Banshee TODO list:
#1: Fix display errors caused by incorrect implementation and find out the
expected behaviour of the undocumented stuff.
#2: Make color depth > 8 work correctly.
#3: Screen resulution change not possible yet (only 640 x 480 works).
#4: Implement remaining 2D commands (line, polyline, polygon fill, stretchBlt).
#5: 3D core usablility, code cleanup, ...
- Implemented ternary raster operations. With 3 binary input values we have 8
possible combinations und the ROP code is the binary representation of the
result table. Instead of implementing cascades of binary operations at byte
level, we simply build a 3-bit value from D, S and P. This is the bit number
of the ROP code to use as the result. This is performed for each bit of color
data. We still keep the existing binary operations at byte level, since the
code is faster for that case.
- Inplemented the "row #0 only" flag handling in pattern fill operations. This
flags seems to have an undocumented effect on the source pitch of
host-to-screen operations. This issue needs to be investigated.
- TODO: Some bits of the srcXY register seem to have an undocumented effect on
host-to-screen operations, too.
- Added basic support for the tiled desktop space. Win95 now starts building
the desktop, but without icons and it is not yet usable in 16 bpp mode.
- Fixed 32 to 24 bpp format conversion in host-to-screen operations.
- Fixed some raster operations.
- Now allocate buffer for launch area writes dynamicly.
- Added error messages for undocumented features.
tile-based updates and screen update locking. Some related changes in the init
and save/restore code are also included. The main update code has been moved
from bx_voodoo_vga_c to bx_voodoo_c and the required parameters are set up
depending on the model. The BX_LOCK/BX_UNLOCK and SET_TILE_UPDATED stuff still
needs to be added in the 3D core. Currently a redraw_area() hack makes the
Voodoo 1/2 screen updates work as expected.
- Fixes and optimizations in the BitBlt code.
- Added mutex for display updates and lock some register writes and the BitBlt
methods during screen update.
- Improved some debug messages.
TODO: Use the render_mutex for the Voodoo1/2 update code, too.
almost usable in 8 bpp mode, but there are still issues (see below).
- Added basic support for pattern fill functions.
- Added some more ROPs (all S/D and P/D ones).
- Now using the ROP functions in rectangle, pattern and host-to-screen code.
- Several other related changes.
TODO:
- Implement remaining binary (P/S) and all of the ternary ROPs.
- Some Windows icons are still black or messy (e.g. start button) and screen
are missing in some cases. Shutdown background is wrong, although the correct
ROP is used.
- Cleanups and optimiations when the code is known to work as expected.
- Moved basic BitBlt stuff from the Cirrus code to a separate header file to
make it usable for the Voodoo Banshee, too.
- Rewrite of the screen-to-screen BitBlt code for the new ROP code.
- Calculate redraw parameters based on the screen start.
- Set "Banshee busy" bit in the status register when 2D is busy.
- Added some error messages for unimplemented features.
- Host-to-screen and screen-to-screen copy operations implemented (ROP 0xcc).
- Rectangle fill command now handles clipping.
- Attempt to fix the redraw_area() location after bitblt operations.
- Update the host palette entries when writing to CLUT (fixes screenshot).
- Added 2D register names for debug messages.
- Fixed hardware cursor bottom/right edge display bug.
- Started splitting 2D code in different methods (no functional changes).
- Added hardware cursor to save/restore list.
TODO #1: Implement bitblt engine similar to the Cirrus one.
TODO #2: Voodoo code reorganization: base class with Voodoo core, one subclass
for Voodoo1/2 and one for Banshee.
in 8 bpp mode, but it's not yet usable (no cursor, no text). List of changes:
- Added AGP/CMD/Misc register handling (CMDFIFO #0 is now usable).
- Started work on 2D register handling ("Rectangle fill" command implemented).
- Added Banshee specific code in redraw_area().
- Writing to VGA DAC registers also modifies the Banshee CLUT.
- Banshee now indicates that no TV-out and video-in device is connected.
- Some other related fixes and improvements.
- Added VCLK array and removed hardcoded values.
- Added new method get_crtc_params() to retrieve htotal and vtotal values.
- Cirrus: Set up VCLK registers with default on reset (TODO: VGABIOS should do
that when setting video mode).
- Voodoo Banshee: Set up VCLK #3 when writing to pllCtrl0 register. VCLK #0 and
#1 are VGA compatible and VCLK #2 is fixed 50 MHz. Fixed maximum resolution
based on PLL section of the spec.
- Banshee Memory layout implemented (AGP/misc and 2D register behaviour not
implemented yet). Now Win95 with Banshee driver can boot in standard VGA mode.
- Fixed memory allocation: Banshee has 16 MB shared LFB / TMU memory.
- Some work on save/restore support.
- Renamed Banshee VGA register handlers.
- tile-based VBE graphics updates implemented
- "half mode" support added (line doubling for low resultions)
- change screen resolution only after re-enabling the video processor
"Voodoo VGA" framework. Most of the work is still to do.
- Added Banshee specific PCI write handler and related stuff. The device is
initialzed properly by the Bochs BIOS (mem, lfb, rom and i/o).
- With Banshee specific VGA BIOS it passes the init stage and boots properly
to DOS. Parts of the init code coming from file "voodoo.h.txt" (part of the
original patch).
- Standard VGA modes are all usable (using the Bochs VGA core).
- VBE graphics modes with 8, 16, 24 and 32 BPP are usable (VBE drawing code is
mostly copy&paste from the Bochs VBE code. DAC 6/8 bit switch is present.
