forked from google/XNNPACK
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathfma3.c.in
141 lines (118 loc) · 6.58 KB
/
fma3.c.in
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
// Copyright 2022 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
$assert CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert PIXEL_TILE == 1
$ABC = "456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <immintrin.h>
#include <xnnpack/common.h>
#include <xnnpack/ibilinear.h>
#include <xnnpack/intrinsics-polyfill.h>
void xnn_f16_ibilinear_ukernel__fma3_c${CHANNEL_TILE}${"" if PIXEL_TILE == 1 else "x%d" % PIXEL_TILE}(
size_t output_pixels,
size_t channels,
const void** restrict input,
size_t input_offset,
const void* restrict weights,
void* restrict output,
size_t output_increment) XNN_OOB_READS
{
assert(output_pixels != 0);
assert(channels != 0);
assert(channels % sizeof(uint16_t) == 0);
uint16_t* o = (uint16_t*) output;
do {
const uint16_t* i0 = (const uint16_t*) ((uintptr_t) input[0] + input_offset);
const uint16_t* i1 = (const uint16_t*) ((uintptr_t) input[1] + input_offset);
const uint16_t* i2 = (const uint16_t*) ((uintptr_t) input[2] + input_offset);
const uint16_t* i3 = (const uint16_t*) ((uintptr_t) input[3] + input_offset);
input += 4;
const __m256 valphahv = _mm256_cvtph_ps(_mm_castps_si128(_mm_broadcast_ss(weights)));
const __m256 valphah = _mm256_permute_ps(valphahv, _MM_SHUFFLE(2, 0, 2, 0));
const __m256 valphav = _mm256_permute_ps(valphahv, _MM_SHUFFLE(3, 1, 3, 1));
weights = (const uint16_t*) weights + 2;
size_t c = channels;
$if CHANNEL_TILE > 8:
for (; c >= ${CHANNEL_TILE} * sizeof(uint16_t); c -= ${CHANNEL_TILE} * sizeof(uint16_t)) {
const __m256 vtl${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
const __m256 vtr${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
const __m256 vbl${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
const __m256 vbr${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
$for C in range(8, CHANNEL_TILE, 8):
const __m256 vtl${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + ${C})));
const __m256 vtr${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + ${C})));
const __m256 vbl${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + ${C})));
const __m256 vbr${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i3 + ${C})));
i0 += ${CHANNEL_TILE};
i1 += ${CHANNEL_TILE};
i2 += ${CHANNEL_TILE};
i3 += ${CHANNEL_TILE};
$for C in range(0, CHANNEL_TILE, 8):
const __m256 vtd${ABC[C:C+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vtr${ABC[C:C+8]}, vtl${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT));
const __m256 vbd${ABC[C:C+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbr${ABC[C:C+8]}, vbl${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT));
$for C in range(0, CHANNEL_TILE, 8):
const __m256 vt${ABC[C:C+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vtd${ABC[C:C+8]}, valphah, vtl${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT));
const __m256 vb${ABC[C:C+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vbd${ABC[C:C+8]}, valphah, vbl${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT));
$for C in range(0, CHANNEL_TILE, 8):
const __m256 vd${ABC[C:C+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vb${ABC[C:C+8]}, vt${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT));
$for C in range(0, CHANNEL_TILE, 8):
const __m128i vo${ABC[C:C+8]} = _mm256_cvtps_ph(_mm256_fmadd_ps(vd${ABC[C:C+8]}, valphav, vt${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT);
_mm_storeu_si128((__m128i*) o, vo${ABC[0:8]});
$for C in range(8, CHANNEL_TILE, 8):
_mm_storeu_si128((__m128i*) (o + ${C}), vo${ABC[C:C+8]});
o += ${CHANNEL_TILE};
}
for (; c >= 8 * sizeof(uint16_t); c -= 8 * sizeof(uint16_t)) {
const __m256 vtl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
i0 += 8;
const __m256 vtr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
i1 += 8;
const __m256 vbl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
i2 += 8;
const __m256 vbr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
i3 += 8;
const __m256 vtd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vtr, vtl), _MM_FROUND_TO_NEAREST_INT));
const __m256 vbd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbr, vbl), _MM_FROUND_TO_NEAREST_INT));
const __m256 vt = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vtd, valphah, vtl), _MM_FROUND_TO_NEAREST_INT));
const __m256 vb = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vbd, valphah, vbl), _MM_FROUND_TO_NEAREST_INT));
const __m256 vd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vb, vt), _MM_FROUND_TO_NEAREST_INT));
const __m128i vo = _mm256_cvtps_ph(_mm256_fmadd_ps(vd, valphav, vt), _MM_FROUND_TO_NEAREST_INT);
_mm_storeu_si128((__m128i*) o, vo);
o += 8;
}
if XNN_UNLIKELY(c != 0) {
const __m256 vtl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
i0 += 8;
const __m256 vtr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
i1 += 8;
const __m256 vbl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
i2 += 8;
const __m256 vbr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
i3 += 8;
const __m256 vtd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vtr, vtl), _MM_FROUND_TO_NEAREST_INT));
const __m256 vbd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbr, vbl), _MM_FROUND_TO_NEAREST_INT));
const __m256 vt = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vtd, valphah, vtl), _MM_FROUND_TO_NEAREST_INT));
const __m256 vb = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vbd, valphah, vbl), _MM_FROUND_TO_NEAREST_INT));
const __m256 vd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vb, vt), _MM_FROUND_TO_NEAREST_INT));
__m128i vo = _mm256_cvtps_ph(_mm256_fmadd_ps(vd, valphav, vt), _MM_FROUND_TO_NEAREST_INT);
if (c & (4 * sizeof(uint16_t))) {
_mm_storel_epi64((__m128i*) o, vo);
vo = _mm_unpackhi_epi64(vo, vo);
o += 4;
}
if (c & (2 * sizeof(uint16_t))) {
_mm_storeu_si32(o, vo);
vo = _mm_srli_epi64(vo, 32);
o += 2;
}
if (c & (1 * sizeof(uint16_t))) {
*o = (uint16_t) _mm_extract_epi16(vo, 0);
o += 1;
}
}
o = (uint16_t*) ((uintptr_t) o + output_increment);
} while (--output_pixels != 0);
}