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
196
197
198
199
200
201
202
|
#include "Halide.h"
#include "halide_benchmark.h"
using namespace Halide;
using namespace Halide::Tools;
using std::vector;
double run_test_1(bool auto_schedule) {
Var x("x"), y("y"), dx("dx"), dy("dy"), c("c");
Func f("f");
f(x, y, dx, dy) = x + y + dx + dy;
int search_area = 7;
RDom dom(-search_area / 2, search_area, -search_area / 2, search_area, "dom");
// If 'f' is inlined into 'r', the only storage layout that the auto scheduler
// needs to care about is that of 'r'.
Func r("r");
r(x, y, c) += f(x, y + 1, dom.x, dom.y) * f(x, y - 1, dom.x, dom.y) * c;
Target target = get_jit_target_from_environment();
Pipeline p(r);
if (auto_schedule) {
// Provide estimates on the pipeline output
r.set_estimates({{0, 1024}, {0, 1024}, {0, 3}});
// Auto-schedule the pipeline
p.auto_schedule(target);
} else {
/*
r.update(0).fuse(c, y, par).parallel(par).reorder(x, dom.x, dom.y).vectorize(x, 4);
r.fuse(c, y, par).parallel(par).vectorize(x, 4); */
// The sequential schedule in this case seems to perform best which is
// odd have to investigate this further.
}
r.print_loop_nest();
// Run the schedule
Buffer<int> out(1024, 1024, 3);
double t = benchmark(3, 10, [&]() {
p.realize(out);
});
return t * 1000;
}
double run_test_2(bool auto_schedule) {
Var x("x"), y("y"), z("z"), c("c");
int W = 1024;
int H = 1920;
Buffer<uint8_t> left_im(W, H, 3);
Buffer<uint8_t> right_im(W, H, 3);
for (int y = 0; y < left_im.height(); y++) {
for (int x = 0; x < left_im.width(); x++) {
for (int c = 0; c < 3; c++) {
left_im(x, y, c) = rand() & 0xfff;
right_im(x, y, c) = rand() & 0xfff;
}
}
}
Func left = BoundaryConditions::repeat_edge(left_im);
Func right = BoundaryConditions::repeat_edge(right_im);
Func diff;
diff(x, y, z, c) = min(absd(left(x, y, c), right(x + 2 * z, y, c)),
absd(left(x, y, c), right(x + 2 * z + 1, y, c)));
Target target = get_jit_target_from_environment();
Pipeline p(diff);
if (auto_schedule) {
// Provide estimates on the pipeline output
diff.set_estimates({{0, left_im.width()}, {0, left_im.height()}, {0, 32}, {0, 3}});
// Auto-schedule the pipeline
p.auto_schedule(target);
} else {
Var t("t");
diff.reorder(c, z).fuse(c, z, t).parallel(t).vectorize(x, 16);
}
diff.print_loop_nest();
// Run the schedule
Buffer<uint8_t> out(left_im.width(), left_im.height(), 32, 3);
double t = benchmark(3, 10, [&]() {
p.realize(out);
});
return t * 1000;
}
double run_test_3(bool auto_schedule) {
Buffer<uint8_t> im(1024, 1028, 14, 14);
Var x("x"), y("y"), dx("dx"), dy("dy"), c("c");
Func f("f");
f(x, y, dx, dy) = im(x, y, dx, dy);
int search_area = 7;
RDom dom(-search_area / 2, search_area, -search_area / 2, search_area, "dom");
Func r("r");
r(x, y, c) += f(x, y + 1, search_area / 2 + dom.x, search_area / 2 + dom.y) *
f(x, y + 2, search_area / 2 + dom.x, search_area / 2 + dom.y) * c;
Target target = get_jit_target_from_environment();
Pipeline p(r);
if (auto_schedule) {
// Provide estimates on the pipeline output
r.set_estimates({{0, 1024}, {0, 1024}, {0, 3}});
// Auto-schedule the pipeline
p.auto_schedule(target);
} else {
Var par("par");
r.update(0).fuse(c, y, par).parallel(par).reorder(x, dom.x, dom.y).vectorize(x, 4);
r.fuse(c, y, par).parallel(par).vectorize(x, 4);
}
r.print_loop_nest();
// Run the schedule
Buffer<int> out(1024, 1024, 3);
double t = benchmark(3, 10, [&]() {
p.realize(out);
});
return t * 1000;
}
int main(int argc, char **argv) {
if (get_jit_target_from_environment().arch == Target::WebAssembly) {
printf("[SKIP] Autoschedulers do not support WebAssembly.\n");
return 0;
}
if (argc != 2) {
fprintf(stderr, "Usage: %s <autoscheduler-lib>\n", argv[0]);
return 1;
}
load_plugin(argv[1]);
const double slowdown_factor = 6.0;
{
std::cout << "Test 1:\n";
double manual_time = run_test_1(false);
double auto_time = run_test_1(true);
std::cout << "======================\n"
<< "Manual time: " << manual_time << "ms\n"
<< "Auto time: " << auto_time << "ms\n"
<< "======================\n";
if (auto_time > manual_time * slowdown_factor) {
fprintf(stderr, "Warning: Auto-scheduler is much much slower than it should be.\n");
}
}
{
std::cout << "Test 2:"
<< "\n";
double manual_time = run_test_2(false);
double auto_time = run_test_2(true);
std::cout << "======================\n"
<< "Manual time: " << manual_time << "ms\n"
<< "Auto time: " << auto_time << "ms\n"
<< "======================\n";
if (auto_time > manual_time * slowdown_factor) {
fprintf(stderr, "Warning: Auto-scheduler is much much slower than it should be.\n");
}
}
{
std::cout << "Test 3:\n";
double manual_time = run_test_3(false);
double auto_time = run_test_3(true);
std::cout << "======================\n"
<< "Manual time: " << manual_time << "ms\n"
<< "Auto time: " << auto_time << "ms\n"
<< "======================\n";
if (auto_time > manual_time * slowdown_factor) {
fprintf(stderr, "Warning: Auto-scheduler is much much slower than it should be.\n");
}
}
printf("Success!\n");
return 0;
}
|
