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
|
#include <chrono>
#include <cstdio>
#include "conv_layer.h"
#include "conv_layer_auto_schedule.h"
#include "HalideBuffer.h"
#include "halide_benchmark.h"
using namespace Halide::Tools;
using namespace Halide::Runtime;
int main(int argc, char **argv) {
const int N = 5, CI = 128, CO = 128, W = 100, H = 80;
Buffer<float, 4> input(CI, W + 2, H + 2, N);
Buffer<float, 4> filter(CO, 3, 3, CI);
Buffer<float, 1> bias(CO);
for (int c = 0; c < input.dim(3).extent(); c++) {
for (int z = 0; z < input.channels(); z++) {
for (int y = 0; y < input.height(); y++) {
for (int x = 0; x < input.width(); x++) {
input(x, y, z, c) = rand();
}
}
}
}
for (int c = 0; c < filter.dim(3).extent(); c++) {
for (int z = 0; z < filter.channels(); z++) {
for (int y = 0; y < filter.height(); y++) {
for (int x = 0; x < filter.width(); x++) {
filter(x, y, z, c) = rand();
}
}
}
}
for (int x = 0; x < bias.width(); x++) {
bias(x) = rand();
}
Buffer<float, 4> output(CO, W, H, N);
// This is necessary to get the PTX compiler to do a good
// job. TODO: This should be a scheduling directive or a runtime
// function.
#ifdef _WIN32
_putenv_s("HL_CUDA_JIT_MAX_REGISTERS", "256");
#else
setenv("HL_CUDA_JIT_MAX_REGISTERS", "256", 1);
#endif
conv_layer(input, filter, bias, output);
// Timing code
// Manually-tuned version
double min_t_manual = benchmark(10, 10, [&]() {
conv_layer(input, filter, bias, output);
output.device_sync();
});
printf("Manually-tuned time: %gms\n", min_t_manual * 1e3);
// Auto-scheduled version
double min_t_auto = benchmark(10, 10, [&]() {
conv_layer_auto_schedule(input, filter, bias, output);
output.device_sync();
});
printf("Auto-scheduled time: %gms\n", min_t_auto * 1e3);
printf("Success!\n");
return 0;
}
|
