/aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/ |
H A D | 1x16c8-minmax-rndnu-neon-mull.c | 126 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local 159 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
|
H A D | 1x16c16-minmax-rndnu-neon-mlal.c | 142 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local 174 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
|
H A D | 1x16c8-minmax-rndnu-neon-mlal.c | 215 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local 248 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
|
H A D | 2x16c16-minmax-rndnu-neon-mlal.c | 213 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local 257 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
|
H A D | 2x16c8-minmax-rndnu-neon-mull.c | 181 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local 226 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
|
H A D | 2x16c8-minmax-rndnu-neon-mlal.c | 320 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local 365 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
|
H A D | 3x16c16-minmax-rndnu-neon-mlal.c | 284 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local 340 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
|
H A D | 3x16c8-minmax-rndnu-neon-mull.c | 236 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local 293 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
|
H A D | 1x16c4s2-minmax-rndnu-neon-mull.c | 126 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local
|
H A D | 4x16c8-minmax-rndnu-neon-mull.c | 291 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local 360 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
|
H A D | 3x16c8-minmax-rndnu-neon-mlal.c | 425 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local 482 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
|
H A D | 4x16c16-minmax-rndnu-neon-mlal.c | 355 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local 423 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
|
H A D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 162 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
/aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/ |
H A D | 1x16c8-minmax-rndnu-neon-mull.c | 140 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local 173 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
|
H A D | 1x16c16-minmax-rndnu-neon-mlal.c | 156 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local 188 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
|
H A D | 1x16c8-minmax-rndnu-neon-mlal.c | 229 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local 262 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
|
H A D | 2x16c8-minmax-rndnu-neon-mull.c | 197 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local 242 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
|
H A D | 2x16c16-minmax-rndnu-neon-mlal.c | 229 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local 273 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
|
H A D | 2x16c8-minmax-rndnu-neon-mlal.c | 336 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local 381 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
|
H A D | 3x16c8-minmax-rndnu-neon-mull.c | 254 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local 311 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
|
H A D | 3x16c16-minmax-rndnu-neon-mlal.c | 302 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local 358 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
|
H A D | 1x16c4s2-minmax-rndnu-neon-mull.c | 140 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local
|
H A D | 4x16c8-minmax-rndnu-neon-mull.c | 311 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local 380 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
|
H A D | 4x16c16-minmax-rndnu-neon-mlal.c | 375 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local 443 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
|
H A D | 3x16c8-minmax-rndnu-neon-mlal.c | 443 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local 500 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
|