Home
last modified time | relevance | path

Searched refs:vprod3x4567c3 (Results 1 – 25 of 40) sorted by relevance

12

/external/XNNPACK/src/qs8-igemm/gen/
D4x8c2s4-minmax-rndnu-neon-mlal.c226 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
231 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
235 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
325 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
329 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
D4x8c2-minmax-rndnu-neon-mlal-dup.c238 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
243 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
247 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
346 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
350 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
D4x8c2-minmax-rndnu-neon-mlal-ld1r.c262 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
267 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
271 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
382 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
386 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
D4x8c2-minmax-rndnu-neon-mlal-ld2r.c246 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
251 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
255 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
358 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
362 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
D4x8c2-minmax-rndnu-neon-mlal-ld4r.c238 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
243 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
247 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
346 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
350 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
D4x16c2s4-minmax-rndnu-neon-mlal.c320 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local
325 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal()
329 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal()
501 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local
505 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal()
D4x8c2s4-minmax-rndnu-neon-mull.c175 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull() local
179 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull()
D4x16c2-minmax-rndnu-neon-mlal-ld2r.c340 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
345 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
349 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
534 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
538 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
D4x16c2-minmax-rndnu-neon-mlal-ld1r.c356 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
361 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
365 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
558 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
562 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
D4x16c2-minmax-rndnu-neon-mlal-ld4r.c332 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
337 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
341 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
522 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
526 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
D4x16c2-minmax-rndnu-neon-mlal-dup.c332 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
337 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
341 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
522 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
526 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
D4x8c2-minmax-rndnu-neon-mull-ld4r.c184 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local
188 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r()
/external/XNNPACK/src/qs8-gemm/gen/
D4x8c2s4-minmax-rndnu-neon-mlal.c209 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
214 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
218 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
308 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
312 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
D4x8c2-minmax-rndnu-neon-mlal-ld4r.c220 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
225 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
229 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
328 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
332 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
D4x8c2-minmax-rndnu-neon-mlal-dup.c220 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
225 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
229 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
328 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
332 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
D4x8c2-minmax-rndnu-neon-mlal-ld1r.c244 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
249 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
253 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
364 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
368 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
D4x8c2-minmax-rndnu-neon-mlal-ld2r.c228 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
233 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
237 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
340 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
344 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
D4x16c2s4-minmax-rndnu-neon-mlal.c303 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local
308 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal()
312 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal()
484 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local
488 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal()
D4x8c2s4-minmax-rndnu-neon-mull.c158 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull() local
162 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull()
D4x16c2-minmax-rndnu-neon-mlal-dup.c314 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
319 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
323 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
504 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
508 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
D4x16c2-minmax-rndnu-neon-mlal-ld2r.c322 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
327 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
331 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
516 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
520 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
D4x16c2-minmax-rndnu-neon-mlal-ld4r.c314 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
319 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
323 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
504 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
508 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
D4x16c2-minmax-rndnu-neon-mlal-ld1r.c338 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
343 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
347 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
540 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
544 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
D4x8c2-minmax-rndnu-neon-mull-ld1r.c179 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local
183 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r()
D4x8c2-minmax-rndnu-neon-mull-ld4r.c167 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local
171 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r()

12