Searched refs:vacc3x15 (Results 1 – 6 of 6) sorted by relevance
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 4x16c8-minmax-neon-mull-padal.c | 122 int32x4_t vacc3x15 = vacc0x15; in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 296 vacc3x15 = vpadalq_s16(vacc3x15, vprod3x15); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() 336 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() 462 const int32x2_t vpsum3xF = vadd_s32(vget_low_s32(vacc3x15), vget_high_s32(vacc3x15)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 122 int32x4_t vacc3x15 = vacc0x15; in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 380 vacc3x15 = vpadalq_s16(vacc3x15, vprod3x15); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() 535 vacc3x15 = vpadalq_s16(vacc3x15, vprod3x15); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() 575 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() 701 const int32x2_t vpsum3xF = vadd_s32(vget_low_s32(vacc3x15), vget_high_s32(vacc3x15)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 122 int32x4_t vacc3x15 = vacc0x15; in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 360 vacc3x15 = vpadalq_s16(vacc3x15, vprod3x15); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() 400 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() 526 const int32x2_t vpsum3xF = vadd_s32(vget_low_s32(vacc3x15), vget_high_s32(vacc3x15)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 4x16c8-minmax-neon-mull-padal.c | 125 int32x4_t vacc3x15 = vacc0x15; in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 279 vacc3x15 = vpadalq_s16(vacc3x15, vprod3x15); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() 316 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() 442 const int32x2_t vpsum3xF = vadd_s32(vget_low_s32(vacc3x15), vget_high_s32(vacc3x15)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 125 int32x4_t vacc3x15 = vacc0x15; in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 363 vacc3x15 = vpadalq_s16(vacc3x15, vprod3x15); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() 518 vacc3x15 = vpadalq_s16(vacc3x15, vprod3x15); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() 555 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() 681 const int32x2_t vpsum3xF = vadd_s32(vget_low_s32(vacc3x15), vget_high_s32(vacc3x15)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 125 int32x4_t vacc3x15 = vacc0x15; in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 343 vacc3x15 = vpadalq_s16(vacc3x15, vprod3x15); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() 380 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() 506 const int32x2_t vpsum3xF = vadd_s32(vget_low_s32(vacc3x15), vget_high_s32(vacc3x15)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|