Frank Barchard | 354cbc6 | 2021-09-27 21:42:41 -0700 | [diff] [blame^] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/qu8-dwconv/unipass-neon-mul8.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2020 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 <arm_neon.h> |
| 13 | |
| 14 | #include <xnnpack/dwconv.h> |
| 15 | |
| 16 | |
| 17 | void xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul8( |
| 18 | size_t channels, |
| 19 | size_t output_width, |
| 20 | const uint8_t** input, |
| 21 | const void* weights, |
| 22 | uint8_t* output, |
| 23 | size_t input_stride, |
| 24 | size_t output_increment, |
| 25 | size_t input_offset, |
| 26 | const uint8_t* zero, |
| 27 | const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN |
| 28 | { |
| 29 | assert(channels != 0); |
| 30 | assert(output_width != 0); |
| 31 | |
| 32 | const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->rndnu_neon.kernel_zero_point); |
| 33 | const uint16x8_t vkernel_zero_point16 = vmovl_u8(vkernel_zero_point); |
| 34 | const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift); |
| 35 | const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier); |
| 36 | const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift); |
| 37 | const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point); |
| 38 | const uint8x8_t voutput_min = vld1_dup_u8(¶ms->rndnu_neon.output_min); |
| 39 | const uint8x8_t voutput_max = vld1_dup_u8(¶ms->rndnu_neon.output_max); |
| 40 | do { |
| 41 | const uint8_t* i0 = input[0]; |
| 42 | assert(i0 != NULL); |
| 43 | if XNN_UNPREDICTABLE(i0 != zero) { |
| 44 | i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset); |
| 45 | } |
| 46 | const uint8_t* i1 = input[1]; |
| 47 | assert(i1 != NULL); |
| 48 | if XNN_UNPREDICTABLE(i1 != zero) { |
| 49 | i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset); |
| 50 | } |
| 51 | const uint8_t* i2 = input[2]; |
| 52 | assert(i2 != NULL); |
| 53 | if XNN_UNPREDICTABLE(i2 != zero) { |
| 54 | i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset); |
| 55 | } |
| 56 | const uint8_t* i3 = input[3]; |
| 57 | assert(i3 != NULL); |
| 58 | if XNN_UNPREDICTABLE(i3 != zero) { |
| 59 | i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset); |
| 60 | } |
| 61 | const uint8_t* i4 = input[4]; |
| 62 | assert(i4 != NULL); |
| 63 | if XNN_UNPREDICTABLE(i4 != zero) { |
| 64 | i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset); |
| 65 | } |
| 66 | const uint8_t* i5 = input[5]; |
| 67 | assert(i5 != NULL); |
| 68 | if XNN_UNPREDICTABLE(i5 != zero) { |
| 69 | i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset); |
| 70 | } |
| 71 | const uint8_t* i6 = input[6]; |
| 72 | assert(i6 != NULL); |
| 73 | if XNN_UNPREDICTABLE(i6 != zero) { |
| 74 | i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset); |
| 75 | } |
| 76 | const uint8_t* i7 = input[7]; |
| 77 | assert(i7 != NULL); |
| 78 | if XNN_UNPREDICTABLE(i7 != zero) { |
| 79 | i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset); |
| 80 | } |
| 81 | const uint8_t* i8 = input[8]; |
| 82 | assert(i8 != NULL); |
| 83 | if XNN_UNPREDICTABLE(i8 != zero) { |
| 84 | i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset); |
| 85 | } |
| 86 | const uint8_t* i9 = input[9]; |
| 87 | assert(i9 != NULL); |
| 88 | if XNN_UNPREDICTABLE(i9 != zero) { |
| 89 | i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset); |
| 90 | } |
| 91 | const uint8_t* i10 = input[10]; |
| 92 | assert(i10 != NULL); |
| 93 | if XNN_UNPREDICTABLE(i10 != zero) { |
| 94 | i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset); |
| 95 | } |
| 96 | const uint8_t* i11 = input[11]; |
| 97 | assert(i11 != NULL); |
| 98 | if XNN_UNPREDICTABLE(i11 != zero) { |
| 99 | i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset); |
| 100 | } |
| 101 | const uint8_t* i12 = input[12]; |
| 102 | assert(i12 != NULL); |
| 103 | if XNN_UNPREDICTABLE(i12 != zero) { |
| 104 | i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset); |
| 105 | } |
| 106 | const uint8_t* i13 = input[13]; |
| 107 | assert(i13 != NULL); |
| 108 | if XNN_UNPREDICTABLE(i13 != zero) { |
| 109 | i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset); |
| 110 | } |
| 111 | const uint8_t* i14 = input[14]; |
| 112 | assert(i14 != NULL); |
| 113 | if XNN_UNPREDICTABLE(i14 != zero) { |
| 114 | i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset); |
| 115 | } |
| 116 | const uint8_t* i15 = input[15]; |
| 117 | assert(i15 != NULL); |
| 118 | if XNN_UNPREDICTABLE(i15 != zero) { |
| 119 | i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset); |
| 120 | } |
| 121 | const uint8_t* i16 = input[16]; |
| 122 | assert(i16 != NULL); |
| 123 | if XNN_UNPREDICTABLE(i16 != zero) { |
| 124 | i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset); |
