S32A_D565_Blend has a bug where it can overflow the final 5/6/5 values. This

S32A_D565_Blend has a bug where it can overflow the final 5/6/5 values. This
happens because when we try to merge an 8-bit and a 5/6/5-bit value into the
same "space", we had been truncing down the 8-bit value. This is wrong...

e.g.

src 8-bit value: 0xF8
dst 5-bit value: 0x1F

If we trunc down, to move everyone into the 5-bit space, we get

src: 0x1F
dst: 0x1F

But now we *think* that our src is fully saturated (which it wasn't.. 0xF8)
and subsequently we've violated our premul requirement that all components
are <= alpha.

The fix is to perform the math in 8-bit space instead, so we promote the 5/6/5
values UP, do the math, and then shift them back down.

[bsalomon, 2012-10-11 16:56:48]

On 2012/10/11 15:56:56, reed1 wrote:

LGTM