Home
last modified time | relevance | path

Searched refs:vprod3xEFc0 (Results 1 – 16 of 16) sorted by relevance

/external/XNNPACK/src/qs8-igemm/gen/
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c252 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
257 vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
261 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
470 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
474 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
629 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
630 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-dup.c244 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
249 vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
253 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
458 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
462 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
617 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
618 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c244 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
249 vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
253 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
458 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
462 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
617 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
618 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D4x16c4s2-minmax-rndnu-neon-mlal.c234 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
239 vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
243 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
441 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
445 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mull-ld2r.c202 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
206 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r()
361 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
362 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r()
D4x16c4-minmax-rndnu-neon-mull-ld1r.c206 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
210 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r()
365 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
366 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r()
D4x16c4-minmax-rndnu-neon-mull-dup.c202 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
206 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup()
361 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
362 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup()
D4x16c4s2-minmax-rndnu-neon-mull.c195 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull() local
199 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull()
/external/XNNPACK/src/qs8-gemm/gen/
D4x16c4-minmax-rndnu-neon-mlal-dup.c227 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
232 vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
236 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
441 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
445 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
600 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
601 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c235 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
240 vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
244 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
453 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
457 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
612 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
613 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c227 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
232 vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
236 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
441 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
445 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
600 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
601 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D4x16c4s2-minmax-rndnu-neon-mlal.c217 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
222 vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
226 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
424 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
428 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mull-ld2r.c185 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
189 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r()
344 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
345 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r()
D4x16c4-minmax-rndnu-neon-mull-dup.c185 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
189 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup()
344 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
345 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup()
D4x16c4-minmax-rndnu-neon-mull-ld1r.c189 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
193 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r()
348 const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
349 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r()
D4x16c4s2-minmax-rndnu-neon-mull.c178 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull() local
182 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull()