The driver appears to work reliably (i.e. lockup-free) with a card based
on this chip. However, I have not explored whether we could get anything
in terms of 3D performance "for free" just by setting some magic bits
in those registers that are still a mystery to us.
Make sure texture base addresses are used correctly.
They are relative to the card's address space, not relative to fbLocation,
which will be important once we program MC_FB/AGP_LOCATION in the way
they are programmed on older Radeon chips.
This patch makes use of two of the new ioctls added in MGA DRM version
3.2. Specifically, the DRM_MGA_SET_FENCE and DRM_MGA_WAIT_FENCE
ioctls are used in mgaWaitForFrameCompletion. As a result the MMIO
register region and the primary DMA region are *not* mapped (if DRM
3.2 is available).
This patch does *not* make use of the new get_param query was added to
differentiate between G4x0 cards and G550 cards. That is left to a
future update.
Xorg bug: 3259
Reviewed by: Eric Anholt
Store compiled vertex program representations in a pointer in the
vertex_program struct.
Allow switching between regular and vertex_program implementations
of fixed function TNL with the MESA_TNL_PROG environment var
(previously this required recompilation).
Ensure program compilation only references program data, not the
wider context. This means that compiled programs only need to be
invalidated when the program string changes, not on other state
changes.
Build vertex programs using state from a 'state key' struct which is
distilled from the current state.
Maintain a list of vertex programs for various different tnl states
indexed by these state keys.
This cuts down the number of vertex programs generated and means that
time spent optimizing vertex programs for execution isn't wasted each
time the state changes.
Runtime generate sse/sse2 code for some vertex programs. Experimental
and only active when _TNL_FIXED_FUNCTION_PROGRAM is defined *and*
the MESA_EXPERIMENTAL environment variable is set...
Works for the arbfplight demo at least.