| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 3x16c8-minmax-neon-mull-padal.c | 357 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
360 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
| D | 3x16c16-minmax-neon-mlal-padal.c | 405 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
408 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
| D | 4x16c8-minmax-neon-mull-padal.c | 426 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
429 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
| D | 4x16c16-minmax-neon-mlal-padal.c | 490 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
493 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
| D | 3x16c8-minmax-neon-mlal-padal.c | 546 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
549 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
| D | 4x16c8-minmax-neon-mlal-padal.c | 665 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
668 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 3x16c8-minmax-neon-mull-padal.c | 339 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
342 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
| D | 3x16c16-minmax-neon-mlal-padal.c | 387 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
390 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
| D | 4x16c8-minmax-neon-mull-padal.c | 406 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
409 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
| D | 3x16c8-minmax-neon-mlal-padal.c | 528 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
531 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
| D | 4x16c16-minmax-neon-mlal-padal.c | 470 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
473 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
| D | 4x16c8-minmax-neon-mlal-padal.c | 645 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
648 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|