We will emit this packet at startup (dirty == ~0), even if we haven't
had the state tracker call into brw_set_blend_color() yet.
This way is a little more efficient also.
I'm emitting this in two places now, to the data presented
for upload and also in the delta field of the reloc struct.
Probably want to remove the delta field and just pull the
background from the key.
Makes it easier to dump as we get all of the information
about the upload in a single hit.
Opens the window to simplification in the driver if these
relocation arrays can be maintained statically rather than
being recreated whenever we check for a new upload.
Still needs some cleanup to avoid uglyness introduced with the
delta values.
The winsys now inserts the presumed offset into referring buffers from
inside of bo_emit_reloc(). Remove the many locally coded places where
this was happening in the driver and eliminate the worry of getting it
wrong.
No longer need to expose offset values to the driver at all, so no need
to worry about what to do in the driver when they change. Just use
zero values wherever we had offsets previously -- the relocations will
fix it all up for us.
Any allocation that may fail should be checked, and propogate the
error upwards. At the highest level we will flush batch and retry.
This is an alternate strategy to what the original DRI driver did of
attempting to flush batch from the lowest levels (eg inside
BEGIN_BATCH). The trouble with that strategy was that flushes could
occur at unexpected times, and additionally there was a need for a
wierd notification mechanism to propogate the 'lost context' state
back up to higher levels.
Propogating the errors directly gives us a lot of flexibility how to
deal with these states, at the expense of a lot more checking in the
code.
Will add some sanity checks later to make sure that out-of-memory
conditions are properly escalated and not lost halfway up the stack.
Keith Whitwell