Home
last modified time | relevance | path

Searched defs:vsum1xAB (Results 1 – 25 of 66) sorted by relevance

123

/aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/
H A D2x16c16-minmax-rndnu-neon-mlal.c220 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
279 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
H A D2x16c8-minmax-rndnu-neon-mull.c188 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
248 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
H A D2x16c8-minmax-rndnu-neon-mlal.c327 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
387 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
H A D3x16c16-minmax-rndnu-neon-mlal.c291 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
362 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
H A D3x16c8-minmax-rndnu-neon-mull.c243 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
315 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
H A D4x16c8-minmax-rndnu-neon-mull.c298 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
382 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
H A D3x16c8-minmax-rndnu-neon-mlal.c432 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
504 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
H A D4x16c16-minmax-rndnu-neon-mlal.c362 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
445 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
H A D2x16c4s2-minmax-rndnu-neon-mull.c188 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local
H A D4x16c8-minmax-rndnu-neon-mlal.c537 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
621 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
H A D2x16c4-minmax-rndnu-neon-mull-ld1r.c245 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local
H A D2x16c4-minmax-rndnu-neon-mull-ld2r.c243 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local
H A D2x16c4-minmax-rndnu-neon-mull-dup.c243 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local
/aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/
H A D2x16c8-minmax-rndnu-neon-mull.c204 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
264 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
H A D2x16c16-minmax-rndnu-neon-mlal.c236 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
295 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
H A D2x16c8-minmax-rndnu-neon-mlal.c343 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
403 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
H A D3x16c8-minmax-rndnu-neon-mull.c261 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
333 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
H A D3x16c16-minmax-rndnu-neon-mlal.c309 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
380 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
H A D4x16c8-minmax-rndnu-neon-mull.c318 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
402 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
H A D4x16c16-minmax-rndnu-neon-mlal.c382 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
465 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
H A D3x16c8-minmax-rndnu-neon-mlal.c450 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
522 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
H A D4x16c8-minmax-rndnu-neon-mlal.c557 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
641 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
H A D2x16c4s2-minmax-rndnu-neon-mull.c204 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local
H A D2x16c4-minmax-rndnu-neon-mull-ld2r.c258 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local
H A D2x16c4-minmax-rndnu-neon-mull-ld1r.c260 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local

123