| 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_6x8__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 <= 6); |
| 34 |
assert(nc != 0); |
| 35 |
assert(kc != 0); |
| 36 |
assert(kc % sizeof(uint16_t) == 0); |
| 37 |
assert(ks != 0); |
| 38 |
assert(ks % (6 * 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 |
uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride); |
| 59 |
if XNN_UNPREDICTABLE(mr < 4) { |
| 60 |
c3 = c2; |
| 61 |
} |
| 62 |
uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride); |
| 63 |
if XNN_UNPREDICTABLE(mr <= 4) { |
| 64 |
c4 = c3; |
| 65 |
} |
| 66 |
uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride); |
| 67 |
if XNN_UNPREDICTABLE(mr != 6) { |
| 68 |
c5 = c4; |
| 69 |
} |
| 70 |
|
| 71 |
do { |
| 72 |
__m256 vacc0x0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w)); |
| 73 |
__m256 vacc1x0 = vacc0x0; |
| 74 |
__m256 vacc2x0 = vacc0x0; |
| 75 |
__m256 vacc3x0 = vacc0x0; |
| 76 |
__m256 vacc4x0 = vacc0x0; |
| 77 |
__m256 vacc5x0 = vacc0x0; |
| 78 |
w = (const xnn_float16*) w + 8; |
| 79 |
|
| 80 |
size_t p = ks; |
| 81 |
do { |
| 82 |
const uint16_t* restrict a0 = (const uint16_t*) a[0]; |
| 83 |
assert(a0 != NULL); |
| 84 |
if XNN_UNPREDICTABLE(a0 != (const uint16_t*) zero) { |
| 85 |
a0 = (const uint16_t*) ((uintptr_t) a0 + a_offset); |
| 86 |
} |
| 87 |
const uint16_t* restrict a1 = (const uint16_t*) a[1]; |
| 88 |
assert(a1 != NULL); |
| 89 |
if XNN_UNPREDICTABLE(a1 != (const uint16_t*) zero) { |
| 90 |
a1 = (const uint16_t*) ((uintptr_t) a1 + a_offset); |
| 91 |
} |
| 92 |
const uint16_t* restrict a2 = (const uint16_t*) a[2]; |
| 93 |
assert(a2 != NULL); |
| 94 |
if XNN_UNPREDICTABLE(a2 != (const uint16_t*) zero) { |
| 95 |
a2 = (const uint16_t*) ((uintptr_t) a2 + a_offset); |
| 96 |
} |
| 97 |
const uint16_t* restrict a3 = (const uint16_t*) a[3]; |
| 98 |
assert(a3 != NULL); |
| 99 |
if XNN_UNPREDICTABLE(a3 != (const uint16_t*) zero) { |
| 100 |
a3 = (const uint16_t*) ((uintptr_t) a3 + a_offset); |
| 101 |
} |
| 102 |
const uint16_t* restrict a4 = (const uint16_t*) a[4]; |
| 103 |
assert(a4 != NULL); |
| 104 |
if XNN_UNPREDICTABLE(a4 != (const uint16_t*) zero) { |
| 105 |
a4 = (const uint16_t*) ((uintptr_t) a4 + a_offset); |
| 106 |
} |
| 107 |
const uint16_t* restrict a5 = (const uint16_t*) a[5]; |
| 108 |
assert(a5 != NULL); |
| 109 |
if XNN_UNPREDICTABLE(a5 != (const uint16_t*) zero) { |
| 110 |
a5 = (const uint16_t*) ((uintptr_t) a5 + a_offset); |
| 111 |
} |
| 112 |
a += 6; |
| 113 |
|
| 114 |
size_t k = kc; |
| 115 |
do { |
| 116 |
const __m256 vb0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w)); |
