Source

Branch

1/ @te42kyfo

@ChipsandCheese9

Is there a reason why you use the texture sampling instructions? How are they different to “normal” loads?

By the way: GCN/CDNA gets 64B/cycle only for 4B or 16B loads (so float or float4/double2). With 8B loads, its only 32B/cycle, because of only 4 addresses per cycle.