| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-dwconv/multipass-avx512skx-mul32.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2023 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/intrinsics-polyfill.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_qs8_dwconv_minmax_fp32_ukernel_5f5m5l32c16s1r__avx512skx_mul32( |
| 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 |
size_t kernel_size, |
| 29 |
int32_t* buffer, |
| 30 |
const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 31 |
{ |
| 32 |
assert(channels != 0); |
| 33 |
assert(output_width != 0); |
| 34 |
assert(kernel_size > 5); |
| 35 |
|
| 36 |
|
| 37 |
const __m512 vscale = _mm512_set1_ps(params->fp32_scalar.scale); |
| 38 |
// XNN_FORCE_REALIZATION(vscale); |
| 39 |
const __m512 voutput_max_less_zero_point = _mm512_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point); |
| 40 |
const __m512i voutput_zero_point = _mm512_set1_epi16((int16_t) params->fp32_scalar.output_zero_point); |
| 41 |
const __m256i voutput_min = _mm256_set1_epi8(params->fp32_scalar.output_min); |
| 42 |
const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 5, 1, 6, 2, 4, 0); |
| 43 |
// XNN_FORCE_REALIZATION(voutput_max_less_zero_point); |
| 44 |
// XNN_FORCE_REALIZATION(voutput_zero_point); |
| 45 |
// XNN_FORCE_REALIZATION(voutput_min); |
| 46 |
|
| 47 |
do { |
| 48 |
const void* w = weights; |
| 49 |
|
| 50 |
// First pass to process 5 inputs. |
| 51 |
{ |
| 52 |
int32_t* b = buffer; |
| 53 |
const int8_t* i0 = input[0]; |
| 54 |
assert(i0 != NULL); |
| 55 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 56 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 57 |
} |
| 58 |
const int8_t* i1 = input[1]; |
| 59 |
assert(i1 != NULL); |
| 60 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 61 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 62 |
} |
| 63 |
const int8_t* i2 = input[2]; |
| 64 |
assert(i2 != NULL); |
| 65 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 66 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 67 |
} |
| 68 |
const int8_t* i3 = input[3]; |
| 69 |
assert(i3 != NULL); |
| 70 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 71 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 72 |
} |
| 73 |
const int8_t* i4 = input[4]; |
| 74 |
assert(i4 != NULL); |
| 75 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 76 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 77 |
} |
| 78 |
input += 5; |
| 79 |
|
| 80 |
size_t c = channels; |
| 81 |
|
| 82 |
for (; c >= 32; c -= 32) { |
| 83 |
__m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w); |
| 84 |
__m512i vaccGHIJKLMNOPQRSTUV = _mm512_loadu_si512((const void*) ((const int32_t*) w + 16)); |
| 85 |
|
| 86 |
const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0)); |
| 87 |
const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t)))); |
| 88 |
const __m512i vi0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i0 + 16))); |
| 89 |
const __m512i vk0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t)))); |
| 90 |
i0 += 32; |
| 91 |
|
| 92 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF)); |
| 93 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV)); |
| 94 |
|
| 95 |
const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1)); |
| 96 |
const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t)))); |
| 97 |
const __m512i vi1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i1 + 16))); |
| 98 |
const __m512i vk1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t)))); |
| 99 |
i1 += 32; |
| 100 |
|
| 101 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF)); |
| 102 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV)); |
| 103 |
|
| 104 |
const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2)); |
| 105 |
const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t)))); |
| 106 |
const __m512i vi2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i2 + 16))); |
| 107 |
const __m512i vk2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t)))); |
| 108 |
i2 += 32; |
| 109 |
|
| 110 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF)); |
| 111 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV)); |
| 112 |
|
| 113 |
const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3)); |
| 114 |
const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t)))); |
| 115 |
const __m512i vi3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i3 + 16))); |
| 116 |
const __m512i vk3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t)))); |
| 117 |
i3 += 32; |
| 118 |
|
| 119 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF)); |
| 120 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV)); |
| 121 |
|
| 122 |
const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); |
| 123 |
const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t)))); |
| 124 |
const __m512i vi4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i4 + 16))); |
| 125 |
const __m512i vk4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t)))); |
