Home
last modified time | relevance | path

Searched refs:vprod1xABc1 (Results 1 – 25 of 48) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D2x16c4s2-minmax-rndnu-neon-mlal.c187 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
190 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
192 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
286 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
288 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
D2x16c4-minmax-rndnu-neon-mlal-ld1r.c198 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
201 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
203 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
304 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
306 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
D2x16c4-minmax-rndnu-neon-mlal-dup.c194 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
197 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
199 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
298 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
300 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
D2x16c4-minmax-rndnu-neon-mlal-ld2r.c194 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
197 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
199 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
298 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
300 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
D3x16c4s2-minmax-rndnu-neon-mlal.c244 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
248 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
251 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
380 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
383 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
D3x16c4-minmax-rndnu-neon-mlal-ld2r.c253 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
257 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
260 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
395 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
398 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-dup.c253 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
257 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
260 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
395 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
398 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
D3x16c4-minmax-rndnu-neon-mlal-ld1r.c259 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
263 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
266 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
404 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
407 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
D4x16c4s2-minmax-rndnu-neon-mlal.c301 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
306 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
310 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
474 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
478 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mlal-dup.c312 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
317 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
321 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
492 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
496 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c320 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
325 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
329 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
504 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
508 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c312 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
317 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
321 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
492 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
496 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D2x16c4s2-minmax-rndnu-neon-mull.c144 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local
146 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull()
/external/XNNPACK/src/qs8-igemm/gen/
D2x16c4s2-minmax-rndnu-neon-mlal.c200 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
203 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
205 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
299 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
301 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
D2x16c4-minmax-rndnu-neon-mlal-dup.c207 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
210 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
212 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
311 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
313 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
D2x16c4-minmax-rndnu-neon-mlal-ld1r.c211 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
214 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
216 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
317 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
319 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
D2x16c4-minmax-rndnu-neon-mlal-ld2r.c207 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
210 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
212 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
311 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
313 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
D3x16c4s2-minmax-rndnu-neon-mlal.c259 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
263 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
266 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
395 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
398 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
D3x16c4-minmax-rndnu-neon-mlal-ld1r.c274 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
278 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
281 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
419 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
422 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
D3x16c4-minmax-rndnu-neon-mlal-ld2r.c268 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
272 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
275 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
410 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
413 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-dup.c268 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
272 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
275 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
410 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
413 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
D4x16c4s2-minmax-rndnu-neon-mlal.c318 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
323 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
327 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
491 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
495 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c337 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
342 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
346 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
521 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
525 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-dup.c329 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
334 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
338 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
509 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
513 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c329 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
334 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
338 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
509 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
513 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()

12