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 <cmath>
#include <cstdint>
#include <cstdio>
#ifdef __SSE2__
#include <emmintrin.h>
#elif __ARM_NEON
#include <arm_neon.h>
#endif
#include "HalideBuffer.h"
#include "halide_benchmark.h"
using namespace Halide::Runtime;
using namespace Halide::Tools;
double t;
Buffer<uint16_t, 2> blur(Buffer<uint16_t, 2> in) {
Buffer<uint16_t, 2> tmp(in.width() - 8, in.height());
Buffer<uint16_t, 2> out(in.width() - 8, in.height() - 2);
t = benchmark(10, 1, [&]() {
for (int y = 0; y < tmp.height(); y++)
for (int x = 0; x < tmp.width(); x++)
tmp(x, y) = (in(x, y) + in(x + 1, y) + in(x + 2, y)) / 3;
for (int y = 0; y < out.height(); y++)
for (int x = 0; x < out.width(); x++)
out(x, y) = (tmp(x, y) + tmp(x, y + 1) + tmp(x, y + 2)) / 3;
});
return out;
}
Buffer<uint16_t, 2> blur_fast(Buffer<uint16_t, 2> in) {
Buffer<uint16_t, 2> out(in.width() - 8, in.height() - 2);
t = benchmark(10, 1, [&]() {
#ifdef __SSE2__
__m128i one_third = _mm_set1_epi16(21846);
#pragma omp parallel for
for (int yTile = 0; yTile < out.height(); yTile += 32) {
__m128i tmp[(128 / 8) * (32 + 2)];
for (int xTile = 0; xTile < out.width(); xTile += 128) {
__m128i *tmpPtr = tmp;
for (int y = 0; y < 32 + 2; y++) {
const uint16_t *inPtr = &(in(xTile, yTile + y));
for (int x = 0; x < 128; x += 8) {
__m128i a = _mm_load_si128((const __m128i *)(inPtr));
__m128i b = _mm_loadu_si128((const __m128i *)(inPtr + 1));
__m128i c = _mm_loadu_si128((const __m128i *)(inPtr + 2));
__m128i sum = _mm_add_epi16(_mm_add_epi16(a, b), c);
__m128i avg = _mm_mulhi_epi16(sum, one_third);
_mm_store_si128(tmpPtr++, avg);
inPtr += 8;
}
}
tmpPtr = tmp;
for (int y = 0; y < 32; y++) {
__m128i *outPtr = (__m128i *)(&(out(xTile, yTile + y)));
for (int x = 0; x < 128; x += 8) {
__m128i a = _mm_load_si128(tmpPtr + (2 * 128) / 8);
__m128i b = _mm_load_si128(tmpPtr + 128 / 8);
__m128i c = _mm_load_si128(tmpPtr++);
__m128i sum = _mm_add_epi16(_mm_add_epi16(a, b), c);
__m128i avg = _mm_mulhi_epi16(sum, one_third);
_mm_store_si128(outPtr++, avg);
}
}
}
}
#elif __ARM_NEON
uint16x4_t one_third = vdup_n_u16(21846);
#pragma omp parallel for
for (int yTile = 0; yTile < out.height(); yTile += 32) {
uint16x8_t tmp[(128 / 8) * (32 + 2)];
for (int xTile = 0; xTile < out.width(); xTile += 128) {
uint16_t *tmpPtr = (uint16_t *)tmp;
for (int y = 0; y < 32 + 2; y++) {
const uint16_t *inPtr = &(in(xTile, yTile + y));
for (int x = 0; x < 128; x += 8) {
uint16x8_t a = vld1q_u16(inPtr);
uint16x8_t b = vld1q_u16(inPtr + 1);
uint16x8_t c = vld1q_u16(inPtr + 2);
uint16x8_t sum = vaddq_u16(vaddq_u16(a, b), c);
uint16x4_t sumlo = vget_low_u16(sum);
uint16x4_t sumhi = vget_high_u16(sum);
uint16x4_t avglo = vshrn_n_u32(vmull_u16(sumlo, one_third), 16);
