1// Copyright 2021 Google LLC 2// 3// This source code is licensed under the BSD-style license found in the 4// LICENSE file in the root directory of this source tree. 5 6$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 7$assert NR % 8 == 0 8$assert 8 <= NR <= 16 9#include <assert.h> 10 11#include <arm_neon.h> 12 13#include <xnnpack/gemm.h> 14#include <xnnpack/math.h> 15 16 17void xnn_qs8_igemm_minmax_ukernel_${MR}x${NR}c2__neon_${"mlal" if MLA else "mull"}_padal_dup( 18 size_t mr, 19 size_t nc, 20 size_t kc, 21 size_t ks, 22 const int8_t** restrict a, 23 const void* restrict w, 24 int8_t* restrict c, 25 size_t cm_stride, 26 size_t cn_stride, 27 size_t a_offset, 28 const int8_t* zero, 29 const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN 30{ 31 assert(mr != 0); 32 assert(mr <= ${MR}); 33 assert(nc != 0); 34 assert(kc != 0); 35 assert(ks != 0); 36 assert(ks % (${MR} * sizeof(void*)) == 0); 37 assert(a_offset % sizeof(int8_t) == 0); 38 assert(a != NULL); 39 assert(w != NULL); 40 assert(c != NULL); 41 42 kc = round_up_po2(kc, 2); 43 int8_t* c0 = c; 44 $for M in range(1, MR): 45 int8_t* c${M} = (int8_t*) ((uintptr_t) c${M-1} + cm_stride); 46 $if M % 2 == 0: 47 if XNN_UNPREDICTABLE(mr <= ${M}) { 48 c${M} = c${M-1}; 49 } 50 $elif M + 1 == MR: 51 if XNN_UNPREDICTABLE(mr != ${M+1}) { 52 c${M} = c${M-1}; 53 } 54 $else: 55 if XNN_UNPREDICTABLE(mr < ${M+1}) { 56 c${M} = c${M-1}; 57 } 58 59 do { 60 $for N in range(0, NR, 4): 61 int32x4_t vacc0x${ABC[N:N+4]} = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); 62 $for M in range(1, MR): 63 $for N in range(0, NR, 4): 64 int32x4_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]}; 65 66 size_t p = ks; 67 do { 68 $for M in range(MR): 69 const int8_t* restrict a${M} = a[${M}]; 70 if XNN_UNPREDICTABLE(a${M} != zero) { 71 a${M} = (const int8_t*) ((uintptr_t) a${M} + a_offset); 72 } 73 a += ${MR}; 74 75 size_t k = kc; 76 77 $if MLA: 78 while (k >= 16 * sizeof(int8_t)) { 79 $for M in range(MR): 80 const int8x8_t va${M}x0 = vld1_s8(a${M}); a${M} += 8; 81 const int8x8_t va${M}x1 = vld1_s8(a${M}); a${M} += 8; 82 83 $for K in range(4): 84 $for N in range(0, NR, 4): 85 const int8x8_t vb${ABC[N:N+4]}c${K}x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); 86 87 $for K in range(4): 88 $for N in range(0, NR, 4): 89 $for M in range(MR): 90 int16x8_t vprod${M}x${ABC[N:N+4]}c${K} = vmull_s8(vb${ABC[N:N+4]}c${K}x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va${M}x0), ${K}))); 91 const int8x8_t vb${ABC[N:N+4]}c${K}x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); 92 $for M in range(MR): 93 vprod${M}x${ABC[N:N+4]}c${K} = vmlal_s8(vprod${M}x${ABC[N:N+4]}c${K}, vb${ABC[N:N+4]}c${K}x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va${M}x1), ${K}))); 94 $for M in range(MR): 95 vacc${M}x${ABC[N:N+4]} = vpadalq_s16(vacc${M}x${ABC[N:N+4]}, vprod${M}x${ABC[N:N+4]}c${K}); 96 97 k -= 16 * sizeof(int8_t); 98 } 99 100 ${"if" if MLA else "while"} (k >= 8 * sizeof(int8_t)) { 101 $for M in range(MR): 102 const int8x8_t va${M} = vld1_s8(a${M}); a${M} += 8; 103 104 $for K in range(4): 105 $for N in range(0, NR, 4): 106 const int8x8_t vb${ABC[N:N+4]}c${K} = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); 107 108 $for M in range(MR): 109 $for N in range(0, NR, 4): 110 $for K in range(4): 111 const int16x8_t vprod${M}x${ABC[N:N+4]}c${K} = vmull_s8(vb${ABC[N:N+4]}c${K}, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va${M}), ${K}))); 112 $for K in range(4): 113 vacc${M}x${ABC[N:N+4]} = vpadalq_s16(vacc${M}x${ABC[N:N+4]}, vprod${M}x${ABC[N:N+4]}c${K}); 114 115 k -= 8 * sizeof(int8_t); 116 } 117 118 if XNN_UNLIKELY(k != 0) { 119 $for M in range(MR): 120 const int8x8_t va${M} = vld1_s8(a${M}); a${M} = (const int8_t*) ((uintptr_t) a${M} + k); 121 122 $for N in range(0, NR, 4): 123 const int8x8_t vb${ABC[N:N+4]}c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); 124 125 $for M in range(MR): 126 $for N in range(0, NR, 4): 127 