| 117 |
w = (const xnn_float16*) w + 8; |
| 118 |
|
| 119 |
const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0)); |
| 120 |
a0 += 1; |
| 121 |
const __m256 va1 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a1)); |
| 122 |
a1 += 1; |
| 123 |
const __m256 va2 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a2)); |
| 124 |
a2 += 1; |
| 125 |
const __m256 va3 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a3)); |
| 126 |
a3 += 1; |
| 127 |
const __m256 va4 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a4)); |
| 128 |
a4 += 1; |
| 129 |
const __m256 va5 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a5)); |
| 130 |
a5 += 1; |
| 131 |
|
| 132 |
vacc0x0 = _mm256_fmadd_ps(va0, vb0, vacc0x0); |
| 133 |
vacc1x0 = _mm256_fmadd_ps(va1, vb0, vacc1x0); |
| 134 |
vacc2x0 = _mm256_fmadd_ps(va2, vb0, vacc2x0); |
| 135 |
vacc3x0 = _mm256_fmadd_ps(va3, vb0, vacc3x0); |
| 136 |
vacc4x0 = _mm256_fmadd_ps(va4, vb0, vacc4x0); |
| 137 |
vacc5x0 = _mm256_fmadd_ps(va5, vb0, vacc5x0); |
| 138 |
|
| 139 |
k -= sizeof(uint16_t); |
| 140 |
} while (k != 0); |
| 141 |
p -= 6 * sizeof(void*); |
| 142 |
} while (p != 0); |
| 143 |
|
| 144 |
vacc0x0 = _mm256_max_ps(vacc0x0, vmin); |
| 145 |
vacc1x0 = _mm256_max_ps(vacc1x0, vmin); |
| 146 |
vacc2x0 = _mm256_max_ps(vacc2x0, vmin); |
| 147 |
vacc3x0 = _mm256_max_ps(vacc3x0, vmin); |
| 148 |
vacc4x0 = _mm256_max_ps(vacc4x0, vmin); |
| 149 |
vacc5x0 = _mm256_max_ps(vacc5x0, vmin); |
| 150 |
|
| 151 |
vacc0x0 = _mm256_min_ps(vacc0x0, vmax); |
| 152 |
vacc1x0 = _mm256_min_ps(vacc1x0, vmax); |
| 153 |
vacc2x0 = _mm256_min_ps(vacc2x0, vmax); |
| 154 |
vacc3x0 = _mm256_min_ps(vacc3x0, vmax); |
| 155 |
vacc4x0 = _mm256_min_ps(vacc4x0, vmax); |
| 156 |
vacc5x0 = _mm256_min_ps(vacc5x0, vmax); |
| 157 |
|
| 158 |
if XNN_LIKELY(nc >= 8) { |
| 159 |
_mm_storeu_si128((__m128i*) c5, _mm256_cvtps_ph(vacc5x0, _MM_FROUND_TO_NEAREST_INT)); |
| 160 |
c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride); |
| 161 |
_mm_storeu_si128((__m128i*) c4, _mm256_cvtps_ph(vacc4x0, _MM_FROUND_TO_NEAREST_INT)); |
| 162 |
c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride); |
| 163 |
_mm_storeu_si128((__m128i*) c3, _mm256_cvtps_ph(vacc3x0, _MM_FROUND_TO_NEAREST_INT)); |
| 164 |
c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride); |
| 165 |
_mm_storeu_si128((__m128i*) c2, _mm256_cvtps_ph(vacc2x0, _MM_FROUND_TO_NEAREST_INT)); |
| 166 |
c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride); |
| 167 |
_mm_storeu_si128((__m128i*) c1, _mm256_cvtps_ph(vacc1x0, _MM_FROUND_TO_NEAREST_INT)); |
| 168 |
c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride); |