uint16x4_t avghi = vshrn_n_u32(vmull_u16(sumhi, one_third), 16);
uint16x8_t avg = vcombine_u16(avglo, avghi);
vst1q_u16(tmpPtr, avg);
tmpPtr += 8;
inPtr += 8;
}
}
tmpPtr = (uint16_t *)tmp;
for (int y = 0; y < 32; y++) {
uint16_t *outPtr = &(out(xTile, yTile + y));
for (int x = 0; x < 128; x += 8) {
uint16x8_t a = vld1q_u16(tmpPtr + (2 * 128));
uint16x8_t b = vld1q_u16(tmpPtr + 128);
uint16x8_t c = vld1q_u16(tmpPtr);
uint16x8_t sum = vaddq_u16(vaddq_u16(a, b), c);
uint16x4_t sumlo = vget_low_u16(sum);
uint16x4_t sumhi = vget_high_u16(sum);
uint16x4_t avglo = vshrn_n_u32(vmull_u16(sumlo, one_third), 16);
uint16x4_t avghi = vshrn_n_u32(vmull_u16(sumhi, one_third), 16);
uint16x8_t avg = vcombine_u16(avglo, avghi);
vst1q_u16(outPtr, avg);
tmpPtr += 8;
outPtr += 8;
}
}
}
}
#else
// No intrinsics enabled, do a naive thing.
for (int y = 0; y < out.height(); y++) {
for (int x = 0; x < out.width(); x++) {
int tmp[3] = {
(in(x, y) + in(x + 1, y) + in(x + 2, y)) / 3,
(in(x, y + 1) + in(x + 1, y + 1) + in(x + 2, y + 1)) / 3,
(in(x, y + 2) + in(x + 1, y + 2) + in(x + 2, y + 2)) / 3,
};
out(x, y) = (tmp[0] + tmp[1] + tmp[2]) / 3;
}
}
#endif
});
return out;
}
#include "halide_blur.h"
Buffer<uint16_t, 2> blur_halide(Buffer<uint16_t, 2> in) {
Buffer<uint16_t, 2> out(in.width() - 8, in.height() - 2);
// Call it once to initialize the halide runtime stuff
halide_blur(in, out);
// Copy-out result if it's device buffer and dirty.
out.copy_to_host();
t = benchmark(10, 1, [&]() {
// Compute the same region of the output as blur_fast (i.e., we're
// still being sloppy with boundary conditions)
halide_blur(in, out);
// Sync device execution if any.
out.device_sync();
});
out.copy_to_host();
return out;
}
int main(int argc, char **argv) {
const auto *md = halide_blur_metadata();
const bool is_hexagon = strstr(md->target, "hvx_128") || strstr(md->target, "hvx_64");
// The Hexagon simulator can't allocate as much memory as the above wants.
const int width = is_hexagon ? 648 : 2568;
const int height = is_hexagon ? 482 : 1922;
Buffer<uint16_t, 2> input(width, height);
for (int y = 0; y < input.height(); y++) {
for (int x = 0; x < input.width(); x++) {
input(x, y) = rand() & 0xfff;
}
}
Buffer<uint16_t, 2> blurry = blur(input);
double slow_time = t;
Buffer<uint16_t, 2> speedy = blur_fast(input);
double fast_time = t;
Buffer<uint16_t, 2> halide = blur_halide(input);
double halide_time = t;
printf("times: %f %f %f\n", slow_time, fast_time, halide_time);
for (int y = 64; y < input.height() - 64; y++) {
for (int x = 64; x < input.width() - 64; x++) {
if (blurry(x, y) != speedy(x, y) || blurry(x, y) != halide(x, y)) {
printf("difference at (%d,%d): %d %d %d\n", x, y, blurry(x, y), speedy(x, y), halide(x, y));
abort();
}
}
}
printf("Success!\n");
return 0;
}
|
