Home
last modified time | relevance | path

Searched refs:vpsum1x5 (Results 1 – 25 of 36) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D2x8c8-minmax-neon-mull-padal.c154 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D2x8c16-minmax-neon-mlal-padal.c170 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D2x8c8-minmax-neon-mlal-padal.c229 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D3x8c8-minmax-neon-mull-padal.c191 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D3x8c16-minmax-neon-mlal-padal.c215 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D4x8c8-minmax-neon-mull-padal.c228 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D2x16c8-minmax-neon-mull-padal.c236 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D3x8c8-minmax-neon-mlal-padal.c292 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D2x16c16-minmax-neon-mlal-padal.c268 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D4x8c16-minmax-neon-mlal-padal.c260 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D4x8c8-minmax-neon-mlal-padal.c355 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D3x16c8-minmax-neon-mull-padal.c303 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D2x16c8-minmax-neon-mlal-padal.c375 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
/external/XNNPACK/src/qs8-igemm/gen/
D2x8c8-minmax-neon-mull-padal.c170 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D2x8c16-minmax-neon-mlal-padal.c186 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D2x8c8-minmax-neon-mlal-padal.c245 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D3x8c8-minmax-neon-mull-padal.c209 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D3x8c16-minmax-neon-mlal-padal.c233 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D4x8c8-minmax-neon-mull-padal.c248 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D2x16c8-minmax-neon-mull-padal.c252 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D3x8c8-minmax-neon-mlal-padal.c310 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D4x8c16-minmax-neon-mlal-padal.c280 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D2x16c16-minmax-neon-mlal-padal.c284 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()
D3x16c8-minmax-neon-mull-padal.c321 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
324 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c391 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(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()

12