| 126 |
i4 += 32; |
| 127 |
|
| 128 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF)); |
| 129 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV)); |
| 130 |
|
| 131 |
w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t)); |
| 132 |
|
| 133 |
_mm512_storeu_si512((__m512i*) b, vacc0123456789ABCDEF); |
| 134 |
_mm512_storeu_si512((__m512i*) (b + 16), vaccGHIJKLMNOPQRSTUV); |
| 135 |
b += 32; |
| 136 |
} |
| 137 |
|
| 138 |
if XNN_UNLIKELY(c != 0) { |
| 139 |
for (; c >= 16; c -= 16) { |
| 140 |
__m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w); |
| 141 |
|
| 142 |
const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0)); |
| 143 |
const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)))); |
| 144 |
i0 += 16; |
| 145 |
|
| 146 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF)); |
| 147 |
|
| 148 |
const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1)); |
| 149 |
const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)))); |
| 150 |
i1 += 16; |
| 151 |
|
| 152 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF)); |
| 153 |
|
| 154 |
const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2)); |
| 155 |
const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)))); |
| 156 |
i2 += 16; |
| 157 |
|
| 158 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF)); |
| 159 |
|
| 160 |
const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3)); |
| 161 |
const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)))); |
| 162 |
i3 += 16; |
| 163 |
|
| 164 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF)); |
| 165 |
|
| 166 |
const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); |
| 167 |
const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)))); |
| 168 |
i4 += 16; |
| 169 |
|
| 170 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF)); |
| 171 |
|
| 172 |
w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)); |
| 173 |
|
| 174 |
_mm512_storeu_si512((__m512i*) b, vacc0123456789ABCDEF); |
| 175 |
b += 16; |
| 176 |
} |
| 177 |
} |
| 178 |
|
| 179 |
if (c != 0) { |
| 180 |
assert(c >= 1); |
| 181 |
assert(c <= 15); |
| 182 |
__m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w); |
| 183 |
|
| 184 |
const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0)); |
| 185 |
const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)))); |
| 186 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF)); |
| 187 |
|
| 188 |
const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1)); |
| 189 |
const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)))); |
| 190 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF)); |
| 191 |
|
| 192 |
const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2)); |
| 193 |
const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)))); |
| 194 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF)); |
| 195 |
|
| 196 |
const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3)); |
| 197 |
const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)))); |
| 198 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF)); |
| 199 |
|
| 200 |
const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); |
| 201 |
const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)))); |
| 202 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF)); |
| 203 |
|
| 204 |
w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)); |
| 205 |
|
| 206 |
_mm512_storeu_si512((__m512i*) b, vacc0123456789ABCDEF); |
| 207 |
} |
| 208 |
} |
| 209 |
|
| 210 |
// Middle pass to process 5 inputs in each iteration. |
| 211 |
for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) { |
| 212 |
int32_t* b = buffer; |
| 213 |
const int8_t* i0 = input[0]; |
| 214 |
assert(i0 != NULL); |
| 215 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 216 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 217 |
} |
| 218 |
const int8_t* i1 = input[1]; |
| 219 |
assert(i1 != NULL); |
| 220 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 221 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 222 |
} |
| 223 |
const int8_t* i2 = input[2]; |
| 224 |
assert(i2 != NULL); |
| 225 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 226 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 227 |
} |
| 228 |
const int8_t* i3 = input[3]; |
| 229 |
assert(i3 != NULL); |
| 230 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 231 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 232 |
} |
| 233 |
const int8_t* i4 = input[4]; |
| 234 |
assert(i4 != NULL); |
| 235 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 236 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 237 |
} |
| 238 |
input += 5; |
| 239 |
|
| 240 |
size_t c = channels; |
| 241 |
|
| 242 |
for (; c >= 32; c -= 32) { |
| 243 |
__m512i vacc0123456789ABCDEF = _mm512_loadu_si512(b); |
| 244 |
__m512i vaccGHIJKLMNOPQRSTUV = _mm512_loadu_si512((const void*) ((const int32_t*) b + 16)); |
| 245 |
|
| 246 |
const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0)); |
| 247 |
