Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/qs8-igemm/c4-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 | |
| 18 | void xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r( |
| 19 | size_t mr, |
| 20 | size_t nc, |
| 21 | size_t kc, |
| 22 | size_t ks, |
| 23 | const int8_t** restrict a, |
| 24 | const void* restrict w, |
| 25 | int8_t* restrict c, |
| 26 | size_t cm_stride, |
| 27 | size_t cn_stride, |
| 28 | size_t a_offset, |
| 29 | const int8_t* zero, |
Marat Dukhan | d541fc0 | 2021-12-14 00:14:27 -0800 | [diff] [blame] | 30 | const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 31 | { |
| 32 | assert(mr != 0); |
| 33 | assert(mr <= 4); |
| 34 | assert(nc != 0); |
| 35 | assert(kc != 0); |
| 36 | assert(ks != 0); |
| 37 | assert(ks % (4 * sizeof(void*)) == 0); |
| 38 | assert(a_offset % sizeof(int8_t) == 0); |
| 39 | assert(a != NULL); |
| 40 | assert(w != NULL); |
| 41 | assert(c != NULL); |
| 42 | |
| 43 | kc = round_up_po2(kc, 4 * sizeof(int8_t)); |
| 44 | int8_t* c0 = c; |
| 45 | int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); |
| 46 | if XNN_UNPREDICTABLE(mr < 2) { |
| 47 | c1 = c0; |
| 48 | } |
| 49 | int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride); |
| 50 | if XNN_UNPREDICTABLE(mr <= 2) { |
| 51 | c2 = c1; |
| 52 | } |
| 53 | int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride); |
| 54 | if XNN_UNPREDICTABLE(mr != 4) { |
| 55 | c3 = c2; |
| 56 | } |
| 57 | |
| 58 | do { |
Marat Dukhan | dc54e12 | 2021-12-29 12:56:08 -0800 | [diff] [blame^] | 59 | int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t)); |
| 60 | int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t)); |
| 61 | int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t)); |
| 62 | int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t)); |
| 63 | int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t)); |
| 64 | int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t)); |
| 65 | int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t)); |
| 66 | int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t)); |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 67 | int32x4_t vacc1x01 = vacc0x01; |
| 68 | int32x4_t vacc1x23 = vacc0x23; |
| 69 | int32x4_t vacc1x45 = vacc0x45; |
| 70 | int32x4_t vacc1x67 = vacc0x67; |
| 71 | int32x4_t vacc1x89 = vacc0x89; |
| 72 | int32x4_t vacc1xAB = vacc0xAB; |
| 73 | int32x4_t vacc1xCD = vacc0xCD; |
| 74 | int32x4_t vacc1xEF = vacc0xEF; |
| 75 | int32x4_t vacc2x01 = vacc0x01; |
| 76 | int32x4_t vacc2x23 = vacc0x23; |
| 77 | int32x4_t vacc2x45 = vacc0x45; |
| 78 | int32x4_t vacc2x67 = vacc0x67; |
| 79 | int32x4_t vacc2x89 = vacc0x89; |
| 80 | int32x4_t vacc2xAB = vacc0xAB; |
| 81 | int32x4_t vacc2xCD = vacc0xCD; |
| 82 | int32x4_t vacc2xEF = vacc0xEF; |
| 83 | int32x4_t vacc3x01 = vacc0x01; |
| 84 | int32x4_t vacc3x23 = vacc0x23; |
| 85 | int32x4_t vacc3x45 = vacc0x45; |
| 86 | int32x4_t vacc3x67 = vacc0x67; |
| 87 | int32x4_t vacc3x89 = vacc0x89; |
| 88 | int32x4_t vacc3xAB = vacc0xAB; |
| 89 | int32x4_t vacc3xCD = vacc0xCD; |
| 90 | int32x4_t vacc3xEF = vacc0xEF; |
| 91 | |
| 92 | size_t p = ks; |
| 93 | do { |
| 94 | const int8_t* restrict a0 = a[0]; |
| 95 | if XNN_UNPREDICTABLE(a0 != zero) { |
| 96 | a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); |
| 97 | } |
| 98 | const int8_t* restrict a1 = a[1]; |
| 99 | if XNN_UNPREDICTABLE(a1 != zero) { |
| 100 | a1 = (const int8_t*) ((uintptr_t) a1 + a_offset); |
| 101 | } |
| 102 | const int8_t* restrict a2 = a[2]; |
| 103 | if XNN_UNPREDICTABLE(a2 != zero) { |
| 104 | a2 = (const int8_t*) ((uintptr_t) a2 + a_offset); |
| 105 | } |
| 106 | const int8_t* restrict a3 = a[3]; |
| 107 | if XNN_UNPREDICTABLE(a3 != zero) { |
| 108 | a3 = (const int8_t*) ((uintptr_t) a3 + a_offset); |
| 109 | } |
| 110 | a += 4; |
| 111 | |
| 112 | size_t k = kc; |
| 113 | |
| 114 | while (k >= 16 * sizeof(int8_t)) { |
| 115 | const int32x2_t va00x0 = vld1_dup_s32((const void*)a0); |
| 116 | const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8; |
| 117 | const int32x2_t va00x1 = vld1_dup_s32((const void*)a0); |
| 118 | const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8; |
| 119 | const int32x2_t va10x0 = vld1_dup_s32((const void*)a1); |
| 120 | const int32x2_t va11x0 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8; |
| 121 | const int32x2_t va10x1 = vld1_dup_s32((const void*)a1); |
| 122 | const int32x2_t va11x1 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8; |
| 123 | const int32x2_t va20x0 = vld1_dup_s32((const void*)a2); |
| 124 | const int32x2_t va21x0 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8; |
| 125 | const int32x2_t va20x1 = vld1_dup_s32((const void*)a2); |
| 126 | const int32x2_t va21x1 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8; |
| 127 | const int32x2_t va30x0 = vld1_dup_s32((const void*)a3); |
| 128 | const int32x2_t va31x0 = vld1_dup_s32((const void*)(a3 + 4)); a3 += 8; |
| 129 | const int32x2_t va30x1 = vld1_dup_s32((const void*)a3); |
| 130 | const int32x2_t va31x1 = vld1_dup_s32((const void*)(a3 + 4)); a3 += 8; |
| 131 | |
| 132 | const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 133 | const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 134 | const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 135 | const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 136 | const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 137 | const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 138 | const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 139 | const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 140 | const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 141 | const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 142 | const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 143 | const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 144 | const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 145 | const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 146 | const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 147 | const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 148 | |
| 149 | const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0); |
| 150 | const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1); |
| 151 | const int8x8_t va1c0x0 = vreinterpret_s8_s32(va10x0); |
| 152 | const int8x8_t va1c0x1 = vreinterpret_s8_s32(va10x1); |
| 153 | const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0); |
| 154 | const int8x8_t va2c0x1 = vreinterpret_s8_s32(va20x1); |
| 155 | const int8x8_t va3c0x0 = vreinterpret_s8_s32(va30x0); |
| 156 | const int8x8_t va3c0x1 = vreinterpret_s8_s32(va30x1); |
| 157 | |
| 158 | int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0); |
