/aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/ |
H A D | 2x8c16-minmax-rndnu-neon-mlal.c | 140 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local 167 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
|
H A D | 2x8c8-minmax-rndnu-neon-mull.c | 124 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local 152 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
|
H A D | 2x8c8-minmax-fp32-neon-mlal.c | 199 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 227 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-rndnu-neon-mlal.c | 199 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local 227 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neonv8-mlal.c | 200 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 228 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
H A D | 3x8c16-minmax-rndnu-neon-mlal.c | 179 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local 212 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
|
H A D | 3x8c8-minmax-rndnu-neon-mull.c | 155 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local 189 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
|
H A D | 4x8c16-minmax-rndnu-neon-mlal.c | 218 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local 257 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
|
H A D | 4x8c8-minmax-rndnu-neon-mull.c | 186 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local 226 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
|
H A D | 3x8c8-minmax-rndnu-neon-mlal.c | 256 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local 290 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
|
H A D | 2x16c16-minmax-rndnu-neon-mlal.c | 216 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local 265 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
|
H A D | 2x16c8-minmax-rndnu-neon-mull.c | 184 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local 234 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); 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 | 140 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local 168 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
|
H A D | 2x8c16-minmax-rndnu-neon-mlal.c | 156 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local 183 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neonv8-mlal.c | 216 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 244 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
H A D | 2x8c8-minmax-rndnu-neon-mlal.c | 215 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local 243 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neon-mlal.c | 215 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 243 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
H A D | 3x8c8-minmax-rndnu-neon-mull.c | 173 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local 207 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
|
H A D | 3x8c16-minmax-rndnu-neon-mlal.c | 197 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local 230 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
|
H A D | 4x8c8-minmax-rndnu-neon-mull.c | 206 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local 246 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
|
H A D | 4x8c16-minmax-rndnu-neon-mlal.c | 238 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local 277 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); 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 | 200 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 228 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
H A D | 2x8c8-minmax-fp32-neon-mlal.c | 199 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 227 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); 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 | 215 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 243 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neonv8-mlal.c | 216 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 244 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|