Home
last modified time | relevance | path

Searched refs:vprod3x45c0 (Results 1 – 25 of 32) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D4x8c4-minmax-rndnu-neon-mlal-ld2r.c138 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
143 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
147 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
265 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
269 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
362 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
363 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
D4x8c4-minmax-rndnu-neon-mlal-dup.c138 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
143 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
147 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
265 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
269 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
362 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
363 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
D4x8c4-minmax-rndnu-neon-mlal-ld1r.c146 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
151 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
155 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
277 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
281 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
374 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
375 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
D4x8c4s2-minmax-rndnu-neon-mlal.c128 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
133 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
137 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
248 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
252 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
D4x8c4-minmax-rndnu-neon-mull-ld2r.c121 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r() local
125 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r()
218 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r() local
219 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r()
D4x8c4-minmax-rndnu-neon-mull-dup.c121 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup() local
125 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup()
218 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup() local
219 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup()
D4x8c4-minmax-rndnu-neon-mull-ld1r.c125 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r() local
129 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r()
222 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r() local
223 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-dup.c162 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
167 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
171 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
401 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
405 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
590 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
591 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c170 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
175 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
179 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
413 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
417 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
602 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
603 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c162 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
167 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
171 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
401 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
405 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
590 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
591 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D4x16c4s2-minmax-rndnu-neon-mlal.c152 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
157 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
161 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
384 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
388 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mull-ld2r.c145 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
149 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r()
334 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
335 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r()
D4x16c4-minmax-rndnu-neon-mull-dup.c145 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
149 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup()
334 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
335 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup()
D4x16c4-minmax-rndnu-neon-mull-ld1r.c149 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
153 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r()
338 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
339 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r()
/external/XNNPACK/src/qs8-igemm/gen/
D4x8c4-minmax-rndnu-neon-mlal-ld1r.c163 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
168 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
172 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
294 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
298 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
391 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
392 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
D4x8c4-minmax-rndnu-neon-mlal-dup.c155 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
160 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
164 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
282 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
286 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
379 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
380 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
D4x8c4-minmax-rndnu-neon-mlal-ld2r.c155 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
160 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
164 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
282 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
286 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
379 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
380 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
D4x8c4s2-minmax-rndnu-neon-mlal.c145 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
150 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
154 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
265 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
269 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
D4x8c4-minmax-rndnu-neon-mull-ld2r.c138 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r() local
142 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r()
235 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r() local
236 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r()
D4x8c4-minmax-rndnu-neon-mull-dup.c138 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup() local
142 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup()
235 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup() local
236 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup()
D4x8c4-minmax-rndnu-neon-mull-ld1r.c142 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r() local
146 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r()
239 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r() local
240 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c187 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
192 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
196 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
430 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
434 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
619 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
620 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-dup.c179 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
184 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
188 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
418 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
422 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
607 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
608 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c179 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
184 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
188 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
418 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
422 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
607 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
608 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D4x16c4s2-minmax-rndnu-neon-mlal.c169 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
174 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
178 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
401 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
405 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()

12