| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-dwconv/unipass-avx2-mul16-vpunpck.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2021 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/dwconv.h" |
| 15 |
#include "xnnpack/unaligned.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_qs8_dwconv_minmax_fp32_ukernel_9p32c__avx2_mul16_add16_vpunpck( |
| 19 |
size_t channels, |
| 20 |
size_t output_width, |
| 21 |
const int8_t** input, |
| 22 |
const void* weights, |
| 23 |
int8_t* output, |
| 24 |
intptr_t input_stride, |
| 25 |
size_t output_increment, |
| 26 |
size_t input_offset, |
| 27 |
const int8_t* zero, |
| 28 |
const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 29 |
{ |
| 30 |
assert(channels != 0); |
| 31 |
assert(output_width != 0); |
| 32 |
|
| 33 |
const __m256 vscale = _mm256_set1_ps(params->fp32_scalar.scale); |
| 34 |
XNN_FORCE_REALIZATION(vscale); |
| 35 |
|
| 36 |
const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point); |
| 37 |
const __m256i voutput_zero_point = _mm256_set1_epi16(params->fp32_scalar.output_zero_point); |
| 38 |
const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min); |
| 39 |
XNN_FORCE_REALIZATION(voutput_max_less_zero_point); |
| 40 |
XNN_FORCE_REALIZATION(voutput_zero_point); |
| 41 |
XNN_FORCE_REALIZATION(voutput_min); |
| 42 |
|
| 43 |
do { |
| 44 |
const int8_t* i0 = input[0]; |
| 45 |
assert(i0 != NULL); |
| 46 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 47 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 48 |
} |
| 49 |
const int8_t* i1 = input[1]; |
| 50 |
assert(i1 != NULL); |
| 51 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 52 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 53 |
} |
| 54 |
const int8_t* i2 = input[2]; |
| 55 |
assert(i2 != NULL); |
| 56 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 57 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 58 |
} |
| 59 |
const int8_t* i3 = input[3]; |
| 60 |
assert(i3 != NULL); |
| 61 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 62 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 63 |
} |
| 64 |
const int8_t* i4 = input[4]; |
| 65 |
assert(i4 != NULL); |
| 66 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 67 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 68 |
} |
| 69 |
const int8_t* i5 = input[5]; |
| 70 |
assert(i5 != NULL); |
| 71 |
if XNN_UNPREDICTABLE(i5 != zero) { |
| 72 |
i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); |
| 73 |
} |
| 74 |
const int8_t* i6 = input[6]; |
| 75 |
assert(i6 != NULL); |
| 76 |
if XNN_UNPREDICTABLE(i6 != zero) { |
| 77 |
i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); |
| 78 |
} |
| 79 |
const int8_t* i7 = input[7]; |
| 80 |
assert(i7 != NULL); |
| 81 |
if XNN_UNPREDICTABLE(i7 != zero) { |
| 82 |
i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); |
| 83 |
} |
| 84 |
const int8_t* i8 = input[8]; |
| 85 |
assert(i8 != NULL); |
| 86 |
if XNN_UNPREDICTABLE(i8 != zero) { |
| 87 |
i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); |
| 88 |
} |
| 89 |
input = (const int8_t**) ((uintptr_t) input + input_stride); |
| 90 |
|
| 91 |
size_t c = channels; |
| 92 |
const void* w = weights; |
| 93 |
for (; c >= 32; c -= 32) { |
| 94 |
__m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w); |
| 95 |
__m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t))); |
| 96 |
__m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 16 * sizeof(int32_t))); |
| 97 |
__m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 24 * sizeof(int32_t))); |
| 98 |
|
| 99 |
__m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1); |
| 100 |
__m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31); |
| 101 |
__m256i vaccGHIJOPQR = _mm256_inserti128_si256(vaccGHIJKLMN, _mm256_castsi256_si128(vaccOPQRSTUV), 1); |
| 102 |
__m256i vaccKLMNSTUV = _mm256_permute2x128_si256(vaccGHIJKLMN, vaccOPQRSTUV, 0x31); |
| 103 |
|
| 104 |
|
| 105 |
const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0)); |
| 106 |
const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t)))); |
| 107 |
const __m256i vi0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i0 + 16))); |
| 108 |
const __m256i vk0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t)))); |
| 109 |
i0 += 32; |
| 110 |
|
| 111 |
__m256i vacc0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF); |
| 112 |
__m256i vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV); |
| 113 |
|
| 114 |
|
| 115 |
const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1)); |
| 116 |
const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t)))); |
| 117 |
const __m256i vi1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i1 + 16))); |
| 118 |
const __m256i vk1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t)))); |
| 119 |
i1 += 32; |
| 120 |
|
| 121 |
vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF)); |
| 122 |
vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV)); |
| 123 |
|
| 124 |
