/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 124 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 130 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() 151 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 153 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 140 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 146 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() 167 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 169 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 213 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 219 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() 240 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 242 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 2x16c8-minmax-neon-mull-padal.c | 179 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local 193 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() 218 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local 220 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 211 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 225 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() 250 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 252 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 234 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 256 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() 285 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 287 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 318 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 332 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() 357 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 359 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 282 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 304 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() 333 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 335 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 289 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 319 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() 352 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 354 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 423 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 445 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() 474 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 476 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 353 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 383 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() 416 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 418 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 528 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 558 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() 591 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 593 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 138 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 144 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() 165 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 167 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 154 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 160 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() 181 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 183 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 227 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 233 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() 254 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 256 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 2x16c8-minmax-neon-mull-padal.c | 195 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 209 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() 234 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 236 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 227 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 241 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() 266 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 268 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 252 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 274 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() 303 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 305 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 334 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 348 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() 373 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 375 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 300 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 322 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() 351 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 353 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 309 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 339 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() 372 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 374 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 373 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 403 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() 436 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 438 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 441 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 463 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() 492 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 494 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 548 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 578 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() 611 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 613 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB ); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|