1 #include "Halide.h"
2 #include <iostream>
3 #include <stdio.h>
4
5 using namespace Halide;
6 using namespace Halide::ConciseCasts;
7 using namespace Halide::Internal;
8
main(int arch,char ** argv)9 int main(int arch, char **argv) {
10 const int W = 256, H = 256;
11
12 Buffer<uint8_t> in(W, H);
13 // Set up the input.
14 for (int y = 0; y < H; y++) {
15 for (int x = 0; x < W; x++) {
16 in(x, y) = rand() & 0xff;
17 }
18 }
19
20 // Define a convolution kernel, and its sum.
21 Buffer<int8_t> kernel(3, 3);
22 kernel.set_min(-1, -1);
23 for (int y = -1; y <= 1; y++) {
24 for (int x = -1; x <= 1; x++) {
25 kernel(x, y) = rand() % 8 - 4;
26 }
27 }
28
29 Var x("x"), y("y"), xi("xi"), yi("yi");
30 RDom r(-1, 3, -1, 3);
31
32 // Boundary condition.
33 Func input = BoundaryConditions::repeat_edge(in);
34 input.compute_root();
35
36 // Test a widening reduction, followed by a narrowing.
37 {
38 Func f;
39 f(x, y) = u8_sat(sum(i16(input(x + r.x, y + r.y)) * kernel(r.x, r.y)) / 16);
40
41 // Schedule.
42 Target target = get_jit_target_from_environment();
43 if (target.has_gpu_feature()) {
44 f.gpu_tile(x, y, xi, yi, 16, 16);
45 } else if (target.features_any_of({Target::HVX_64, Target::HVX_128})) {
46 f.hexagon().vectorize(x, 128);
47 } else {
48 f.vectorize(x, target.natural_vector_size<uint8_t>());
49 }
50
51 // Run the pipeline and verify the results are correct.
52 Buffer<uint8_t> out = f.realize(W, H, target);
53
54 for (int y = 1; y < H - 1; y++) {
55 for (int x = 1; x < W - 1; x++) {
56 int16_t correct = 0;
57 for (int ry = -1; ry <= 1; ry++) {
58 for (int rx = -1; rx <= 1; rx++) {
59 correct += static_cast<int16_t>(in(x + rx, y + ry)) * kernel(rx, ry);
60 }
61 }
62 correct = std::min(std::max(correct / 16, 0), 255);
63 if (correct != out(x, y)) {
64 std::cout << "out(" << x << ", " << y << ") = " << (int)out(x, y) << " instead of " << correct << "\n";
65 return -1;
66 }
67 }
68 }
69 }
70
71 // Test a tuple reduction with widening, followed by narrowing the result.
72 {
73 Func f;
74 f(x, y) = {i16(0), i8(0)};
75 f(x, y) = {
76 f(x, y)[0] + i16(input(x + r.x, y + r.y)) * kernel(r.x, r.y),
77 f(x, y)[1] + kernel(r.x, r.y),
78 };
79
80 Func g;
81 g(x, y) = u8_sat((f(x, y)[0] + f(x, y)[1]) / 16);
82
83 // Schedule.
84 Target target = get_jit_target_from_environment();
85 if (target.has_gpu_feature()) {
86 g.gpu_tile(x, y, xi, yi, 16, 16);
87 } else if (target.features_any_of({Target::HVX_64, Target::HVX_128})) {
88 g.hexagon().vectorize(x, 128);
89 } else {
90 g.vectorize(x, target.natural_vector_size<uint8_t>());
91 }
92
93 // Run the pipeline and verify the results are correct.
94 Buffer<uint8_t> out = g.realize(W, H, target);
95
96 for (int y = 1; y < H - 1; y++) {
97 for (int x = 1; x < W - 1; x++) {
98 int16_t correct = 0;
99 for (int ry = -1; ry <= 1; ry++) {
100 for (int rx = -1; rx <= 1; rx++) {
101 correct += static_cast<int16_t>(in(x + rx, y + ry)) * kernel(rx, ry);
102 correct += kernel(rx, ry);
103 }
104 }
105 correct = std::min(std::max(correct / 16, 0), 255);
106 if (correct != out(x, y)) {
107 std::cout << "out(" << x << ", " << y << ") = " << (int)out(x, y) << " instead of " << correct << "\n";
108 return -1;
109 }
110 }
111 }
112 }
113
114 // Test a widening, followed by a narrowing reduction with an
115 // unaligned output. This triggered a bug in EliminateInterleaves
116 // on Hexagon.
117 {
118 Func f;
119 f(x, y) = i16(input(x, y));
120
121 Func g;
122 g(x, y) = u8_sat((f(x, y) + f(x + 1, y)) / 2);
123
124 // Schedule.
125 Target target = get_jit_target_from_environment();
126 if (target.has_gpu_feature()) {
127 g.gpu_tile(x, y, xi, yi, 16, 16);
128 } else if (target.features_any_of({Target::HVX_64, Target::HVX_128})) {
129 g.hexagon().vectorize(x, 128);
130 f.compute_at(g, y).vectorize(x, 128, TailStrategy::RoundUp);
131 } else {
132 g.vectorize(x, target.natural_vector_size<uint8_t>());
133 }
134
135 g.output_buffer().dim(0).set_min(0).set_extent(W - 2);
136 g.output_buffer().dim(1).set_min(0).set_extent(H);
137
138 // Run the pipeline and verify the results are correct.
139 Buffer<uint8_t> out = g.realize(W - 2, H, target);
140
141 for (int y = 1; y < H - 1; y++) {
142 for (int x = 0; x < W - 3; x++) {
143 uint8_t correct = (static_cast<int16_t>(in(x, y)) + in(x + 1, y)) / 2;
144 if (correct != out(x, y)) {
145 std::cout << "out(" << x << ", " << y << ") = " << (int)out(x, y) << " instead of " << (int)correct << "\n";
146 return -1;
147 }
148 }
149 }
150 }
151
152 std::cout << "Success!\n";
153 return 0;
154 }
155