__m256i vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15); |
| 125 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF)); |
| 126 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF)); |
| 127 |
__m256i vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15); |
| 128 |
vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV)); |
| 129 |
vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV)); |
| 130 |
|
| 131 |
const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2)); |
| 132 |
const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t)))); |
| 133 |
const __m256i vi2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i2 + 16))); |
| 134 |
const __m256i vk2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t)))); |
| 135 |
i2 += 32; |
| 136 |
|
| 137 |
vacc0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF); |
| 138 |
vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV); |
| 139 |
|
| 140 |
|
| 141 |
const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3)); |
| 142 |
const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t)))); |
| 143 |
const __m256i vi3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i3 + 16))); |
| 144 |
const __m256i vk3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t)))); |
| 145 |
i3 += 32; |
| 146 |
|
| 147 |
vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF)); |
| 148 |
vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV)); |
| 149 |
|
| 150 |
vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15); |
| 151 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF)); |
| 152 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF)); |
| 153 |
vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15); |
| 154 |
vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV)); |
| 155 |
vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV)); |
| 156 |
|
| 157 |
const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4)); |
| 158 |
const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t)))); |
| 159 |
const __m256i vi4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i4 + 16))); |
| 160 |
const __m256i vk4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t)))); |
| 161 |
i4 += 32; |
| 162 |
|
| 163 |
vacc0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF); |
| 164 |
vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV); |
| 165 |
|
| 166 |
|
| 167 |
const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5)); |
| 168 |
const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t)))); |
| 169 |
const __m256i vi5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i5 + 16))); |
| 170 |
const __m256i vk5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t)))); |
| 171 |
i5 += 32; |
| 172 |
|
| 173 |
vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF)); |
| 174 |
vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV)); |
| 175 |
|
| 176 |
vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15); |
| 177 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF)); |
| 178 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF)); |
| 179 |
vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15); |
| 180 |
vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV)); |
| 181 |
vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV)); |
| 182 |
|
| 183 |
const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6)); |
| 184 |
const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t)))); |
| 185 |
const __m256i vi6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i6 + 16))); |
| 186 |
const __m256i vk6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t)))); |
| 187 |
i6 += 32; |
| 188 |
|
| 189 |
vacc0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF); |
| 190 |
vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV); |
| 191 |
|
| 192 |
|
| 193 |
const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7)); |
| 194 |
const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t)))); |
| 195 |
const __m256i vi7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i7 + 16))); |
| 196 |
const __m256i vk7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t)))); |
| 197 |
i7 += 32; |
| 198 |
|
| 199 |
vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF)); |
| 200 |
vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV)); |
| 201 |
|
| 202 |
vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15); |
| 203 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF)); |
| 204 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF)); |
| 205 |
vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15); |
| 206 |
vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV)); |
| 207 |
vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV)); |
| 208 |
|
| 209 |
const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); |
| 210 |
const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t)))); |
| 211 |
const __m256i vi8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i8 + 16))); |
| 212 |
const __m256i vk8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(int8_t)))); |
| 213 |
