1// Copyright 2020 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$assert 8 <= NR <= 16 9#include <assert.h> 10 11#include <arm_neon.h> 12 13#include <xnnpack/igemm.h> 14#include <xnnpack/math.h> 15 16 17void xnn_qs8_igemm_minmax_ukernel_${MR}x${NR}c4__neondot( 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, 4); 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 // Inner accumulation loop along the ${NR} columns. 76 size_t k = kc; 77 // 2x partial unrolled loop to load 8 bytes at a time. 78 while (k >= 8 * sizeof(int8_t)) { 79 // Load a ${MR}x8 block of activations. 80 $for M in range(MR): 81 const int8x8_t va${M}x01234567 = vld1_s8(a${M}); a${M} += 8; 82 83 // Load a 8x${NR} block of weights. 84 $for K in range(0, 8, 4): 85 $for N in range(0, NR, 4): 86 const int8x16_t vb${ABC[K:K+4]}x${ABC[N:N+4]} = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); 87 88 // Multiply-accumulate: ${MR}x8 * 8x${NR} --> ${MR}x${NR}. 89 $for K in range(0, 8, 4): 90 $for M in range(MR): 91 $for N in range(0, NR, 4): 92 vacc${M}x${ABC[N:N+4]} = vdotq_lane_s32(vacc${M}x${ABC[N:N+4]}, vb${ABC[K:K+4]}x${ABC[N:N+4]}, va${M}x01234567, ${K/4}); 93 94 k -= 8 * sizeof(int8_t); 95 } 96 // Handle up to 4 final positions of `k` 97 if XNN_UNLIKELY(k != 0) { 98 // Load a ${MR}x4 block of activations. 99 $for M in range(MR): 100 const int8x8_t va${M}x01234567 = vld1_s8(a${M}); 101 102 // Load a 4x${NR} block of weights. 103 $for N in range(0, NR, 4): 104 const int8x16_t vb0123x${ABC[N:N+4]} = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); 105 106 // Multiply-accumulate: ${MR}x4 * 4x${NR} --> ${MR}x${NR}. 107 $for M in range(MR): 108 $for N in range(0, NR, 4): 109 vacc${M}x${ABC[N:N+4]} = vdotq_lane_s32(vacc${M}x${ABC[N:N+4]}, vb0123x${ABC[N:N+4]}, va${M}x01234567, 0); 110 } 111 p -= ${MR} * sizeof(void*); 112 } while (p != 0); 113 114 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->neon.multiplier); 115 $for M in range(MR): 116 $for N in range(0, NR, 4): 117 vacc${M}x${ABC[N:N+4]} = vqrdmulhq_s32(vacc${M}x${ABC[N:N+4]}, vmultiplier); 118 119 const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift); 120 const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0))); 121 $for M in range(MR): 122 $for N in range(0, NR, 4): 123 vacc${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); 124 125 $for M in range(MR): 126 $for N in range(0, NR, 4): 127 vacc${M}x${ABC[N:N+4]} = vrshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_shift); 128 129 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point); 130#if XNN_ARCH_ARM64 131 $for M in range(MR): 132 $for N in range(0, NR, 8): 133 const int16x8_t vacc${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); 134 135 $for M in range(MR): 136 $for N in range(0, NR, 16): 137 $if N + 8 < NR: 138 int8x16_t vout${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]}); 139 $elif M % 2 == 1: 140 int8x16_t vout${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]}); 141 $elif M + 1 == MR: 142 int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]}); 143#else 144 $for M in range(MR): 145 $for N in range(0, NR, 8): 146 const int16x8_t vacc${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); 147 148 $for M in range(MR): 149 $for N in range(0, NR, 16): 150 $if N + 8 < NR: 151 int8x16_t vout${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]})); 152 $elif M % 2 == 1: 153 int8x16_t vout${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]})); 154 $elif M + 1 == MR: 155 int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]}); 156#endif 157 $if NR == 8 and MR == 1: 158 const int8x8_t voutput_min = vld1_dup_s8(¶ms->neon.output_min); 159 const int8x8_t voutput_max = vld1_dup_s8(¶ms->neon.output_max); 160 $else: 161 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min); 162 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max); 163 164 $for M in reversed(range(MR)): 165 $for N in range(0, NR, 16): 166 $if N + 8 < NR: 167 vout${M}x${ABC[N:N+16]} = vmaxq_s8(vout${M}x${ABC[N:N+16]}, voutput_min); 