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
|
#include "Halide.h"
#include "halide_benchmark.h"
using namespace Halide;
int main(int argc, char **argv) {
Target t = get_jit_target_from_environment();
if (t.arch == Target::WebAssembly) {
printf("[SKIP] Performance tests are meaningless and/or misleading under WebAssembly interpreter.\n");
return 0;
}
if (!(t.has_feature(Target::CUDA) ||
t.has_feature(Target::Metal))) {
printf("[SKIP] No GPU target enabled supporting half-precision.\n");
return 0;
}
std::vector<Target::Feature> cuda_capabilities{Target::CUDACapability30, Target::CUDACapability32, Target::CUDACapability35, Target::CUDACapability50, Target::CUDACapability61};
if (t.has_feature(Target::CUDA) && !(t.features_any_of(cuda_capabilities))) {
printf("[SKIP] Need CUDA Capability 30 or greater.\n");
return 0;
}
std::cout << t.to_string() << "\n";
// Test three variants, in increasing order of speed.
// 1) Store as float, math as float
// 2) Store as half, math as float
// 3) Store as half, math as half
const int size = 1024 * 1024 * 10;
const int step = 1024;
Buffer<float> f32_in(size + step);
Buffer<float16_t> f16_in(size + step);
f32_in.fill(2.0f);
f16_in.fill(float16_t(2.0f));
Buffer<float> f32_out(size);
Buffer<float16_t> f16_out(size);
Func f1, f2, f3;
Var x;
f1(x) = f32_in(x) * f32_in(x + step);
f2(x) = f16_in(x) * f16_in(x + step);
Var xi;
f1.bound(x, 0, size).vectorize(x, 2).gpu_tile(x, xi, 32);
f2.bound(x, 0, size).vectorize(x, 4).gpu_tile(x, xi, 32);
f1.compile_jit(t);
f2.compile_jit(t);
double t1 = Tools::benchmark([&]() {f1.realize(f32_out); f32_out.device_sync(); });
double t2 = Tools::benchmark([&]() {f2.realize(f16_out); f16_out.device_sync(); });
printf("Times: %f %f\n", t1, t2);
printf("Speed-up from using half type: %f x\n", t1 / t2);
if (t2 > t1) {
printf("Half should not have been slower than float\n");
return 1;
}
f32_out.copy_to_host();
f16_out.copy_to_host();
for (int i = 0; i < size; i++) {
if (f32_out(i) != 4.0f) {
printf("f32_out(%d) = %f instead of 4\n", i, f32_out(i));
return 1;
}
if (f16_out(i) != float16_t(4.0f)) {
printf("f16_out(%d) = %f instead of 4\n", i, (float)f16_out(i));
return 1;
}
}
printf("Success!\n");
return 0;
}
|