| 159 | int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); |
| 160 | int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0); |
| 161 | int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3c0x0); |
| 162 | const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 163 | vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1); |
| 164 | vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); |
| 165 | vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2c0x1); |
| 166 | vprod3x01c0 = vmlal_s8(vprod3x01c0, vb01c0x1, va3c0x1); |
| 167 | vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0); |
| 168 | vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); |
| 169 | vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0); |
| 170 | vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0); |
| 171 | int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0); |
| 172 | int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); |
| 173 | int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); |
| 174 | int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3c0x0); |
| 175 | const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 176 | vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1); |
| 177 | vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1); |
| 178 | vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); |
| 179 | vprod3x23c0 = vmlal_s8(vprod3x23c0, vb23c0x1, va3c0x1); |
| 180 | vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0); |
| 181 | vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0); |
| 182 | vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); |
| 183 | vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0); |
| 184 | int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0); |
| 185 | int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0); |
| 186 | int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0); |
| 187 | int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); |
| 188 | const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 189 | vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1); |
| 190 | vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1); |
| 191 | vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2c0x1); |
| 192 | vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); |
| 193 | vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0); |
| 194 | vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0); |
| 195 | vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0); |
| 196 | vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); |
| 197 | int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0); |
| 198 | int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); |
| 199 | int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); |
| 200 | int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3c0x0); |
| 201 | const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 202 | vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1); |
| 203 | vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1); |
| 204 | vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); |
| 205 | vprod3x67c0 = vmlal_s8(vprod3x67c0, vb67c0x1, va3c0x1); |
| 206 | vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0); |
| 207 | vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0); |
| 208 | vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); |
| 209 | vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0); |
| 210 | int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0); |
| 211 | int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); |
| 212 | int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); |
| 213 | int16x8_t vprod3x89c0 = vmull_s8(vb89c0x0, va3c0x0); |
| 214 | const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 215 | vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); |
| 216 | vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); |
| 217 | vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); |
| 218 | vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3c0x1); |
| 219 | vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0); |
| 220 | vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0); |
| 221 | vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0); |
| 222 | vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0); |
| 223 | int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); |
| 224 | int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1c0x0); |
| 225 | int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); |
| 226 | int16x8_t vprod3xABc0 = vmull_s8(vbABc0x0, va3c0x0); |
| 227 | const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 228 | vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1); |
| 229 | vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1c0x1); |
| 230 | vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); |
| 231 | vprod3xABc0 = vmlal_s8(vprod3xABc0, vbABc0x1, va3c0x1); |
| 232 | vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0); |
| 233 | vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0); |
| 234 | vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); |
| 235 | vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0); |
| 236 | int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); |
| 237 | int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1c0x0); |
| 238 | int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); |
| 239 | int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0x0, va3c0x0); |
| 240 | const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 241 | vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1); |
| 242 | vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1c0x1); |
| 243 | vprod2xCDc0 = vmlal_s8(vprod2xCDc0, vbCDc0x1, va2c0x1); |
| 244 | vprod3xCDc0 = vmlal_s8(vprod3xCDc0, vbCDc0x1, va3c0x1); |
| 245 | vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0); |
| 246 | vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0); |
| 247 | vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0); |
| 248 | vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0); |
| 249 | int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0); |
| 250 | int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1c0x0); |
| 251 | int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0); |
| 252 | int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3c0x0); |
| 253 | const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 254 | vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1); |
| 255 | vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1c0x1); |
| 256 | vprod2xEFc0 = vmlal_s8(vprod2xEFc0, vbEFc0x1, va2c0x1); |
| 257 | vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3c0x1); |
