Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/qs8-gemm/c2-neon-mull-padal-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/common.h> |
| 15 | #include <xnnpack/gemm.h> |
| 16 | |
| 17 | |
| 18 | void xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup( |
| 19 | size_t mr, |
| 20 | size_t nc, |
| 21 | size_t kc, |
| 22 | const int8_t* restrict a, |
| 23 | size_t a_stride, |
| 24 | const void* restrict w, |
| 25 | int8_t* restrict c, |
| 26 | size_t cm_stride, |
| 27 | size_t cn_stride, |
| 28 | const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN |
| 29 | { |
| 30 | assert(mr != 0); |
| 31 | assert(mr <= 3); |
| 32 | assert(nc != 0); |
| 33 | assert(kc != 0); |
| 34 | assert(kc % sizeof(int8_t) == 0); |
| 35 | assert(a != NULL); |
| 36 | assert(w != NULL); |
| 37 | assert(c != NULL); |
| 38 | |
| 39 | const int8_t* a0 = a; |
| 40 | int8_t* c0 = c; |
| 41 | const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); |
| 42 | int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); |
| 43 | if XNN_UNPREDICTABLE(mr < 2) { |
| 44 | a1 = a0; |
| 45 | c1 = c0; |
| 46 | } |
| 47 | const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride); |
| 48 | int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride); |
| 49 | if XNN_UNPREDICTABLE(mr <= 2) { |
| 50 | a2 = a1; |
| 51 | c2 = c1; |
| 52 | } |
| 53 | |
| 54 | do { |
| 55 | int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| 56 | int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| 57 | int32x4_t vacc1x0123 = vacc0x0123; |
| 58 | int32x4_t vacc1x4567 = vacc0x4567; |
| 59 | int32x4_t vacc2x0123 = vacc0x0123; |
| 60 | int32x4_t vacc2x4567 = vacc0x4567; |
| 61 | |
| 62 | size_t k = kc; |
Frank Barchard | 7e1f371 | 2021-02-05 11:52:23 -0800 | [diff] [blame^] | 63 | |
| 64 | |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 65 | while (k >= 8 * sizeof(int8_t)) { |
| 66 | const int8x8_t va0 = vld1_s8(a0); a0 += 8; |
| 67 | const int8x8_t va1 = vld1_s8(a1); a1 += 8; |
| 68 | const int8x8_t va2 = vld1_s8(a2); a2 += 8; |
| 69 | |
| 70 | const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 71 | const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 72 | const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 73 | const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 74 | const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 75 | const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 76 | const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 77 | const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 78 | |
Frank Barchard | 7e1f371 | 2021-02-05 11:52:23 -0800 | [diff] [blame^] | 79 | const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0))); |
| 80 | const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1))); |
| 81 | const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2))); |
| 82 | const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3))); |
| 83 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0); |
| 84 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1); |
| 85 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2); |
| 86 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3); |
| 87 | const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0))); |
| 88 | const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1))); |
| 89 | const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2))); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 90 | const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3))); |
Frank Barchard | 7e1f371 | 2021-02-05 11:52:23 -0800 | [diff] [blame^] | 91 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0); |
| 92 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); |
| 93 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 94 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3); |
Frank Barchard | 7e1f371 | 2021-02-05 11:52:23 -0800 | [diff] [blame^] | 95 | const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0))); |
| 96 | const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1))); |
| 97 | const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2))); |
| 98 | const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3))); |
| 99 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0); |
| 100 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1); |
| 101 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2); |
| 102 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3); |
| 103 | const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0))); |
| 104 | const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1))); |
| 105 | const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2))); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 106 | const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3))); |
Frank Barchard | 7e1f371 | 2021-02-05 11:52:23 -0800 | [diff] [blame^] | 107 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0); |
| 108 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1); |
| 109 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 110 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3); |
Frank Barchard | 7e1f371 | 2021-02-05 11:52:23 -0800 | [diff] [blame^] | 111 | const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0))); |
| 112 | const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1))); |
| 113 | const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2))); |
| 114 | const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3))); |
| 115 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0); |
| 116 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); |
| 117 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2); |
| 118 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3); |
| 119 | const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0))); |
| 120 | const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1))); |
| 121 | const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2))); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 122 | const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3))); |
Frank Barchard | 7e1f371 | 2021-02-05 11:52:23 -0800 | [diff] [blame^] | 123 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0); |
| 124 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1); |
| 125 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2); |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 126 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3); |
| 127 | |
| 128 | k -= 8 * sizeof(int8_t); |
| 129 | } |
Frank Barchard | 7e1f371 | 2021-02-05 11:52:23 -0800 | [diff] [blame^] | 130 | |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 131 | if XNN_UNLIKELY(k != 0) { |
| 132 | const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k); |
| 133 | const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k); |
| 134 | const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k); |
| 135 | |
| 136 | const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 137 | const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 138 | |
| 139 | const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0))); |
| 140 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0); |
| 141 | const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0))); |
| 142 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0); |
| 143 | const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0))); |
| 144 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0); |
| 145 | const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0))); |
| 146 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0); |
| 147 | const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0))); |
| 148 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0); |