| 125 | } |
| 126 | const uint8_t* i17 = input[17]; |
| 127 | assert(i17 != NULL); |
| 128 | if XNN_UNPREDICTABLE(i17 != zero) { |
| 129 | i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset); |
| 130 | } |
| 131 | const uint8_t* i18 = input[18]; |
| 132 | assert(i18 != NULL); |
| 133 | if XNN_UNPREDICTABLE(i18 != zero) { |
| 134 | i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset); |
| 135 | } |
| 136 | const uint8_t* i19 = input[19]; |
| 137 | assert(i19 != NULL); |
| 138 | if XNN_UNPREDICTABLE(i19 != zero) { |
| 139 | i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset); |
| 140 | } |
| 141 | const uint8_t* i20 = input[20]; |
| 142 | assert(i20 != NULL); |
| 143 | if XNN_UNPREDICTABLE(i20 != zero) { |
| 144 | i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset); |
| 145 | } |
| 146 | const uint8_t* i21 = input[21]; |
| 147 | assert(i21 != NULL); |
| 148 | if XNN_UNPREDICTABLE(i21 != zero) { |
| 149 | i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset); |
| 150 | } |
| 151 | const uint8_t* i22 = input[22]; |
| 152 | assert(i22 != NULL); |
| 153 | if XNN_UNPREDICTABLE(i22 != zero) { |
| 154 | i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset); |
| 155 | } |
| 156 | const uint8_t* i23 = input[23]; |
| 157 | assert(i23 != NULL); |
| 158 | if XNN_UNPREDICTABLE(i23 != zero) { |
| 159 | i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset); |
| 160 | } |
| 161 | const uint8_t* i24 = input[24]; |
| 162 | assert(i24 != NULL); |
| 163 | if XNN_UNPREDICTABLE(i24 != zero) { |
| 164 | i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset); |
| 165 | } |
| 166 | input = (const uint8_t**) ((uintptr_t) input + input_stride); |
| 167 | |
| 168 | |
| 169 | size_t c = channels; |
| 170 | const void* w = weights; |
| 171 | for (; c >= 8; c -= 8) { |
| 172 | int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4); |
| 173 | int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4); |
| 174 | |
| 175 | |
| 176 | const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; |
| 177 | const uint8x8_t vk0x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 178 | |
| 179 | uint16x8_t vprod01234567 = vmull_u8(vi0x01234567, vk0x01234567); |
| 180 | |
| 181 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 182 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 183 | const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; |
| 184 | const uint8x8_t vk1x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 185 | |
| 186 | vprod01234567 = vmull_u8(vi1x01234567, vk1x01234567); |
| 187 | uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); |
| 188 | |
| 189 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 190 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 191 | const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; |
| 192 | const uint8x8_t vk2x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 193 | |
| 194 | vprod01234567 = vmull_u8(vi2x01234567, vk2x01234567); |
| 195 | vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); |
| 196 | |
| 197 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 198 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 199 | const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8; |
| 200 | const uint8x8_t vk3x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 201 | |
| 202 | vprod01234567 = vmull_u8(vi3x01234567, vk3x01234567); |
| 203 | vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567); |
| 204 | |
| 205 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 206 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 207 | const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8; |
| 208 | const uint8x8_t vk4x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 209 | |
| 210 | vprod01234567 = vmull_u8(vi4x01234567, vk4x01234567); |
| 211 | vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567); |
| 212 | |
| 213 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 214 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 215 | const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8; |
| 216 | const uint8x8_t vk5x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 217 | |
| 218 | vprod01234567 = vmull_u8(vi5x01234567, vk5x01234567); |
| 219 | vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567); |
| 220 | |
| 221 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 222 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 223 | const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8; |
| 224 | const uint8x8_t vk6x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 225 | |
| 226 | vprod01234567 = vmull_u8(vi6x01234567, vk6x01234567); |
| 227 | vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567); |
| 228 | |
