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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
|
#include "Halide.h"
#include <stdio.h>
using namespace Halide;
int main(int argc, char **argv) {
const bool use_gpu = get_jit_target_from_environment().has_gpu_feature();
// An internal Func that produces multiple values.
{
Func f, g;
Var x, xi;
f(x) = {x, sin(x)};
f.compute_root();
Tuple t = f(x);
g(x) = t[0] + t[1];
if (use_gpu) {
g.gpu_tile(x, xi, 8);
}
g.realize({100});
}
// Now try a reduction where the pipeline returns that tuple value.
{
Func f, g;
Var x, y;
f(x, y) = sin(x * y);
f.compute_root();
// Find argmax of f over [0, 100]^2
RDom r(0, 100, 0, 100);
g() = Tuple(0, 0, f(0, 0));
Expr best_x = g()[0], best_y = g()[1], best_so_far = g()[2];
Expr next_value = f(r.x, r.y);
g() = select(next_value > best_so_far,
{r.x, r.y, next_value},
{best_x, best_y, best_so_far});
if (use_gpu) {
g.gpu_single_thread();
}
Realization result = g.realize();
// int result_x = Buffer<int>(result[0])(0);
// int result_y = Buffer<int>(result[1])(0);
float result_val = Buffer<float>(result[2])(0);
if (result_val < 0.9999) {
printf("Argmax of sin(x*y) is underwhelming: %f. We expected it to be closer to one.\n", result_val);
return 1;
}
}
// Now multiple output Funcs with different sizes
{
Func f, g;
Var x, xi;
f(x) = 100 * x;
g(x) = x;
if (use_gpu) {
f.gpu_tile(x, xi, 8);
g.gpu_tile(x, xi, 8);
}
Buffer<int> f_im(100);
Buffer<int> g_im(10);
Pipeline({f, g}).realize({f_im, g_im});
if (use_gpu) {
assert(f_im.device_dirty() && g_im.device_dirty());
f_im.copy_to_host();
g_im.copy_to_host();
}
for (int x = 0; x < f_im.width(); x++) {
if (f_im(x) != 100 * x) {
printf("f(%d) = %d instead of %d\n", x, f_im(x), 100 * x);
}
}
for (int x = 0; x < g_im.width(); x++) {
if (g_im(x) != x) {
printf("g(%d) = %d instead of %d\n", x, g_im(x), x);
return 1;
}
}
}
// Now multiple output Funcs via inferred Realization
{
Func f, g;
Var x, xi;
f(x) = cast<float>(100 * x);
g(x) = Tuple(cast<uint8_t>(x), cast<int16_t>(x + 1));
if (use_gpu) {
f.gpu_tile(x, xi, 8);
g.gpu_tile(x, xi, 8);
}
Realization r = Pipeline({f, g}).realize({100});
Buffer<float> f_im = r[0];
Buffer<uint8_t> g0_im = r[1];
Buffer<int16_t> g1_im = r[2];
for (int x = 0; x < f_im.width(); x++) {
if (f_im(x) != 100 * x) {
printf("f(%d) = %f instead of %f\n", x, f_im(x), (float)100 * x);
}
}
for (int x = 0; x < g0_im.width(); x++) {
if (g0_im(x) != x) {
printf("g0(%d) = %d instead of %d\n", x, (int)g0_im(x), x);
return 1;
}
}
for (int x = 0; x < g1_im.width(); x++) {
if (g1_im(x) != x + 1) {
printf("g1(%d) = %d instead of %d\n", x, (int)g1_im(x), x + 1);
return 1;
}
}
}
// Multiple output Funcs of different dimensionalities that call each other and some of them are Tuples.
{
Func f, g, h;
Var x, y, xi, yi;
f(x) = x;
h(x) = {f(x) + 17, f(x) - 17};
g(x, y) = {f(x + y) * 2, h(x)[0] * y, h(x)[1] - 2};
if (get_jit_target_from_environment().has_gpu_feature()) {
g.gpu_tile(x, y, xi, yi, 1, 1);
}
Buffer<int> f_im(100), g_im0(20, 20), g_im1(20, 20), g_im2(20, 20), h_im0(50), h_im1(50);
Pipeline({h, g, f}).realize({h_im0, h_im1, g_im0, g_im1, g_im2, f_im});
if (use_gpu) {
// g should have been written on the device
assert(g_im0.device_dirty() &&
g_im1.device_dirty() &&
g_im2.device_dirty());
// f and h should have been copied to the device for g to read
assert(f_im.has_device_allocation() &&
h_im0.has_device_allocation() &&
h_im1.has_device_allocation());
g_im0.copy_to_host();
g_im1.copy_to_host();
g_im2.copy_to_host();
}
for (int x = 0; x < 100; x++) {
if (f_im(x) != x) {
printf("f(%d) = %d instead of %d\n", x, f_im(x), x);
return 1;
}
if (x < 50) {
int c0 = f_im(x) + 17;
int c1 = f_im(x) - 17;
if (h_im0(x) != c0 || h_im1(x) != c1) {
printf("h(%d) = {%d, %d} instead of {%d, %d}\n",
x, h_im0(x), h_im1(x), c0, c1);
return 1;
}
}
if (x < 20) {
for (int y = 0; y < 20; y++) {
int c0 = f_im(x + y) * 2;
int c1 = h_im0(x) * y;
int c2 = h_im1(x) - 2;
if (g_im0(x, y) != c0 || g_im1(x, y) != c1 || g_im2(x, y) != c2) {
printf("g(%d) = {%d, %d, %d} instead of {%d, %d, %d}\n",
x, g_im0(x, y), g_im1(x, y), g_im2(x, y), c0, c1, c2);
return 1;
}
}
}
}
}
printf("Success!\n");
return 0;
}
|
