/aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/ |
H A D | 2x8c16-minmax-rndnu-neon-mlal.c | 139 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local 166 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
|
H A D | 2x8c8-minmax-rndnu-neon-mull.c | 123 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local 151 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
|
H A D | 2x8c8-minmax-fp32-neon-mlal.c | 198 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 226 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-rndnu-neon-mlal.c | 198 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local 226 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neonv8-mlal.c | 199 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 227 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
H A D | 3x8c16-minmax-rndnu-neon-mlal.c | 178 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local 211 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
|
H A D | 3x8c8-minmax-rndnu-neon-mull.c | 154 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local 188 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
|
H A D | 4x8c16-minmax-rndnu-neon-mlal.c | 217 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local 256 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
|
H A D | 4x8c8-minmax-rndnu-neon-mull.c | 185 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local 225 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
|
H A D | 3x8c8-minmax-rndnu-neon-mlal.c | 255 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local 289 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
|
H A D | 2x16c16-minmax-rndnu-neon-mlal.c | 215 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local 264 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
|
H A D | 2x16c8-minmax-rndnu-neon-mull.c | 183 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local 233 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); 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 | 139 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local 167 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
|
H A D | 2x8c16-minmax-rndnu-neon-mlal.c | 155 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local 182 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neonv8-mlal.c | 215 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 243 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
H A D | 2x8c8-minmax-rndnu-neon-mlal.c | 214 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local 242 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neon-mlal.c | 214 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 242 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
H A D | 3x8c8-minmax-rndnu-neon-mull.c | 172 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local 206 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
|
H A D | 3x8c16-minmax-rndnu-neon-mlal.c | 196 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local 229 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
|
H A D | 4x8c8-minmax-rndnu-neon-mull.c | 205 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local 245 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
|
H A D | 4x8c16-minmax-rndnu-neon-mlal.c | 237 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local 276 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); 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 | 199 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 227 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
H A D | 2x8c8-minmax-fp32-neon-mlal.c | 198 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 226 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); 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 | 214 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 242 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neonv8-mlal.c | 215 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 243 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|