Home
last modified time | relevance | path

Searched refs:vprod1x89c1 (Results 1 – 25 of 48) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D2x16c4s2-minmax-rndnu-neon-mlal.c180 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
183 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
185 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
282 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
284 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
D2x16c4-minmax-rndnu-neon-mlal-ld1r.c191 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
194 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
196 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
300 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
302 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
D2x16c4-minmax-rndnu-neon-mlal-dup.c187 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
190 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
192 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
294 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
296 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
D2x16c4-minmax-rndnu-neon-mlal-ld2r.c187 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
190 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
192 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
294 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
296 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
D3x16c4s2-minmax-rndnu-neon-mlal.c234 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
238 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
241 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
374 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
377 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
D3x16c4-minmax-rndnu-neon-mlal-ld2r.c243 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
247 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
250 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
389 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
392 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-dup.c243 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
247 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
250 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
389 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
392 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
D3x16c4-minmax-rndnu-neon-mlal-ld1r.c249 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
253 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
256 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
398 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
401 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
D4x16c4s2-minmax-rndnu-neon-mlal.c288 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
293 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
297 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
466 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
470 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mlal-dup.c299 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
304 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
308 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
484 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
488 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c307 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
312 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
316 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
496 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
500 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c299 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
304 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
308 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
484 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
488 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D2x16c4s2-minmax-rndnu-neon-mull.c140 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local
142 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull()
/external/XNNPACK/src/qs8-igemm/gen/
D2x16c4s2-minmax-rndnu-neon-mlal.c193 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
196 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
198 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
295 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
297 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
D2x16c4-minmax-rndnu-neon-mlal-dup.c200 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
203 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
205 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
307 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
309 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
D2x16c4-minmax-rndnu-neon-mlal-ld1r.c204 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
207 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
209 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
313 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
315 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
D2x16c4-minmax-rndnu-neon-mlal-ld2r.c200 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
203 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
205 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
307 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
309 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
D3x16c4s2-minmax-rndnu-neon-mlal.c249 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
253 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
256 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
389 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
392 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
D3x16c4-minmax-rndnu-neon-mlal-ld1r.c264 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
268 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
271 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
413 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
416 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
D3x16c4-minmax-rndnu-neon-mlal-ld2r.c258 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
262 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
265 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
404 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
407 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-dup.c258 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
262 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
265 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
404 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
407 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
D4x16c4s2-minmax-rndnu-neon-mlal.c305 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
310 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
314 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
483 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
487 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c324 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
329 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
333 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
513 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
517 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-dup.c316 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
321 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
325 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
501 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
505 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c316 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
321 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
325 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
501 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
505 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()

12