Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
Marat Dukhan | 66ccf64 | 2020-09-28 16:23:42 -0700 | [diff] [blame] | 2 | // Template: src/qs8-gemm/MRxNRc4-neondot.c.in |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2020 Google LLC |
| 6 | // |
| 7 | // This source code is licensed under the BSD-style license found in the |
| 8 | // LICENSE file in the root directory of this source tree. |
| 9 | |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 10 | #include <assert.h> |
| 11 | |
| 12 | #include <arm_neon.h> |
| 13 | |
| 14 | #include <xnnpack/gemm.h> |
Frank Barchard | 6d8ca7d | 2021-03-01 11:05:08 -0800 | [diff] [blame] | 15 | #include <xnnpack/math.h> |
Benoit Jacob | 12c5777 | 2020-08-21 20:11:21 -0700 | [diff] [blame] | 16 | |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 17 | |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 18 | void xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot( |
| 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 | assert(mr != 0); |
| 30 | assert(mr <= 8); |
| 31 | assert(nc != 0); |
| 32 | assert(kc != 0); |
Frank Barchard | 6d8ca7d | 2021-03-01 11:05:08 -0800 | [diff] [blame] | 33 | assert(kc % sizeof(int8_t) == 0); |
| 34 | assert(a != NULL); |
| 35 | assert(w != NULL); |
| 36 | assert(c != NULL); |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 37 | |
Frank Barchard | 6d8ca7d | 2021-03-01 11:05:08 -0800 | [diff] [blame] | 38 | kc = round_up_po2(kc, 4); |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 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 | const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride); |
| 54 | int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride); |
| 55 | if XNN_UNPREDICTABLE(mr < 4) { |
| 56 | a3 = a2; |
| 57 | c3 = c2; |
| 58 | } |
| 59 | const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride); |
| 60 | int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride); |
| 61 | if XNN_UNPREDICTABLE(mr <= 4) { |
| 62 | a4 = a3; |
| 63 | c4 = c3; |
| 64 | } |
| 65 | const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride); |
| 66 | int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride); |
| 67 | if XNN_UNPREDICTABLE(mr < 6) { |
| 68 | a5 = a4; |
| 69 | c5 = c4; |
| 70 | } |
| 71 | const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride); |
| 72 | int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride); |
| 73 | if XNN_UNPREDICTABLE(mr <= 6) { |
| 74 | a6 = a5; |
| 75 | c6 = c5; |
| 76 | } |
| 77 | const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride); |
| 78 | int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride); |
| 79 | if XNN_UNPREDICTABLE(mr != 8) { |
| 80 | a7 = a6; |
| 81 | c7 = c6; |
| 82 | } |
| 83 | |
| 84 | // Loop over groups of 8 columns. |
| 85 | do { |
| 86 | // Initialize accumulators with bias. 8 bias values are loaded from the |
| 87 | // weight matrix, at the start of the group of 8 columns. |
| 88 | int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| 89 | int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| 90 | int32x4_t vacc1x0123 = vacc0x0123; |
| 91 | int32x4_t vacc1x4567 = vacc0x4567; |
| 92 | int32x4_t vacc2x0123 = vacc0x0123; |
| 93 | int32x4_t vacc2x4567 = vacc0x4567; |
| 94 | int32x4_t vacc3x0123 = vacc0x0123; |
| 95 | int32x4_t vacc3x4567 = vacc0x4567; |
| 96 | int32x4_t vacc4x0123 = vacc0x0123; |
| 97 | int32x4_t vacc4x4567 = vacc0x4567; |
| 98 | int32x4_t vacc5x0123 = vacc0x0123; |
| 99 | int32x4_t vacc5x4567 = vacc0x4567; |
| 100 | int32x4_t vacc6x0123 = vacc0x0123; |
| 101 | int32x4_t vacc6x4567 = vacc0x4567; |
| 102 | int32x4_t vacc7x0123 = vacc0x0123; |
| 103 | int32x4_t vacc7x4567 = vacc0x4567; |
| 104 | |
| 105 | // Inner accumulation loop along the 8 columns. |
Benoit Jacob | 12c5777 | 2020-08-21 20:11:21 -0700 | [diff] [blame] | 106 | size_t k = kc; |
| 107 | // 2x partial unrolled loop to load 8 bytes at a time. |
| 108 | while (k >= 8 * sizeof(int8_t)) { |
| 109 | // Load a 8x8 block of activations. |
| 110 | const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; |
| 111 | const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8; |
| 112 | const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8; |
| 113 | const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8; |
| 114 | const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8; |
| 115 | const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8; |
