Home
last modified time | relevance | path

Searched refs:vprod3x4567c2 (Results 1 – 25 of 40) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D4x8c2-minmax-rndnu-neon-mlal-ld4r.c185 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
190 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
194 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
307 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
311 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
414 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
415 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
D4x8c2-minmax-rndnu-neon-mlal-dup.c185 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
190 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
194 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
307 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
311 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
414 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
415 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
D4x8c2-minmax-rndnu-neon-mlal-ld1r.c209 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
214 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
218 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
343 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
347 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
450 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
451 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
D4x8c2-minmax-rndnu-neon-mlal-ld2r.c193 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
198 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
202 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
319 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
323 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
426 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
427 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
D4x8c2s4-minmax-rndnu-neon-mlal.c175 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
180 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
184 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
288 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
292 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
D4x8c2-minmax-rndnu-neon-mull-ld1r.c158 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local
162 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r()
265 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local
266 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r()
D4x8c2-minmax-rndnu-neon-mull-ld4r.c146 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local
150 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r()
253 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local
254 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r()
D4x8c2-minmax-rndnu-neon-mull-dup.c146 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup() local
150 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup()
253 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup() local
254 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup()
D4x8c2-minmax-rndnu-neon-mull-ld2r.c150 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r() local
154 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r()
257 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r() local
258 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r()
D4x16c2-minmax-rndnu-neon-mlal-dup.c253 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
258 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
262 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
467 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
471 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
656 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
657 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
D4x16c2-minmax-rndnu-neon-mlal-ld2r.c261 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
266 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
270 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
479 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
483 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
668 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
669 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
D4x16c2-minmax-rndnu-neon-mlal-ld4r.c253 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
258 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
262 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
467 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
471 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
656 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
657 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
/external/XNNPACK/src/qs8-igemm/gen/
D4x8c2-minmax-rndnu-neon-mlal-dup.c203 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
208 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
212 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
325 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
329 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
432 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
433 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
D4x8c2-minmax-rndnu-neon-mlal-ld1r.c227 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
232 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
236 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
361 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
365 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
468 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
469 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
D4x8c2-minmax-rndnu-neon-mlal-ld2r.c211 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
216 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
220 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
337 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
341 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
444 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
445 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
D4x8c2-minmax-rndnu-neon-mlal-ld4r.c203 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
208 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
212 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
325 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
329 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
432 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
433 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
D4x8c2s4-minmax-rndnu-neon-mlal.c192 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
197 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
201 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
305 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
309 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
D4x8c2-minmax-rndnu-neon-mull-ld4r.c163 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local
167 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r()
270 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local
271 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r()
D4x8c2-minmax-rndnu-neon-mull-ld2r.c167 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r() local
171 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r()
274 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r() local
275 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r()
D4x8c2-minmax-rndnu-neon-mull-ld1r.c175 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local
179 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r()
282 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local
283 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r()
D4x8c2-minmax-rndnu-neon-mull-dup.c163 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup() local
167 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup()
270 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup() local
271 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup()
D4x16c2-minmax-rndnu-neon-mlal-ld2r.c279 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
284 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
288 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
497 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
501 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
686 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
687 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
D4x16c2-minmax-rndnu-neon-mlal-ld1r.c295 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
300 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
304 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
521 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
525 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
710 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
711 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
D4x16c2-minmax-rndnu-neon-mlal-ld4r.c271 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
276 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
280 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
485 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
489 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
674 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
675 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
D4x16c2-minmax-rndnu-neon-mlal-dup.c271 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
276 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
280 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
485 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
489 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
674 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
675 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()

12