| 258 | vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0); |
| 259 | vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0); |
| 260 | vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0); |
| 261 | vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); |
| 262 | const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0); |
| 263 | const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1); |
| 264 | const int8x8_t va1c1x0 = vreinterpret_s8_s32(va11x0); |
| 265 | const int8x8_t va1c1x1 = vreinterpret_s8_s32(va11x1); |
| 266 | const int8x8_t va2c1x0 = vreinterpret_s8_s32(va21x0); |
| 267 | const int8x8_t va2c1x1 = vreinterpret_s8_s32(va21x1); |
| 268 | const int8x8_t va3c1x0 = vreinterpret_s8_s32(va31x0); |
| 269 | const int8x8_t va3c1x1 = vreinterpret_s8_s32(va31x1); |
| 270 | |
| 271 | int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0); |
| 272 | int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0); |
| 273 | int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2c1x0); |
| 274 | int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3c1x0); |
| 275 | const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 276 | vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1); |
| 277 | vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1); |
| 278 | vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2c1x1); |
| 279 | vprod3x01c1 = vmlal_s8(vprod3x01c1, vb01c1x1, va3c1x1); |
| 280 | vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1); |
| 281 | vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1); |
| 282 | vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1); |
| 283 | vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1); |
| 284 | int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0); |
| 285 | int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0); |
| 286 | int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2c1x0); |
| 287 | int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3c1x0); |
| 288 | const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 289 | vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1); |
| 290 | vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1); |
| 291 | vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2c1x1); |
| 292 | vprod3x23c1 = vmlal_s8(vprod3x23c1, vb23c1x1, va3c1x1); |
| 293 | vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1); |
| 294 | vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1); |
| 295 | vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); |
| 296 | vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1); |
| 297 | int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0); |
| 298 | int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0); |
| 299 | int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2c1x0); |
| 300 | int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3c1x0); |
| 301 | const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 302 | vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1); |
| 303 | vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1); |
| 304 | vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2c1x1); |
| 305 | vprod3x45c1 = vmlal_s8(vprod3x45c1, vb45c1x1, va3c1x1); |
| 306 | vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1); |
| 307 | vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1); |
| 308 | vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1); |
| 309 | vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1); |
| 310 | int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0); |
| 311 | int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0); |
| 312 | int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2c1x0); |
| 313 | int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3c1x0); |
| 314 | const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 315 | vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1); |
| 316 | vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1); |
| 317 | vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2c1x1); |
| 318 | vprod3x67c1 = vmlal_s8(vprod3x67c1, vb67c1x1, va3c1x1); |
| 319 | vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1); |
| 320 | vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1); |
| 321 | vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1); |
| 322 | vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1); |
| 323 | int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0); |
| 324 | int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); |
| 325 | int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2c1x0); |
| 326 | int16x8_t vprod3x89c1 = vmull_s8(vb89c1x0, va3c1x0); |
| 327 | const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 328 | vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1); |
| 329 | vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); |
| 330 | vprod2x89c1 = vmlal_s8(vprod2x89c1, vb89c1x1, va2c1x1); |
| 331 | vprod3x89c1 = vmlal_s8(vprod3x89c1, vb89c1x1, va3c1x1); |
| 332 | vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1); |
| 333 | vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); |
| 334 | vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1); |
| 335 | vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1); |
| 336 | int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0); |
| 337 | int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); |
| 338 | int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2c1x0); |
| 339 | int16x8_t vprod3xABc1 = vmull_s8(vbABc1x0, va3c1x0); |
| 340 | const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 341 | vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1); |
| 342 | vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); |
| 343 | vprod2xABc1 = vmlal_s8(vprod2xABc1, vbABc1x1, va2c1x1); |
| 344 | vprod3xABc1 = vmlal_s8(vprod3xABc1, vbABc1x1, va3c1x1); |
| 345 | vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1); |
| 346 | vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); |
| 347 | vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1); |
| 348 | vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1); |
| 349 | int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0); |
| 350 | int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); |
| 351 | int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2c1x0); |
| 352 | int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3c1x0); |
| 353 | const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 354 | vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1); |
| 355 | vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1c1x1); |
| 356 | vprod2xCDc1 = vmlal_s8(vprod2xCDc1, vbCDc1x1, va2c1x1); |
