| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 8x16c4-minmax-neondot.c | 395 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
396 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
397 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
398 …const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
399 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
400 …const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
401 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
402 …const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
403 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
404 …const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
[all …]
|
| D | 6x16c4-minmax-neondot.c | 319 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
320 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
321 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
322 …const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
323 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
324 …const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
325 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
326 …const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
327 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
328 …const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
[all …]
|
| D | 4x16c4-minmax-neondot.c | 243 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
244 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
245 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
246 …const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
247 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
248 …const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
249 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
250 …const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), v… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
257 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
258 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0x… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
[all …]
|
| D | 8x8c4-minmax-neondot.c | 277 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
278 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
279 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
280 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
281 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
282 …const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
283 …const int16x8_t vacc6x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x0123), vacc6x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
284 …const int16x8_t vacc7x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x0123), vacc7x4567), v… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
291 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
292 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
[all …]
|
| D | 6x8c4-minmax-neondot.c | 229 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
230 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
231 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
232 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
233 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
234 …const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), v… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
240 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
241 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
242 …const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
243 …const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x… in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
[all …]
|
| D | 4x8c4-minmax-neondot.c | 181 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
182 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
183 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
184 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
189 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
190 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
191 …const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
192 …const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x… in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
|
| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 8x16c4-minmax-neondot.c | 371 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
372 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
373 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
374 …const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
375 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
376 …const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
377 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
378 …const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
379 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
380 …const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
[all …]
|
| D | 6x16c4-minmax-neondot.c | 299 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
300 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
301 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
302 …const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
303 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
304 …const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
305 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
306 …const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
307 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
308 …const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
[all …]
|
| D | 4x16c4-minmax-neondot.c | 227 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
228 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
229 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
230 …const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
231 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
232 …const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
233 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
234 …const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
241 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
242 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
[all …]
|
| D | 8x8c4-minmax-neondot.c | 253 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
254 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
255 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
256 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
257 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
258 …const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
259 …const int16x8_t vacc6x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x0123), vacc6x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
260 …const int16x8_t vacc7x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x0123), vacc7x4567), v… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
267 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
268 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
[all …]
|
| D | 6x8c4-minmax-neondot.c | 209 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
210 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
211 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
212 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
213 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
214 …const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), v… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
220 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
221 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
222 …const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
223 …const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x… in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
[all …]
|
| D | 4x8c4-minmax-neondot.c | 165 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
166 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
167 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
168 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
173 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
