Frank Barchard | 42f5c50 | 2021-11-16 10:04:21 -0800 | [diff] [blame^] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/qs8-igemm/c2-neon-mull-dup.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 <arm_neon.h> |
| 13 | |
| 14 | #include <xnnpack/gemm.h> |
| 15 | #include <xnnpack/math.h> |
| 16 | |
| 17 | void xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r( |
| 18 | size_t mr, |
| 19 | size_t nc, |
| 20 | size_t kc, |
| 21 | size_t ks, |
| 22 | const int8_t** restrict a, |
| 23 | const void* restrict w, |
| 24 | int8_t* restrict c, |
| 25 | size_t cm_stride, |
| 26 | size_t cn_stride, |
| 27 | size_t a_offset, |
| 28 | const int8_t* zero, |
| 29 | const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN |
| 30 | { |
| 31 | assert(mr != 0); |
| 32 | assert(mr <= 3); |
| 33 | assert(nc != 0); |
| 34 | assert(kc != 0); |
| 35 | assert(ks != 0); |
| 36 | assert(ks % (3 * sizeof(void*)) == 0); |
| 37 | assert(a_offset % sizeof(int8_t) == 0); |
| 38 | assert(a != NULL); |
| 39 | assert(w != NULL); |
| 40 | assert(c != NULL); |
| 41 | |
| 42 | kc = round_up_po2(kc, 2 * sizeof(int8_t)); |
| 43 | int8_t* c0 = c; |
| 44 | int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); |
| 45 | if XNN_UNPREDICTABLE(mr < 2) { |
| 46 | c1 = c0; |
| 47 | } |
| 48 | int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride); |
| 49 | if XNN_UNPREDICTABLE(mr <= 2) { |
| 50 | c2 = c1; |
| 51 | } |
| 52 | |
| 53 | do { |
| 54 | int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| 55 | int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| 56 | int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| 57 | int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| 58 | int32x4_t vacc1x0123 = vacc0x0123; |
| 59 | int32x4_t vacc1x4567 = vacc0x4567; |
| 60 | int32x4_t vacc1x89AB = vacc0x89AB; |
| 61 | int32x4_t vacc1xCDEF = vacc0xCDEF; |
| 62 | int32x4_t vacc2x0123 = vacc0x0123; |
| 63 | int32x4_t vacc2x4567 = vacc0x4567; |
| 64 | int32x4_t vacc2x89AB = vacc0x89AB; |
| 65 | int32x4_t vacc2xCDEF = vacc0xCDEF; |
| 66 | |
| 67 | size_t p = ks; |
| 68 | do { |
| 69 | const int8_t* restrict a0 = a[0]; |
| 70 | if XNN_UNPREDICTABLE(a0 != zero) { |
| 71 | a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); |
| 72 | } |
| 73 | const int8_t* restrict a1 = a[1]; |
| 74 | if XNN_UNPREDICTABLE(a1 != zero) { |
| 75 | a1 = (const int8_t*) ((uintptr_t) a1 + a_offset); |
| 76 | } |
| 77 | const int8_t* restrict a2 = a[2]; |
| 78 | if XNN_UNPREDICTABLE(a2 != zero) { |
| 79 | a2 = (const int8_t*) ((uintptr_t) a2 + a_offset); |
| 80 | } |
| 81 | a += 3; |
| 82 | |
| 83 | size_t k = kc; |
| 84 | |
| 85 | |
| 86 | while (k >= 8 * sizeof(int8_t)) { |
| 87 | const int16x4x4_t va0 = vld4_dup_s16((const void*)a0); a0 += 8; |
| 88 | const int16x4x4_t va1 = vld4_dup_s16((const void*)a1); a1 += 8; |
| 89 | const int16x4x4_t va2 = vld4_dup_s16((const void*)a2); a2 += 8; |
| 90 | |
| 91 | const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 92 | const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 93 | const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 94 | const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 95 | const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 96 | const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 97 | const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 98 | const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 99 | const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 100 | const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 101 | const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 102 | const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 103 | const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 104 | const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 105 | const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 106 | const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 107 | |
| 108 | const int8x8_t va0c0 = vreinterpret_s8_s16(va0.val[0]); |
| 109 | const int8x8_t va1c0 = vreinterpret_s8_s16(va1.val[0]); |
| 110 | const int8x8_t va2c0 = vreinterpret_s8_s16(va2.val[0]); |
| 111 | |
| 112 | const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0); |
| 113 | const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0); |
| 114 | const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0); |
| 115 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0); |
| 116 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0); |
| 117 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0); |
| 118 | const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0); |
| 119 | const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0); |
| 120 | const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); |
| 121 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0); |
| 122 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0); |
| 123 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0); |
| 124 | const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0); |
| 125 | const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); |
| 126 | const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0); |
| 127 | vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0); |
| 128 | vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); |
| 129 | vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0); |
| 130 | const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); |
| 131 | const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); |
| 132 | const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); |
| 133 | vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0); |
| 134 | vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0); |
| 135 | vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0); |
| 136 | const int8x8_t va0c1 = vreinterpret_s8_s16(va0.val[1]); |
| 137 | const int8x8_t va1c1 = vreinterpret_s8_s16(va1.val[1]); |
| 138 | const int8x8_t va2c1 = vreinterpret_s8_s16(va2.val[1]); |
| 139 | |
| 140 | const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1); |
| 141 | const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1); |
| 142 | const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); |
| 143 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1); |
| 144 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1); |
| 145 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); |
| 146 | const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); |
| 147 | const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1); |
| 148 | const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1); |
| 149 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); |
| 150 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1); |
| 151 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1); |
| 152 | const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); |
| 153 | const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); |
| 154 | const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); |
| 155 | vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1); |
| 156 | vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1); |
| 157 | vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1); |
| 158 | const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); |
| 159 | const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); |
| 160 | const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1); |
| 161 | vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1); |
| 162 | vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); |
| 163 | vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1); |
| 164 | const int8x8_t va0c2 = vreinterpret_s8_s16(va0.val[2]); |
| 165 | const int8x8_t va1c2 = vreinterpret_s8_s16(va1.val[2]); |
| 166 | const int8x8_t va2c2 = vreinterpret_s8_s16(va2.val[2]); |
| 167 | |
| 168 | const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2); |
| 169 | const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); |
| 170 | const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2); |
| 171 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2); |
| 172 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2); |
| 173 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2); |
| 174 | const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2); |
| 175 | const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); |
