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$assert 8 <= NR <= 32 9$assert REQUANTIZATION in ["FP32", "RNDNU"] 10#include <assert.h> 11 12#include <arm_neon.h> 13 14#include <xnnpack/gemm.h> 15$if REQUANTIZATION == "FP32": 16 #include <xnnpack/intrinsics-polyfill.h> 17#include <xnnpack/math.h> 18 19 20$PARAMS_STRUCT = "fp32_neonv8" if REQUANTIZATION == "FP32" else REQUANTIZATION.lower() + "_neon" 21void xnn_qu8_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c4__neondot( 22 size_t mr, 23 size_t nc, 24 size_t kc, 25 const uint8_t* restrict a, 26 size_t a_stride, 27 const void* restrict w, 28 uint8_t* restrict c, 29 size_t cm_stride, 30 size_t cn_stride, 31 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 32{ 33 assert(mr != 0); 34 assert(mr <= ${MR}); 35 assert(nc != 0); 36 assert(kc != 0); 37 assert(kc % sizeof(uint8_t) == 0); 38 assert(a != NULL); 39 assert(w != NULL); 40 assert(c != NULL); 41 42 kc = round_up_po2(kc, 4 * sizeof(uint8_t)); 43 const uint8_t* a0 = a; 44 uint8_t* c0 = c; 45 $for M in range(1, MR): 46 const uint8_t* a${M} = (const uint8_t*) ((uintptr_t) a${M-1} + a_stride); 47 uint8_t* c${M} = (uint8_t*) ((uintptr_t) c${M-1} + cm_stride); 48 $if M % 2 == 0: 49 if XNN_UNPREDICTABLE(mr <= ${M}) { 50 a${M} = a${M-1}; 51 c${M} = c${M-1}; 52 } 53 $elif M + 1 == MR: 54 if XNN_UNPREDICTABLE(mr != ${M+1}) { 55 a${M} = a${M-1}; 56 c${M} = c${M-1}; 57 } 58 $else: 59 if XNN_UNPREDICTABLE(mr < ${M+1}) { 60 a${M} = a${M-1}; 61 c${M} = c${M-1}; 62 } 63 64 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->${PARAMS_STRUCT}.kernel_zero_point[0]); 65 66 // Loop over groups of ${NR} columns. 67 do { 68 // Initialize accumulators with bias. ${NR} bias values are loaded from the 69 // weight matrix, at the start of the group of ${NR} columns. 70 $for N in range(0, NR, 4): 71 uint32x4_t vpacc0x${ABC[N:N+4]} = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); 72 $for M in range(1, MR): 73 $for N in range(0, NR, 4): 74 uint32x4_t vpacc${M}x${ABC[N:N+4]} = vpacc0x${ABC[N:N+4]}; 75 $for M in range(0, MR): 76 uint32x2_t vnacc${M} = vmov_n_u32(0); 77 78 // Inner accumulation loop along the ${NR} columns. 79 size_t k = kc; 80 // 2x partial unrolled loop to load 8 bytes at a time. 81 while (k >= 8 * sizeof(uint8_t)) { 82 // Load a ${MR}x8 block of activations. 83 $for M in range(MR): 84 const uint8x8_t va${M}x01234567 = vld1_u8(a${M}); a${M} += 8; 85 86 // Load a 8x${NR} block of weights. 87 $for K in range(0, 8, 4): 88 $for N in range(0, NR, 4): 89 const uint8x16_t vb${ABC[K:K+4]}x${ABC[N:N+4]} = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); 90 91 // Multiply-accumulate: ${MR}x8 * 8x${NR} --> ${MR}x${NR}. 