| 116 | const int8x8_t va6x01234567 = vld1_s8(a6); a6 += 8; |
| 117 | const int8x8_t va7x01234567 = vld1_s8(a7); a7 += 8; |
| 118 | |
| 119 | // Load a 8x8 block of weights. |
Frank Barchard | fe14b85 | 2021-02-13 10:14:10 -0800 | [diff] [blame] | 120 | const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
| 121 | const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
| 122 | const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
| 123 | const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
Benoit Jacob | 12c5777 | 2020-08-21 20:11:21 -0700 | [diff] [blame] | 124 | |
| 125 | // Multiply-accumulate: 8x8 * 8x8 --> 8x8. |
| 126 | vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); |
| 127 | vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); |
| 128 | vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0); |
| 129 | vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0); |
| 130 | vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0); |
| 131 | vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0); |
| 132 | vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0); |
| 133 | vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0); |
| 134 | vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0); |
| 135 | vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0); |
| 136 | vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0); |
| 137 | vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0); |
| 138 | vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0); |
| 139 | vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0); |
| 140 | vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0); |
| 141 | vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0); |
| 142 | vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); |
| 143 | vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); |
| 144 | vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); |
| 145 | vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1); |
| 146 | vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); |
| 147 | vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1); |
| 148 | vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); |
| 149 | vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1); |
| 150 | vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1); |
| 151 | vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1); |
| 152 | vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1); |
| 153 | vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1); |
| 154 | vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1); |
| 155 | vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb4567x4567, va6x01234567, 1); |
| 156 | vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1); |
| 157 | vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb4567x4567, va7x01234567, 1); |
| 158 | |
| 159 | k -= 8 * sizeof(int8_t); |
| 160 | } |
Frank Barchard | 618d85d | 2021-03-02 12:59:46 -0800 | [diff] [blame] | 161 | // Handle up to 4 final positions of `k` |
Benoit Jacob | 12c5777 | 2020-08-21 20:11:21 -0700 | [diff] [blame] | 162 | if XNN_UNLIKELY(k != 0) { |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 163 | // Load a 8x4 block of activations. |
Frank Barchard | 618d85d | 2021-03-02 12:59:46 -0800 | [diff] [blame] | 164 | const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4; |
| 165 | const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4; |
| 166 | const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4; |
| 167 | const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4; |
| 168 | const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 4; |
| 169 | const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 4; |
| 170 | const int8x8_t va6x01234567 = vld1_s8(a6); a6 += 4; |
| 171 | const int8x8_t va7x01234567 = vld1_s8(a7); a7 += 4; |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 172 | |
| 173 | // Load a 4x8 block of weights. |
Frank Barchard | fe14b85 | 2021-02-13 10:14:10 -0800 | [diff] [blame] | 174 | const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
| 175 | const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 176 | |
| 177 | // Multiply-accumulate: 8x4 * 4x8 --> 8x8. |
Benoit Jacob | 12c5777 | 2020-08-21 20:11:21 -0700 | [diff] [blame] | 178 | vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); |
| 179 | vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); |