i8 += 32; |
| 214 |
|
| 215 |
vacc0123456789ABCDEF = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF); |
| 216 |
vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV); |
| 217 |
|
| 218 |
vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15); |
| 219 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF)); |
| 220 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF)); |
| 221 |
vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15); |
| 222 |
vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV)); |
| 223 |
vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV)); |
| 224 |
|
| 225 |
w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t)); |
| 226 |
|
| 227 |
vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1); |
| 228 |
vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31); |
| 229 |
vaccGHIJKLMN = _mm256_inserti128_si256(vaccGHIJOPQR, _mm256_castsi256_si128(vaccKLMNSTUV), 1); |
| 230 |
vaccOPQRSTUV = _mm256_permute2x128_si256(vaccGHIJOPQR, vaccKLMNSTUV, 0x31); |
| 231 |
|
| 232 |
__m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567); |
| 233 |
__m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF); |
| 234 |
__m256 vfpaccGHIJKLMN = _mm256_cvtepi32_ps(vaccGHIJKLMN); |
| 235 |
__m256 vfpaccOPQRSTUV = _mm256_cvtepi32_ps(vaccOPQRSTUV); |
| 236 |
|
| 237 |
vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale); |
| 238 |
vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale); |
| 239 |
vfpaccGHIJKLMN = _mm256_mul_ps(vfpaccGHIJKLMN, vscale); |
| 240 |
vfpaccOPQRSTUV = _mm256_mul_ps(vfpaccOPQRSTUV, vscale); |
| 241 |
|
| 242 |
vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point); |
| 243 |
vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point); |
| 244 |
vfpaccGHIJKLMN = _mm256_min_ps(vfpaccGHIJKLMN, voutput_max_less_zero_point); |
| 245 |
vfpaccOPQRSTUV = _mm256_min_ps(vfpaccOPQRSTUV, voutput_max_less_zero_point); |
| 246 |
|
| 247 |
vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567); |
| 248 |
vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF); |
| 249 |
vaccGHIJKLMN = _mm256_cvtps_epi32(vfpaccGHIJKLMN); |
| 250 |
vaccOPQRSTUV = _mm256_cvtps_epi32(vfpaccOPQRSTUV); |
| 251 |
|
| 252 |
const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point); |
| 253 |
const __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point); |
| 254 |
|
| 255 |
__m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0)); |
| 256 |
__m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0)); |
| 257 |
|
| 258 |
vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min); |
| 259 |
voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min); |
| 260 |
|
| 261 |
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF); |
| 262 |
_mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV); |
| 263 |
output += 32; |
| 264 |
} |
| 265 |
if XNN_UNLIKELY(c != 0) { |
| 266 |
const int8_t* k = (const int8_t*) ((uintptr_t) w + 32 * sizeof(int32_t)); |
| 267 |
do { |
| 268 |
__m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w); |
| 269 |
__m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t))); |
| 270 |
|
| 271 |
__m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1); |
| 272 |
__m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31); |
| 273 |
|
| 274 |
|
| 275 |
const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0)); |
| 276 |
const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) k)); |
| 277 |
i0 += 16; |
| 278 |
|
| 279 |
const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF); |
| 280 |
const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15); |
| 281 |
|
| 282 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi)); |
| 283 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi)); |
| 284 |
|
| 285 |
const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1)); |
| 286 |
const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 32))); |
| 287 |
i1 += 16; |
| 288 |
|
| 289 |
const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF); |
| 290 |
const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15); |
| 291 |
|
| 292 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi)); |
| 293 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi)); |
| 294 |
|
| 295 |
const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2)); |
| 296 |
const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 64))); |
| 297 |
i2 += 16; |
| 298 |
|
| 299 |
const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF); |
| 300 |
const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15); |
| 301 |
|
| 302 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi)); |
| 303 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi)); |
| 304 |
|
| 305 |
const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3)); |
| 306 |
const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 96))); |
| 307 |
i3 += 16; |
| 308 |
|
| 309 |
const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF); |
| 310 |
const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15); |
| 311 |