| 357 | vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3c1x1); |
| 358 | vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1); |
| 359 | vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1); |
| 360 | vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1); |
| 361 | vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); |
| 362 | int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0); |
| 363 | int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1c1x0); |
| 364 | int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2c1x0); |
| 365 | int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1x0, va3c1x0); |
| 366 | const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 367 | vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1); |
| 368 | vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1c1x1); |
| 369 | vprod2xEFc1 = vmlal_s8(vprod2xEFc1, vbEFc1x1, va2c1x1); |
| 370 | vprod3xEFc1 = vmlal_s8(vprod3xEFc1, vbEFc1x1, va3c1x1); |
| 371 | vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1); |
| 372 | vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1); |
| 373 | vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1); |
| 374 | vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1); |
| 375 | |
| 376 | k -= 16 * sizeof(int8_t); |
| 377 | } |
| 378 | |
| 379 | if (k >= 8 * sizeof(int8_t)) { |
| 380 | const int32x2_t va00 = vld1_dup_s32((const void*)a0); |
| 381 | const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8; |
| 382 | const int32x2_t va10 = vld1_dup_s32((const void*)a1); |
| 383 | const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8; |
| 384 | const int32x2_t va20 = vld1_dup_s32((const void*)a2); |
| 385 | const int32x2_t va21 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8; |
| 386 | const int32x2_t va30 = vld1_dup_s32((const void*)a3); |
| 387 | const int32x2_t va31 = vld1_dup_s32((const void*)(a3 + 4)); a3 += 8; |
| 388 | |
| 389 | const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 390 | const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 391 | const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 392 | const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 393 | const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 394 | const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 395 | const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 396 | const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 397 | const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 398 | const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 399 | const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 400 | const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 401 | const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 402 | const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 403 | const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 404 | const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 405 | |
| 406 | const int8x8_t va0c0 = vreinterpret_s8_s32(va00); |
| 407 | const int8x8_t va1c0 = vreinterpret_s8_s32(va10); |
| 408 | const int8x8_t va2c0 = vreinterpret_s8_s32(va20); |
| 409 | const int8x8_t va3c0 = vreinterpret_s8_s32(va30); |
| 410 | |
| 411 | const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0); |
| 412 | const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); |
| 413 | const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0); |
| 414 | const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0); |
| 415 | vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0); |
| 416 | vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); |
| 417 | vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0); |
| 418 | vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0); |
| 419 | const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0); |
| 420 | const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); |
| 421 | const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); |
| 422 | const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0); |
| 423 | vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0); |
| 424 | vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0); |
| 425 | vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); |
| 426 | vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0); |
| 427 | const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0); |
| 428 | const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0); |
| 429 | const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0); |
| 430 | const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); |
| 431 | vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0); |
| 432 | vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0); |
| 433 | vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0); |
| 434 | vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); |
| 435 | const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0); |
| 436 | const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); |
| 437 | const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); |
| 438 | const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0); |
| 439 | vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0); |
| 440 | vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0); |
| 441 | vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); |
| 442 | vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0); |
| 443 | const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0); |
| 444 | const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); |
| 445 | const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); |
| 446 | const int16x8_t vprod3x89c0 = vmull_s8(vb89c0, va3c0); |
| 447 | vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0); |
| 448 | vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0); |
| 449 | vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0); |
| 450 | vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0); |
| 451 | const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); |
| 452 | const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0); |
| 453 | const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); |
| 454 | const int16x8_t vprod3xABc0 = vmull_s8(vbABc0, va3c0); |
| 455 | vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0); |
| 456 | vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0); |
| 457 | vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); |
| 458 | vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0); |
| 459 | const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); |
| 460 | const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0); |
| 461 | const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); |