168 $elif M % 2 == 1: 169 vout${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); 170 $elif M + 1 == MR: 171 $if NR == 8 and MR == 1: 172 vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, voutput_min); 173 $else: 174 vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_min)); 175 176 $for M in reversed(range(MR)): 177 $for N in range(0, NR, 16): 178 $if N + 8 < NR: 179 vout${M}x${ABC[N:N+16]} = vminq_s8(vout${M}x${ABC[N:N+16]}, voutput_max); 180 $elif M % 2 == 1: 181 vout${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); 182 $elif M + 1 == MR: 183 $if NR == 8 and MR == 1: 184 vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, voutput_max); 185 $else: 186 vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_max)); 187 188 if (nc >= ${NR}) { 189 $for M in reversed(range(MR)): 190 $for N in range(0, NR, 16): 191 $if N + 8 < NR: 192 vst1q_s8(c${M} + ${N}, vout${M}x${ABC[N:N+16]}); 193 $elif M % 2 == 1: 194 vst1_s8(c${M} + ${N}, vget_high_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]})); 195 vst1_s8(c${M-1} + ${N}, vget_low_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]})); 196 $elif M + 1 == MR: 197 vst1_s8(c${M} + ${N}, vout${M}x${ABC[N:N+8]}); 198 199 $for M in reversed(range(MR)): 200 c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride); 201 202 a = (const int8_t**restrict) ((uintptr_t) a - ks); 203 204 nc -= ${NR}; 205 } else { 206 $if NR == 16: 207 $for M in range(MR): 208 $if M % 2 == 1: 209 int8x16_t vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_low_s8(vout${M-1}x0123456789ABCDEF), vget_low_s8(vout${M}x0123456789ABCDEF)); 210 $elif M + 1 == MR: 211 int8x8_t vout${M}x01234567 = vget_low_s8(vout${M}x0123456789ABCDEF); 212 if (nc & 8) { 213 $for M in reversed(range(MR)): 214 $if M % 2 == 1: 215 vst1_s8(c${M}, vget_high_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]})); c${M} += 8; 216 vst1_s8(c${M-1}, vget_low_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]})); c${M-1} += 8; 217 $elif M + 1 == MR: 218 vst1_s8(c${M}, vout${M}x${ABC[N:N+8]}); c${M} += 8; 219 $for M in reversed(range(MR)): 220 $if M % 2 == 1: 221 vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_high_s8(vout${M-1}x0123456789ABCDEF), vget_high_s8(vout${M}x0123456789ABCDEF)); 222 $elif M + 1 == MR: 223 vout${M}x01234567 = vget_high_s8(vout${M}x0123456789ABCDEF); 224 } 225 if (nc & 4) { 226 $for M in reversed(range(MR)): 227 $if M % 2 == 1: 228 vst1q_lane_u32(__builtin_assume_aligned(c${M}, 1), vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 2); c${M} += 4; 229 vst1q_lane_u32(__builtin_assume_aligned(c${M-1}, 1), vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 4; 230 $elif M + 1 == MR: 231 vst1_lane_u32(__builtin_assume_aligned(c${M}, 1), vreinterpret_u32_s8(vout${M}x01234567), 0); c${M} += 4; 232 $for M in reversed(range(MR)): 233 $if M % 2 == 1: 234 vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 4); 235 $elif M + 1 == MR: 236 vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 4); 237 } 238 if (nc & 2) { 239 $for M in reversed(range(MR)): 240 $if M % 2 == 1: 241 vst1q_lane_u16(__builtin_assume_aligned(c${M}, 1), vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 4); c${M} += 2; 242 vst1q_lane_u16(__builtin_assume_aligned(c${M-1}, 1), vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 2; 243 $elif M + 1 == MR: 244 vst1_lane_u16(__builtin_assume_aligned(c${M}, 1), vreinterpret_u16_s8(vout${M}x01234567), 0); c${M} += 2; 245 $for M in reversed(range(MR)): 246 $if M % 2 == 1: 247 vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 2); 248 $elif M + 1 == MR: 249 vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 2); 250 } 251 if (nc & 1) { 252 $for M in reversed(range(MR)): 253 $if M % 2 == 1: 254 vst1q_lane_s8(c${M}, vout${M-1}x01234567_${M}x01234567, 8); 255 vst1q_lane_s8(c${M-1}, vout${M-1}x01234567_${M}x01234567, 0); 256 $elif M + 1 == MR: 257 vst1_lane_s8(c${M}, vout${M}x01234567, 0); 258 } 259 260 nc = 0; 261 } 262 } while (nc != 0); 263} 264