| 180 | vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0); |
| 181 | vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0); |
| 182 | vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0); |
| 183 | vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0); |
| 184 | vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0); |
| 185 | vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0); |
| 186 | vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0); |
| 187 | vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0); |
| 188 | vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0); |
| 189 | vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0); |
| 190 | vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0); |
| 191 | vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0); |
| 192 | vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0); |
| 193 | vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0); |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 194 | } |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 195 | |
| 196 | // Post-accumulation work |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 197 | const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift); |
| 198 | const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0))); |
| 199 | |
| 200 | const int32x4_t vproduct0x0123 = vqrdmulhq_n_s32(vacc0x0123, params->neon.multiplier); |
| 201 | const int32x4_t vproduct0x4567 = vqrdmulhq_n_s32(vacc0x4567, params->neon.multiplier); |
| 202 | const int32x4_t vproduct1x0123 = vqrdmulhq_n_s32(vacc1x0123, params->neon.multiplier); |
| 203 | const int32x4_t vproduct1x4567 = vqrdmulhq_n_s32(vacc1x4567, params->neon.multiplier); |
| 204 | const int32x4_t vproduct2x0123 = vqrdmulhq_n_s32(vacc2x0123, params->neon.multiplier); |
| 205 | const int32x4_t vproduct2x4567 = vqrdmulhq_n_s32(vacc2x4567, params->neon.multiplier); |
| 206 | const int32x4_t vproduct3x0123 = vqrdmulhq_n_s32(vacc3x0123, params->neon.multiplier); |
| 207 | const int32x4_t vproduct3x4567 = vqrdmulhq_n_s32(vacc3x4567, params->neon.multiplier); |
| 208 | const int32x4_t vproduct4x0123 = vqrdmulhq_n_s32(vacc4x0123, params->neon.multiplier); |
| 209 | const int32x4_t vproduct4x4567 = vqrdmulhq_n_s32(vacc4x4567, params->neon.multiplier); |
| 210 | const int32x4_t vproduct5x0123 = vqrdmulhq_n_s32(vacc5x0123, params->neon.multiplier); |
| 211 | const int32x4_t vproduct5x4567 = vqrdmulhq_n_s32(vacc5x4567, params->neon.multiplier); |
| 212 | const int32x4_t vproduct6x0123 = vqrdmulhq_n_s32(vacc6x0123, params->neon.multiplier); |
| 213 | const int32x4_t vproduct6x4567 = vqrdmulhq_n_s32(vacc6x4567, params->neon.multiplier); |
| 214 | const int32x4_t vproduct7x0123 = vqrdmulhq_n_s32(vacc7x0123, params->neon.multiplier); |
| 215 | const int32x4_t vproduct7x4567 = vqrdmulhq_n_s32(vacc7x4567, params->neon.multiplier); |
| 216 | |
| 217 | vacc0x0123 = vsraq_n_s32(vproduct0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31); |
| 218 | vacc0x4567 = vsraq_n_s32(vproduct0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31); |
| 219 | vacc1x0123 = vsraq_n_s32(vproduct1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31); |
| 220 | vacc1x4567 = vsraq_n_s32(vproduct1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31); |
| 221 | vacc2x0123 = vsraq_n_s32(vproduct2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31); |
| 222 | vacc2x4567 = vsraq_n_s32(vproduct2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31); |
| 223 | vacc3x0123 = vsraq_n_s32(vproduct3x0123, vbicq_s32(vacc3x0123, vzero_shift_mask), 31); |
| 224 | vacc3x4567 = vsraq_n_s32(vproduct3x4567, vbicq_s32(vacc3x4567, vzero_shift_mask), 31); |
| 225 | vacc4x0123 = vsraq_n_s32(vproduct4x0123, vbicq_s32(vacc4x0123, vzero_shift_mask), 31); |
| 226 | vacc4x4567 = vsraq_n_s32(vproduct4x4567, vbicq_s32(vacc4x4567, vzero_shift_mask), 31); |
| 227 | vacc5x0123 = vsraq_n_s32(vproduct5x0123, vbicq_s32(vacc5x0123, vzero_shift_mask), 31); |
| 228 | vacc5x4567 = vsraq_n_s32(vproduct5x4567, vbicq_s32(vacc5x4567, vzero_shift_mask), 31); |
| 229 | vacc6x0123 = vsraq_n_s32(vproduct6x0123, vbicq_s32(vacc6x0123, vzero_shift_mask), 31); |
| 230 | vacc6x4567 = vsraq_n_s32(vproduct6x4567, vbicq_s32(vacc6x4567, vzero_shift_mask), 31); |
| 231 | vacc7x0123 = vsraq_n_s32(vproduct7x0123, vbicq_s32(vacc7x0123, vzero_shift_mask), 31); |
