In the event these features are utilized, execution may still continue,
albeit with the likely possibility graphical problems of varying degree.
This patch replaces hard assertions with debug print statements when
configured to do so.
Cromwell sets these registers on video initialization and
guest OS should be able to read them in order to obtain
current screen resolution set up by firmware.
This patch sets the class id, subsystem id, and subsystem vendor id
in the PCI device header to match what the NV2A hardware advertises.
This resolves an issue with PCI device enumeration failing to detect and load
the driver for the NV2A device in some operating systems, namely ReactOS
version 0.2.5-xbox (custom build), which depend on these fields being
programmed correctly.
Vertex attribute values provided via NV097_SET_VERTEX_DATA2S are
apparently two 16-bit signed integers, packed into 32 bits, which are
then to be directly mapped to floating point values in the range
[-32768.0, 32767.0].
Halo:CE uses this format to provide texture (u,v) coordinates for
render-to-texture techniques, including weapon scope, dynamic shadows,
and radar beacon--all of which are now working as expected with this
patch.
* Port espes's fifo work to the split up nv2a code
This patch ports over the following commits from the XQEMU 1.x tree
(available via tag archive-xqemu-1.x) to the refactored nv2a code:
- 4d9107e8 (HEAD -> xbox, upstream/xbox) Merge branch 'fix-fifo' into xbox
- c33f4ab2 cleanups
- d3f83d93 cleanups
- faca5dd0 fix
- 7c62d7c1 fixes
- 8ad239b0 wip
- 45ed3c31 wip
- c006d5e8 wip
However, in its current form, this patch causes some regressions and
needs further investigation.
* nv2a: basic pattern object
Fixes several games which utilize clipping, such as Halo:CE which
uses clipping in multiplayer to render multiple player screens.
Research and original PoC work done by Jannik Vogel:
- JayFoxRox/xqemu-espes@c05d91b
- JayFoxRox/xqemu-espes@a254830
This patch adds an additional "retired" queue in which FIFO command
entry objects are placed after execution. This queue of objects is then
returned to the pusher's new "available" queue for re-use.
This improves the performance of the system by avoiding the costly
overhead associated with the general-purpose use of `malloc` and `free`
for previous allocation of FIFO command queue objects.
This patch does the following:
- Fixes up things for Qemu 2.x compat
- Factors out the high-level NV2A blocks into separate files
- Updates g-lru-cache for latest glib compat (github.com/chergert/glrucache@c10af24)
- Changes texture hashing algorithm from FNV to xxH v0.6.5