92 $for M in range(MR): 93 vnacc${M} = vdot_u32(vnacc${M}, va_zero_point, va${M}x01234567); 94 $for K in range(0, 8, 4): 95 $for N in range(0, NR, 4): 96 vpacc${M}x${ABC[N:N+4]} = vdotq_lane_u32(vpacc${M}x${ABC[N:N+4]}, vb${ABC[K:K+4]}x${ABC[N:N+4]}, va${M}x01234567, ${K//4}); 97 98 k -= 8 * sizeof(uint8_t); 99 } 100 // Handle up to 4 final positions of `k` 101 if XNN_UNLIKELY(k != 0) { 102 // Load a ${MR}x4 block of activations. 103 $for M in range(MR): 104 const uint8x8_t va${M}x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a${M}, vmov_n_u32(0), 0)); a${M} += 4; 105 106 // Load a 4x${NR} block of weights. 107 $for N in range(0, NR, 4): 108 const uint8x16_t vb0123x${ABC[N:N+4]} = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); 109 110 // Multiply-accumulate: ${MR}x4 * 4x${NR} --> ${MR}x${NR}. 111 $for M in range(MR): 112 vnacc${M} = vdot_u32(vnacc${M}, va_zero_point, va${M}x01234567); 113 $for N in range(0, NR, 4): 114 vpacc${M}x${ABC[N:N+4]} = vdotq_lane_u32(vpacc${M}x${ABC[N:N+4]}, vb0123x${ABC[N:N+4]}, va${M}x01234567, 0); 115 } 116 117 // Subtract zero point from accumulators. 118 $for M in range(0, MR): 119 vnacc${M} = vpadd_u32(vnacc${M}, vnacc${M}); 120 const uint32x4_t vnacc${M}x0123 = vcombine_u32(vnacc${M}, vnacc${M}); 121 $for N in range(0, NR, 4): 122 int32x4_t vacc${M}x${ABC[N:N+4]} = vreinterpretq_s32_u32(vsubq_u32(vpacc${M}x${ABC[N:N+4]}, vnacc${M}x0123)); 123 124 $if REQUANTIZATION == "RNDNU": 125 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->${PARAMS_STRUCT}.right_pre_shift); 126 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->${PARAMS_STRUCT}.multiplier); 127 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->${PARAMS_STRUCT}.right_post_shift); 128 129 $for M in range(MR): 130 $for N in range(0, NR, 4): 131 vacc${M}x${ABC[N:N+4]} = vshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_pre_shift); 132 133 $for M in range(MR): 134 $for N in range(0, NR, 4): 135 vacc${M}x${ABC[N:N+4]} = vqdmulhq_s32(vacc${M}x${ABC[N:N+4]}, vmultiplier); 136 137 $for M in range(MR): 138 $for N in range(0, NR, 4): 139 vacc${M}x${ABC[N:N+4]} = vrshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_post_shift); 140 $elif REQUANTIZATION == "FP32": 141 $for M in range(MR): 142 $for N in range(0, NR, 4): 143 float32x4_t vfpacc${M}x${ABC[N:N+4]} = vcvtq_f32_s32(vacc${M}x${ABC[N:N+4]}); 144 145 const float32x4_t vscale = vld1q_dup_f32(¶ms->${PARAMS_STRUCT}.scale); 146 $for M in range(MR): 147 $for N in range(0, NR, 4): 148 vfpacc${M}x${ABC[N:N+4]} = vmulq_f32(vfpacc${M}x${ABC[N:N+4]}, vscale); 149 150 $for M in range(MR): 151 $for N in range(0, NR, 4): 152 vacc${M}x${ABC[N:N+4]} = vcvtnq_s32_f32(vfpacc${M}x${ABC[N:N+4]}); 153 154 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->${PARAMS_STRUCT}.output_zero_point); 155#if XNN_ARCH_ARM64 156 $for M in range(MR): 157 $for N in range(0, NR, 8): 158 const int16x8_t vacc${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); 159 160 $for M in range(MR): 161 $for N in range(0, NR, 16): 162 $if N + 8 < NR: 163 uint8x16_t