const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)))); |
| 248 |
const __m512i vi0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i0 + 16))); |
| 249 |
const __m512i vk0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)))); |
| 250 |
i0 += 32; |
| 251 |
|
| 252 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF)); |
| 253 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV)); |
| 254 |
|
| 255 |
const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1)); |
| 256 |
const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)))); |
| 257 |
const __m512i vi1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i1 + 16))); |
| 258 |
const __m512i vk1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)))); |
| 259 |
i1 += 32; |
| 260 |
|
| 261 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF)); |
| 262 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV)); |
| 263 |
|
| 264 |
const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2)); |
| 265 |
const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)))); |
| 266 |
const __m512i vi2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i2 + 16))); |
| 267 |
const __m512i vk2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t)))); |
| 268 |
i2 += 32; |
| 269 |
|
| 270 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF)); |
| 271 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV)); |
| 272 |
|
| 273 |
const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3)); |
| 274 |
const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t)))); |
| 275 |
const __m512i vi3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i3 + 16))); |
| 276 |
const __m512i vk3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t)))); |
| 277 |
i3 += 32; |
| 278 |
|
| 279 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF)); |
| 280 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV)); |
| 281 |
|
| 282 |
const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); |
| 283 |
const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 128 * sizeof(int8_t)))); |
| 284 |
const __m512i vi4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i4 + 16))); |
| 285 |
const __m512i vk4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 144 * sizeof(int8_t)))); |
| 286 |
i4 += 32; |
| 287 |
|
| 288 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF)); |
| 289 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV)); |
| 290 |
|
| 291 |
w = (const void*) ((uintptr_t) w + 160 * sizeof(int8_t)); |
| 292 |
|
| 293 |
_mm512_storeu_si512((__m512i*) b, vacc0123456789ABCDEF); |
| 294 |
_mm512_storeu_si512((__m512i*) (b + 16), vaccGHIJKLMNOPQRSTUV); |
| 295 |
b += 32; |
| 296 |
} |
| 297 |
|
| 298 |
if XNN_UNLIKELY(c != 0) { |
| 299 |
for (; c >= 16; c -= 16) { |
| 300 |
__m512i vacc0123456789ABCDEF = _mm512_loadu_si512(b); |
| 301 |
|
| 302 |
const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0)); |
| 303 |
const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)))); |
| 304 |
i0 += 16; |
| 305 |
|
| 306 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF)); |
| 307 |
|
| 308 |
const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1)); |
| 309 |
const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)))); |
| 310 |
i1 += 16; |
| 311 |
|
| 312 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF)); |
| 313 |
|
| 314 |
const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2)); |
| 315 |
const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)))); |
| 316 |
i2 += 16; |
| 317 |
|
| 318 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF)); |
| 319 |
|
| 320 |
const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3)); |
| 321 |
const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)))); |
| 322 |
i3 += 16; |
| 323 |
|
| 324 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF)); |
| 325 |
|
| 326 |
const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); |
| 327 |
const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)))); |
| 328 |
i4 += 16; |
| 329 |
|
| 330 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF)); |
| 331 |
|
| 332 |
w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t)); |
| 333 |
|
| 334 |
_mm512_storeu_si512((__m512i*) b, vacc0123456789ABCDEF); |
| 335 |
b += 16; |
| 336 |
} |
| 337 |
} |
| 338 |
|
| 339 |
if (c != 0) { |
| 340 |
assert(c >= 1); |
| 341 |
assert(c <= 15); |
| 342 |
__m512i vacc0123456789ABCDEF = _mm512_loadu_si512(b); |
| 343 |
|
| 344 |
const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0)); |
| 345 |
const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)))); |
| 346 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF)); |
| 347 |
|
| 348 |
const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1)); |
| 349 |
const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)))); |
| 350 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF)); |
| 351 |
|
| 352 |
const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2)); |
| 353 |
const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)))); |