| 169 |
_mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x0, _MM_FROUND_TO_NEAREST_INT)); |
| 170 |
c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride); |
| 171 |
|
| 172 |
a = (const xnn_float16**restrict) ((uintptr_t) a - ks); |
| 173 |
nc -= 8; |
| 174 |
} else { |
| 175 |
__m128i vh5x0 = _mm256_cvtps_ph(vacc5x0, _MM_FROUND_TO_NEAREST_INT); |
| 176 |
__m128i vh4x0 = _mm256_cvtps_ph(vacc4x0, _MM_FROUND_TO_NEAREST_INT); |
| 177 |
__m128i vh3x0 = _mm256_cvtps_ph(vacc3x0, _MM_FROUND_TO_NEAREST_INT); |
| 178 |
__m128i vh2x0 = _mm256_cvtps_ph(vacc2x0, _MM_FROUND_TO_NEAREST_INT); |
| 179 |
__m128i vh1x0 = _mm256_cvtps_ph(vacc1x0, _MM_FROUND_TO_NEAREST_INT); |
| 180 |
__m128i vh0x0 = _mm256_cvtps_ph(vacc0x0, _MM_FROUND_TO_NEAREST_INT); |
| 181 |
if (nc & 4) { |
| 182 |
_mm_storel_epi64((__m128i*) c5, vh5x0); |
| 183 |
_mm_storel_epi64((__m128i*) c4, vh4x0); |
| 184 |
_mm_storel_epi64((__m128i*) c3, vh3x0); |
| 185 |
_mm_storel_epi64((__m128i*) c2, vh2x0); |
| 186 |
_mm_storel_epi64((__m128i*) c1, vh1x0); |
| 187 |
_mm_storel_epi64((__m128i*) c0, vh0x0); |
| 188 |
|
| 189 |
vh5x0 = _mm_unpackhi_epi64(vh5x0, vh5x0); |
| 190 |
vh4x0 = _mm_unpackhi_epi64(vh4x0, vh4x0); |
| 191 |
vh3x0 = _mm_unpackhi_epi64(vh3x0, vh3x0); |
| 192 |
vh2x0 = _mm_unpackhi_epi64(vh2x0, vh2x0); |
| 193 |
vh1x0 = _mm_unpackhi_epi64(vh1x0, vh1x0); |
| 194 |
vh0x0 = _mm_unpackhi_epi64(vh0x0, vh0x0); |
| 195 |
|
| 196 |
c5 += 4; |
| 197 |
c4 += 4; |
| 198 |
c3 += 4; |
| 199 |
c2 += 4; |
| 200 |
c1 += 4; |
| 201 |
c0 += 4; |
| 202 |
} |
| 203 |
if (nc & 2) { |
| 204 |
_mm_storeu_si32(c5, vh5x0); |
| 205 |
_mm_storeu_si32(c4, vh4x0); |
| 206 |
_mm_storeu_si32(c3, vh3x0); |
| 207 |
_mm_storeu_si32(c2, vh2x0); |
| 208 |
_mm_storeu_si32(c1, vh1x0); |
| 209 |
_mm_storeu_si32(c0, vh0x0); |
| 210 |
|
| 211 |
vh5x0 = _mm_srli_epi64(vh5x0, 32); |
| 212 |
vh4x0 = _mm_srli_epi64(vh4x0, 32); |
| 213 |
vh3x0 = _mm_srli_epi64(vh3x0, 32); |
| 214 |
vh2x0 = _mm_srli_epi64(vh2x0, 32); |
| 215 |
vh1x0 = _mm_srli_epi64(vh1x0, 32); |
| 216 |
vh0x0 = _mm_srli_epi64(vh0x0, 32); |
| 217 |
|
| 218 |
c5 += 2; |
| 219 |
c4 += 2; |
| 220 |
c3 += 2; |
| 221 |
c2 += 2; |
| 222 |
c1 += 2; |
| 223 |
c0 += 2; |
| 224 |
} |
| 225 |
if (nc & 1) { |
| 226 |
*c5 = _mm_extract_epi16(vh5x0, 0); |
| 227 |
*c4 = _mm_extract_epi16(vh4x0, 0); |
| 228 |
*c3 = _mm_extract_epi16(vh3x0, 0); |
| 229 |
*c2 = _mm_extract_epi16(vh2x0, 0); |
| 230 |
*c1 = _mm_extract_epi16(vh1x0, 0); |
| 231 |
*c0 = _mm_extract_epi16(vh0x0, 0); |
| 232 |
} |
| 233 |
|
| 234 |
nc = 0; |
| 235 |
} |
| 236 |
} while (nc != 0); |
| 237 |
} |