vout${M}x${ABC[N:N+16]} = vqmovun_high_s16(vqmovun_s16(vacc${M}x${ABC[N:N+8]}), vacc${M}x${ABC[N+8:N+16]}); 164 $elif M % 2 == 1: 165 uint8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vqmovun_high_s16(vqmovun_s16(vacc${M-1}x${ABC[N:N+8]}), vacc${M}x${ABC[N:N+8]}); 166 $elif M + 1 == MR: 167 uint8x8_t vout${M}x${ABC[N:N+8]} = vqmovun_s16(vacc${M}x${ABC[N:N+8]}); 168#else 169 $for M in range(MR): 170 $for N in range(0, NR, 8): 171 const int16x8_t vacc${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); 172 173 $for M in range(MR): 174 $for N in range(0, NR, 16): 175 $if N + 8 < NR: 176 uint8x16_t vout${M}x${ABC[N:N+16]} = vcombine_u8(vqmovun_s16(vacc${M}x${ABC[N:N+8]}), vqmovun_s16(vacc${M}x${ABC[N+8:N+16]})); 177 $elif M % 2 == 1: 178 uint8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vcombine_u8(vqmovun_s16(vacc${M-1}x${ABC[N:N+8]}), vqmovun_s16(vacc${M}x${ABC[N:N+8]})); 179 $elif M + 1 == MR: 180 uint8x8_t vout${M}x${ABC[N:N+8]} = vqmovun_s16(vacc${M}x${ABC[N:N+8]}); 181#endif 182 $if NR == 8 and MR == 1: 183 const uint8x8_t voutput_min = vld1_dup_u8(¶ms->${PARAMS_STRUCT}.output_min); 184 const uint8x8_t voutput_max = vld1_dup_u8(¶ms->${PARAMS_STRUCT}.output_max); 185 $else: 186 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->${PARAMS_STRUCT}.output_min); 187 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->${PARAMS_STRUCT}.output_max); 188 189 $for M in range(MR): 190 $for N in range(0, NR, 16): 191 $if N + 8 < NR: 192 vout${M}x${ABC[N:N+16]} = vmaxq_u8(vout${M}x${ABC[N:N+16]}, voutput_min); 193 $elif M % 2 == 1: 194 vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vmaxq_u8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}, voutput_min); 195 $elif M + 1 == MR: 196 $if NR == 8 and MR == 1: 197 vout${M}x${ABC[N:N+8]} = vmax_u8(vout${M}x${ABC[N:N+8]}, voutput_min); 198 $else: 199 vout${M}x${ABC[N:N+8]} = vmax_u8(vout${M}x${ABC[N:N+8]}, vget_low_u8(voutput_min)); 200 201 $for M in range(MR): 202 $for N in range(0, NR, 16): 203 $if N + 8 < NR: 204 vout${M}x${ABC[N:N+16]} = vminq_u8(vout${M}x${ABC[N:N+16]}, voutput_max); 205 $elif M % 2 == 1: 206 vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vminq_u8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}, voutput_max); 207 $elif M + 1 == MR: 208 $if NR == 8 and MR == 1: 209 vout${M}x${ABC[N:N+8]} = vmin_u8(vout${M}x${ABC[N:N+8]}, voutput_max); 210 $else: 211 vout${M}x${ABC[N:N+8]} = vmin_u8(vout${M}x${ABC[N:N+8]}, vget_low_u8(voutput_max)); 212 213 if (nc >= ${NR}) { 214 $for M in range(MR): 215 $for N in range(0, NR, 16): 216 $if N + 8 < NR: 217 vst1q_u8(c${M} + ${N}, vout${M}x${ABC[N:N+16]}); 218 $elif M % 2 == 1: 219 vst1_u8(c${M-1} + ${N}, vget_low_u8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]})); 220 vst1_u8(c${M} + ${N}, vget_high_u8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]})); 221 $elif M + 1 == MR: 222 vst1_u8(c${M} + ${N}, vout${M}x${ABC[N:N+8]}); 223 224 $for M in