| 176 | const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2); |
| 177 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2); |
| 178 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2); |
| 179 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2); |
| 180 | const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2); |
| 181 | const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); |
| 182 | const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2); |
| 183 | vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2); |
| 184 | vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2); |
| 185 | vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2); |
| 186 | const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); |
| 187 | const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); |
| 188 | const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2); |
| 189 | vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2); |
| 190 | vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2); |
| 191 | vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2); |
| 192 | const int8x8_t va0c3 = vreinterpret_s8_s16(va0.val[3]); |
| 193 | const int8x8_t va1c3 = vreinterpret_s8_s16(va1.val[3]); |
| 194 | const int8x8_t va2c3 = vreinterpret_s8_s16(va2.val[3]); |
| 195 | |
| 196 | const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3); |
| 197 | const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3); |
| 198 | const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3); |
| 199 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3); |
| 200 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3); |
| 201 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3); |
| 202 | const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3); |
| 203 | const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3); |
| 204 | const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3); |
| 205 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3); |
| 206 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3); |
| 207 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3); |
| 208 | const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3); |
| 209 | const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3); |
| 210 | const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); |
| 211 | vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3); |
| 212 | vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3); |
| 213 | vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3); |
| 214 | const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); |
| 215 | const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3); |
| 216 | const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3); |
| 217 | vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3); |
| 218 | vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3); |
| 219 | vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3); |
| 220 | |
| 221 | k -= 8 * sizeof(int8_t); |
| 222 | } |
| 223 | |
| 224 | if XNN_UNLIKELY(k != 0) { |
| 225 | const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k); |
| 226 | const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k); |
| 227 | const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k); |
| 228 | |
| 229 | const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 230 | const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 231 | const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 232 | const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 233 | |
| 234 | const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)); |
| 235 | const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0); |
| 236 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0); |
| 237 | const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0); |
| 238 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0); |
| 239 | const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0); |
| 240 | vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0); |
| 241 | const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); |
| 242 | vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0); |
| 243 | const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)); |
| 244 | const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0); |
| 245 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0); |
| 246 | const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0); |
| 247 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0); |
| 248 | const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); |
| 249 | vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); |
| 250 | const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); |
| 251 | vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0); |
| 252 | const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)); |
| 253 | const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0); |
| 254 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0); |
| 255 | const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); |
| 256 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0); |
| 257 | const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0); |
| 258 | vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0); |
| 259 | const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); |
| 260 | vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0); |
| 261 | |
| 262 | if (k > 2 * sizeof(int8_t)) { |
| 263 | const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 264 | const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 265 | const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 266 | const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 267 | |
| 268 | const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)); |
| 269 | const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1); |
| 270 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1); |
| 271 | const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); |
| 272 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); |
| 273 | const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); |
| 274 | vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1); |
| 275 | const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); |
| 276 | vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1); |
| 277 | const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)); |
| 278 | const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1); |
| 279 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1); |
| 280 | const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1); |
| 281 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1); |
| 282 | const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); |
| 283 | vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1); |
| 284 | const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); |
| 285 | vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); |
| 286 | const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)); |
| 287 | const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); |
| 288 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); |
| 289 | const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1); |
| 290 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1); |
| 291 | const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); |
| 292 | vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1); |
| 293 | const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1); |
| 294 | vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1); |
| 295 | |
| 296 | if (k > 4 * sizeof(int8_t)) { |
| 297 | const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 298 | const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 299 | const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 300 | const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 301 | |
| 302 | const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)); |
| 303 | const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2); |
| 304 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2); |
| 305 | const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2); |
| 306 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2); |
| 307 | const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2); |
| 308 | vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2); |
