Frank Barchard | da78da1 | 2021-03-02 14:28:00 -0800 | [diff] [blame^] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/qs8-igemm/c8-neon-mull-padal.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/igemm.h> |
| 15 | #include <xnnpack/math.h> |
| 16 | |
| 17 | |
| 18 | void xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal( |
| 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, |
| 30 | const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN |
| 31 | { |
| 32 | assert(mr != 0); |
| 33 | assert(mr <= 1); |
| 34 | assert(nc != 0); |
| 35 | assert(kc != 0); |
| 36 | assert(ks != 0); |
| 37 | assert(ks % (1 * 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, 8); |
| 44 | int8_t* c0 = c; |
| 45 | |
| 46 | do { |
| 47 | int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 48 | int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 49 | int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 50 | int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 51 | int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 52 | int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 53 | int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 54 | int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 55 | int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 56 | int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 57 | int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 58 | int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 59 | int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 60 | int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 61 | int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 62 | int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t)); |
| 63 | |
| 64 | size_t p = ks; |
| 65 | do { |
| 66 | const int8_t* restrict a0 = a[0]; |
| 67 | if XNN_UNPREDICTABLE(a0 != zero) { |
| 68 | a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); |
| 69 | } |
| 70 | a += 1; |
| 71 | |
| 72 | size_t k = kc; |
| 73 | // 2x partial unrolled loop to load 16 bytes at a time using MLA. |
| 74 | while (k >= 16 * sizeof(int8_t)) { |
| 75 | const int8x8_t va0x0 = vld1_s8(a0); a0 += 8; |
| 76 | const int8x8_t va0x1 = vld1_s8(a0); a0 += 8; |
| 77 | |
| 78 | const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 79 | const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 80 | const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 81 | const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 82 | const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 83 | const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 84 | const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 85 | const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 86 | const int8x8_t vb8x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 87 | const int8x8_t vb9x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 88 | const int8x8_t vb10x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 89 | const int8x8_t vb11x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 90 | const int8x8_t vb12x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 91 | const int8x8_t vb13x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 92 | const int8x8_t vb14x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 93 | const int8x8_t vb15x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 94 | |
| 95 | const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 96 | int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); |
| 97 | vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); |
| 98 | vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); |
| 99 | const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 100 | int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); |
| 101 | vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); |
| 102 | vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); |
| 103 | const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 104 | int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0); |
| 105 | vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1); |
| 106 | vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2); |
| 107 | const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 108 | int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0); |
| 109 | vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1); |
| 110 | vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3); |
| 111 | const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 112 | int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); |
| 113 | vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); |
| 114 | vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); |
| 115 | const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 116 | int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); |
| 117 | vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); |
| 118 | vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); |
| 119 | const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 120 | int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); |
| 121 | vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); |
| 122 | vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); |
| 123 | const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 124 | int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); |
| 125 | vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); |
| 126 | vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); |
| 127 | const int8x8_t vb8x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 128 | int16x8_t vprod0x8 = vmull_s8(vb8x0, va0x0); |
| 129 | vprod0x8 = vmlal_s8(vprod0x8, vb8x1, va0x1); |
| 130 | vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8); |
| 131 | const int8x8_t vb9x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 132 | int16x8_t vprod0x9 = vmull_s8(vb9x0, va0x0); |
| 133 | vprod0x9 = vmlal_s8(vprod0x9, vb9x1, va0x1); |
| 134 | vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9); |
| 135 | const int8x8_t vb10x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 136 | int16x8_t vprod0x10 = vmull_s8(vb10x0, va0x0); |
| 137 | vprod0x10 = vmlal_s8(vprod0x10, vb10x1, va0x1); |
| 138 | vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10); |
| 139 | const int8x8_t vb11x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 140 | int16x8_t vprod0x11 = vmull_s8(vb11x0, va0x0); |
| 141 | vprod0x11 = vmlal_s8(vprod0x11, vb11x1, va0x1); |
| 142 | vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11); |
| 143 | const int8x8_t vb12x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 144 | int16x8_t vprod0x12 = vmull_s8(vb12x0, va0x0); |
| 145 | vprod0x12 = vmlal_s8(vprod0x12, vb12x1, va0x1); |
| 146 | vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12); |
| 147 | const int8x8_t vb13x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 148 | int16x8_t vprod0x13 = vmull_s8(vb13x0, va0x0); |
| 149 | vprod0x13 = vmlal_s8(vprod0x13, vb13x1, va0x1); |
| 150 | vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13); |
| 151 | const int8x8_t vb14x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 152 | int16x8_t vprod0x14 = vmull_s8(vb14x0, va0x0); |
| 153 | vprod0x14 = vmlal_s8(vprod0x14, vb14x1, va0x1); |
| 154 | vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14); |
| 155 | const int8x8_t vb15x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t)); |
| 156 | int16x8_t vprod0x15 = vmull_s8(vb15x0, va0x0); |
| 157 | vprod0x15 = vmlal_s8(vprod0x15, vb15x1, va0x1); |
| 158 | vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15); |
| 159 | |
| 160 | k -= 16 * sizeof(int8_t); |
| 161 | } |
| 162 | |
| 163 | // Handle 8 bytes at a time using MUL. |
| 164 | if (k > 0) { |
| 165 | const int8x8_t va0 = vld1_s8(a0); a0 += 8; |
| 166 | |
| 167 | const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 168 | const int16x8_t vprod0x0 = vmull_s8(vb0, va0); |
| 169 | vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); |
| 170 | const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 171 | const int16x8_t vprod0x1 = vmull_s8(vb1, va0); |
| 172 | vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); |
| 173 | const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 174 | const int16x8_t vprod0x2 = vmull_s8(vb2, va0); |
| 175 | vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2); |
| 176 | const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 177 | const int16x8_t vprod0x3 = vmull_s8(vb3, va0); |