|
| 312 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi)); |
| 313 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi)); |
| 314 |
|
| 315 |
const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4)); |
| 316 |
const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 128))); |
| 317 |
i4 += 16; |
| 318 |
|
| 319 |
const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF); |
| 320 |
const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15); |
| 321 |
|
| 322 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi)); |
| 323 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi)); |
| 324 |
|
| 325 |
const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5)); |
| 326 |
const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 160))); |
| 327 |
i5 += 16; |
| 328 |
|
| 329 |
const __m256i vprod5x0123456789ABCDEFlo = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF); |
| 330 |
const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15); |
| 331 |
|
| 332 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi)); |
| 333 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi)); |
| 334 |
|
| 335 |
const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6)); |
| 336 |
const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 192))); |
| 337 |
i6 += 16; |
| 338 |
|
| 339 |
const __m256i vprod6x0123456789ABCDEFlo = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF); |
| 340 |
const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15); |
| 341 |
|
| 342 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi)); |
| 343 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi)); |
| 344 |
|
| 345 |
const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7)); |
| 346 |
const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 224))); |
| 347 |
i7 += 16; |
| 348 |
|
| 349 |
const __m256i vprod7x0123456789ABCDEFlo = _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF); |
| 350 |
const __m256i vprod7x0123456789ABCDEFhi = _mm256_srai_epi16(vprod7x0123456789ABCDEFlo, 15); |
| 351 |
|
| 352 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi)); |
| 353 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi)); |
| 354 |
|
| 355 |
const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); |
| 356 |
const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 256))); |
| 357 |
i8 += 16; |
| 358 |
|
| 359 |
const __m256i vprod8x0123456789ABCDEFlo = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF); |
| 360 |
const __m256i vprod8x0123456789ABCDEFhi = _mm256_srai_epi16(vprod8x0123456789ABCDEFlo, 15); |
| 361 |
|
| 362 |
vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi)); |
| 363 |
vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi)); |
| 364 |
|
| 365 |
vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1); |
| 366 |
vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31); |
| 367 |
|
| 368 |
k += 16; |
| 369 |
|
| 370 |
__m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567); |
| 371 |
__m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF); |
| 372 |
|
| 373 |
vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale); |
| 374 |
vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale); |
| 375 |
|
| 376 |
vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point); |
| 377 |
vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point); |
| 378 |
|
| 379 |
vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567); |
| 380 |
vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF); |
| 381 |
|
| 382 |
w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t)); |
| 383 |
|
| 384 |
__m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point)); |
| 385 |
__m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), _mm256_castsi256_si128(voutput_zero_point)); |
| 386 |
|
| 387 |
__m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF); |
| 388 |
vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min); |
| 389 |
|
| 390 |
if XNN_LIKELY(c >= 16) { |
| 391 |
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF); |
| 392 |
output += 16; |
| 393 |
c -= 16; |
| 394 |
} else { |
| 395 |
if (c & 8) { |
| 396 |
_mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF); |
| 397 |
vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF); |
| 398 |
output += 8; |
| 399 |
} |
| 400 |
if (c & 4) { |
| 401 |
unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF)); |
| 402 |
vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32); |
| 403 |
output += 4; |
| 404 |
} |
| 405 |
if (c & 2) { |
| 406 |
unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0)); |
| 407 |
vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16); |
| 408 |
output += 2; |
| 409 |
} |
| 410 |
if (c & 1) { |
| 411 |
*output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0); |
| 412 |
output += 1; |
| 413 |
} |
| 414 |
c = 0; |
| 415 |
} |
| 416 |
} while (c != 0); |
| 417 |
} |
| 418 |
|
| 419 |
output = (int8_t*) ((uintptr_t) output + output_increment); |
| 420 |
} while (--output_width != 0); |
| 421 |
} |