| 229 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 230 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 231 | const uint8x8_t vi7x01234567 = vld1_u8(i7); i7 += 8; |
| 232 | const uint8x8_t vk7x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 233 | |
| 234 | vprod01234567 = vmull_u8(vi7x01234567, vk7x01234567); |
| 235 | vsum01234567 = vaddw_u8(vsum01234567, vi7x01234567); |
| 236 | |
| 237 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 238 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 239 | const uint8x8_t vi8x01234567 = vld1_u8(i8); i8 += 8; |
| 240 | const uint8x8_t vk8x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 241 | |
| 242 | vprod01234567 = vmull_u8(vi8x01234567, vk8x01234567); |
| 243 | vsum01234567 = vaddw_u8(vsum01234567, vi8x01234567); |
| 244 | |
| 245 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 246 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 247 | const uint8x8_t vi9x01234567 = vld1_u8(i9); i9 += 8; |
| 248 | const uint8x8_t vk9x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 249 | |
| 250 | vprod01234567 = vmull_u8(vi9x01234567, vk9x01234567); |
| 251 | vsum01234567 = vaddw_u8(vsum01234567, vi9x01234567); |
| 252 | |
| 253 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 254 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 255 | const uint8x8_t vi10x01234567 = vld1_u8(i10); i10 += 8; |
| 256 | const uint8x8_t vk10x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 257 | |
| 258 | vprod01234567 = vmull_u8(vi10x01234567, vk10x01234567); |
| 259 | vsum01234567 = vaddw_u8(vsum01234567, vi10x01234567); |
| 260 | |
| 261 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 262 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 263 | const uint8x8_t vi11x01234567 = vld1_u8(i11); i11 += 8; |
| 264 | const uint8x8_t vk11x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 265 | |
| 266 | vprod01234567 = vmull_u8(vi11x01234567, vk11x01234567); |
| 267 | vsum01234567 = vaddw_u8(vsum01234567, vi11x01234567); |
| 268 | |
| 269 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 270 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 271 | const uint8x8_t vi12x01234567 = vld1_u8(i12); i12 += 8; |
| 272 | const uint8x8_t vk12x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 273 | |
| 274 | vprod01234567 = vmull_u8(vi12x01234567, vk12x01234567); |
| 275 | vsum01234567 = vaddw_u8(vsum01234567, vi12x01234567); |
| 276 | |
| 277 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 278 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 279 | const uint8x8_t vi13x01234567 = vld1_u8(i13); i13 += 8; |
| 280 | const uint8x8_t vk13x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 281 | |
| 282 | vprod01234567 = vmull_u8(vi13x01234567, vk13x01234567); |
| 283 | vsum01234567 = vaddw_u8(vsum01234567, vi13x01234567); |
| 284 | |
| 285 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 286 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 287 | const uint8x8_t vi14x01234567 = vld1_u8(i14); i14 += 8; |
| 288 | const uint8x8_t vk14x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 289 | |
| 290 | vprod01234567 = vmull_u8(vi14x01234567, vk14x01234567); |
| 291 | vsum01234567 = vaddw_u8(vsum01234567, vi14x01234567); |
| 292 | |
| 293 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 294 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 295 | const uint8x8_t vi15x01234567 = vld1_u8(i15); i15 += 8; |
| 296 | const uint8x8_t vk15x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 297 | |
| 298 | vprod01234567 = vmull_u8(vi15x01234567, vk15x01234567); |
| 299 | vsum01234567 = vaddw_u8(vsum01234567, vi15x01234567); |
| 300 | |
| 301 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 302 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 303 | const uint8x8_t vi16x01234567 = vld1_u8(i16); i16 += 8; |
| 304 | const uint8x8_t vk16x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 305 | |
| 306 | vprod01234567 = vmull_u8(vi16x01234567, vk16x01234567); |
| 307 | vsum01234567 = vaddw_u8(vsum01234567, vi16x01234567); |
| 308 | |
| 309 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 310 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 311 | const uint8x8_t vi17x01234567 = vld1_u8(i17); i17 += 8; |
| 312 | const uint8x8_t vk17x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 313 | |
| 314 | vprod01234567 = vmull_u8(vi17x01234567, vk17x01234567); |
| 315 | vsum01234567 = vaddw_u8(vsum01234567, vi17x01234567); |
| 316 | |
| 317 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 318 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 319 | const uint8x8_t vi18x01234567 = vld1_u8(i18); i18 += 8; |
| 320 | const uint8x8_t vk18x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 321 | |