| 178 | vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3); |
| 179 | const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 180 | const int16x8_t vprod0x4 = vmull_s8(vb4, va0); |
| 181 | vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); |
| 182 | const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 183 | const int16x8_t vprod0x5 = vmull_s8(vb5, va0); |
| 184 | vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); |
| 185 | const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 186 | const int16x8_t vprod0x6 = vmull_s8(vb6, va0); |
| 187 | vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); |
| 188 | const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 189 | const int16x8_t vprod0x7 = vmull_s8(vb7, va0); |
| 190 | vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); |
| 191 | const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 192 | const int16x8_t vprod0x8 = vmull_s8(vb8, va0); |
| 193 | vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8); |
| 194 | const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 195 | const int16x8_t vprod0x9 = vmull_s8(vb9, va0); |
| 196 | vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9); |
| 197 | const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 198 | const int16x8_t vprod0x10 = vmull_s8(vb10, va0); |
| 199 | vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10); |
| 200 | const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 201 | const int16x8_t vprod0x11 = vmull_s8(vb11, va0); |
| 202 | vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11); |
| 203 | const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 204 | const int16x8_t vprod0x12 = vmull_s8(vb12, va0); |
| 205 | vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12); |
| 206 | const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 207 | const int16x8_t vprod0x13 = vmull_s8(vb13, va0); |
| 208 | vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13); |
| 209 | const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 210 | const int16x8_t vprod0x14 = vmull_s8(vb14, va0); |
| 211 | vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14); |
| 212 | const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 213 | const int16x8_t vprod0x15 = vmull_s8(vb15, va0); |
| 214 | vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15); |
| 215 | |
| 216 | k -= 8 * sizeof(int8_t); |
| 217 | } |
| 218 | |
| 219 | p -= 1 * sizeof(void*); |
| 220 | } while (p != 0); |
| 221 | |
| 222 | #if XNN_ARCH_ARM64 |
| 223 | const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); |
| 224 | const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); |
| 225 | const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); |
| 226 | const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); |
| 227 | const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); |
| 228 | const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); |
| 229 | const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); |
| 230 | const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); |
| 231 | int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); |
| 232 | int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); |
| 233 | int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); |
| 234 | int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); |
| 235 | #else |
| 236 | const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); |
| 237 | const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); |
| 238 | const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); |
| 239 | const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); |
| 240 | const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); |
| 241 | const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); |
| 242 | int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); |
| 243 | const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); |
| 244 | const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); |
| 245 | const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); |
| 246 | const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); |
| 247 | const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); |
| 248 | const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); |
| 249 | int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); |
| 250 | const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8)); |
| 251 | const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); |
| 252 | const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10)); |
| 253 | const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); |
| 254 | const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); |
| 255 | const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); |
| 256 | int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); |
| 257 | const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); |
| 258 | const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13)); |
| 259 | const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14)); |
| 260 | const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15)); |
| 261 | const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); |
| 262 | const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); |
| 263 | int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); |
| 264 | #endif |
| 265 | |
| 266 | const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->neon.multiplier); |
| 267 | vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier); |
| 268 | vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier); |
| 269 | vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier); |
| 270 | vacc0xCDEF = vqrdmulhq_s32(vacc0xCDEF, vmultiplier); |
| 271 | |
| 272 | const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift); |
| 273 | const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0))); |
| 274 | vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31); |
| 275 | vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31); |
| 276 | vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31); |
| 277 | vacc0xCDEF = vsraq_n_s32(vacc0xCDEF, vbicq_s32(vacc0xCDEF, vzero_shift_mask), 31); |
| 278 | |
| 279 | vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift); |
| 280 | vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift); |
| 281 | vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift); |
| 282 | vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_shift); |
| 283 | |
| 284 | const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point); |
| 285 | #if XNN_ARCH_ARM64 |
| 286 | const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point); |
| 287 | const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point); |
| 288 | int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); |
| 289 | #else |
| 290 | const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point); |
| 291 | const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point); |
| 292 | |
| 293 | int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF)); |
| 294 | #endif |
| 295 | const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min); |
| 296 | const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max); |
| 297 | |
| 298 | vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min); |
| 299 | |
| 300 | vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max); |
| 301 | |
| 302 | if (nc >= 16) { |
| 303 | vst1q_s8(c0 + 0, vout0x0123456789ABCDEF); |
| 304 | |
| 305 | c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 306 | |
| 307 | a = (const int8_t**restrict) ((uintptr_t) a - ks); |
| 308 | |
| 309 | nc -= 16; |
| 310 | } else { |
| 311 | int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF); |
| 312 | if (nc & 8) { |
| 313 | vst1_s8(c0, vout0x01234567); c0 += 8; |
| 314 | vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF); |
| 315 | } |
| 316 | if (nc & 4) { |
| 317 | vst1_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4; |
| 318 | vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4); |
| 319 | } |
| 320 | if (nc & 2) { |
| 321 | vst1_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2; |
| 322 | vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2); |
| 323 | } |
| 324 | if (nc & 1) { |
| 325 | vst1_lane_s8(c0, vout0x01234567, 0); |
| 326 | } |
| 327 | |
| 328 | nc = 0; |
| 329 | } |
| 330 | } while (nc != 0); |
| 331 | } |