| 462 | const int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0, va3c0); |
| 463 | vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0); |
| 464 | vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0); |
| 465 | vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0); |
| 466 | vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0); |
| 467 | const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0); |
| 468 | const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0); |
| 469 | const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0); |
| 470 | const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); |
| 471 | vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0); |
| 472 | vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0); |
| 473 | vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0); |
| 474 | vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); |
| 475 | const int8x8_t va0c1 = vreinterpret_s8_s32(va01); |
| 476 | const int8x8_t va1c1 = vreinterpret_s8_s32(va11); |
| 477 | const int8x8_t va2c1 = vreinterpret_s8_s32(va21); |
| 478 | const int8x8_t va3c1 = vreinterpret_s8_s32(va31); |
| 479 | |
| 480 | const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1); |
| 481 | const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1); |
| 482 | const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1); |
| 483 | const int16x8_t vprod3x01c1 = vmull_s8(vb01c1, va3c1); |
| 484 | vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1); |
| 485 | vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1); |
| 486 | vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1); |
| 487 | vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1); |
| 488 | const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1); |
| 489 | const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1); |
| 490 | const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1); |
| 491 | const int16x8_t vprod3x23c1 = vmull_s8(vb23c1, va3c1); |
| 492 | vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1); |
| 493 | vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1); |
| 494 | vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); |
| 495 | vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1); |
| 496 | const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1); |
| 497 | const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1); |
| 498 | const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1); |
| 499 | const int16x8_t vprod3x45c1 = vmull_s8(vb45c1, va3c1); |
| 500 | vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1); |
| 501 | vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1); |
| 502 | vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1); |
| 503 | vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1); |
| 504 | const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1); |
| 505 | const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1); |
| 506 | const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1); |
| 507 | const int16x8_t vprod3x67c1 = vmull_s8(vb67c1, va3c1); |
| 508 | vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1); |
| 509 | vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1); |
| 510 | vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1); |
| 511 | vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1); |
| 512 | const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1); |
| 513 | const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); |
| 514 | const int16x8_t vprod2x89c1 = vmull_s8(vb89c1, va2c1); |
| 515 | const int16x8_t vprod3x89c1 = vmull_s8(vb89c1, va3c1); |
| 516 | vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1); |
| 517 | vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); |
| 518 | vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1); |
| 519 | vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1); |
| 520 | const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1); |
| 521 | const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); |
| 522 | const int16x8_t vprod2xABc1 = vmull_s8(vbABc1, va2c1); |
| 523 | const int16x8_t vprod3xABc1 = vmull_s8(vbABc1, va3c1); |
| 524 | vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1); |
| 525 | vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); |
| 526 | vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1); |
| 527 | vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1); |
| 528 | const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1); |
| 529 | const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); |
| 530 | const int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1, va2c1); |
| 531 | const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); |
| 532 | vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1); |
| 533 | vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1); |
| 534 | vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1); |
| 535 | vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); |
| 536 | const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1); |
| 537 | const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1); |
| 538 | const int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1, va2c1); |
| 539 | const int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1, va3c1); |
| 540 | vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1); |
| 541 | vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1); |
| 542 | vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1); |
| 543 | vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1); |
| 544 | |
| 545 | k -= 8 * sizeof(int8_t); |
| 546 | } |
| 547 | |
| 548 | if XNN_UNLIKELY(k != 0) { |
| 549 | const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k); |
| 550 | const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k); |
| 551 | const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k); |
| 552 | const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k); |
| 553 | |
| 554 | const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 555 | const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 556 | const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 557 | const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 558 | const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 559 | const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 560 | const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 561 | const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 562 | |
| 563 | const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0)); |
| 564 | const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0); |
| 565 | vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0); |
