/aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/ |
H A D | 2x8c16-minmax-rndnu-neon-mlal.c | 142 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local 174 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
|
H A D | 2x8c8-minmax-rndnu-neon-mull.c | 126 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local 159 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
|
H A D | 2x8c8-minmax-fp32-neon-mlal.c | 201 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 234 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-rndnu-neon-mlal.c | 201 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local 234 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neonv8-mlal.c | 202 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 235 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
H A D | 3x8c16-minmax-rndnu-neon-mlal.c | 181 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local 219 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
|
H A D | 3x8c8-minmax-rndnu-neon-mull.c | 157 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local 196 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
|
H A D | 4x8c16-minmax-rndnu-neon-mlal.c | 220 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local 264 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
|
H A D | 4x8c8-minmax-rndnu-neon-mull.c | 188 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local 233 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
|
H A D | 3x8c8-minmax-rndnu-neon-mlal.c | 258 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local 297 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
|
H A D | 2x16c16-minmax-rndnu-neon-mlal.c | 218 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local 272 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
|
H A D | 2x16c8-minmax-rndnu-neon-mull.c | 186 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local 241 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
|
/aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/ |
H A D | 2x8c8-minmax-rndnu-neon-mull.c | 142 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local 175 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
|
H A D | 2x8c16-minmax-rndnu-neon-mlal.c | 158 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local 190 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neonv8-mlal.c | 218 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 251 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
H A D | 2x8c8-minmax-rndnu-neon-mlal.c | 217 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local 250 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neon-mlal.c | 217 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 250 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
H A D | 3x8c8-minmax-rndnu-neon-mull.c | 175 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local 214 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
|
H A D | 3x8c16-minmax-rndnu-neon-mlal.c | 199 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local 237 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
|
H A D | 4x8c8-minmax-rndnu-neon-mull.c | 208 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local 253 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
|
H A D | 4x8c16-minmax-rndnu-neon-mlal.c | 240 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local 284 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
|
/aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/ |
H A D | 2x8c8-minmax-fp32-neonv8-mlal.c | 202 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 235 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
H A D | 2x8c8-minmax-fp32-neon-mlal.c | 201 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 234 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
/aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/ |
H A D | 2x8c8-minmax-fp32-neon-mlal.c | 217 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 250 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neonv8-mlal.c | 218 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 251 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|