| 149 | const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0))); |
| 150 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0); |
| 151 | |
| 152 | if (k > 2 * sizeof(int8_t)) { |
| 153 | const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 154 | const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 155 | |
| 156 | const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1))); |
| 157 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1); |
| 158 | const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1))); |
| 159 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); |
| 160 | const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1))); |
| 161 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1); |
| 162 | const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1))); |
| 163 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1); |
| 164 | const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1))); |
| 165 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); |
| 166 | const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1))); |
| 167 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1); |
| 168 | |
| 169 | if (k > 4 * sizeof(int8_t)) { |
| 170 | const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 171 | const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 172 | |
| 173 | const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2))); |
| 174 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2); |
| 175 | const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2))); |
| 176 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2); |
| 177 | const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2))); |
| 178 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2); |
| 179 | const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2))); |
| 180 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2); |
| 181 | const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2))); |
| 182 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2); |
| 183 | const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2))); |
| 184 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2); |
| 185 | |
| 186 | if (k > 6 * sizeof(int8_t)) { |
| 187 | const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 188 | const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| 189 | |
| 190 | const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3))); |
| 191 | vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3); |
| 192 | const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3))); |
| 193 | vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3); |
| 194 | const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3))); |
| 195 | vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3); |
| 196 | const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3))); |
| 197 | vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3); |
| 198 | const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3))); |
| 199 | vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3); |
| 200 | const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3))); |
| 201 | vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3); |
| 202 | } |
| 203 | } |
| 204 | } |
| 205 | } |
| 206 | const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->neon.multiplier); |
| 207 | vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier); |
| 208 | vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier); |
| 209 | vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier); |
| 210 | vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier); |
| 211 | vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier); |
| 212 | vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier); |
| 213 | |
| 214 | const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift); |
| 215 | const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0))); |
| 216 | vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31); |
| 217 | vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31); |
| 218 | vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31); |
| 219 | vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31); |
| 220 | vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31); |
| 221 | vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31); |
| 222 | |
| 223 | vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift); |
| 224 | vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift); |
| 225 | vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift); |
| 226 | vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift); |
| 227 | vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift); |
| 228 | vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift); |
| 229 | |
| 230 | const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point); |
| 231 | #if XNN_ARCH_ARM64 |
| 232 | const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point); |
| 233 | const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point); |
| 234 | const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point); |
| 235 | |
| 236 | int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567); |
| 237 | int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567); |
| 238 | #else |
| 239 | const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point); |
| 240 | const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point); |
| 241 | const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point); |
| 242 | |
| 243 | int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567)); |
| 244 | int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567); |
| 245 | #endif |
| 246 | const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min); |
| 247 | const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max); |
| 248 | |
| 249 | vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min); |
| 250 | vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min)); |
| 251 | |
| 252 | vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max); |
| 253 | vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max)); |
| 254 | |
| 255 | if (nc >= 8) { |
| 256 | vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567)); |
| 257 | vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567)); |
| 258 | vst1_s8(c2 + 0, vout2x01234567); |
| 259 | |
| 260 | c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 261 | c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); |
| 262 | c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); |
| 263 | |
| 264 | a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 265 | a1 = (const int8_t*) ((uintptr_t) a1 - kc); |
| 266 | a2 = (const int8_t*) ((uintptr_t) a2 - kc); |
| 267 | |
| 268 | nc -= 8; |
| 269 | } else { |
| 270 | if (nc & 4) { |
| 271 | vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4; |
| 272 | vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4; |
| 273 | vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4; |
| 274 | vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4); |
| 275 | vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4); |
| 276 | } |
| 277 | if (nc & 2) { |
| 278 | vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2; |
| 279 | vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2; |
| 280 | vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2; |
| 281 | vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2); |
| 282 | vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2); |
| 283 | } |
| 284 | if (nc & 1) { |
| 285 | vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0); |
| 286 | vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8); |
| 287 | vst1_lane_s8(c2, vout2x01234567, 0); |
| 288 | } |
| 289 | |
| 290 | nc = 0; |
| 291 | } |
| 292 | } while (nc != 0); |
| 293 | } |