const int16x8_t vprod${M}x${ABC[N:N+4]}c0 = vmull_s8(vb${ABC[N:N+4]}c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va${M}), 0))); 128 vacc${M}x${ABC[N:N+4]} = vpadalq_s16(vacc${M}x${ABC[N:N+4]}, vprod${M}x${ABC[N:N+4]}c0); 129 130 if (k > 2 * sizeof(int8_t)) { 131 $for N in range(0, NR, 4): 132 const int8x8_t vb${ABC[N:N+4]}c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); 133 134 $for M in range(MR): 135 $for N in range(0, NR, 4): 136 const int16x8_t vprod${M}x${ABC[N:N+4]}c1 = vmull_s8(vb${ABC[N:N+4]}c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va${M}), 1))); 137 vacc${M}x${ABC[N:N+4]} = vpadalq_s16(vacc${M}x${ABC[N:N+4]}, vprod${M}x${ABC[N:N+4]}c1); 138 139 if (k > 4 * sizeof(int8_t)) { 140 $for N in range(0, NR, 4): 141 const int8x8_t vb${ABC[N:N+4]}c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); 142 143 $for M in range(MR): 144 $for N in range(0, NR, 4): 145 const int16x8_t vprod${M}x${ABC[N:N+4]}c2 = vmull_s8(vb${ABC[N:N+4]}c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va${M}), 2))); 146 vacc${M}x${ABC[N:N+4]} = vpadalq_s16(vacc${M}x${ABC[N:N+4]}, vprod${M}x${ABC[N:N+4]}c2); 147 } 148 } 149 } 150 p -= ${MR} * sizeof(void*); 151 } while (p != 0); 152 153 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->neon.multiplier); 154 $for M in range(MR): 155 $for N in range(0, NR, 4): 156 vacc${M}x${ABC[N:N+4]} = vqrdmulhq_s32(vacc${M}x${ABC[N:N+4]}, vmultiplier); 157 158 const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift); 159 const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0))); 160 $for M in range(MR): 161 $for N in range(0, NR, 4): 162 vacc${M}x${ABC[N:N+4]} = vsraq_n_s32(vacc${M}x${ABC[N:N+4]}, vbicq_s32(vacc${M}x${ABC[N:N+4]}, vzero_shift_mask), 31); 163 164 $for M in range(MR): 165 $for N in range(0, NR, 4): 166 vacc${M}x${ABC[N:N+4]} = vrshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_shift); 167 168 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point); 169#if XNN_ARCH_ARM64 170 $for M in range(MR): 171 $for N in range(0, NR, 8): 172 const int16x8_t vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vacc${M}x${ABC[N+4:N+8]}), voutput_zero_point); 173 174 $for M in range(MR): 175 $for N in range(0, NR, 16): 176 $if N + 8 < NR: 177 int8x16_t vout${M}x${ABC[N:N+16]} = vqmovn_high_s16(vqmovn_s16(vacc${M}x${ABC[N:N+8]}), vacc${M}x${ABC[N+8:N+16]}); 178 $elif M % 2 == 1: 179 int8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vqmovn_high_s16(vqmovn_s16(vacc${M-1}x${ABC[N:N+8]}), vacc${M}x${ABC[N:N+8]}); 180 $elif M + 1 == MR: 181 int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]}); 182#else 183 $for M in range(MR): 184 $for N in range(0, NR, 8): 185 const int16x8_t vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vqmovn_s32(vacc${M}x${ABC[N+4:N+8]})), voutput_zero_point); 186 187 $for M in range(MR): 188 $for N in range(0, NR, 16): 189 $if N + 8 < NR: 190 int8x16_t vout${M}x${ABC[N:N+16]} = vcombine_s8(vqmovn_s16(vacc${M}x${ABC[N:N+8]}), vqmovn_s16(vacc${M}x${ABC[N+8:N+16]})); 191 $elif M % 2 == 1: 192 int8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vcombine_s8(vqmovn_s16(vacc${M-1}x${ABC[N:N+8]}), vqmovn_s16(vacc${M}x${ABC[N:N+8]})); 193 $elif M + 1 == MR: 194 int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]}); 195#endif 196 $if NR == 8 and MR == 1: 197 const int8x8_t voutput_min = vld1_dup_s8(¶ms->neon.output_min); 198 const int8x8_t voutput_max = vld1_dup_s8(¶ms->neon.output_max); 199 $else: 200 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min); 201 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max); 202 203 $for M in reversed(range(MR)): 204 $for N in range(0, NR, 16): 205 $if N + 8 < NR: 206 vout${M}x${ABC[N:N+16]} = vmaxq_s8(vout${M}x${ABC[N:N+16]}, voutput_min); 207 $elif M % 2 == 1: 208 vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vmaxq_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}, voutput_min); 209 $elif M + 1 == MR: 210 $if NR == 8 and MR == 1: 211 vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, voutput_min); 212 $else: 213 vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_min)); 214 215 $for M in reversed(range(MR)): 216 $for N in range(0, NR, 16): 217 $if N + 8 < NR: 218 vout${M}x${ABC[N:N+16]} = vminq_s8(vout${M}x${ABC[N:N+16]}, voutput_max); 219 $elif M % 2 == 1: 220 vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vminq_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}, voutput_max); 221 $elif M + 1 == MR: 222 $if NR == 8 and MR == 1: 223 vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, voutput_max); 224 $else: 225 vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_max)); 226 227 if (nc >= ${NR}) { 228 $for M in reversed(range(MR)): 229 $for N in range(0, NR, 16): 230 $if N + 8 < NR: 231 vst1q_s8(c${M} + ${N}, vout${M}x${ABC[N:N+16]}); 232 $elif M % 2 == 1: 233 vst1_s8(c${M} + ${N}, vget_high_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]})); 234 vst1_s8(c${M-1} + ${N}, vget_low_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]})); 235 $elif M + 1 == MR: 236 vst1_s8(c${M} + ${N}, vout${M}x${ABC[N:N+8]}); 237 238 $for M in reversed(range(MR)): 239 c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride); 240 241 a = (const int8_t**restrict) ((uintptr_t) a - ks); 242 243 nc -= ${NR}; 244 } else { 245 $if NR == 16: 246 $for M in range(MR): 247 $if M % 2 == 1: 248 int8x16_t vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_low_s8(vout${M-1}x0123456789ABCDEF), vget_low_s8(vout${M}x0123456789ABCDEF)); 249 $elif M + 1 == MR: 250 int8x8_t vout${M}x01234567 = vget_low_s8(vout${M}x0123456789ABCDEF); 251 if (nc & 8) { 252 $for M in reversed(range(MR)): 253 $if M % 2 == 1: 254 vst1_s8(c${M}, vget_high_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]})); c${M} += 8; 255 vst1_s8(c${M-1}, vget_low_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]})); c${M-1} += 8; 256 $elif M + 1 == MR: 257 vst1_s8(c${M}, vout${M}x${ABC[N:N+8]}); c${M} += 8; 258 $for M in reversed(range(MR)): 259 $if M % 2 == 1: 260 vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_high_s8(vout${M-1}x0123456789ABCDEF), vget_high_s8(vout${M}x0123456789ABCDEF)); 261 $elif M + 1 == MR: 262 vout${M}x01234567 = vget_high_s8(vout${M}x0123456789ABCDEF); 263 } 264 if (nc & 4) { 265 $for M in reversed(range(MR)): 266 $if M % 2 == 1: 267 vst1q_lane_u32(__builtin_assume_aligned(c${M}, 1), vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 2); c${M} += 4; 268 vst1q_lane_u32(__builtin_assume_aligned(c${M-1}, 1), vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 4; 269 $elif M + 1 == MR: 270 vst1_lane_u32(__builtin_assume_aligned(c${M}, 1), vreinterpret_u32_s8(vout${M}x01234567), 0); c${M} += 4; 271 $for M in reversed(range(MR)): 272 $if M % 2 == 1: 273 vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 4); 274 $elif M + 1 == MR: 275 vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 4); 276 } 277 if (nc & 2) { 278 $for M in reversed(range(MR)): 279 $if M % 2 == 1: 280 vst1q_lane_u16(__builtin_assume_aligned(c${M}, 1), vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 4); c${M} += 2; 281 vst1q_lane_u16(__builtin_assume_aligned(c${M-1}, 1), vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 2; 282 $elif M + 1 == MR: 283 vst1_lane_u16(__builtin_assume_aligned(c${M}, 1), vreinterpret_u16_s8(vout${M}x01234567), 0); c${M} += 2; 284 $for M in reversed(range(MR)): 285 $if M % 2 == 1: 286 vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 2); 287 $elif M + 1 == MR: 288 vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 2); 289 } 290 if (nc & 1) { 291 $for M in reversed(range(MR)): 292 $if M % 2 == 1: 293 vst1q_lane_s8(c${M}, vout${M-1}x01234567_${M}x01234567, 8); 294 vst1q_lane_s8(c${M-1}, vout${M-1}x01234567_${M}x01234567, 0); 295 $elif M + 1 == MR: 296 vst1_lane_s8(c${M}, vout${M}x01234567, 0); 297 } 298 299 nc = 0; 300 } 301 } while (nc != 0); 302} 303