| 232 | vacc7x4567 = vsraq_n_s32(vproduct7x4567, vbicq_s32(vacc7x4567, vzero_shift_mask), 31); |
| 233 | |
| 234 | vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift); |
| 235 | vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift); |
| 236 | vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift); |
| 237 | vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift); |
| 238 | vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift); |
| 239 | vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift); |
| 240 | vacc3x0123 = vrshlq_s32(vacc3x0123, vright_shift); |
| 241 | vacc3x4567 = vrshlq_s32(vacc3x4567, vright_shift); |
| 242 | vacc4x0123 = vrshlq_s32(vacc4x0123, vright_shift); |
| 243 | vacc4x4567 = vrshlq_s32(vacc4x4567, vright_shift); |
| 244 | vacc5x0123 = vrshlq_s32(vacc5x0123, vright_shift); |
| 245 | vacc5x4567 = vrshlq_s32(vacc5x4567, vright_shift); |
| 246 | vacc6x0123 = vrshlq_s32(vacc6x0123, vright_shift); |
| 247 | vacc6x4567 = vrshlq_s32(vacc6x4567, vright_shift); |
| 248 | vacc7x0123 = vrshlq_s32(vacc7x0123, vright_shift); |
| 249 | vacc7x4567 = vrshlq_s32(vacc7x4567, vright_shift); |
| 250 | |
| 251 | const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point); |
| 252 | #if XNN_ARCH_ARM64 |
| 253 | const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point); |
| 254 | const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point); |
| 255 | const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point); |
| 256 | const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point); |
| 257 | const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point); |
| 258 | const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point); |
| 259 | const int16x8_t vacc6x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x0123), vacc6x4567), voutput_zero_point); |
| 260 | const int16x8_t vacc7x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x0123), vacc7x4567), voutput_zero_point); |
| 261 | |
| 262 | int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567); |
| 263 | int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567); |
| 264 | int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567); |
| 265 | int8x16_t vout6x01234567_7x01234567 = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc7x01234567); |
| 266 | #else |
| 267 | const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point); |
| 268 | const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point); |
| 269 | const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point); |
| 270 | const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point); |
| 271 | const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point); |
| 272 | const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point); |
| 273 | const int16x8_t vacc6x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x0123), vqmovn_s32(vacc6x4567)), voutput_zero_point); |
| 274 | const int16x8_t vacc7x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc7x0123), vqmovn_s32(vacc7x4567)), voutput_zero_point); |
| 275 | |
| 276 | int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567)); |
| 277 | int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567)); |
| 278 | int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x01234567)); |
| 279 | int8x16_t vout6x01234567_7x01234567 = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc7x01234567)); |
| 280 | #endif |
| 281 | const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min); |
| 282 | const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max); |
| 283 | |
| 284 | vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min); |
| 285 | vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min); |
| 286 | vout4x01234567_5x01234567 = vmaxq_s8(vout4x01234567_5x01234567, voutput_min); |
| 287 | vout6x01234567_7x01234567 = vmaxq_s8(vout6x01234567_7x01234567, voutput_min); |
| 288 | |
| 289 | vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max); |
| 290 | vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max); |
| 291 | vout4x01234567_5x01234567 = vminq_s8(vout4x01234567_5x01234567, voutput_max); |
| 292 | vout6x01234567_7x01234567 = vminq_s8(vout6x01234567_7x01234567, voutput_max); |
| 293 | |
| 294 | if (nc >= 8) { |
| 295 | // Main case where there the 8 columns fit in the destination. |
| 296 | vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567)); |
| 297 | vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567)); |