| 566 | const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0); |
| 567 | vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0); |
| 568 | const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0); |
| 569 | vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0); |
| 570 | const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0); |
| 571 | vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0); |
| 572 | const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0); |
| 573 | vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0); |
| 574 | const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); |
| 575 | vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0); |
| 576 | const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); |
| 577 | vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0); |
| 578 | const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0); |
| 579 | vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0); |
| 580 | const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0)); |
| 581 | const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); |
| 582 | vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); |
| 583 | const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); |
| 584 | vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0); |
| 585 | const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0); |
| 586 | vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0); |
| 587 | const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); |
| 588 | vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0); |
| 589 | const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); |
| 590 | vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0); |
| 591 | const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0); |
| 592 | vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0); |
| 593 | const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0); |
| 594 | vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0); |
| 595 | const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0); |
| 596 | vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0); |
| 597 | const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0)); |
| 598 | const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0); |
| 599 | vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0); |
| 600 | const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); |
| 601 | vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); |
| 602 | const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0); |
| 603 | vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0); |
| 604 | const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); |
| 605 | vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); |
| 606 | const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); |
| 607 | vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0); |
| 608 | const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); |
| 609 | vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); |
| 610 | const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); |
| 611 | vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0); |
| 612 | const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0); |
| 613 | vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0); |
| 614 | const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0)); |
| 615 | const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0); |
| 616 | vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0); |
| 617 | const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0); |
| 618 | vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0); |
| 619 | const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); |
| 620 | vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); |
| 621 | const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0); |
| 622 | vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0); |
| 623 | const int16x8_t vprod3x89c0 = vmull_s8(vb89c0, va3c0); |
| 624 | vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0); |
| 625 | const int16x8_t vprod3xABc0 = vmull_s8(vbABc0, va3c0); |
| 626 | vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0); |
| 627 | const int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0, va3c0); |
| 628 | vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0); |
| 629 | const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); |
| 630 | vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); |
| 631 | } |
| 632 | p -= 4 * sizeof(void*); |
| 633 | } while (p != 0); |
| 634 | |
| 635 | #if XNN_ARCH_ARM64 |
| 636 | int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23); |
| 637 | int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67); |
| 638 | int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB); |
| 639 | int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF); |
| 640 | int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23); |
| 641 | int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67); |
| 642 | int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB); |
| 643 | int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF); |
| 644 | int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); |
| 645 | int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67); |
| 646 | int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB); |
| 647 | int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF); |
| 648 | int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23); |
| 649 | int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67); |
| 650 | int32x4_t vacc3x89AB = vpaddq_s32(vacc3x89, vacc3xAB); |
| 651 | int32x4_t vacc3xCDEF = vpaddq_s32(vacc3xCD, vacc3xEF); |
| 652 | #else |
| 653 | const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01)); |
| 654 | const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23)); |
| 655 | int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23); |
| 656 | const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45)); |
| 657 | const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67)); |
| 658 | int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67); |
| 659 | const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89)); |
| 660 | const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB)); |
| 661 | int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB); |
| 662 | const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); |
| 663 | const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF)); |
| 664 | int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); |
| 665 | const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01)); |
| 666 | const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23)); |
