/aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/ |
H A D | 1x8c8-minmax-rndnu-neon-mull.c | 90 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mull() local 107 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mull() local
|
H A D | 1x8c16-minmax-rndnu-neon-mlal.c | 98 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local 114 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local
|
H A D | 1x8c8-minmax-fp32-neonv8-mlal.c | 140 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local 157 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
|
H A D | 1x8c8-minmax-rndnu-neon-mlal.c | 139 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local 156 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
|
H A D | 1x8c8-minmax-fp32-neon-mlal.c | 139 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local 156 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
|
H A D | 2x8c16-minmax-rndnu-neon-mlal.c | 137 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local 159 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
|
H A D | 2x8c8-minmax-rndnu-neon-mull.c | 121 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local 144 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
|
H A D | 1x16c8-minmax-rndnu-neon-mull.c | 122 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local 145 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
|
H A D | 1x16c16-minmax-rndnu-neon-mlal.c | 138 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local 160 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neon-mlal.c | 196 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local 219 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
|
/aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/ |
H A D | 1x8c8-minmax-rndnu-neon-mull.c | 104 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mull() local 121 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mull() local
|
H A D | 1x8c16-minmax-rndnu-neon-mlal.c | 112 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local 128 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local
|
H A D | 1x8c8-minmax-fp32-neonv8-mlal.c | 154 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local 171 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
|
H A D | 1x8c8-minmax-fp32-neon-mlal.c | 153 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local 170 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
|
H A D | 1x8c8-minmax-rndnu-neon-mlal.c | 153 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local 170 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
|
H A D | 2x8c8-minmax-rndnu-neon-mull.c | 137 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local 160 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
|
H A D | 2x8c16-minmax-rndnu-neon-mlal.c | 153 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local 175 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
|
H A D | 1x16c8-minmax-rndnu-neon-mull.c | 136 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local 159 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
|
H A D | 1x16c16-minmax-rndnu-neon-mlal.c | 152 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local 174 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
|
H A D | 2x8c8-minmax-fp32-neonv8-mlal.c | 213 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 236 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
/aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/ |
H A D | 1x8c8-minmax-fp32-neon-mlal.c | 139 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local 156 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
|
H A D | 1x8c8-minmax-fp32-neonv8-mlal.c | 140 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local 157 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
|
H A D | 2x8c8-minmax-fp32-neonv8-mlal.c | 197 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local 220 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
|
/aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/ |
H A D | 1x8c8-minmax-fp32-neon-mlal.c | 153 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local 170 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
|
H A D | 1x8c8-minmax-fp32-neonv8-mlal.c | 154 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local 171 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
|