/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c8-minmax-rndnu-neon-mull.c | 125 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local 153 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
|
D | 1x16c16-minmax-rndnu-neon-mlal.c | 141 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local 168 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
|
D | 1x16c8-minmax-rndnu-neon-mlal.c | 214 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local 242 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
|
D | 2x16c16-minmax-rndnu-neon-mlal.c | 212 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local 251 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
|
D | 2x16c8-minmax-rndnu-neon-mull.c | 180 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local 220 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
|
D | 1x16c4s2-minmax-rndnu-neon-mull.c | 124 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local
|
D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 160 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld2r.c | 160 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld1r.c | 161 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
|
D | 3x16c16-minmax-rndnu-neon-mlal.c | 283 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local 334 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
|
D | 3x16c8-minmax-rndnu-neon-mull.c | 235 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local 287 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
|
D | 2x16c8-minmax-rndnu-neon-mlal.c | 319 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local 359 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
|
D | 1x16c4s2-minmax-rndnu-neon-mlal.c | 213 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c16-minmax-rndnu-neon-mlal.c | 155 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local 182 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
|
D | 1x16c8-minmax-rndnu-neon-mull.c | 139 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local 167 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
|
D | 1x16c8-minmax-rndnu-neon-mlal.c | 228 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local 256 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
|
D | 2x16c16-minmax-rndnu-neon-mlal.c | 228 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local 267 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
|
D | 2x16c8-minmax-rndnu-neon-mull.c | 196 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local 236 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
|
D | 1x16c4s2-minmax-rndnu-neon-mull.c | 138 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local
|
D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 173 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld2r.c | 173 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld1r.c | 174 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
|
D | 2x16c8-minmax-rndnu-neon-mlal.c | 335 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local 375 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
|
D | 3x16c8-minmax-rndnu-neon-mull.c | 253 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local 305 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
|
D | 3x16c16-minmax-rndnu-neon-mlal.c | 301 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local 352 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
|