Home
last modified time | relevance | path

Searched refs:vprod3x0123c1 (Results 1 – 25 of 40) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D4x8c2-minmax-rndnu-neon-mlal-ld4r.c137 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
142 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
146 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
278 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
282 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
387 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
388 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
D4x8c2-minmax-rndnu-neon-mlal-dup.c137 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
142 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
146 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
278 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
282 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
387 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
388 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
D4x8c2-minmax-rndnu-neon-mlal-ld1r.c161 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
166 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
170 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
314 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
318 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
423 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
424 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
D4x8c2-minmax-rndnu-neon-mlal-ld2r.c145 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
150 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
154 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
290 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
294 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
399 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
400 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
D4x8c2s4-minmax-rndnu-neon-mlal.c128 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
133 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
137 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
260 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
264 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
D4x8c2-minmax-rndnu-neon-mull-ld1r.c129 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local
133 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r()
238 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local
239 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r()
D4x8c2-minmax-rndnu-neon-mull-ld4r.c117 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local
121 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r()
226 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local
227 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r()
D4x8c2-minmax-rndnu-neon-mull-dup.c117 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup() local
121 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup()
226 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup() local
227 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup()
D4x8c2-minmax-rndnu-neon-mull-ld2r.c121 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r() local
125 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r()
230 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r() local
231 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r()
D4x16c2-minmax-rndnu-neon-mlal-dup.c179 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
184 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
188 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
422 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
426 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
611 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
612 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
D4x16c2-minmax-rndnu-neon-mlal-ld2r.c187 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
192 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
196 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
434 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
438 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
623 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
624 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
D4x16c2-minmax-rndnu-neon-mlal-ld4r.c179 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
184 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
188 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
422 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
426 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
611 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
612 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
/external/XNNPACK/src/qs8-igemm/gen/
D4x8c2-minmax-rndnu-neon-mlal-dup.c155 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
160 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
164 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
296 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
300 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
405 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
406 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
D4x8c2-minmax-rndnu-neon-mlal-ld1r.c179 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
184 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
188 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
332 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
336 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
441 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
442 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
D4x8c2-minmax-rndnu-neon-mlal-ld2r.c163 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
168 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
172 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
308 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
312 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
417 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
418 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
D4x8c2-minmax-rndnu-neon-mlal-ld4r.c155 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
160 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
164 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
296 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
300 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
405 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
406 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
D4x8c2s4-minmax-rndnu-neon-mlal.c145 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
150 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
154 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
277 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
281 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
D4x8c2-minmax-rndnu-neon-mull-ld4r.c134 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local
138 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r()
243 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local
244 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r()
D4x8c2-minmax-rndnu-neon-mull-ld2r.c138 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r() local
142 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r()
247 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r() local
248 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r()
D4x8c2-minmax-rndnu-neon-mull-ld1r.c146 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local
150 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r()
255 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local
256 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r()
D4x8c2-minmax-rndnu-neon-mull-dup.c134 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup() local
138 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup()
243 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup() local
244 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup()
D4x16c2-minmax-rndnu-neon-mlal-ld2r.c205 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
210 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
214 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
452 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
456 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
641 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
642 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r()
D4x16c2-minmax-rndnu-neon-mlal-ld1r.c221 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
226 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
230 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
476 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
480 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
665 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
666 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r()
D4x16c2-minmax-rndnu-neon-mlal-ld4r.c197 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
202 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
206 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
440 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
444 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
629 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
630 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r()
D4x16c2-minmax-rndnu-neon-mlal-dup.c197 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
202 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
206 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
440 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
444 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()
629 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
630 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup()

12