| 354 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF)); |
| 355 |
|
| 356 |
const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3)); |
| 357 |
const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)))); |
| 358 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF)); |
| 359 |
|
| 360 |
const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); |
| 361 |
const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)))); |
| 362 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF)); |
| 363 |
|
| 364 |
w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t)); |
| 365 |
|
| 366 |
_mm512_storeu_si512((__m512i*) b, vacc0123456789ABCDEF); |
| 367 |
} |
| 368 |
} |
| 369 |
|
| 370 |
// Last pass to process up to 5 inputs. |
| 371 |
{ |
| 372 |
const int32_t* b = buffer; |
| 373 |
const int8_t* i0 = input[0]; |
| 374 |
assert(i0 != NULL); |
| 375 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 376 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 377 |
} |
| 378 |
const int8_t* i1 = input[1]; |
| 379 |
assert(i1 != NULL); |
| 380 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 381 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 382 |
} |
| 383 |
const int8_t* i2 = input[2]; |
| 384 |
assert(i2 != NULL); |
| 385 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 386 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 387 |
} |
| 388 |
const int8_t* i3 = input[3]; |
| 389 |
assert(i3 != NULL); |
| 390 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 391 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 392 |
} |
| 393 |
const int8_t* i4 = input[4]; |
| 394 |
assert(i4 != NULL); |
| 395 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 396 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 397 |
} |
| 398 |
|
| 399 |
size_t c = channels; |
| 400 |
|
| 401 |
for (; c >= 32; c -= 32) { |
| 402 |
__m512i vacc0123456789ABCDEF = _mm512_loadu_si512(b); |
| 403 |
__m512i vaccGHIJKLMNOPQRSTUV = _mm512_loadu_si512((const void*) ((const int32_t*) b + 16)); |
| 404 |
b += 32; |
| 405 |
|
| 406 |
const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0)); |
| 407 |
const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)))); |
| 408 |
const __m512i vi0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i0 + 16))); |
| 409 |
const __m512i vk0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)))); |
| 410 |
i0 += 32; |
| 411 |
|
| 412 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF)); |
| 413 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV)); |
| 414 |
|
| 415 |
const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1)); |
| 416 |
const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)))); |
| 417 |
const __m512i vi1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i1 + 16))); |
| 418 |
const __m512i vk1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)))); |
| 419 |
i1 += 32; |
| 420 |
|
| 421 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF)); |
| 422 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV)); |
| 423 |
|
| 424 |
const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2)); |
| 425 |
const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)))); |
| 426 |
const __m512i vi2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i2 + 16))); |
| 427 |
const __m512i vk2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t)))); |
| 428 |
i2 += 32; |
| 429 |
|
| 430 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF)); |
| 431 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV)); |
| 432 |
|
| 433 |
const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3)); |
| 434 |
const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t)))); |
| 435 |
const __m512i vi3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i3 + 16))); |
| 436 |
const __m512i vk3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t)))); |
| 437 |
i3 += 32; |
| 438 |
|
| 439 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF)); |
| 440 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV)); |
| 441 |
|
| 442 |
const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); |
| 443 |
const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 128 * sizeof(int8_t)))); |
| 444 |
const __m512i vi4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i4 + 16))); |
| 445 |
const __m512i vk4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 144 * sizeof(int8_t)))); |
| 446 |
i4 += 32; |
| 447 |
|
| 448 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF)); |
| 449 |
vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV)); |
| 450 |
|
| 451 |
w = (const void*) ((uintptr_t) w + 160 * sizeof(int8_t)); |
| 452 |
|
| 453 |
__m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF); |
| 454 |
__m512 vscaledGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vaccGHIJKLMNOPQRSTUV); |
| 455 |
|
| 456 |
vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale); |
| 457 |
vscaledGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaledGHIJKLMNOPQRSTUV, vscale); |
