Lines Matching refs:int32x2_t
315 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
316 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
317 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
318 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
319 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
320 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
322 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
323 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
324 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
325 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
326 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
327 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
329 const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
330 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
331 const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
332 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
333 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
334 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
336 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
337 const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
338 const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
339 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
340 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
341 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
343 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
344 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
345 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
346 const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
347 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
348 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
350 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
351 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
352 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
353 const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
354 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
355 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
357 const int32x2_t vpsum1x8 = vadd_s32(vget_low_s32(vacc1x8), vget_high_s32(vacc1x8)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
358 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
359 const int32x2_t vpsum1xA = vadd_s32(vget_low_s32(vacc1x10), vget_high_s32(vacc1x10)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
360 const int32x2_t vpsum1xB = vadd_s32(vget_low_s32(vacc1x11), vget_high_s32(vacc1x11)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
361 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
362 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
364 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
365 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
366 const int32x2_t vpsum1xE = vadd_s32(vget_low_s32(vacc1x14), vget_high_s32(vacc1x14)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
367 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
368 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
369 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
371 const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
372 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
373 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
374 const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
375 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
376 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
378 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
379 const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
380 const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
381 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
382 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
383 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
385 const int32x2_t vpsum2x8 = vadd_s32(vget_low_s32(vacc2x8), vget_high_s32(vacc2x8)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
386 const int32x2_t vpsum2x9 = vadd_s32(vget_low_s32(vacc2x9), vget_high_s32(vacc2x9)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
387 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
388 const int32x2_t vpsum2xB = vadd_s32(vget_low_s32(vacc2x11), vget_high_s32(vacc2x11)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
389 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
390 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
392 const int32x2_t vpsum2xC = vadd_s32(vget_low_s32(vacc2x12), vget_high_s32(vacc2x12)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
393 const int32x2_t vpsum2xD = vadd_s32(vget_low_s32(vacc2x13), vget_high_s32(vacc2x13)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
394 const int32x2_t vpsum2xE = vadd_s32(vget_low_s32(vacc2x14), vget_high_s32(vacc2x14)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
395 const int32x2_t vpsum2xF = vadd_s32(vget_low_s32(vacc2x15), vget_high_s32(vacc2x15)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
396 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
397 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()