| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 3x8c2-minmax-rndnu-neon-mlal-ld2r.c | 122 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r() local
126 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r()
129 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r()
237 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r() local
240 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r()
322 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r() local
323 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r()
|
| D | 3x8c2-minmax-rndnu-neon-mlal-ld1r.c | 134 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r() local
138 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r()
141 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r()
255 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r() local
258 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r()
340 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r() local
341 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r()
|
| D | 3x8c2-minmax-rndnu-neon-mlal-dup.c | 116 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup() local
120 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup()
123 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup()
228 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup() local
231 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup()
313 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup() local
314 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup()
|
| D | 3x8c2-minmax-rndnu-neon-mlal-ld4r.c | 116 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r() local
120 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r()
123 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r()
228 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r() local
231 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r()
313 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r() local
314 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r()
|
| D | 3x8c2s4-minmax-rndnu-neon-mlal.c | 109 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal() local
113 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal()
116 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal()
213 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal() local
216 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal()
|
| D | 4x8c2-minmax-rndnu-neon-mlal-ld4r.c | 136 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
141 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
145 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
277 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
281 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
382 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
383 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
|
| D | 4x8c2-minmax-rndnu-neon-mlal-dup.c | 136 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
141 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
145 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
277 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
281 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
382 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
383 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
|
| D | 4x8c2-minmax-rndnu-neon-mlal-ld1r.c | 160 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
165 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
169 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
313 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
317 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
418 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
419 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
|
| D | 4x8c2-minmax-rndnu-neon-mlal-ld2r.c | 144 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
149 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
153 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
289 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
293 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
394 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
395 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
|
| D | 4x8c2s4-minmax-rndnu-neon-mlal.c | 127 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
132 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
136 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
259 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
263 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
|
| D | 3x8c2-minmax-rndnu-neon-mull-ld1r.c | 110 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld1r() local
113 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld1r()
195 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld1r() local
196 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld1r()
|
| D | 3x8c2-minmax-rndnu-neon-mull-ld4r.c | 101 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r() local
104 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r()
186 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r() local
187 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r()
|
| D | 3x8c2-minmax-rndnu-neon-mull-dup.c | 101 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup() local
104 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup()
186 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup() local
187 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup()
|
| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 3x8c2-minmax-rndnu-neon-mlal-ld1r.c | 150 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r() local
154 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r()
157 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r()
271 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r() local
274 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r()
356 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r() local
357 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r()
|
| D | 3x8c2-minmax-rndnu-neon-mlal-ld2r.c | 138 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r() local
142 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r()
145 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r()
253 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r() local
256 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r()
338 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r() local
339 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r()
|
| D | 3x8c2-minmax-rndnu-neon-mlal-ld4r.c | 132 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r() local
136 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r()
139 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r()
244 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r() local
247 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r()
329 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r() local
330 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r()
|
| D | 3x8c2-minmax-rndnu-neon-mlal-dup.c | 132 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup() local
136 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup()
139 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup()
244 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup() local
247 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup()
329 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup() local
330 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup()
|
| D | 3x8c2s4-minmax-rndnu-neon-mlal.c | 124 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal() local
128 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal()
131 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal()
228 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal() local
231 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal()
|
| D | 4x8c2-minmax-rndnu-neon-mlal-dup.c | 154 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
159 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
163 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
295 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
299 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
400 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
401 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup()
|
| D | 4x8c2-minmax-rndnu-neon-mlal-ld1r.c | 178 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
183 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
187 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
331 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
335 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
436 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
437 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r()
|
| D | 4x8c2-minmax-rndnu-neon-mlal-ld2r.c | 162 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
167 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
171 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
307 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
311 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
412 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
413 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r()
|
| D | 4x8c2-minmax-rndnu-neon-mlal-ld4r.c | 154 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
159 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
163 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
295 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
299 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
400 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
401 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r()
|
| D | 4x8c2s4-minmax-rndnu-neon-mlal.c | 144 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
149 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
153 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
276 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
280 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
|
| D | 3x8c2-minmax-rndnu-neon-mull-dup.c | 116 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup() local
119 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup()
201 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup() local
202 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup()
|
| D | 3x8c2-minmax-rndnu-neon-mull-ld2r.c | 119 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld2r() local
122 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld2r()
204 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld2r() local
205 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld2r()
|