/external/XNNPACK/src/qu8-gemm/gen/ |
D | 3x32c4-minmax-rndnu-neondot.c | 240 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot() local 241 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot() 242 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot() 243 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot() 244 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot() 245 int32x4_t vacc2xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc2xGHIJ, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot() 246 int32x4_t vacc2xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc2xKLMN, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot() 247 int32x4_t vacc2xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc2xOPQR, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot() 248 int32x4_t vacc2xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc2xSTUV, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot()
|
D | 3x16c4-minmax-rndnu-neondot.c | 172 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot() local 173 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot() 174 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot() 175 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot() 176 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot()
|
D | 3x8c4-minmax-rndnu-neondot.c | 138 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot() local 139 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot() 140 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot()
|
D | 4x16c4-minmax-fp32-neondot.c | 200 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot() local 201 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot() 202 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot() 203 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot() 204 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot()
|
D | 4x16c4-minmax-rndnu-neondot.c | 199 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot() local 200 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot() 201 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot() 202 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot() 203 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot()
|
D | 5x16c4-minmax-rndnu-neondot.c | 226 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot() local 227 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot() 228 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot() 229 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot() 230 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot()
|
D | 4x8c4-minmax-rndnu-neondot.c | 157 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot() local 158 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot() 159 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
|
D | 5x8c4-minmax-rndnu-neondot.c | 176 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot() local 177 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot() 178 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
|
D | 6x16c4-minmax-rndnu-neondot.c | 253 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot() local 254 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot() 255 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot() 256 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot() 257 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot()
|
D | 6x8c4-minmax-rndnu-neondot.c | 195 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot() local 196 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot() 197 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
|
D | 8x16c4-minmax-rndnu-neondot.c | 307 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x16c4__neondot() local 308 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x16c4__neondot() 309 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x16c4__neondot() 310 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x16c4__neondot() 311 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
|
D | 8x8c4-minmax-rndnu-neondot.c | 233 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot() local 234 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot() 235 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
|
/external/XNNPACK/src/qu8-igemm/gen/ |
D | 3x32c4-minmax-rndnu-neondot.c | 256 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot() local 257 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot() 258 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot() 259 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot() 260 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot() 261 int32x4_t vacc2xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc2xGHIJ, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot() 262 int32x4_t vacc2xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc2xKLMN, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot() 263 int32x4_t vacc2xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc2xOPQR, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot() 264 int32x4_t vacc2xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc2xSTUV, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot()
|
D | 3x16c4-minmax-rndnu-neondot.c | 188 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot() local 189 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot() 190 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot() 191 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot() 192 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot()
|
D | 3x8c4-minmax-rndnu-neondot.c | 154 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot() local 155 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot() 156 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot()
|
D | 4x16c4-minmax-fp32-neondot.c | 218 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot() local 219 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot() 220 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot() 221 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot() 222 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot()
|
D | 4x16c4-minmax-rndnu-neondot.c | 217 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot() local 218 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot() 219 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot() 220 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot() 221 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot()
|
D | 5x16c4-minmax-rndnu-neondot.c | 246 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot() local 247 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot() 248 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot() 249 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot() 250 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot()
|
D | 4x8c4-minmax-rndnu-neondot.c | 175 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot() local 176 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot() 177 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
|
D | 5x8c4-minmax-rndnu-neondot.c | 196 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot() local 197 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot() 198 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
|
D | 6x16c4-minmax-rndnu-neondot.c | 275 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x16c4__neondot() local 276 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x16c4__neondot() 277 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x16c4__neondot() 278 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x16c4__neondot() 279 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x16c4__neondot()
|
D | 6x8c4-minmax-rndnu-neondot.c | 217 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot() local 218 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot() 219 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
|
D | 8x16c4-minmax-rndnu-neondot.c | 333 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x16c4__neondot() local 334 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x16c4__neondot() 335 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x16c4__neondot() 336 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x16c4__neondot() 337 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x16c4__neondot()
|
D | 8x8c4-minmax-rndnu-neondot.c | 259 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot() local 260 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot() 261 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
|