| 322 | vprod01234567 = vmull_u8(vi18x01234567, vk18x01234567); |
| 323 | vsum01234567 = vaddw_u8(vsum01234567, vi18x01234567); |
| 324 | |
| 325 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 326 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 327 | const uint8x8_t vi19x01234567 = vld1_u8(i19); i19 += 8; |
| 328 | const uint8x8_t vk19x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 329 | |
| 330 | vprod01234567 = vmull_u8(vi19x01234567, vk19x01234567); |
| 331 | vsum01234567 = vaddw_u8(vsum01234567, vi19x01234567); |
| 332 | |
| 333 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 334 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 335 | const uint8x8_t vi20x01234567 = vld1_u8(i20); i20 += 8; |
| 336 | const uint8x8_t vk20x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 337 | |
| 338 | vprod01234567 = vmull_u8(vi20x01234567, vk20x01234567); |
| 339 | vsum01234567 = vaddw_u8(vsum01234567, vi20x01234567); |
| 340 | |
| 341 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 342 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 343 | const uint8x8_t vi21x01234567 = vld1_u8(i21); i21 += 8; |
| 344 | const uint8x8_t vk21x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 345 | |
| 346 | vprod01234567 = vmull_u8(vi21x01234567, vk21x01234567); |
| 347 | vsum01234567 = vaddw_u8(vsum01234567, vi21x01234567); |
| 348 | |
| 349 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 350 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 351 | const uint8x8_t vi22x01234567 = vld1_u8(i22); i22 += 8; |
| 352 | const uint8x8_t vk22x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 353 | |
| 354 | vprod01234567 = vmull_u8(vi22x01234567, vk22x01234567); |
| 355 | vsum01234567 = vaddw_u8(vsum01234567, vi22x01234567); |
| 356 | |
| 357 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 358 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 359 | const uint8x8_t vi23x01234567 = vld1_u8(i23); i23 += 8; |
| 360 | const uint8x8_t vk23x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 361 | |
| 362 | vprod01234567 = vmull_u8(vi23x01234567, vk23x01234567); |
| 363 | vsum01234567 = vaddw_u8(vsum01234567, vi23x01234567); |
| 364 | |
| 365 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 366 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 367 | const uint8x8_t vi24x01234567 = vld1_u8(i24); i24 += 8; |
| 368 | const uint8x8_t vk24x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); |
| 369 | |
| 370 | vprod01234567 = vmull_u8(vi24x01234567, vk24x01234567); |
| 371 | vsum01234567 = vaddw_u8(vsum01234567, vi24x01234567); |
| 372 | |
| 373 | vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567))); |
| 374 | vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567))); |
| 375 | |
| 376 | vacc0123 = vreinterpretq_s32_u32(vmlsl_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567), vget_low_u16(vkernel_zero_point16))); |
| 377 | vacc4567 = vreinterpretq_s32_u32(vmlsl_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567), vget_high_u16(vkernel_zero_point16))); |
| 378 | |
| 379 | vacc0123 = vshlq_s32(vacc0123, vright_pre_shift); |
| 380 | vacc4567 = vshlq_s32(vacc4567, vright_pre_shift); |
| 381 | |
| 382 | vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier); |
| 383 | vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier); |
| 384 | |
| 385 | vacc0123 = vrshlq_s32(vacc0123, vright_post_shift); |
| 386 | vacc4567 = vrshlq_s32(vacc4567, vright_post_shift); |
| 387 | |
| 388 | #if XNN_ARCH_ARM64 |
| 389 | const int16x8_t vacc01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567), voutput_zero_point); |
| 390 | |
| 391 | uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); |
| 392 | #else |
| 393 | const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point); |
| 394 | |
| 395 | uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); |
| 396 | #endif |
| 397 | |
| 398 | vout01234567 = vmax_u8(vout01234567, voutput_min); |
| 399 | |
| 400 | vout01234567 = vmin_u8(vout01234567, voutput_max); |
| 401 | |
| 402 | vst1_u8(output, vout01234567); output += 8; |
| 403 | } |
| 404 | if XNN_UNLIKELY(c != 0) { |
| 405 | { |
| 406 | int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4); |
| 407 | int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4); |
| 408 | |
| 409 | const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); |
| 410 | const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); |
| 411 | |
| 412 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567)); |
| 413 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567)); |
| 414 | const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); |
| 415 | const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 8)), vkernel_zero_point)); |
| 416 | |
