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
|
#include "Halide.h"
#include <stdio.h>
using namespace Halide;
int main(int argc, char **argv) {
Target target = get_jit_target_from_environment();
if (1) {
// Test a tuple reduction on the gpu
Func f;
Var x, y, xo, yo, xi, yi;
f(x, y) = Tuple(x + y, x - y);
// Updates to a reduction are atomic.
f(x, y) = Tuple(f(x, y)[1] * 2, f(x, y)[0] * 2);
// now equals ((x - y)*2, (x + y)*2)
if (target.has_gpu_feature()) {
f.gpu_tile(x, y, xo, yo, xi, yi, 16, 16);
f.update().gpu_tile(x, y, xo, yo, xi, yi, 16, 16);
} else if (target.has_feature(Target::HVX)) {
f.hexagon(y).vectorize(x, 32);
f.update().hexagon(y).vectorize(x, 32);
}
Realization result = f.realize({1024, 1024});
Buffer<int> a = result[0], b = result[1];
for (int y = 0; y < a.height(); y++) {
for (int x = 0; x < a.width(); x++) {
int correct_a = (x - y) * 2;
int correct_b = (x + y) * 2;
if (a(x, y) != correct_a || b(x, y) != correct_b) {
printf("result(%d, %d) = (%d, %d) instead of (%d, %d)\n",
x, y, a(x, y), b(x, y), correct_a, correct_b);
return 1;
}
}
}
}
if (1) {
// Now test one that alternates between cpu and gpu per update step
Func f;
Var x, y, xo, yo, xi, yi;
f(x, y) = Tuple(x + y, x - y);
for (size_t i = 0; i < 10; i++) {
// Swap the tuple elements and increment both
f(x, y) = Tuple(f(x, y)[1] + 1, f(x, y)[0] + 1);
}
// Schedule the pure step and the odd update steps on the gpu
if (target.has_gpu_feature()) {
f.gpu_tile(x, y, xo, yo, xi, yi, 16, 16);
} else if (target.has_feature(Target::HVX)) {
f.hexagon(y).vectorize(x, 32);
}
for (int i = 0; i < 10; i++) {
f.update(i).unscheduled();
if (i & 1) {
if (target.has_gpu_feature()) {
f.update(i).gpu_tile(x, y, xo, yo, xi, yi, 16, 16);
} else if (target.has_feature(Target::HVX)) {
f.update(i).hexagon(y).vectorize(x, 32);
}
}
}
Realization result = f.realize({1024, 1024});
Buffer<int> a = result[0], b = result[1];
for (int y = 0; y < a.height(); y++) {
for (int x = 0; x < a.width(); x++) {
int correct_a = (x + y) + 10;
int correct_b = (x - y) + 10;
if (a(x, y) != correct_a || b(x, y) != correct_b) {
printf("result(%d, %d) = (%d, %d) instead of (%d, %d)\n",
x, y, a(x, y), b(x, y), correct_a, correct_b);
return 1;
}
}
}
}
if (1) {
// Same as above, but switches which steps are gpu and cpu
Func f;
Var x, y, xo, yo, xi, yi;
f(x, y) = Tuple(x + y, x - y);
for (size_t i = 0; i < 10; i++) {
// Swap the tuple elements and increment both
f(x, y) = Tuple(f(x, y)[1] + 1, f(x, y)[0] + 1);
}
// Schedule the even update steps on the gpu
for (int i = 0; i < 10; i++) {
f.update(i).unscheduled();
if (i & 1) {
if (target.has_gpu_feature()) {
f.update(i).gpu_tile(x, y, xo, yo, xi, yi, 16, 16);
} else if (target.has_feature(Target::HVX)) {
f.update(i).hexagon(y).vectorize(x, 32);
}
}
}
Realization result = f.realize({1024, 1024});
Buffer<int> a = result[0], b = result[1];
for (int y = 0; y < a.height(); y++) {
for (int x = 0; x < a.width(); x++) {
int correct_a = (x + y) + 10;
int correct_b = (x - y) + 10;
if (a(x, y) != correct_a || b(x, y) != correct_b) {
printf("result(%d, %d) = (%d, %d) instead of (%d, %d)\n",
x, y, a(x, y), b(x, y), correct_a, correct_b);
return 1;
}
}
}
}
if (1) {
// In this one, each step only uses one of the tuple elements
// of the previous step, so only that buffer should get copied
// back to host or copied to device.
Func f;
Var x, y, xo, yo, xi, yi;
f(x, y) = Tuple(x + y - 1000, x - y + 1000);
for (size_t i = 0; i < 10; i++) {
f(x, y) = Tuple(f(x, y)[1] - 1, f(x, y)[1] + 1);
}
// Schedule the even update steps on the gpu
for (int i = 0; i < 10; i++) {
f.update(i).unscheduled();
if ((i & 1) == 0) {
if (target.has_gpu_feature()) {
f.update(i).gpu_tile(x, y, xo, yo, xi, yi, 16, 16);
} else if (target.has_feature(Target::HVX)) {
f.update(i).hexagon(y).vectorize(x, 32);
}
}
}
Realization result = f.realize({1024, 1024});
Buffer<int> a = result[0], b = result[1];
for (int y = 0; y < a.height(); y++) {
for (int x = 0; x < a.width(); x++) {
int correct_a = (x - y + 1000) + 8;
int correct_b = (x - y + 1000) + 10;
if (a(x, y) != correct_a || b(x, y) != correct_b) {
printf("result(%d, %d) = (%d, %d) instead of (%d, %d)\n",
x, y, a(x, y), b(x, y), correct_a, correct_b);
return 1;
}
}
}
}
printf("Success!\n");
return 0;
}
|
