/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x8c16-minmax-neon-mlal-padal.c | 185 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 233 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mull-padal.c | 161 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local 209 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
|
D | 3x8c8-minmax-neon-mlal-padal.c | 262 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 310 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mull-padal.c | 192 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local 246 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
|
D | 4x8c16-minmax-neon-mlal-padal.c | 224 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local 278 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mlal-padal.c | 319 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 373 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 249 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 335 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 297 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 383 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 368 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 466 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 304 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 402 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 438 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 524 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 543 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 641 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x8c8-minmax-neon-mull-padal.c | 179 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local 227 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 203 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 251 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mull-padal.c | 212 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local 266 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
|
D | 4x8c16-minmax-neon-mlal-padal.c | 244 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local 298 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mlal-padal.c | 280 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 328 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mlal-padal.c | 339 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 393 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 267 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 353 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 315 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 401 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 324 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 422 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 388 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 486 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 456 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 542 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 563 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 661 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|