/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x8c8-minmax-neon-mull-padal.c | 125 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local 157 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
|
D | 2x8c16-minmax-neon-mlal-padal.c | 141 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local 173 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
|
D | 2x8c8-minmax-neon-mlal-padal.c | 200 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 232 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 180 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 218 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mull-padal.c | 156 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local 194 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
|
D | 2x16c16-minmax-neon-mlal-padal.c | 217 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 271 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mlal-padal.c | 257 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 295 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mull-padal.c | 187 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local 231 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
|
D | 2x16c8-minmax-neon-mull-padal.c | 185 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local 239 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
|
D | 4x8c16-minmax-neon-mlal-padal.c | 219 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local 263 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mlal-padal.c | 314 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 358 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 240 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 306 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 2x16c8-minmax-neon-mlal-padal.c | 324 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 378 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 2x8c16-minmax-neon-mlal-padal.c | 157 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local 189 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
|
D | 2x8c8-minmax-neon-mull-padal.c | 141 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local 173 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
|
D | 2x8c8-minmax-neon-mlal-padal.c | 216 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 248 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mull-padal.c | 174 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local 212 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 198 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 236 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mull-padal.c | 207 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local 251 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
|
D | 4x8c16-minmax-neon-mlal-padal.c | 239 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local 283 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mlal-padal.c | 275 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 313 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
D | 2x16c8-minmax-neon-mull-padal.c | 201 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 255 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
|
D | 2x16c16-minmax-neon-mlal-padal.c | 233 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 287 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mlal-padal.c | 334 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 378 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
|
D | 2x16c8-minmax-neon-mlal-padal.c | 340 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 394 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
|