| 667 | int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23); |
| 668 | const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45)); |
| 669 | const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67)); |
| 670 | int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67); |
| 671 | const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89)); |
| 672 | const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB)); |
| 673 | int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB); |
| 674 | const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD)); |
| 675 | const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF)); |
| 676 | int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF); |
| 677 | const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01)); |
| 678 | const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); |
| 679 | int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23); |
| 680 | const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45)); |
| 681 | const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67)); |
| 682 | int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67); |
| 683 | const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89)); |
| 684 | const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB)); |
| 685 | int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB); |
| 686 | const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD)); |
| 687 | const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); |
| 688 | int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF); |
| 689 | const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01)); |
| 690 | const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23)); |
| 691 | int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23); |
| 692 | const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45)); |
| 693 | const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67)); |
| 694 | int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67); |
| 695 | const int32x2_t vsum3x89 = vpadd_s32(vget_low_s32(vacc3x89), vget_high_s32(vacc3x89)); |
| 696 | const int32x2_t vsum3xAB = vpadd_s32(vget_low_s32(vacc3xAB), vget_high_s32(vacc3xAB)); |
| 697 | int32x4_t vacc3x89AB = vcombine_s32(vsum3x89, vsum3xAB); |
| 698 | const int32x2_t vsum3xCD = vpadd_s32(vget_low_s32(vacc3xCD), vget_high_s32(vacc3xCD)); |
| 699 | const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); |
| 700 | int32x4_t vacc3xCDEF = vcombine_s32(vsum3xCD, vsum3xEF); |
| 701 | #endif |
| 702 | |
| 703 | const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift); |
| 704 | const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier); |
| 705 | const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift); |
| 706 | |
Marat Dukhan | 8978ac2 | 2021-12-06 13:24:53 -0800 | [diff] [blame] | 707 | vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift); |
| 708 | vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift); |
| 709 | vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift); |
| 710 | vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift); |
| 711 | vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift); |
| 712 | vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift); |
| 713 | vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift); |
| 714 | vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift); |
| 715 | vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift); |
| 716 | vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift); |
| 717 | vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift); |
| 718 | vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift); |
| 719 | vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift); |
| 720 | vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift); |
| 721 | vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift); |
| 722 | vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift); |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 723 | |
| 724 | vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier); |
| 725 | vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier); |
| 726 | vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier); |
| 727 | vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier); |
| 728 | vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier); |
| 729 | vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier); |
| 730 | vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier); |
| 731 | vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier); |
| 732 | vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier); |
| 733 | vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier); |
| 734 | vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier); |
| 735 | vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier); |
| 736 | vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier); |
| 737 | vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier); |
| 738 | vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier); |
| 739 | vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier); |
| 740 | |
| 741 | vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift); |
| 742 | vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift); |
| 743 | vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift); |
| 744 | vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift); |
| 745 | vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift); |
| 746 | vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift); |
| 747 | vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift); |
| 748 | vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift); |
| 749 | vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift); |
| 750 | vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift); |
| 751 | vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift); |
| 752 | vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift); |
| 753 | vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift); |
| 754 | vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift); |
| 755 | vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift); |
| 756 | vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift); |
| 757 | |
| 758 | const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point); |
| 759 | #if XNN_ARCH_ARM64 |
Marat Dukhan | 5a31dc6 | 2021-12-07 00:54:02 -0800 | [diff] [blame] | 760 | int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567); |
| 761 | int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF); |
| 762 | int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567); |
| 763 | int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF); |
