Home
last modified time | relevance | path

Searched refs:vsum1x67 (Results 1 – 25 of 172) sorted by relevance

1234567

/external/XNNPACK/src/qs8-gemm/gen/
D2x8c8-minmax-rndnu-neon-mull.c126 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
131 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull()
159 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
160 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull()
D2x8c16-minmax-rndnu-neon-mlal.c142 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
146 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
174 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
175 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
D3x8c8-minmax-rndnu-neon-mull.c157 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
166 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
196 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
197 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D2x8c8-minmax-fp32-neonv8-mlal.c202 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
207 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
235 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
236 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c201 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
206 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
234 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
235 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c201 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
206 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
234 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
235 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D3x8c16-minmax-rndnu-neon-mlal.c181 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
189 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
219 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
220 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D4x8c8-minmax-rndnu-neon-mull.c188 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
201 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
233 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
234 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D2x16c8-minmax-rndnu-neon-mull.c186 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
197 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
241 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
242 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
D2x8c4s2-minmax-rndnu-neon-mull.c127 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4s2__neon_mull() local
128 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4s2__neon_mull()
D3x8c8-minmax-rndnu-neon-mlal.c258 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
267 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
297 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
298 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
/external/XNNPACK/src/qs8-igemm/gen/
D2x8c8-minmax-rndnu-neon-mull.c142 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
147 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull()
175 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
176 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull()
D2x8c16-minmax-rndnu-neon-mlal.c158 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
162 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
190 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
191 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c217 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
222 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
250 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
251 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neon-mlal.c217 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
222 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
250 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
251 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c218 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
223 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
251 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
252 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D3x8c8-minmax-rndnu-neon-mull.c175 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
184 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
214 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
215 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D3x8c16-minmax-rndnu-neon-mlal.c199 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
207 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
237 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
238 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D2x16c8-minmax-rndnu-neon-mull.c202 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
213 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
257 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
258 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
D4x8c8-minmax-rndnu-neon-mull.c208 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
221 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
253 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
254 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D2x8c4s2-minmax-rndnu-neon-mull.c143 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mull() local
144 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mull()
/external/XNNPACK/src/qc8-gemm/gen/
D2x8c8-minmax-fp32-neon-mlal.c201 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
206 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
234 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
235 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c202 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
207 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
235 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
236 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
/external/XNNPACK/src/qc8-igemm/gen/
D2x8c8-minmax-fp32-neonv8-mlal.c218 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
223 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
251 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
252 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c217 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
222 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
250 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
251 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 ); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()

1234567