range(MR): 225 c${M} = (uint8_t*) ((uintptr_t) c${M} + cn_stride); 226 227 $for M in range(MR): 228 a${M} = (const uint8_t*) ((uintptr_t) a${M} - kc); 229 230 nc -= ${NR}; 231 } else { 232 $if NR == 32: 233 if (nc & 16) { 234 $for M in range(MR): 235 vst1q_u8(c${M}, vout${M}x${ABC[0:16]}); c${M} += 16; 236 237 $for M in range(MR): 238 vout${M}x${ABC[0:16]} = vout${M}x${ABC[16:32]}; 239 } 240 $if NR >= 16: 241 $for M in range(MR): 242 $if M % 2 == 1: 243 uint8x16_t vout${M-1}x01234567_${M}x01234567 = vcombine_u8(vget_low_u8(vout${M-1}x0123456789ABCDEF), vget_low_u8(vout${M}x0123456789ABCDEF)); 244 $elif M + 1 == MR: 245 uint8x8_t vout${M}x01234567 = vget_low_u8(vout${M}x0123456789ABCDEF); 246 if (nc & 8) { 247 $for M in range(MR): 248 $if M % 2 == 1: 249 vst1_u8(c${M-1}, vget_low_u8(vout${M-1}x01234567_${M}x01234567)); c${M-1} += 8; 250 vst1_u8(c${M}, vget_high_u8(vout${M-1}x01234567_${M}x01234567)); c${M} += 8; 251 $elif M + 1 == MR: 252 vst1_u8(c${M}, vout${M}x01234567); c${M} += 8; 253 $for M in range(MR): 254 $if M % 2 == 1: 255 vout${M-1}x01234567_${M}x01234567 = vcombine_u8(vget_high_u8(vout${M-1}x0123456789ABCDEF), vget_high_u8(vout${M}x0123456789ABCDEF)); 256 $elif M + 1 == MR: 257 vout${M}x01234567 = vget_high_u8(vout${M}x0123456789ABCDEF); 258 } 259 if (nc & 4) { 260 $for M in range(MR): 261 $if M % 2 == 1: 262 vst1q_lane_u32((void*) c${M-1}, vreinterpretq_u32_u8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 4; 263 vst1q_lane_u32((void*) c${M}, vreinterpretq_u32_u8(vout${M-1}x01234567_${M}x01234567), 2); c${M} += 4; 264 $elif M + 1 == MR: 265 vst1_lane_u32((void*) c${M}, vreinterpret_u32_u8(vout${M}x01234567), 0); c${M} += 4; 266 $for M in range(MR): 267 $if M % 2 == 1: 268 vout${M-1}x01234567_${M}x01234567 = vextq_u8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 4); 269 $elif M + 1 == MR: 270 vout${M}x01234567 = vext_u8(vout${M}x01234567, vout${M}x01234567, 4); 271 } 272 if (nc & 2) { 273 $for M in range(MR): 274 $if M % 2 == 1: 275 vst1q_lane_u16((void*) c${M-1}, vreinterpretq_u16_u8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 2; 276 vst1q_lane_u16((void*) c${M}, vreinterpretq_u16_u8(vout${M-1}x01234567_${M}x01234567), 4); c${M} += 2; 277 $elif M + 1 == MR: 278 vst1_lane_u16((void*) c${M}, vreinterpret_u16_u8(vout${M}x01234567), 0); c${M} += 2; 279 $for M in range(MR): 280 $if M % 2 == 1: 281 vout${M-1}x01234567_${M}x01234567 = vextq_u8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 2); 282 $elif M + 1 == MR: 283 vout${M}x01234567 = vext_u8(vout${M}x01234567, vout${M}x01234567, 2); 284 } 285 if (nc & 1) { 286 $for M in range(MR): 287 $if M % 2 == 1: 288 vst1q_lane_u8(c${M-1}, vout${M-1}x01234567_${M}x01234567, 0); 289 vst1q_lane_u8(c${M}, vout${M-1}x01234567_${M}x01234567, 8); 290 $elif M + 1 == MR: 291 vst1_lane_u8(c${M}, vout${M}x01234567, 0); 292 } 293 294 nc = 0; 295 } 296 } while (nc != 0); 297} 298