Home
last modified time | relevance | path

Searched refs:vout1x0123456789ABCDEF (Results 1 – 25 of 48) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D4x16c4-minmax-neondot.c237 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() local
251 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() local
259 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
264 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
271 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
289 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
296 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
D2x16c2-minmax-neon-mull-padal-dup.c263 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() local
271 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() local
277 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
280 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
284 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
294 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
298 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
D2x16c8-minmax-neon-mull-padal.c295 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
303 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
309 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
312 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
316 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
326 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
330 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
D2x16c16-minmax-neon-mlal-padal.c327 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
335 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
341 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
344 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
348 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
358 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
362 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D2x16-minmax-neon-mlal-lane.c336 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane() local
344 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane() local
350 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
353 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
357 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
367 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
371 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
D2x16-minmax-neon-mull-addw-dup.c360 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mull_addw_dup() local
368 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mull_addw_dup() local
374 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mull_addw_dup()
377 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mull_addw_dup()
381 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mull_addw_dup()
391 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mull_addw_dup()
395 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mull_addw_dup()
D6x16c4-minmax-neondot.c313 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() local
333 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() local
343 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
350 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
359 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
383 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
393 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
D3x16c2-minmax-neon-mull-padal-dup.c345 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup() local
356 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup() local
363 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
367 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
372 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
385 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
391 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
D3x16c8-minmax-neon-mull-padal.c404 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
415 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
422 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
426 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
431 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
444 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
450 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c434 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
442 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
448 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
451 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
455 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
465 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
469 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D2x16c2-minmax-neon-mlal-padal-dup.c401 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup() local
409 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup() local
415 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
418 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
422 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
432 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
436 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
D8x16c4-minmax-neondot.c389 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() local
415 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() local
427 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
436 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
447 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
477 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
490 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
/external/XNNPACK/src/qs8-igemm/gen/
D4x16c4-minmax-neondot.c253 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() local
267 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() local
276 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
281 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
287 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
299 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
307 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
D2x16c2-minmax-neon-mull-padal-dup.c279 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() local
287 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() local
292 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
295 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
299 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
309 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
313 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
D2x16c8-minmax-neon-mull-padal.c311 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
319 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
324 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
327 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
331 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
341 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
345 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
D6x16c4-minmax-neondot.c333 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() local
353 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() local
366 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
373 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
381 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
395 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
407 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
D2x16-minmax-neon-mlal-lane.c351 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane() local
359 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane() local
364 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
367 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
371 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
381 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
385 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
D2x16-minmax-neon-mull-addw-dup.c376 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mull_addw_dup() local
384 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mull_addw_dup() local
389 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mull_addw_dup()
392 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mull_addw_dup()
396 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mull_addw_dup()
406 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mull_addw_dup()
410 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mull_addw_dup()
D2x16c16-minmax-neon-mlal-padal.c343 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
351 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
356 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
359 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
363 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
373 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
377 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c2-minmax-neon-mull-padal-dup.c363 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup() local
374 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup() local
381 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
385 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
390 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
401 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
408 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
D8x16c4-minmax-neondot.c413 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() local
439 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() local
456 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
465 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
475 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
491 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
507 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
D3x16c8-minmax-neon-mull-padal.c422 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
433 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
440 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
444 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
449 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
460 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
467 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c450 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
458 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
463 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
466 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
470 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
480 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
484 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D3x16-minmax-neon-mull-addw-dup.c494 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup() local
505 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup() local
512 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
516 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
521 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
532 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
539 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mull_addw_dup()
D3x16-minmax-neon-mlal-lane.c441 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane() local
452 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane() local
459 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
463 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
468 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
479 …1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
486 …01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()

12