistmatches++;
}
cout << endl << endl << "Mistmatches " << mistmatches << endl;
cout << "Exec time " << elapsed << endl << endl;
OCL_UNMAP_BUFFER(1);
return BANDWIDTH(WG_GLOBAL_SIZE * WG_LOOP_COUNT, elapsed);
}
/*
* Benchmark workgroup reduce add
*/
double benchmark_workgroup_reduce_add_int(void)
{
cl_int *input = NULL;
cl_int *expected = NULL;
OCL_CREATE_KERNEL_FROM_FILE("bench_workgroup",
"bench_workgroup_reduce_add_int");
return benchmark_generic(WG_REDUCE_ADD, input, expected);
}
MAKE_BENCHMARK_FROM_FUNCTION(benchmark_workgroup_reduce_add_int, "GB/S");
double benchmark_workgroup_reduce_add_long(void)
{
cl_long *input = NULL;
cl_long *expected = NULL;
OCL_CREATE_KERNEL_FROM_FILE("bench_workgroup",
"bench_workgroup_reduce_add_long");
return benchmark_generic(WG_REDUCE_ADD, input, expected);
}
MAKE_BENCHMARK_FROM_FUNCTION(benchmark_workgroup_reduce_add_long, "GB/S");
/*
* Benchmark workgroup reduce min
*/
double benchmark_workgroup_reduce_min_int(void)
{
cl_int *input = NULL;
cl_int *expected = NULL;
OCL_CREATE_KERNEL_FROM_FILE("bench_workgroup",
"bench_workgroup_reduce_min_int");
return benchmark_generic(WG_REDUCE_MIN, input, expected);
}
MAKE_BENCHMARK_FROM_FUNCTION(benchmark_workgroup_reduce_min_int, "GB/S");
double benchmark_workgroup_reduce_min_long(void)
{
cl_long *input = NULL;
cl_long *expected = NULL;
OCL_CREATE_KERNEL_FROM_FILE("bench_workgroup",
"bench_workgroup_reduce_min_long");
return benchmark_generic(WG_REDUCE_MIN, input, expected);
}
MAKE_BENCHMARK_FROM_FUNCTION(benchmark_workgroup_reduce_min_long, "GB/S");
/*
* Benchmark workgroup scan inclusive add
*/
double benchmark_workgroup_scan_inclusive_add_int(void)
{
cl_int *input = NULL;
cl_int *expected = NULL;
OCL_CREATE_KERNEL_FROM_FILE("bench_workgroup",
"bench_workgroup_scan_inclusive_add_int");
return benchmark_generic(WG_SCAN_INCLUSIVE_ADD, input, expected);
}
MAKE_BENCHMARK_FROM_FUNCTION(benchmark_workgroup_scan_inclusive_add_int, "GB/S");
double benchmark_workgroup_scan_inclusive_add_long(void)
{
cl_long *input = NULL;
cl_long *expected = NULL;
OCL_CREATE_KERNEL_FROM_FILE("bench_workgroup",
"bench_workgroup_scan_inclusive_add_long");
return benchmark_generic(WG_SCAN_INCLUSIVE_ADD, input, expected);
}
MAKE_BENCHMARK_FROM_FUNCTION(benchmark_workgroup_scan_inclusive_add_long, "GB/S");
/*
* Benchmark workgroup scan inclusive min
*/
double benchmark_workgroup_scan_inclusive_min_int(void)
{
cl_int *input = NULL;
cl_int *expected = NULL;
OCL_CREATE_KERNEL_FROM_FILE("bench_workgroup",
"bench_workgroup_scan_inclusive_min_int");
return benchmark_generic(WG_SCAN_INCLUSIVE_MIN, input, expected);
}
MAKE_BENCHMARK_FROM_FUNCTION(benchmark_workgroup_scan_inclusive_min_int, "GB/S");
double benchmark_workgroup_scan_inclusive_min_long(void)
{
cl_long *input = NULL;
cl_long *expected = NULL;
OCL_CREATE_KERNEL_FROM_FILE("bench_workgroup",
"bench_workgroup_scan_inclusive_min_long");
return benchmark_generic(WG_SCAN_INCLUSIVE_MIN, input, expected);
}
MAKE_BENCHMARK_FROM_FUNCTION(benchmark_workgroup_scan_inclusive_min_long, "GB/S");
|
