Philipp Hancke c1f2a326d8 AVX2 column kernel for ScaleUVBox at 3x ... The box path costs +34% over point sampling, all of it the scalar column pass. A UV pair is 32 bits, so vpaddw keeps U and V in separate halves and vpblendd + vpermd compacts every third pair. 96 bytes in, 16 out. The +34% perf regression from the previous CL becomes +3%. Box width 3 only, AVX2 only; other widths need their own permute constants. Bug: libyuv:564129034 Change-Id: Ic20c798645be94548f74cb450e04b2e28c6dc9b2		2026-10-16 17:30:13 -07:00
..
libyuv	AVX2 column kernel for ScaleUVBox at 3x	2026-10-16 17:30:13 -07:00
libyuv.h	libyuv.h: Add scale_rgb.h and remove scale_row.h	2026-08-26 19:06:22 -07:00