Home
last modified time | relevance | path

Searched refs:vpsum2x1 (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D3x8c8-minmax-rndnu-neon-mull.c199 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
202 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D3x8c16-minmax-rndnu-neon-mlal.c222 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
225 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D4x8c8-minmax-rndnu-neon-mull.c236 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
239 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D3x8c8-minmax-rndnu-neon-mlal.c300 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
303 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
D4x8c16-minmax-rndnu-neon-mlal.c267 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
270 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
D4x8c8-minmax-rndnu-neon-mlal.c363 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mlal() local
366 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mlal()
D3x16c8-minmax-rndnu-neon-mull.c325 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
328 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
D3x16c16-minmax-rndnu-neon-mlal.c372 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
375 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
D4x16c8-minmax-rndnu-neon-mull.c392 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
395 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
D3x16c8-minmax-rndnu-neon-mlal.c514 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
517 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
D4x16c16-minmax-rndnu-neon-mlal.c455 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
458 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
D4x16c8-minmax-rndnu-neon-mlal.c631 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
634 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
/external/XNNPACK/src/qs8-igemm/gen/
D3x8c8-minmax-rndnu-neon-mull.c217 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
220 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D3x8c16-minmax-rndnu-neon-mlal.c240 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
243 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D4x8c8-minmax-rndnu-neon-mull.c256 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
259 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D4x8c16-minmax-rndnu-neon-mlal.c287 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
290 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
D3x8c8-minmax-rndnu-neon-mlal.c318 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
321 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
D4x8c8-minmax-rndnu-neon-mlal.c383 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mlal() local
386 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mlal()
D3x16c8-minmax-rndnu-neon-mull.c343 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
346 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
D3x16c16-minmax-rndnu-neon-mlal.c390 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
393 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
D4x16c8-minmax-rndnu-neon-mull.c412 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
415 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
D3x16c8-minmax-rndnu-neon-mlal.c532 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
535 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
D4x16c16-minmax-rndnu-neon-mlal.c475 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
478 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
D4x16c8-minmax-rndnu-neon-mlal.c651 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
654 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()