Home
last modified time | relevance | path

Searched refs:vpsum2x7 (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D3x8c8-minmax-rndnu-neon-mull.c208 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
210 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D3x8c16-minmax-rndnu-neon-mlal.c231 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
233 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D4x8c8-minmax-rndnu-neon-mull.c245 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
247 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D3x8c8-minmax-rndnu-neon-mlal.c309 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
311 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
D4x8c16-minmax-rndnu-neon-mlal.c276 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
278 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
D4x8c8-minmax-rndnu-neon-mlal.c372 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mlal() local
374 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mlal()
D3x16c8-minmax-rndnu-neon-mull.c334 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
336 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
D3x16c16-minmax-rndnu-neon-mlal.c381 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
383 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
D4x16c8-minmax-rndnu-neon-mull.c401 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
403 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
D3x16c8-minmax-rndnu-neon-mlal.c523 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
525 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
D4x16c16-minmax-rndnu-neon-mlal.c464 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
466 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
D4x16c8-minmax-rndnu-neon-mlal.c640 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
642 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
/external/XNNPACK/src/qs8-igemm/gen/
D3x8c8-minmax-rndnu-neon-mull.c226 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
228 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D3x8c16-minmax-rndnu-neon-mlal.c249 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
251 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D4x8c8-minmax-rndnu-neon-mull.c265 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
267 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D4x8c16-minmax-rndnu-neon-mlal.c296 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
298 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
D3x8c8-minmax-rndnu-neon-mlal.c327 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
329 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
D4x8c8-minmax-rndnu-neon-mlal.c392 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mlal() local
394 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mlal()
D3x16c8-minmax-rndnu-neon-mull.c352 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
354 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
D3x16c16-minmax-rndnu-neon-mlal.c399 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
401 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
D4x16c8-minmax-rndnu-neon-mull.c421 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
423 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
D3x16c8-minmax-rndnu-neon-mlal.c541 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
543 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
D4x16c16-minmax-rndnu-neon-mlal.c484 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
486 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
D4x16c8-minmax-rndnu-neon-mlal.c660 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
662 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()