Home
last modified time | relevance | path

Searched defs:vsum1x01 (Results 1 – 25 of 36) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D2x8c8-minmax-neon-mull-padal.c123 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
150 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
D2x8c16-minmax-neon-mlal-padal.c139 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
166 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
D2x8c8-minmax-neon-mlal-padal.c198 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
225 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
D3x8c16-minmax-neon-mlal-padal.c178 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
211 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
D3x8c8-minmax-neon-mull-padal.c154 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
187 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
D2x16c16-minmax-neon-mlal-padal.c215 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
264 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
D3x8c8-minmax-neon-mlal-padal.c255 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
288 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
D4x8c8-minmax-neon-mull-padal.c185 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
224 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
D2x16c8-minmax-neon-mull-padal.c183 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
232 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
D4x8c16-minmax-neon-mlal-padal.c217 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
256 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
D4x8c8-minmax-neon-mlal-padal.c312 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
351 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
D3x16c8-minmax-neon-mull-padal.c238 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
299 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
D2x16c8-minmax-neon-mlal-padal.c322 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
371 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
/external/XNNPACK/src/qs8-igemm/gen/
D2x8c16-minmax-neon-mlal-padal.c155 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
182 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
D2x8c8-minmax-neon-mull-padal.c139 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
166 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
D2x8c8-minmax-neon-mlal-padal.c214 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
241 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
D3x8c8-minmax-neon-mull-padal.c172 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
205 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
D3x8c16-minmax-neon-mlal-padal.c196 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
229 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
D4x8c8-minmax-neon-mull-padal.c205 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
244 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
D4x8c16-minmax-neon-mlal-padal.c237 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
276 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
D3x8c8-minmax-neon-mlal-padal.c273 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
306 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
D2x16c8-minmax-neon-mull-padal.c199 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
248 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
D2x16c16-minmax-neon-mlal-padal.c231 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
280 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
D4x8c8-minmax-neon-mlal-padal.c332 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
371 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
D2x16c8-minmax-neon-mlal-padal.c338 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
387 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local

12