If only parameters change (like wrapping or clut, etc.) we don't need to
rehash the data - we know it hasn't changed.
Should reduce the distance between lazy texture hashing on and off.
This will become really powerful if we add some code to the vertex decoder
to check for non-full alpha in the vertices, and set gstate_c.vertexFullAlpha if none
is found (probably want to do the reverse, set it to true and clear if any non-255 alpha is found).
Alpha testing is a performance killer on many mobile GPUs so big efforts to
avoid it can be worth it.
It's actually already pretty decent (unlike the softgpu), but there were a
few places it could use a bit of help. Speeds up things with hardware
transform off, or areas that need to use software transform.
Seems like anything from 4 is supported in 8888, most likely it just needs
to align to 16 bytes. Values above 1024 work, but e.g. 2044 seems buggy.
Fixes the map on Hexyz Force (rendered at 80 stride.)
The bezier/spline code will no longer need to handle morph and splines,
when it's finally written. This is done by pre-skinning in software and
pre-decoding to get rid of the morph.
That is, if there's no depth testing/writing, and the alpha equation will produce the same result as the alpha test (when the alpha test is >= 0 or != 0 and the blend func is safe), we turn off the alpha test.
Turns out it's one hell of a lot and we really should fix this by
checking both vertices and textures to deduce when we can just skip
alpha testing safely, probably quite often.
I care about this because alpha testing drawing is very expensive
on some mobile chips.
This is to get an idea if it would be beneficial to compile and cache display lists. Looks like it would be for some games, for example in GTA it's often 20k/80k/20k commands per frame - that is 100k executed commands per frame that could potentially just be executed instead of interpreted. Likewise in Wipeout. Of course, this means yet another cache with invalidation issues etc..
