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
|
#include "Halide.h"
#include <stdio.h>
using namespace Halide;
int main(int argc, char **argv) {
Halide::Target target = get_jit_target_from_environment();
if (!target.has_gpu_feature()) {
printf("[SKIP] No GPU target enabled.\n");
return 0;
}
if (target.has_feature(Target::Vulkan) && ((target.os == Target::IOS) || target.os == Target::OSX)) {
printf("[SKIP] Skipping test for Vulkan on iOS/OSX (MoltenVK doesn't support dynamically allocated shared mem)!\n");
return 0;
}
{
Func f("f"), g("g"), h("h");
Var x("x"), y("y");
Param<bool> use_gpu;
f(x, y) = x + y;
g(x, y) = f(x - 1, y + 1) + f(x + 1, y - 1) + x;
h(x, y) = g(x + 1, y - 1) + g(x - 1, y + 1) + y;
// Specialize is a little tricky for producer-consumer pairs: the
// compute_at must be the same in either case, which means you
// must have a matching var name in either case.
// Compute h in tiles either on the cpu or gpu
Var xo("xo"), yo("yo"), xi("xi"), yi("yi"), t("t");
h.compute_root().specialize(use_gpu).gpu_tile(x, y, xi, yi, 4, 4);
h.tile(x, y, xo, yo, xi, yi, 8, 8).fuse(xo, yo, t).parallel(t);
// Peel off a size-1 loop from blockidx to make a scheduling point
// that matches the cpu case. We need to mark it as serial,
// because by default when you split up a parallel loop both the
// inside and outside are parallel.
h.specialize(use_gpu).split(x, x, t, 1).serial(t);
// Because t exists in both version of h, we can compute g at it.
g.compute_at(h, t);
// If we're on the gpu, we should map g's x and y to thread ids
g.specialize(use_gpu).gpu_threads(x, y);
// We want f compute_at g, x, so do the same trick to g;
g.specialize(use_gpu).split(x, x, xi, 1).serial(xi);
g.rename(x, xi);
f.compute_at(g, xi);
use_gpu.set(get_jit_target_from_environment().has_gpu_feature());
Buffer<int> out1 = h.realize({1024, 1024});
use_gpu.set(false);
Buffer<int> out2 = h.realize({1024, 1024});
for (int y = 0; y < out1.height(); y++) {
for (int x = 0; x < out1.width(); x++) {
int correct = 6 * x + 5 * y;
if (out1(x, y) != correct) {
printf("out1(%d, %d) = %d instead of %d\n",
x, y, out1(x, y), correct);
return 1;
}
if (out2(x, y) != correct) {
printf("out2(%d, %d) = %d instead of %d\n",
x, y, out2(x, y), correct);
return 1;
}
}
}
}
{
Func f("f"), g("g"), h("h");
Var x("x"), y("y");
Param<bool> p;
f(x, y) = x + y;
g(x, y) = f(x, y) + x;
Var xo("xo"), yo("yo"), xi("xi"), yi("yi");
f.specialize(p).tile(x, y, xi, yi, 4, 4).gpu_threads(x, y);
f.tile(x, y, xo, yo, xi, yi, 8, 8).gpu_threads(xo, yo);
f.compute_at(g, x);
g.tile(x, y, xi, yi, 2, 2).gpu_blocks(x, y);
p.set(true);
Buffer<int> out = g.realize({32, 32});
for (int y = 0; y < out.height(); y++) {
for (int x = 0; x < out.width(); x++) {
int correct = 2 * x + y;
if (out(x, y) != correct) {
printf("out(%d, %d) = %d instead of %d\n",
x, y, out(x, y), correct);
return 1;
}
}
}
}
printf("Success!\n");
return 0;
}
|
