Lines Matching refs:vpaddq_s32
349 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
350 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
351 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
352 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
353 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
354 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
355 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
356 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
357 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
358 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
359 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
360 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
361 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
362 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
363 const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
364 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, vacc1x15); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
365 const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
366 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
367 const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
368 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
369 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
370 const int32x4_t vsum2xAB = vpaddq_s32(vacc2x10, vacc2x11); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
371 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
372 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
373 const int32x4_t vsum3x01 = vpaddq_s32(vacc3x0, vacc3x1); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
374 const int32x4_t vsum3x23 = vpaddq_s32(vacc3x2, vacc3x3); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
375 const int32x4_t vsum3x45 = vpaddq_s32(vacc3x4, vacc3x5); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
376 const int32x4_t vsum3x67 = vpaddq_s32(vacc3x6, vacc3x7); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
377 const int32x4_t vsum3x89 = vpaddq_s32(vacc3x8, vacc3x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
378 const int32x4_t vsum3xAB = vpaddq_s32(vacc3x10, vacc3x11); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
379 const int32x4_t vsum3xCD = vpaddq_s32(vacc3x12, vacc3x13); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
380 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
381 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
382 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
383 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
384 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
385 int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
386 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
387 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
388 int32x4_t vacc1xCDEF = vpaddq_s32(vsum1xCD, vsum1xEF); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
389 int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
390 int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
391 int32x4_t vacc2x89AB = vpaddq_s32(vsum2x89, vsum2xAB); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
392 int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
393 int32x4_t vacc3x0123 = vpaddq_s32(vsum3x01, vsum3x23); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
394 int32x4_t vacc3x4567 = vpaddq_s32(vsum3x45, vsum3x67); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
395 int32x4_t vacc3x89AB = vpaddq_s32(vsum3x89, vsum3xAB); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
396 int32x4_t vacc3xCDEF = vpaddq_s32(vsum3xCD, vsum3xEF); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()