| 417 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567)); |
| 418 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567)); |
| 419 | const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); |
| 420 | const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 16)), vkernel_zero_point)); |
| 421 | |
| 422 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567)); |
| 423 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567)); |
| 424 | const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); |
| 425 | const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 24)), vkernel_zero_point)); |
| 426 | |
| 427 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567)); |
| 428 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567)); |
| 429 | const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); |
| 430 | const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 32)), vkernel_zero_point)); |
| 431 | |
| 432 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567)); |
| 433 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567)); |
| 434 | const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); |
| 435 | const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 40)), vkernel_zero_point)); |
| 436 | |
| 437 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567)); |
| 438 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567)); |
| 439 | const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); |
| 440 | const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 48)), vkernel_zero_point)); |
| 441 | |
| 442 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567)); |
| 443 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567)); |
| 444 | const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); |
| 445 | const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 56)), vkernel_zero_point)); |
| 446 | |
| 447 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567)); |
| 448 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567)); |
| 449 | const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); |
| 450 | const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 64)), vkernel_zero_point)); |
| 451 | |
| 452 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567)); |
| 453 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567)); |
| 454 | const int16x8_t vi9x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i9))); |
| 455 | const int16x8_t vk9x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 72)), vkernel_zero_point)); |
| 456 | |
| 457 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi9x01234567), vget_low_s16(vk9x01234567)); |
| 458 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567)); |
| 459 | const int16x8_t vi10x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i10))); |
| 460 | const int16x8_t vk10x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 80)), vkernel_zero_point)); |
| 461 | |
| 462 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567)); |
| 463 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567)); |
| 464 | const int16x8_t vi11x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i11))); |
| 465 | const int16x8_t vk11x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 88)), vkernel_zero_point)); |
| 466 | |
| 467 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567)); |
| 468 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567)); |
| 469 | const int16x8_t vi12x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i12))); |
| 470 | const int16x8_t vk12x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 96)), vkernel_zero_point)); |
| 471 | |
| 472 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567)); |
| 473 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567)); |
| 474 | const int16x8_t vi13x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i13))); |
| 475 | const int16x8_t vk13x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 104)), vkernel_zero_point)); |
| 476 | |
| 477 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567)); |
| 478 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567)); |
| 479 | const int16x8_t vi14x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i14))); |
| 480 | const int16x8_t vk14x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 112)), vkernel_zero_point)); |
| 481 | |
| 482 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567)); |
| 483 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567)); |
| 484 | const int16x8_t vi15x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i15))); |
| 485 | const int16x8_t vk15x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 120)), vkernel_zero_point)); |
| 486 | |
| 487 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567)); |
| 488 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567)); |