| 764 | int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567); |
| 765 | int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF); |
| 766 | int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567); |
| 767 | int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF); |
| 768 | |
| 769 | vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point); |
| 770 | vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point); |
| 771 | vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point); |
| 772 | vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point); |
| 773 | vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point); |
| 774 | vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point); |
| 775 | vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point); |
| 776 | vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point); |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 777 | |
| 778 | int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); |
| 779 | int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); |
| 780 | int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF); |
| 781 | int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF); |
| 782 | #else |
Marat Dukhan | 5a31dc6 | 2021-12-07 00:54:02 -0800 | [diff] [blame] | 783 | int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)); |
| 784 | int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)); |
| 785 | int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)); |
| 786 | int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)); |
| 787 | int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)); |
| 788 | int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)); |
| 789 | int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)); |
| 790 | int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)); |
| 791 | |
| 792 | vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point); |
| 793 | vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point); |
| 794 | vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point); |
| 795 | vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point); |
| 796 | vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point); |
| 797 | vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point); |
| 798 | vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point); |
| 799 | vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point); |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 800 | |
| 801 | int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF)); |
| 802 | int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF)); |
| 803 | int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF)); |
| 804 | int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF)); |
| 805 | #endif |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 806 | |
Marat Dukhan | 03efa0f | 2021-12-07 10:53:39 -0800 | [diff] [blame] | 807 | const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min); |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 808 | vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min); |
Marat Dukhan | 5a31dc6 | 2021-12-07 00:54:02 -0800 | [diff] [blame] | 809 | vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); |
| 810 | vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min); |
| 811 | vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min); |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 812 | |
Marat Dukhan | 5a31dc6 | 2021-12-07 00:54:02 -0800 | [diff] [blame] | 813 | const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max); |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 814 | vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max); |
Marat Dukhan | 5a31dc6 | 2021-12-07 00:54:02 -0800 | [diff] [blame] | 815 | vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); |
| 816 | vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max); |
| 817 | vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max); |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 818 | |
| 819 | if (nc >= 16) { |
| 820 | vst1q_s8(c3 + 0, vout3x0123456789ABCDEF); |
| 821 | vst1q_s8(c2 + 0, vout2x0123456789ABCDEF); |
| 822 | vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); |
| 823 | vst1q_s8(c0 + 0, vout0x0123456789ABCDEF); |
| 824 | |
| 825 | c3 = (int8_t*) ((uintptr_t) c3 + cn_stride); |
| 826 | c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); |
| 827 | c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); |
| 828 | c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 829 | |
| 830 | a = (const int8_t**restrict) ((uintptr_t) a - ks); |
| 831 | |
| 832 | nc -= 16; |
| 833 | } else { |
| 834 | int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF)); |
| 835 | int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); |
| 836 | if (nc & 8) { |
| 837 | vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8; |
| 838 | vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8; |
| 839 | vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8; |
| 840 | vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8; |
| 841 | vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF)); |
| 842 | vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); |
| 843 | } |
| 844 | if (nc & 4) { |
Marat Dukhan | 5f7cf55 | 2021-11-25 17:37:03 -0800 | [diff] [blame] | 845 | vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4; |
| 846 | vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4; |
| 847 | vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4; |
| 848 | vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4; |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 849 | vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4); |
| 850 | vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4); |
| 851 | } |
| 852 | if (nc & 2) { |
Marat Dukhan | 5f7cf55 | 2021-11-25 17:37:03 -0800 | [diff] [blame] | 853 | vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2; |
| 854 | vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2; |
| 855 | vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2; |
| 856 | vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2; |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 857 | vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2); |
| 858 | vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2); |
| 859 | } |
| 860 | if (nc & 1) { |
| 861 | vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8); |
| 862 | vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0); |
| 863 | vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8); |
| 864 | vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0); |
| 865 | } |
| 866 | |
| 867 | nc = 0; |
| 868 | } |
| 869 | } while (nc != 0); |
| 870 | } |