| 458 |
|
| 459 |
vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point); |
| 460 |
vscaledGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaledGHIJKLMNOPQRSTUV, voutput_max_less_zero_point); |
| 461 |
|
| 462 |
vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF); |
| 463 |
vaccGHIJKLMNOPQRSTUV = _mm512_cvtps_epi32(vscaledGHIJKLMNOPQRSTUV); |
| 464 |
|
| 465 |
__m512i vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV = _mm512_adds_epi16(_mm512_packs_epi32(vacc0123456789ABCDEF, vaccGHIJKLMNOPQRSTUV), voutput_zero_point); |
| 466 |
__m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vaccGHIJKLMNOPQRSTUV), _mm512_extracti32x8_epi32(vaccGHIJKLMNOPQRSTUV, 1)), _mm512_castsi512_si256(voutput_zero_point)); |
| 467 |
|
| 468 |
const __m256i vout0123GHIJ4567KLMN = _mm512_castsi512_si256(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV); |
| 469 |
const __m256i vout89ABOPQRCDEFSTUV = _mm512_extracti32x8_epi32(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV, 1); |
| 470 |
const __m256i vout0123GHIJ89ABOPQR4567KLMNCDEFSTUV = _mm256_packs_epi16(vout0123GHIJ4567KLMN, vout89ABOPQRCDEFSTUV); |
| 471 |
__m256i vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_permutevar8x32_epi32(vout0123GHIJ89ABOPQR4567KLMNCDEFSTUV, vpermute_mask); |
| 472 |
const __m128i voutGHIJOPQR = _mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV); |
| 473 |
const __m128i voutKLMNSTUV = _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1); |
| 474 |
__m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(voutGHIJOPQR, voutKLMNSTUV), _MM_SHUFFLE(3, 1, 2, 0)); |
| 475 |
|
| 476 |
vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_max_epi8(vout0123456789ABCDEFGHIJKLMNOPQRSTUV, voutput_min); |
| 477 |
voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, _mm256_castsi256_si128(voutput_min)); |
| 478 |
|
| 479 |
_mm256_storeu_si256((__m256i*) output, vout0123456789ABCDEFGHIJKLMNOPQRSTUV); |
| 480 |
_mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV); |
| 481 |
output += 32; |
| 482 |
} |
| 483 |
|
| 484 |
if XNN_UNLIKELY(c != 0) { |
| 485 |
// Prepare mask for valid 8-bit elements (depends on nc). |
| 486 |
const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << (c & 15)) - UINT32_C(1))); |
| 487 |
do { |
| 488 |
__m512i vacc0123456789ABCDEF = _mm512_loadu_si512(b); |
| 489 |
b += 16; |
| 490 |
|
| 491 |
const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0)); |
| 492 |
const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)))); |
| 493 |
i0 += 16; |
| 494 |
|
| 495 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF)); |
| 496 |
|
| 497 |
const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1)); |
| 498 |
const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)))); |
| 499 |
i1 += 16; |
| 500 |
|
| 501 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF)); |
| 502 |
|
| 503 |
const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2)); |
| 504 |
const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)))); |
| 505 |
i2 += 16; |
| 506 |
|
| 507 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF)); |
| 508 |
|
| 509 |
const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3)); |
| 510 |
const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)))); |
| 511 |
i3 += 16; |
| 512 |
|
| 513 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF)); |
| 514 |
|
| 515 |
const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); |
| 516 |
const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)))); |
| 517 |
i4 += 16; |
| 518 |
|
| 519 |
vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF)); |
| 520 |
|
| 521 |
__m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF); |
| 522 |
vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale); |
| 523 |
vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point); |
| 524 |
vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF); |
| 525 |
|
| 526 |
w = (void*) ((uintptr_t) w + 80 * sizeof(int8_t)); |
| 527 |
|
| 528 |
__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), _mm512_castsi512_si256(voutput_zero_point)); |
| 529 |
|
| 530 |
const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF); |
| 531 |
const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1); |
| 532 |
__m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0)); |
| 533 |
vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, _mm256_castsi256_si128(voutput_min)); |
| 534 |
|
| 535 |
|
| 536 |
if XNN_LIKELY(c >= 16) { |
| 537 |
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF); |
| 538 |
output += 16; |
| 539 |
c -= 16; |
| 540 |
} else { |
| 541 |
_mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF); |
| 542 |
output = (int8_t*) ((uintptr_t) output + c); |
| 543 |
c = 0; |
| 544 |
} |
| 545 |
} while (c != 0); |
| 546 |
} |
| 547 |
} |
| 548 |
|
| 549 |
input = (const int8_t**) ((uintptr_t) input + input_stride); |
| 550 |
output = (int8_t*) ((uintptr_t) output + output_increment); |
| 551 |
} while (--output_width != 0); |
| 552 |
} |