| 489 | const int16x8_t vi16x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i16))); |
| 490 | const int16x8_t vk16x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 128)), vkernel_zero_point)); |
| 491 | |
| 492 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567)); |
| 493 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567)); |
| 494 | const int16x8_t vi17x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i17))); |
| 495 | const int16x8_t vk17x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 136)), vkernel_zero_point)); |
| 496 | |
| 497 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567)); |
| 498 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567)); |
| 499 | const int16x8_t vi18x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); |
| 500 | const int16x8_t vk18x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 144)), vkernel_zero_point)); |
| 501 | |
| 502 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567)); |
| 503 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567)); |
| 504 | const int16x8_t vi19x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i19))); |
| 505 | const int16x8_t vk19x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 152)), vkernel_zero_point)); |
| 506 | |
| 507 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567)); |
| 508 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567)); |
| 509 | const int16x8_t vi20x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i20))); |
| 510 | const int16x8_t vk20x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 160)), vkernel_zero_point)); |
| 511 | |
| 512 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567)); |
| 513 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567)); |
| 514 | const int16x8_t vi21x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); |
| 515 | const int16x8_t vk21x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 168)), vkernel_zero_point)); |
| 516 | |
| 517 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567)); |
| 518 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567)); |
| 519 | const int16x8_t vi22x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i22))); |
| 520 | const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 176)), vkernel_zero_point)); |
| 521 | |
| 522 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); |
| 523 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); |
| 524 | const int16x8_t vi23x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i23))); |
| 525 | const int16x8_t vk23x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 184)), vkernel_zero_point)); |
| 526 | |
| 527 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567)); |
| 528 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567)); |
| 529 | const int16x8_t vi24x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i24))); |
| 530 | const int16x8_t vk24x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 192)), vkernel_zero_point)); |
| 531 | |
| 532 | vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567)); |
| 533 | vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567)); |
| 534 | |
| 535 | vacc0123 = vrshlq_s32(vacc0123, vright_pre_shift); |
| 536 | vacc4567 = vrshlq_s32(vacc4567, vright_pre_shift); |
| 537 | |
| 538 | vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier); |
| 539 | vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier); |
| 540 | |
| 541 | vacc0123 = vrshlq_s32(vacc0123, vright_post_shift); |
| 542 | vacc4567 = vrshlq_s32(vacc4567, vright_post_shift); |
| 543 | |
| 544 | #if XNN_ARCH_ARM64 |
| 545 | const int16x8_t vacc01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567), voutput_zero_point); |
| 546 | uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); |
| 547 | #else |
| 548 | const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point); |
| 549 | uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); |
| 550 | #endif |
| 551 | |
| 552 | vout01234567 = vmax_u8(vout01234567, voutput_min); |
| 553 | vout01234567 = vmin_u8(vout01234567, voutput_max); |
| 554 | |
| 555 | if (c & 4) { |
| 556 | vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_u8(vout01234567), 0); output += 4; |
| 557 | vout01234567 = vext_u8(vout01234567, vout01234567, 4); |
| 558 | } |
| 559 | if (c & 2) { |
| 560 | vst1_lane_u16(__builtin_assume_aligned(output, 1), vreinterpret_u16_u8(vout01234567), 0); output += 2; |
| 561 | vout01234567 = vext_u8(vout01234567, vout01234567, 2); |
| 562 | } |
| 563 | if (c & 1) { |
| 564 | vst1_lane_u8(output, vout01234567, 0); output += 1; |
| 565 | } |
| 566 | } |
| 567 | } |
| 568 | |
| 569 | output = (uint8_t*) ((uintptr_t) output + output_increment); |
| 570 | } while (--output_width != 0); |
| 571 | } |