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
|
#include "Halide.h"
#include "halide_benchmark.h"
using namespace Halide;
// Return zero, slowly
Expr expensive_zero(Expr x, Expr y, Expr t, int n) {
// Count how many Fermat's last theorem counterexamples we can find using n trials.
RDom r(0, n);
Func a, b, c;
Var z;
a(x, y, t, z) = random_int() % 1024 + 5;
b(x, y, t, z) = random_int() % 1024 + 5;
c(x, y, t, z) = random_int() % 1024 + 5;
return sum(select(pow(a(x, y, t, r), 3) + pow(b(x, y, t, r), 3) == pow(c(x, y, t, r), 3), 1, 0));
}
int main(int argc, char **argv) {
if (!get_jit_target_from_environment().has_gpu_feature()) {
printf("[SKIP] No GPU target enabled.\n");
return 0;
}
// Compute frames on GPU/CPU, and then sum then on
// CPU/GPU. async() lets us overlap the CPU computation with the
// copies.
Var x, y, t, xo, yo;
const int N = 16;
RDom r(0, N);
for (int i = 0; i < 4; i++) {
Func frames;
frames(x, y, t) = expensive_zero(x, y, t, 1) + ((x + y) % 8) + t;
Func avg;
avg(x, y) += frames(x, y, r);
if (i == 0) {
// Synchronously GPU -> CPU
avg.compute_root().update().reorder(x, y, r).vectorize(x, 8);
frames.store_root().compute_at(frames.in(), Var::outermost()).gpu_tile(x, y, xo, yo, x, y, 16, 16);
frames.in().store_root().compute_at(avg, r).copy_to_host();
} else if (i == 1) {
// Asynchronously GPU -> CPU, via a double-buffer
avg.compute_root().update().reorder(x, y, r).vectorize(x, 8);
frames.store_root().compute_at(frames.in(), Var::outermost()).gpu_tile(x, y, xo, yo, x, y, 16, 16);
frames.in().store_root().compute_at(avg, r).copy_to_host().fold_storage(t, 2).async();
} else if (i == 2) {
// Synchronously CPU -> GPU
avg.compute_root().gpu_tile(x, y, xo, yo, x, y, 16, 16).update().reorder(x, y, r).gpu_tile(x, y, xo, yo, x, y, 16, 16);
frames.store_root().compute_at(avg, r).vectorize(x, 8).fold_storage(t, 2).parallel(y);
frames.in().store_root().compute_at(avg, r).copy_to_device();
} else if (i == 3) {
// Asynchronously CPU -> GPU, via a double-buffer
avg.compute_root().gpu_tile(x, y, xo, yo, x, y, 16, 16).update().reorder(x, y, r).gpu_tile(x, y, xo, yo, x, y, 16, 16);
frames.store_root().compute_at(avg, r).vectorize(x, 8).fold_storage(t, 2).async().parallel(y);
frames.in().store_root().compute_at(avg, r).copy_to_device();
}
Buffer<int> out = avg.realize({1024, 1024});
for (int y = 0; y < out.height(); y++) {
for (int x = 0; x < out.width(); x++) {
int correct = ((x + y) % 8) * N + (N * (N - 1)) / 2;
int actual = out(x, y);
if (correct != actual) {
printf("out(%d, %d) = %d instead of %d\n",
x, y, actual, correct);
return 1;
}
}
}
// Report a benchmark, but don't assert anything about it. Not
// sure how to tune the relative cost of the two stages to
// make the async version reliably better than the non-async
// version.
double t = Halide::Tools::benchmark(3, 3, [&]() {avg.realize(out); out.device_sync(); });
printf("Case %d: %f\n", i, t);
}
printf("Success!\n");
return 0;
}
|
