Home
last modified time | relevance | path

Searched refs:vprod2xABc0 (Results 1 – 25 of 32) sorted by relevance

12

/external/XNNPACK/src/qs8-igemm/gen/
D3x16c4-minmax-rndnu-neon-mlal-ld1r.c188 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
192 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
195 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
368 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
371 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
499 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
500 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
D3x16c4-minmax-rndnu-neon-mlal-ld2r.c182 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
186 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
189 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
359 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
362 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
490 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
491 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-dup.c182 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
186 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
189 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
359 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
362 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
490 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
491 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
D3x16c4s2-minmax-rndnu-neon-mlal.c174 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
178 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
181 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
345 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
348 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c225 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
230 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
234 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
453 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
457 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
608 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
609 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-dup.c217 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
222 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
226 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
441 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
445 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
596 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
597 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c217 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
222 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
226 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
441 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
445 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
596 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
597 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mull-dup.c157 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local
160 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup()
288 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local
289 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup()
D3x16c4-minmax-rndnu-neon-mull-ld2r.c157 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local
160 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r()
288 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local
289 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r()
D3x16c4-minmax-rndnu-neon-mull-ld1r.c160 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local
163 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r()
291 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local
292 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r()
D4x16c4s2-minmax-rndnu-neon-mlal.c207 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
212 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
216 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
424 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
428 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
/external/XNNPACK/src/qs8-gemm/gen/
D3x16c4-minmax-rndnu-neon-mlal-ld2r.c167 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
171 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
174 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
344 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
347 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
475 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
476 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-dup.c167 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
171 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
174 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
344 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
347 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
475 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
476 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
D3x16c4-minmax-rndnu-neon-mlal-ld1r.c173 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
177 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
180 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
353 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
356 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
484 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
485 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
D3x16c4s2-minmax-rndnu-neon-mlal.c159 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
163 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
166 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
330 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
333 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mlal-dup.c200 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
205 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
209 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
424 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
428 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
579 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
580 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c208 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
213 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
217 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
436 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
440 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
591 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
592 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c200 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
205 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
209 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
424 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
428 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
579 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
580 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mull-ld1r.c145 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local
148 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r()
276 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local
277 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r()
D3x16c4-minmax-rndnu-neon-mull-ld2r.c142 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local
145 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r()
273 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local
274 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r()
D3x16c4-minmax-rndnu-neon-mull-dup.c142 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local
145 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup()
273 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local
274 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup()
D4x16c4s2-minmax-rndnu-neon-mlal.c190 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
195 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
199 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
407 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
411 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mull-ld2r.c168 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
172 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r()
323 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
324 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r()
D4x16c4-minmax-rndnu-neon-mull-dup.c168 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
172 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup()
323 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
324 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup()
D4x16c4-minmax-rndnu-neon-mull-ld1r.c172 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
176 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r()
327 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
328 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r()

12