| 298 | vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567)); |
| 299 | vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567)); |
| 300 | vst1_s8(c4 + 0, vget_low_s8(vout4x01234567_5x01234567)); |
| 301 | vst1_s8(c5 + 0, vget_high_s8(vout4x01234567_5x01234567)); |
| 302 | vst1_s8(c6 + 0, vget_low_s8(vout6x01234567_7x01234567)); |
| 303 | vst1_s8(c7 + 0, vget_high_s8(vout6x01234567_7x01234567)); |
| 304 | |
| 305 | // Advance to the next 8 columns. |
| 306 | c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 307 | c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); |
| 308 | c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); |
| 309 | c3 = (int8_t*) ((uintptr_t) c3 + cn_stride); |
| 310 | c4 = (int8_t*) ((uintptr_t) c4 + cn_stride); |
| 311 | c5 = (int8_t*) ((uintptr_t) c5 + cn_stride); |
| 312 | c6 = (int8_t*) ((uintptr_t) c6 + cn_stride); |
| 313 | c7 = (int8_t*) ((uintptr_t) c7 + cn_stride); |
| 314 | |
Frank Barchard | 6d8ca7d | 2021-03-01 11:05:08 -0800 | [diff] [blame] | 315 | a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 316 | a1 = (const int8_t*) ((uintptr_t) a1 - kc); |
| 317 | a2 = (const int8_t*) ((uintptr_t) a2 - kc); |
| 318 | a3 = (const int8_t*) ((uintptr_t) a3 - kc); |
| 319 | a4 = (const int8_t*) ((uintptr_t) a4 - kc); |
| 320 | a5 = (const int8_t*) ((uintptr_t) a5 - kc); |
| 321 | a6 = (const int8_t*) ((uintptr_t) a6 - kc); |
| 322 | a7 = (const int8_t*) ((uintptr_t) a7 - kc); |
| 323 | |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 324 | nc -= 8; |
| 325 | } else { |
| 326 | // Final case where not all of the 8 columns fit in the destination. |
| 327 | if (nc & 4) { |
| 328 | vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4; |
| 329 | vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4; |
| 330 | vst1q_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4; |
| 331 | vst1q_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4; |
| 332 | vst1q_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4; |
| 333 | vst1q_lane_u32(__builtin_assume_aligned(c5, 1), vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4; |
| 334 | vst1q_lane_u32(__builtin_assume_aligned(c6, 1), vreinterpretq_u32_s8(vout6x01234567_7x01234567), 0); c6 += 4; |
| 335 | vst1q_lane_u32(__builtin_assume_aligned(c7, 1), vreinterpretq_u32_s8(vout6x01234567_7x01234567), 2); c7 += 4; |
| 336 | vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4); |
| 337 | vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4); |
| 338 | vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4); |
| 339 | vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 4); |
| 340 | } |
| 341 | if (nc & 2) { |
| 342 | vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2; |
| 343 | vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2; |
| 344 | vst1q_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2; |
| 345 | vst1q_lane_u16(__builtin_assume_aligned(c3, 1), vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2; |
| 346 | vst1q_lane_u16(__builtin_assume_aligned(c4, 1), vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2; |
| 347 | vst1q_lane_u16(__builtin_assume_aligned(c5, 1), vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2; |
| 348 | vst1q_lane_u16(__builtin_assume_aligned(c6, 1), vreinterpretq_u16_s8(vout6x01234567_7x01234567), 0); c6 += 2; |
| 349 | vst1q_lane_u16(__builtin_assume_aligned(c7, 1), vreinterpretq_u16_s8(vout6x01234567_7x01234567), 4); c7 += 2; |
| 350 | vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2); |
| 351 | vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2); |
| 352 | vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2); |
| 353 | vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 2); |
| 354 | } |
| 355 | if (nc & 1) { |
| 356 | vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0); |
| 357 | vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8); |
| 358 | vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0); |
| 359 | vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8); |
| 360 | vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0); |
| 361 | vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8); |
| 362 | vst1q_lane_s8(c6, vout6x01234567_7x01234567, 0); |
| 363 | vst1q_lane_s8(c7, vout6x01234567_7x01234567, 8); |
| 364 | } |
| 365 | |
| 366 | nc = 0; |
| 367 | } |
| 368 | } while (nc != 0); |
| 369 | } |