- TODO list:
- Voodoo Banshee memory layout (registers, 2D, 3D, LFB, textures)
- 2D graphics engine must be written from scratch
- interaction between new device and existing 3D core
- non-VGA mode (Voodoo model "banshee" without VGA extension "voodoo")
- save/restore support, move Banshee stuff to separate file, ...
- Fixed missing bx_gui->flush() call after non-VGA update().
- Fixed a BX_DEBUG message in the Voodoo code.
- Re-define LOG_THIS for messages in the "Voodoo VGA" code.
currently only implements an ISA standard VGA adapter with the selected Voodoo
model as it's extension. The methods of the "Voodoo VGA" device only call the
required methods of the Voodoo or VGA core.
TODO: Start implementing the Voodoo Banshee chipset based on this code.
- Moved VGA timer handler to the VGA core and handle the non-VGA case there.
It calls the specific update() code and finally bx_gui->flush().
- Moved the VGA update frequency parameter handler and the refresh_disply()
method to the VGA core.
- Store the VGA extension parameter in the VGA core.
- Voodoo: Fixed graphics snapshot.
- voodoo_init() must be called after initializing FIFO mutex.
- The 'op_pending' counter must be reset after resetting FIFOs.
- Fixed an MSVC warning and removed obsolete comment.
- FIXME: There are still some conditions that can freeze the Voodoo display.
it's 64 entries is still not present, since it makes the emulation a few
percent slower due to the overhead of the FIFO code. The memory FIFO has 64k
entries and makes a speedup of around 75 % possible.
- Now using float type for the 'clk0_freq' variable.
- DAC register 6 can switch to half pixel clock (ported from PCem).
- Writing value 0xf8 to PLL register 0x0e completes clock #0 setup.
non-WIN32 platforms instead of BX_MSLEEP(). The donut demo is once again a
little bit slower, but on other tests the IPS values are still okay. Cleaned
up the code a little bit and added save/restore stuff for the new code.
packets and then always sleep for 1 millisecond.
TODO: The first tests with pthread_cond_* stuff showed a slowdown of approx.
10 % compared with the IPS value of the current code.
command and execute it when there is enough data present (code partly ported
from voodoo.h.txt).
TODO: Use "wait for event / condition" functions instead of msleep() if there
is a performance improvement with it.
- Modified win32 thread macros to use the handle instead of the id.
- Added new macro BX_THREAD_KILL to terminate threads.
Voodoo thread changes:
- Let the cmdfifo_thread sleep while the CMDFIFO is disabled.
- Kill the cmdfifo_thread in destructor and removed old thread control code.
- Added capability to save/restore float type values in paramtree / siminterface.
- Added Voodoo2 triangle generator data to save/restore list.
- Implemented bitBLT function "SGRAM fill".
- Minor other changes and cleanups.
- TODO: remaining bitBLT functions, thread handling improvements.
- Added CMDFIFO thread similar to the sound output thread.
- split register write function in 2 functions: register_w_common() handles
the CMDFIFO writes and valid registers in CMDFIFO mode. It forwards write
accesses to other registers to register_w() (Voodoo1 and non-CMDFIFO mode).
The CMDFIFO thread also calls register_w() when processing data.
- The CMDFIFO thread processes all data, but it does not yet execute type 3
requests (extended triangle generator missing).
- TODO #1: implement Voodoo2 specific triangle functions (for CMDFIFO).
- TODO #2: implement 2D bitblt engine.
- Zero out voodoo device state in init().
- Added the CMDFIFO state to save/restore list.
- Added CMDFIFO register handling in read/write functions.
- Added new function cmdfifo_put().
- TODO: CMDFIFO thread.
valid. The hRetrace value can be set now, since Bochs no longer crashes.
TODO: Implement enough of the FIFO stuff to make the Win9x driver for the
Voodoo2 work.
- Voodoo: Use the VGA/Cirrus timers to drive the gui screen update.
The vertical retrace emulation is still driven by a separate timer.
- VGA/Cirrus: Since we are using the virtual timer, we also have to read
time_usec() from it.
- vgacore: removed unused timer methods.
- prepared hRetrace value implementation (cannot be enabled yet, since the
driver passes this test and gets confused in next stage (segfault).
- Detect CMDFIFO access (not present yet).
- Report disabling Voodoo graphics mode.
- TODO #1: Fix Voodoo2 to make it usable.
- TODO #2: Split update code and drive gui screen updates from the VGA timer
(finally remove new realtime option again).
- Since we are using the virtual timer, so we have to read time_usec()
from it, too (TODO: realtime mode is hardcoded, maybe add an option).
- Voodoo1 and Voodoo2 have different timing register widths.
- Fixed vertical retrace value returned by get_retrace().
- Prepared separate hvRetrace register (TODO: hRetrace).
- Store the Voodoo model in structure bx_voodoo_t.
- The PCI initEnable register of the Voodoo2 model has a read-only secondary
revision ID in bits 12..15. Now the Voodoo2 is correctly detected by the
Windows driver, but display is not yet working (blank screen).
- Reset now disables VGA override.
- PCI: Moved PCI device base class methods to the end of file devices.cc and
re-define LOG_THIS to use the correct log module and prefix.
- Since the pci_read_handler() method is identical in most devices, move it
to the base class to reduce code duplication. Only the 'pcidev' device has
it's own implementation (NOTE: it is not maintained yet).
- Minor other fixes and cleanups in some PCI devices.
