/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x8c16-minmax-rndnu-neon-mlal.c | 141 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local 173 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
|
D | 2x8c8-minmax-rndnu-neon-mull.c | 125 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local 158 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
|
D | 2x8c8-minmax-fp32-neonv8-mlal.c | 201 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 234 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
D | 2x8c8-minmax-fp32-neon-mlal.c | 200 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 233 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
D | 2x8c8-minmax-rndnu-neon-mlal.c | 200 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local 233 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
|
D | 3x8c8-minmax-rndnu-neon-mull.c | 156 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local 195 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
|
D | 3x8c16-minmax-rndnu-neon-mlal.c | 180 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local 218 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
|
D | 2x16c16-minmax-rndnu-neon-mlal.c | 217 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local 271 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
|
D | 4x8c8-minmax-rndnu-neon-mull.c | 187 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local 232 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
|
D | 3x8c8-minmax-rndnu-neon-mlal.c | 257 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local 296 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
|
D | 2x16c8-minmax-rndnu-neon-mull.c | 185 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local 240 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
|
D | 4x8c16-minmax-rndnu-neon-mlal.c | 219 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local 263 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 2x8c8-minmax-rndnu-neon-mull.c | 141 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local 174 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
|
D | 2x8c16-minmax-rndnu-neon-mlal.c | 157 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local 189 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
|
D | 2x8c8-minmax-rndnu-neon-mlal.c | 216 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local 249 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
|
D | 2x8c8-minmax-fp32-neonv8-mlal.c | 217 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 250 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
D | 2x8c8-minmax-fp32-neon-mlal.c | 216 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 249 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
D | 3x8c16-minmax-rndnu-neon-mlal.c | 198 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local 236 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
|
D | 3x8c8-minmax-rndnu-neon-mull.c | 174 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local 213 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
|
D | 4x8c16-minmax-rndnu-neon-mlal.c | 239 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local 283 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
|
D | 4x8c8-minmax-rndnu-neon-mull.c | 207 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local 252 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
|
/external/XNNPACK/src/qc8-gemm/gen/ |
D | 2x8c8-minmax-fp32-neon-mlal.c | 200 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 233 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
D | 2x8c8-minmax-fp32-neonv8-mlal.c | 201 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 234 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
/external/XNNPACK/src/qc8-igemm/gen/ |
D | 2x8c8-minmax-fp32-neonv8-mlal.c | 217 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 250 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
D | 2x8c8-minmax-fp32-neon-mlal.c | 216 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 249 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|