174 …const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
175 …const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
176 …const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x… in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
|
| D | 4x16c2-minmax-neon-mull-padal-dup.c | 417 …const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
418 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
419 …const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
420 …const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
421 …const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
422 …const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
423 …const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), v… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
424 …const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), v… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
431 …const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
432 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0x… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
[all …]
|
| /external/libhevc/common/arm/ |
| D | ihevc_quant_iquant_ssd_neon_intr.c | 203 q0 = vqmovn_s32(qtmp_0); in ihevc_quant_iquant_ssd_flat_scale_mat_neon()
204 q1 = vqmovn_s32(qtmp_1); in ihevc_quant_iquant_ssd_flat_scale_mat_neon()
205 q2 = vqmovn_s32(qtmp_2); in ihevc_quant_iquant_ssd_flat_scale_mat_neon()
206 q3 = vqmovn_s32(qtmp_3); in ihevc_quant_iquant_ssd_flat_scale_mat_neon()
261 iq0 = vqmovn_s32(iqtmp_0); in ihevc_quant_iquant_ssd_flat_scale_mat_neon()
262 iq1 = vqmovn_s32(iqtmp_1); in ihevc_quant_iquant_ssd_flat_scale_mat_neon()
263 iq2 = vqmovn_s32(iqtmp_2); in ihevc_quant_iquant_ssd_flat_scale_mat_neon()
264 iq3 = vqmovn_s32(iqtmp_3); in ihevc_quant_iquant_ssd_flat_scale_mat_neon()
480 q_00 = vcombine_s16(vqmovn_s32(q_tmp0), vqmovn_s32(q_tmp1)); in ihevc_q_iq_ssd_flat_scale_mat_var_rnd_fact_neon()
481 q_01 = vcombine_s16(vqmovn_s32(q_tmp2), vqmovn_s32(q_tmp3)); in ihevc_q_iq_ssd_flat_scale_mat_var_rnd_fact_neon()
[all …]
|
| /external/XNNPACK/src/qs8-vaddc/gen/ |
| D | minmax-neon-ld64-x32.c | 77 …const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32()
78 …const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32()
79 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32()
80 …const int16x8_t vaccOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32()
109 …const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32()
|
| D | minmax-neon-ld64-x24.c | 69 …const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x24()
70 …const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x24()
71 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x24()
100 …const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x24()
|
| D | minmax-neon-ld64-x16.c | 61 …const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16()
62 …const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16()
87 …const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16()
|
| /external/XNNPACK/src/qu8-vadd/ |
| D | minmax-neon.c | 87 …const int16x8_t vacc0 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0_lo), vacc0_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon()
88 …const int16x8_t vacc1 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1_lo), vacc1_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon()
89 …const int16x8_t vacc2 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2_lo), vacc2_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon()
90 …const int16x8_t vacc3 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3_lo), vacc3_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon()
140 …const int16x8_t vacc0 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0_lo), vqmovn_s32(vacc0_hi)), vy_ze… in xnn_qu8_vadd_minmax_ukernel__neon()
141 …const int16x8_t vacc1 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1_lo), vqmovn_s32(vacc1_hi)), vy_ze… in xnn_qu8_vadd_minmax_ukernel__neon()
182 const int16x8_t vacc = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc_lo), vacc_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon()
184 …const int16x8_t vacc = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi)), vy_zero_… in xnn_qu8_vadd_minmax_ukernel__neon()
225 const int16x8_t vacc = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc_lo), vacc_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon()
227 …const int16x8_t vacc = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi)), vy_zero_… in xnn_qu8_vadd_minmax_ukernel__neon()
|
| /external/XNNPACK/src/qu8-requantization/ |
| D | q31-neon.c | 83 …const int16x8_t xy_packed = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(x_scaled), y_scaled), vzero_poin… in xnn_qu8_requantize_q31__neon()
84 …const int16x8_t zw_packed = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(z_scaled), w_scaled), vzero_poin… in xnn_qu8_requantize_q31__neon()
87 …const int16x8_t xy_packed = vqaddq_s16(vcombine_s16(vqmovn_s32(x_scaled), vqmovn_s32(y_scaled)), v… in xnn_qu8_requantize_q31__neon()
88 …const int16x8_t zw_packed = vqaddq_s16(vcombine_s16(vqmovn_s32(z_scaled), vqmovn_s32(w_scaled)), v… in xnn_qu8_requantize_q31__neon()
|
| /external/XNNPACK/src/qs8-requantization/ |
| D | q31-neon.c | 83 …const int16x8_t xy_packed = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(x_scaled), y_scaled), vzero_poin… in xnn_qs8_requantize_q31__neon()
84 …const int16x8_t zw_packed = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(z_scaled), w_scaled), vzero_poin… in xnn_qs8_requantize_q31__neon()
87 …const int16x8_t xy_packed = vqaddq_s16(vcombine_s16(vqmovn_s32(x_scaled), vqmovn_s32(y_scaled)), v… in xnn_qs8_requantize_q31__neon()
88 …const int16x8_t zw_packed = vqaddq_s16(vcombine_s16(vqmovn_s32(z_scaled), vqmovn_s32(w_scaled)), v… in xnn_qs8_requantize_q31__neon()
|
| /external/XNNPACK/src/qs8-vadd/gen/ |
| D | minmax-neon-ld64-x32.c | 89 …const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32()
90 …const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32()
91 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32()
92 …const int16x8_t vaccOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32()
126 …const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32()
|
| D | minmax-neon-ld64-x24.c | 77 …const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24()
78 …const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24()
79 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24()
113 …const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24()
|
| D | minmax-neon-ld64-x16.c | 65 …const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x16()
66 …const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x16()
96 …const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x16()
|
| /external/gemmlowp/internal/ |
| D | output_neon.h | 38 int16x4_t res_16 = vqmovn_s32(input.reg[0]);
58 vcombine_s16(vqmovn_s32(input.reg[0]), vqmovn_s32(input.reg[1]));
77 vcombine_s16(vqmovn_s32(input.reg[0]), vqmovn_s32(input.reg[1]));
79 vcombine_s16(vqmovn_s32(input.reg[2]), vqmovn_s32(input.reg[3]));
100 res_16[i] = vcombine_s16(vqmovn_s32(input.reg[2 * i]),
101 vqmovn_s32(input.reg[2 * i + 1]));
122 int16x4_t res_16 = vqmovn_s32(input.reg[0]);
142 vcombine_s16(vqmovn_s32(input.reg[0]), vqmovn_s32(input.reg[1]));
161 vcombine_s16(vqmovn_s32(input.reg[0]), vqmovn_s32(input.reg[1]));
163 vcombine_s16(vqmovn_s32(input.reg[2]), vqmovn_s32(input.reg[3]));
[all …]
|
| /external/tensorflow/tensorflow/lite/kernels/internal/optimized/integer_ops/ |
| D | mul.h | 108 const auto p1_narrowed = vqmovn_s32(p1); in MulElementwise()
109 const auto p2_narrowed = vqmovn_s32(p2); in MulElementwise()
110 const auto p3_narrowed = vqmovn_s32(p3); in MulElementwise()
111 const auto p4_narrowed = vqmovn_s32(p4); in MulElementwise()
201 const auto p1_narrowed = vqmovn_s32(p1); in MulSimpleBroadcast()
202 const auto p2_narrowed = vqmovn_s32(p2); in MulSimpleBroadcast()
203 const auto p3_narrowed = vqmovn_s32(p3); in MulSimpleBroadcast()
204 const auto p4_narrowed = vqmovn_s32(p4); in MulSimpleBroadcast()
|