| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/f16-igemm/avx2-broadcast.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2022 Google LLC |
| 6 |
// |
| 7 |
// This source code is licensed under the BSD-style license found in the |
| 8 |
// LICENSE file in the root directory of this source tree. |
| 9 |
|
| 10 |
#include <assert.h> |
| 11 |
|
| 12 |
#include <immintrin.h> |
| 13 |
|
| 14 |
#include "xnnpack/igemm.h" |
| 15 |
#include "xnnpack/intrinsics-polyfill.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_f16_f32acc_igemm_minmax_ukernel_3x16__avx2_broadcast( |
| 19 |
size_t mr, |
| 20 |
size_t nc, |
| 21 |
size_t kc, |
| 22 |
size_t ks, |
| 23 |
const xnn_float16** restrict a, |
| 24 |
const xnn_float16* restrict w, |
| 25 |
xnn_float16* restrict c, |
| 26 |
size_t cm_stride, |
| 27 |
size_t cn_stride, |
| 28 |
size_t a_offset, |
| 29 |
const xnn_float16* zero, |
| 30 |
const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 31 |
{ |
| 32 |
assert(mr != 0); |
| 33 |
assert(mr <= 3); |
| 34 |
assert(nc != 0); |
| 35 |
assert(kc != 0); |
| 36 |
assert(kc % sizeof(uint16_t) == 0); |
| 37 |
assert(ks != 0); |
| 38 |
assert(ks % (3 * sizeof(void*)) == 0); |
| 39 |
assert(a_offset % sizeof(uint16_t) == 0); |
| 40 |
assert(a != NULL); |
| 41 |
assert(w != NULL); |
| 42 |
assert(c != NULL); |
| 43 |
|
| 44 |
const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) ¶ms->scalar.min)); |
| 45 |
const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) ¶ms->scalar.max)); |
| 46 |
XNN_FORCE_REALIZATION(vmin); |
| 47 |
XNN_FORCE_REALIZATION(vmax); |
| 48 |
|
| 49 |
uint16_t* c0 = (uint16_t*) c; |
| 50 |
uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride); |
| 51 |
if XNN_UNPREDICTABLE(mr < 2) { |
| 52 |
c1 = c0; |
| 53 |
} |
| 54 |
uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride); |
| 55 |
if XNN_UNPREDICTABLE(mr <= 2) { |
| 56 |
c2 = c1; |
| 57 |
} |
| 58 |
|
| 59 |
do { |
| 60 |
__m256 vacc0x0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w)); |
| 61 |
__m256 vacc0x1 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8))); |
| 62 |
__m256 vacc1x0 = vacc0x0; |
| 63 |
__m256 vacc1x1 = vacc0x1; |
| 64 |
__m256 vacc2x0 = vacc0x0; |
| 65 |
__m256 vacc2x1 = vacc0x1; |
| 66 |
w = (const xnn_float16*) w + 16; |
| 67 |
|
| 68 |
size_t p = ks; |
| 69 |
do { |
| 70 |
const uint16_t* restrict a0 = (const uint16_t*) a[0]; |
| 71 |
assert(a0 != NULL); |
| 72 |
if XNN_UNPREDICTABLE(a0 != (const uint16_t*) zero) { |
| 73 |
a0 = (const uint16_t*) ((uintptr_t) a0 + a_offset); |
| 74 |
} |
| 75 |
const uint16_t* restrict a1 = (const uint16_t*) a[1]; |
| 76 |
assert(a1 != NULL); |
| 77 |
if XNN_UNPREDICTABLE(a1 != (const uint16_t*) zero) { |
| 78 |
a1 = (const uint16_t*) ((uintptr_t) a1 + a_offset); |
| 79 |
} |
| 80 |
const uint16_t* restrict a2 = (const uint16_t*) a[2]; |
| 81 |
assert(a2 != NULL); |
| 82 |
if XNN_UNPREDICTABLE(a2 != (const uint16_t*) zero) { |
| 83 |
a2 = (const uint16_t*) ((uintptr_t) a2 + a_offset); |
| 84 |
} |
| 85 |
a += 3; |
| 86 |
|
| 87 |
size_t k = kc; |
| 88 |
do { |
| 89 |
const __m256 vb0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w)); |
| 90 |
const __m256 vb1 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + 8))); |
| 91 |
w = (const xnn_float16*) w + 16; |
| 92 |
|
| 93 |
const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0)); |
| 94 |
a0 += 1; |
| 95 |
const __m256 va1 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a1)); |
| 96 |
a1 += 1; |
| 97 |
const __m256 va2 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a2)); |
| 98 |
a2 += 1; |
| 99 |
|
| 100 |
vacc0x0 = _mm256_fmadd_ps(va0, vb0, vacc0x0); |
| 101 |
vacc0x1 = _mm256_fmadd_ps(va0, vb1, vacc0x1); |
