/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c16-minmax-rndnu-neon-mlal.c | 154 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local 181 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
|
D | 1x16c8-minmax-rndnu-neon-mull.c | 138 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local 166 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
|
D | 1x16c8-minmax-rndnu-neon-mlal.c | 227 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local 255 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
|
D | 2x16c16-minmax-rndnu-neon-mlal.c | 227 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local 266 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
|
D | 2x16c8-minmax-rndnu-neon-mull.c | 195 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local 235 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
|
D | 1x16c4s2-minmax-rndnu-neon-mull.c | 137 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld2r.c | 172 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld1r.c | 173 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 172 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
D | 2x16c8-minmax-rndnu-neon-mlal.c | 334 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local 374 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
|
D | 3x16c8-minmax-rndnu-neon-mull.c | 252 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local 304 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
|
D | 3x16c16-minmax-rndnu-neon-mlal.c | 300 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local 351 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c16-minmax-rndnu-neon-mlal.c | 140 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local 167 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
|
D | 1x16c8-minmax-rndnu-neon-mull.c | 124 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local 152 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
|
D | 1x16c8-minmax-rndnu-neon-mlal.c | 213 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local 241 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
|
D | 2x16c16-minmax-rndnu-neon-mlal.c | 211 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local 250 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
|
D | 2x16c8-minmax-rndnu-neon-mull.c | 179 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local 219 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
|
D | 1x16c4s2-minmax-rndnu-neon-mull.c | 123 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld1r.c | 160 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld2r.c | 159 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 159 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
D | 3x16c8-minmax-rndnu-neon-mull.c | 234 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local 286 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
|
D | 3x16c16-minmax-rndnu-neon-mlal.c | 282 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local 333 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
|
D | 2x16c8-minmax-rndnu-neon-mlal.c | 318 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local 358 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
|
D | 1x16c4s2-minmax-rndnu-neon-mlal.c | 212 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local
|