Home
last modified time | relevance | path

Searched refs:vprod2x67c0 (Results 1 – 25 of 64) sorted by relevance

123

/external/XNNPACK/src/qs8-gemm/gen/
D3x8c4-minmax-rndnu-neon-mlal-ld2r.c127 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
131 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
134 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
224 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
227 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
295 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
296 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
D3x8c4-minmax-rndnu-neon-mlal-ld1r.c133 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
137 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
140 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
233 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
236 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
304 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
305 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
D3x8c4-minmax-rndnu-neon-mlal-dup.c127 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
131 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
134 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
224 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
227 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
295 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
296 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
D3x8c4s2-minmax-rndnu-neon-mlal.c119 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local
123 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
126 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
210 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local
213 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
D4x8c4-minmax-rndnu-neon-mlal-ld2r.c150 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
155 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
159 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
272 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
276 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
355 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
356 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
D4x8c4-minmax-rndnu-neon-mlal-dup.c150 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
155 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
159 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
272 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
276 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
355 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
356 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
D4x8c4-minmax-rndnu-neon-mlal-ld1r.c158 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
163 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
167 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
284 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
288 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
367 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
368 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
D4x8c4s2-minmax-rndnu-neon-mlal.c140 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
145 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
149 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
255 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
259 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
D3x8c4-minmax-rndnu-neon-mull-ld1r.c113 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local
116 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
184 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local
185 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
D3x8c4-minmax-rndnu-neon-mull-dup.c110 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local
113 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
181 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local
182 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
D3x8c4-minmax-rndnu-neon-mull-ld2r.c110 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local
113 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
181 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local
182 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
/external/XNNPACK/src/qs8-igemm/gen/
D3x8c4-minmax-rndnu-neon-mlal-dup.c142 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
146 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
149 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
239 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
242 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
310 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
311 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
D3x8c4-minmax-rndnu-neon-mlal-ld1r.c148 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
152 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
155 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
248 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
251 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
319 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
320 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
D3x8c4-minmax-rndnu-neon-mlal-ld2r.c142 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
146 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
149 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
239 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
242 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
310 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
311 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
D3x8c4s2-minmax-rndnu-neon-mlal.c134 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local
138 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
141 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
225 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local
228 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
D4x8c4-minmax-rndnu-neon-mlal-ld1r.c175 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
180 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
184 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
301 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
305 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
384 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
385 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
D4x8c4-minmax-rndnu-neon-mlal-dup.c167 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
172 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
176 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
289 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
293 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
372 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
373 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
D4x8c4-minmax-rndnu-neon-mlal-ld2r.c167 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
172 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
176 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
289 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
293 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
372 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
373 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
D4x8c4s2-minmax-rndnu-neon-mlal.c157 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
162 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
166 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
272 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
276 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
D3x8c4-minmax-rndnu-neon-mull-ld1r.c128 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local
131 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
199 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local
200 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
D3x8c4-minmax-rndnu-neon-mull-dup.c125 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local
128 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
196 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local
197 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
D3x8c4-minmax-rndnu-neon-mull-ld2r.c125 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local
128 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
196 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local
197 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-ld1r.c168 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
172 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
175 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
356 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
359 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
495 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
496 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
D3x16c4-minmax-rndnu-neon-mlal-ld2r.c162 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
166 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
169 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
347 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
350 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
486 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
487 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-dup.c162 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
166 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
169 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
347 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
350 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
486 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
487 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()

123