| 102 |
vacc1x0 = _mm256_fmadd_ps(va1, vb0, vacc1x0); |
| 103 |
vacc1x1 = _mm256_fmadd_ps(va1, vb1, vacc1x1); |
| 104 |
vacc2x0 = _mm256_fmadd_ps(va2, vb0, vacc2x0); |
| 105 |
vacc2x1 = _mm256_fmadd_ps(va2, vb1, vacc2x1); |
| 106 |
|
| 107 |
k -= sizeof(uint16_t); |
| 108 |
} while (k != 0); |
| 109 |
p -= 3 * sizeof(void*); |
| 110 |
} while (p != 0); |
| 111 |
|
| 112 |
vacc0x0 = _mm256_max_ps(vacc0x0, vmin); |
| 113 |
vacc1x0 = _mm256_max_ps(vacc1x0, vmin); |
| 114 |
vacc2x0 = _mm256_max_ps(vacc2x0, vmin); |
| 115 |
vacc0x1 = _mm256_max_ps(vacc0x1, vmin); |
| 116 |
vacc1x1 = _mm256_max_ps(vacc1x1, vmin); |
| 117 |
vacc2x1 = _mm256_max_ps(vacc2x1, vmin); |
| 118 |
|
| 119 |
vacc0x0 = _mm256_min_ps(vacc0x0, vmax); |
| 120 |
vacc1x0 = _mm256_min_ps(vacc1x0, vmax); |
| 121 |
vacc2x0 = _mm256_min_ps(vacc2x0, vmax); |
| 122 |
vacc0x1 = _mm256_min_ps(vacc0x1, vmax); |
| 123 |
vacc1x1 = _mm256_min_ps(vacc1x1, vmax); |
| 124 |
vacc2x1 = _mm256_min_ps(vacc2x1, vmax); |
| 125 |
|
| 126 |
if XNN_LIKELY(nc >= 16) { |
| 127 |
_mm_storeu_si128((__m128i*) c2, _mm256_cvtps_ph(vacc2x0, _MM_FROUND_TO_NEAREST_INT)); |
| 128 |
_mm_storeu_si128((__m128i*) (c2 + 8), _mm256_cvtps_ph(vacc2x1, _MM_FROUND_TO_NEAREST_INT)); |
| 129 |
c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride); |
| 130 |
_mm_storeu_si128((__m128i*) c1, _mm256_cvtps_ph(vacc1x0, _MM_FROUND_TO_NEAREST_INT)); |
| 131 |
_mm_storeu_si128((__m128i*) (c1 + 8), _mm256_cvtps_ph(vacc1x1, _MM_FROUND_TO_NEAREST_INT)); |
| 132 |
c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride); |
| 133 |
_mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x0, _MM_FROUND_TO_NEAREST_INT)); |
| 134 |
_mm_storeu_si128((__m128i*) (c0 + 8), _mm256_cvtps_ph(vacc0x1, _MM_FROUND_TO_NEAREST_INT)); |
| 135 |
c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride); |
| 136 |
|
| 137 |
a = (const xnn_float16**restrict) ((uintptr_t) a - ks); |
| 138 |
nc -= 16; |
| 139 |
} else { |
| 140 |
__m128i vh2x0 = _mm256_cvtps_ph(vacc2x0, _MM_FROUND_TO_NEAREST_INT); |
| 141 |
__m128i vh1x0 = _mm256_cvtps_ph(vacc1x0, _MM_FROUND_TO_NEAREST_INT); |
| 142 |
__m128i vh0x0 = _mm256_cvtps_ph(vacc0x0, _MM_FROUND_TO_NEAREST_INT); |
| 143 |
if (nc & 8) { |
| 144 |
_mm_storeu_si128((__m128i*) c2, vh2x0); |
| 145 |
_mm_storeu_si128((__m128i*) c1, vh1x0); |
| 146 |
_mm_storeu_si128((__m128i*) c0, vh0x0); |
| 147 |
|
| 148 |
vh2x0 = _mm256_cvtps_ph(vacc2x1, _MM_FROUND_TO_NEAREST_INT); |
| 149 |
vh1x0 = _mm256_cvtps_ph(vacc1x1, _MM_FROUND_TO_NEAREST_INT); |
| 150 |
vh0x0 = _mm256_cvtps_ph(vacc0x1, _MM_FROUND_TO_NEAREST_INT); |
| 151 |
|
| 152 |
c2 += 8; |
| 153 |
c1 += 8; |
| 154 |
c0 += 8; |
| 155 |
} |
| 156 |
if (nc & 4) { |
| 157 |
_mm_storel_epi64((__m128i*) c2, vh2x0); |
| 158 |
_mm_storel_epi64((__m128i*) c1, vh1x0); |
| 159 |
_mm_storel_epi64((__m128i*) c0, vh0x0); |
| 160 |
|
| 161 |
vh2x0 = _mm_unpackhi_epi64(vh2x0, vh2x0); |
| 162 |
vh1x0 = _mm_unpackhi_epi64(vh1x0, vh1x0); |
| 163 |
vh0x0 = _mm_unpackhi_epi64(vh0x0, vh0x0); |
| 164 |
|
| 165 |
c2 += 4; |
| 166 |
c1 += 4; |
| 167 |
c0 += 4; |
| 168 |
} |
| 169 |
if (nc & 2) { |
| 170 |
_mm_storeu_si32(c2, vh2x0); |
| 171 |
_mm_storeu_si32(c1, vh1x0); |
| 172 |
_mm_storeu_si32(c0, vh0x0); |
| 173 |
|
| 174 |
vh2x0 = _mm_srli_epi64(vh2x0, 32); |
| 175 |
vh1x0 = _mm_srli_epi64(vh1x0, 32); |
| 176 |
vh0x0 = _mm_srli_epi64(vh0x0, 32); |
| 177 |
|
| 178 |
c2 += 2; |
| 179 |
c1 += 2; |
| 180 |
c0 += 2; |
| 181 |
} |
| 182 |
if (nc & 1) { |
| 183 |
*c2 = _mm_extract_epi16(vh2x0, 0); |
| 184 |
*c1 = _mm_extract_epi16(vh1x0, 0); |
| 185 |
*c0 = _mm_extract_epi16(vh0x0, 0); |
| 186 |
} |
| 187 |
|
| 188 |
nc = 0; |
| 189 |
} |
| 190 |
} while (nc != 0); |
| 191 |
} |