/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x8c8-minmax-neon-mull-padal.c | 91 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local 93 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() 107 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local 108 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal()
|
D | 1x8c16-minmax-neon-mlal-padal.c | 99 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local 101 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() 115 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local 116 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
|
D | 1x8c8-minmax-neon-mlal-padal.c | 140 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local 142 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() 156 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local 157 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
|
D | 2x8c8-minmax-neon-mull-padal.c | 122 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local 128 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() 144 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local 145 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal()
|
D | 1x16c8-minmax-neon-mull-padal.c | 123 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 129 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() 145 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 146 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 2x8c16-minmax-neon-mlal-padal.c | 138 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local 144 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() 160 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local 161 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 139 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 145 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() 161 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 162 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 2x8c8-minmax-neon-mlal-padal.c | 197 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 203 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() 219 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 220 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
|
D | 3x8c8-minmax-neon-mull-padal.c | 153 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local 163 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() 181 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local 182 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
|
D | 3x8c16-minmax-neon-mlal-padal.c | 177 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 187 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() 205 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 206 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 212 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 218 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() 234 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 235 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 4x8c8-minmax-neon-mull-padal.c | 184 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local 198 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() 218 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local 219 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x8c8-minmax-neon-mull-padal.c | 105 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local 107 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() 121 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local 122 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal()
|
D | 1x8c16-minmax-neon-mlal-padal.c | 113 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local 115 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() 129 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local 130 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
|
D | 1x8c8-minmax-neon-mlal-padal.c | 154 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local 156 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() 170 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local 171 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
|
D | 2x8c8-minmax-neon-mull-padal.c | 138 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local 144 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() 160 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local 161 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal()
|
D | 1x16c8-minmax-neon-mull-padal.c | 137 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 143 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() 159 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 160 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 2x8c16-minmax-neon-mlal-padal.c | 154 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local 160 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() 176 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local 177 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 153 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 159 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() 175 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 176 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 2x8c8-minmax-neon-mlal-padal.c | 213 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 219 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() 235 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 236 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
|
D | 3x8c8-minmax-neon-mull-padal.c | 171 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local 181 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() 199 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local 200 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 226 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 232 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() 248 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 249 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 3x8c16-minmax-neon-mlal-padal.c | 195 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 205 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() 223 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 224 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
|
D | 4x8c8-minmax-neon-mull-padal.c | 204 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local 218 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() 238 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local 239 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mull-padal.c | 194 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 208 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() 228 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 229 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
|