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
|
#include "Halide.h"
#include <iostream>
#include <stdio.h>
using namespace Halide;
using namespace Halide::ConciseCasts;
using namespace Halide::Internal;
int main(int arch, char **argv) {
const int W = 256, H = 256;
Buffer<uint8_t> in(W, H);
// Set up the input.
for (int y = 0; y < H; y++) {
for (int x = 0; x < W; x++) {
in(x, y) = rand() & 0xff;
}
}
// Define a convolution kernel, and its sum.
Buffer<int8_t> kernel(3, 3);
kernel.set_min(-1, -1);
for (int y = -1; y <= 1; y++) {
for (int x = -1; x <= 1; x++) {
kernel(x, y) = rand() % 8 - 4;
}
}
Var x("x"), y("y"), xi("xi"), yi("yi");
RDom r(-1, 3, -1, 3);
// Boundary condition.
Func input = BoundaryConditions::repeat_edge(in);
input.compute_root();
// Test a widening reduction, followed by a narrowing.
{
Func f;
f(x, y) = u8_sat(sum(i16(input(x + r.x, y + r.y)) * kernel(r.x, r.y)) / 16);
// Schedule.
Target target = get_jit_target_from_environment();
if (target.has_gpu_feature()) {
f.gpu_tile(x, y, xi, yi, 16, 16);
} else if (target.has_feature(Target::HVX)) {
f.hexagon().vectorize(x, 128);
} else {
f.vectorize(x, target.natural_vector_size<uint8_t>());
}
// Run the pipeline and verify the results are correct.
Buffer<uint8_t> out = f.realize({W, H}, target);
for (int y = 1; y < H - 1; y++) {
for (int x = 1; x < W - 1; x++) {
int16_t correct = 0;
for (int ry = -1; ry <= 1; ry++) {
for (int rx = -1; rx <= 1; rx++) {
correct += static_cast<int16_t>(in(x + rx, y + ry)) * kernel(rx, ry);
}
}
correct = std::min(std::max(correct / 16, 0), 255);
if (correct != out(x, y)) {
std::cout << "out(" << x << ", " << y << ") = " << (int)out(x, y) << " instead of " << correct << "\n";
return -1;
}
}
}
}
// Test a tuple reduction with widening, followed by narrowing the result.
{
Func f;
f(x, y) = {i16(0), i8(0)};
f(x, y) = {
f(x, y)[0] + i16(input(x + r.x, y + r.y)) * kernel(r.x, r.y),
f(x, y)[1] + kernel(r.x, r.y),
};
Func g;
g(x, y) = u8_sat((f(x, y)[0] + f(x, y)[1]) / 16);
// Schedule.
Target target = get_jit_target_from_environment();
if (target.has_gpu_feature()) {
g.gpu_tile(x, y, xi, yi, 16, 16);
} else if (target.has_feature(Target::HVX)) {
g.hexagon().vectorize(x, 128);
} else {
g.vectorize(x, target.natural_vector_size<uint8_t>());
}
// Run the pipeline and verify the results are correct.
Buffer<uint8_t> out = g.realize({W, H}, target);
for (int y = 1; y < H - 1; y++) {
for (int x = 1; x < W - 1; x++) {
int16_t correct = 0;
for (int ry = -1; ry <= 1; ry++) {
for (int rx = -1; rx <= 1; rx++) {
correct += static_cast<int16_t>(in(x + rx, y + ry)) * kernel(rx, ry);
correct += kernel(rx, ry);
}
}
correct = std::min(std::max(correct / 16, 0), 255);
if (correct != out(x, y)) {
std::cout << "out(" << x << ", " << y << ") = " << (int)out(x, y) << " instead of " << correct << "\n";
return -1;
}
}
}
}
// Test a widening, followed by a narrowing reduction with an
// unaligned output. This triggered a bug in EliminateInterleaves
// on Hexagon.
{
Func f;
f(x, y) = i16(input(x, y));
Func g;
g(x, y) = u8_sat((f(x, y) + f(x + 1, y)) / 2);
// Schedule.
Target target = get_jit_target_from_environment();
if (target.has_gpu_feature()) {
g.gpu_tile(x, y, xi, yi, 16, 16);
} else if (target.has_feature(Target::HVX)) {
g.hexagon().vectorize(x, 128);
f.compute_at(g, y).vectorize(x, 128, TailStrategy::RoundUp);
} else {
g.vectorize(x, target.natural_vector_size<uint8_t>());
}
g.output_buffer().dim(0).set_min(0).set_extent(W - 2);
g.output_buffer().dim(1).set_min(0).set_extent(H);
// Run the pipeline and verify the results are correct.
Buffer<uint8_t> out = g.realize({W - 2, H}, target);
for (int y = 1; y < H - 1; y++) {
for (int x = 0; x < W - 3; x++) {
uint8_t correct = (static_cast<int16_t>(in(x, y)) + in(x + 1, y)) / 2;
if (correct != out(x, y)) {
std::cout << "out(" << x << ", " << y << ") = " << (int)out(x, y) << " instead of " << (int)correct << "\n";
return -1;
}
}
}
}
std::cout << "Success!\n";
return 0;
}
|