| 309 | const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); |
| 310 | vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2); |
| 311 | const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); |
| 312 | const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); |
| 313 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2); |
| 314 | const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); |
| 315 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2); |
| 316 | const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); |
| 317 | vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2); |
| 318 | const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); |
| 319 | vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2); |
| 320 | const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)); |
| 321 | const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2); |
| 322 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2); |
| 323 | const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2); |
| 324 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2); |
| 325 | const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2); |
| 326 | vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2); |
| 327 | const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2); |
| 328 | vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2); |
| 329 | } |
| 330 | } |
| 331 | } |
| 332 | p -= 3 * sizeof(void*); |
| 333 | } while (p != 0); |
| 334 | |
| 335 | const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift); |
| 336 | const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier); |
| 337 | const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift); |
| 338 | |
| 339 | vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift); |
| 340 | vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift); |
| 341 | vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift); |
| 342 | vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift); |
| 343 | vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift); |
| 344 | vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift); |
| 345 | vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift); |
| 346 | vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift); |
| 347 | vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift); |
| 348 | vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift); |
| 349 | vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift); |
| 350 | vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift); |
| 351 | |
| 352 | vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier); |
| 353 | vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier); |
| 354 | vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier); |
| 355 | vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier); |
| 356 | vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier); |
| 357 | vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier); |
| 358 | vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier); |
| 359 | vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier); |
| 360 | vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier); |
| 361 | vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier); |
| 362 | vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier); |
| 363 | vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier); |
| 364 | |
| 365 | vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift); |
| 366 | vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift); |
| 367 | vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift); |
| 368 | vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift); |
| 369 | vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift); |
| 370 | vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift); |
| 371 | vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift); |
| 372 | vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift); |
| 373 | vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift); |
| 374 | vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift); |
| 375 | vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift); |
| 376 | vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift); |
| 377 | |
| 378 | const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point); |
| 379 | #if XNN_ARCH_ARM64 |
| 380 | const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point); |
| 381 | const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point); |
| 382 | const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point); |
| 383 | const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point); |
| 384 | const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point); |
| 385 | const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point); |
| 386 | |
| 387 | int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); |
| 388 | int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); |
| 389 | int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF); |
| 390 | #else |
| 391 | const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point); |
| 392 | const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point); |
| 393 | const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point); |
| 394 | const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point); |
| 395 | const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point); |
| 396 | const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point); |
| 397 | |
| 398 | int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF)); |
| 399 | int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF)); |
| 400 | int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF)); |
| 401 | #endif |
| 402 | const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min); |
| 403 | const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max); |
| 404 | |
| 405 | vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min); |
| 406 | vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); |
| 407 | vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min); |
| 408 | |
| 409 | vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max); |
| 410 | vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); |
| 411 | vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max); |
| 412 | |
| 413 | if (nc >= 16) { |
| 414 | vst1q_s8(c2 + 0, vout2x0123456789ABCDEF); |
| 415 | vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); |
| 416 | vst1q_s8(c0 + 0, vout0x0123456789ABCDEF); |
| 417 | |
| 418 | c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); |
| 419 | c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); |
| 420 | c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 421 | |
| 422 | a = (const int8_t**restrict) ((uintptr_t) a - ks); |
| 423 | |
| 424 | nc -= 16; |
| 425 | } else { |
| 426 | int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF); |
| 427 | int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); |
| 428 | if (nc & 8) { |
| 429 | vst1_s8(c2, vout2x01234567); c2 += 8; |
| 430 | vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8; |
| 431 | vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8; |
| 432 | vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF); |
| 433 | vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); |
| 434 | } |
| 435 | if (nc & 4) { |
| 436 | vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4; |
| 437 | vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4; |
| 438 | vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4; |
| 439 | vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4); |
| 440 | vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4); |
| 441 | } |
| 442 | if (nc & 2) { |
| 443 | vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2; |
| 444 | vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2; |
| 445 | vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2; |
| 446 | vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2); |
| 447 | vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2); |
| 448 | } |
| 449 | if (nc & 1) { |
| 450 | vst1_lane_s8(c2, vout2x01234567, 0); |
| 451 | vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8); |
| 452 | vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0); |
| 453 | } |
| 454 | |
| 455 | nc = 0; |
| 456 | } |
| 457 | } while (nc != 0); |
| 458 | } |