Home
last modified time | relevance | path

Searched refs:vprod1xCDEFc1 (Results 1 – 25 of 60) sorted by relevance

123

/external/XNNPACK/src/qs8-igemm/gen/
D2x16c2-minmax-rndnu-neon-mlal-ld1r.c168 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
171 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
173 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
306 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
308 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
401 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
402 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
D2x16c2-minmax-rndnu-neon-mlal-ld4r.c156 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
159 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
161 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
288 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
290 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
383 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
384 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
D2x16c2-minmax-rndnu-neon-mlal-dup.c156 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
159 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
161 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
288 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
290 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
383 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
384 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
D2x16c2-minmax-rndnu-neon-mlal-ld2r.c160 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
163 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
165 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
294 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
296 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
389 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
390 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
D2x16c2s4-minmax-rndnu-neon-mlal.c150 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local
153 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal()
155 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal()
275 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local
277 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal()
D3x16c2-minmax-rndnu-neon-mlal-ld4r.c195 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
199 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
202 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
375 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
378 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
500 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
501 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
D3x16c2-minmax-rndnu-neon-mlal-dup.c195 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
199 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
202 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
375 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
378 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
500 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
501 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
D3x16c2-minmax-rndnu-neon-mlal-ld1r.c213 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
217 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
220 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
402 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
405 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
527 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
528 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
D3x16c2-minmax-rndnu-neon-mlal-ld2r.c201 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
205 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
208 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
384 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
387 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
509 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
510 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
D3x16c2s4-minmax-rndnu-neon-mlal.c187 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local
191 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal()
194 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal()
359 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local
362 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal()
D2x16c2-minmax-rndnu-neon-mull-ld1r.c136 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
138 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
231 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
232 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
D2x16c2-minmax-rndnu-neon-mull-dup.c130 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
132 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
225 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
226 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
D2x16c2-minmax-rndnu-neon-mull-ld4r.c130 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
132 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
225 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
226 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
/external/XNNPACK/src/qs8-gemm/gen/
D2x16c2-minmax-rndnu-neon-mlal-ld2r.c146 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
149 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
151 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
280 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
282 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
375 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
376 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
D2x16c2-minmax-rndnu-neon-mlal-ld1r.c154 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
157 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
159 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
292 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
294 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
387 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
388 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
D2x16c2-minmax-rndnu-neon-mlal-dup.c142 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
145 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
147 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
274 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
276 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
369 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
370 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
D2x16c2-minmax-rndnu-neon-mlal-ld4r.c142 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
145 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
147 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
274 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
276 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
369 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
370 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
D2x16c2s4-minmax-rndnu-neon-mlal.c137 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local
140 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal()
142 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal()
262 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local
264 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal()
D3x16c2-minmax-rndnu-neon-mlal-dup.c179 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
183 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
186 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
359 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
362 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
484 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
485 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
D3x16c2-minmax-rndnu-neon-mlal-ld4r.c179 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
183 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
186 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
359 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
362 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
484 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
485 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
D3x16c2-minmax-rndnu-neon-mlal-ld2r.c185 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
189 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
192 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
368 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
371 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
493 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
494 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
D3x16c2-minmax-rndnu-neon-mlal-ld1r.c197 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
201 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
204 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
386 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
389 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
511 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
512 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
D3x16c2s4-minmax-rndnu-neon-mlal.c172 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local
176 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal()
179 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal()
344 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local
347 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal()
D2x16c2-minmax-rndnu-neon-mull-dup.c117 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
119 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
212 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
213 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
D2x16c2-minmax-rndnu-neon-mull-ld1r.c123 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
125 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
218 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
219 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()

123