Home
last modified time | relevance | path

Searched refs:vprod3xCDc1 (Results 1 – 16 of 16) sorted by relevance

/external/XNNPACK/src/qs8-igemm/gen/
D4x16c4s2-minmax-rndnu-neon-mlal.c333 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
338 vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
342 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
501 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
505 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c352 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
357 vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
361 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
531 const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
535 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-dup.c344 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
349 vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
353 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
519 const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
523 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c344 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
349 vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
353 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
519 const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
523 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D4x16c4s2-minmax-rndnu-neon-mull.c255 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull() local
259 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull()
D4x16c4-minmax-rndnu-neon-mull-ld2r.c263 const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
267 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r()
D4x16c4-minmax-rndnu-neon-mull-ld1r.c267 const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
271 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r()
D4x16c4-minmax-rndnu-neon-mull-dup.c263 const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
267 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup()
/external/XNNPACK/src/qs8-gemm/gen/
D4x16c4s2-minmax-rndnu-neon-mlal.c316 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
321 vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
325 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
484 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
488 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mlal-dup.c327 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
332 vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
336 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
502 const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
506 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c335 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
340 vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
344 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
514 const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
518 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c327 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
332 vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
336 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
502 const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
506 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D4x16c4s2-minmax-rndnu-neon-mull.c238 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull() local
242 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull()
D4x16c4-minmax-rndnu-neon-mull-ld2r.c246 const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
250 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r()
D4x16c4-minmax-rndnu-neon-mull-dup.c246 const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
250 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup()
D4x16c4-minmax-rndnu-neon-mull-ld1r.c250 const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
254 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r()