Reply 40 of 49, by 360alaska2
I can't believe it, somehow Google leapfrogged Anthopic during their IPO?
I'm using opus 5.5 to code this and using Sol, Astra, Fable, another Opus and now neon as advisors.
I can't believe it, somehow Google leapfrogged Anthopic during their IPO?
I'm using opus 5.5 to code this and using Sol, Astra, Fable, another Opus and now neon as advisors.
This is the next release to github, it has only been tested on 855, so I won't put it on github until I test on 800 and 900.
Changelog:
Here's a draft of the release notes for **.706**. It covers everything changed since .637 and is written in the same style as the v637 release. I built it from the per-build notes in the source, the .645 handoff and the speedbook's status log.
---
**Build .706**
Windows 98 display driver for the VIA VX800, VX855 and VX900 integrated graphics (Chrome9 HC3 / HCM / HD).
**Tested on the VX855 / VX875. VX800 and VX900 checks are still to come.** The shared changes below have not yet been run on those two chips; `tools\REGRESS.TXT` lists each one with the test that covers it.
## Downloads
- `VIA_Chrome9_706.zip`: the driver (five files, installed through Device Manager; see the README).
- `VIA_Chrome9_706_tools.zip`: optional test tools with `ALL.BAT` and `REGRESS.TXT`.
## New since .637
**Direct3D: the driver's own transform and lighting is much faster**
- Lighting is done four vertices at a time for every common case:
- vertex-coloured materials (Direct3D 8's default);
- point lights and spot lights;
- specular lighting.
Before, many of these went one vertex at a time.
- Cheaper maths on the VIA CPU: reciprocal estimates with a refinement step replace square roots and divides in the light, specular and colour code. The results are unchanged to the last bit.
- Less repeated set-up work:
- set-up is kept across draws when only the world matrix changes;
- the normal matrix is built only when a draw needs it;
- back faces are culled before they are lit.
- Index lists are copied eight at a time. Meshes that sit wholly on screen go to the engine in one block.
- **Guard band (VX855 / VX875):** triangles that cross the screen edge are clipped by the 3D engine instead of the CPU. That cuts CPU-clipped triangles by about 70%. The GRID scenes are 2–10% faster, and A-B-A-B runs on the same session repeat the result.
- The VXTLP grid went from 40.3 ms to 35.7 ms a frame (.640 → .644), level with Direct3D's own T&L. 3DMark 2001's 8-light high-polygon test now completes. It used to time out.
**Direct3D: fixes**
- The driver's T&L drew nothing in these cases, now fixed:
- untransformed points and lines;
- draws larger than its scratch buffer;
- indices spanning more than 16,384 vertices;
- immediate-mode commands;
- XYZB (blended) vertices.
- Vertices at infinite depth are drawn, held to the depth range. Before, they were dropped; GTA3 sends thousands of them.
- More than 16 lights set in a scene, and a third stage asking for a third coordinate set, are now handled.
- Projected textures: the divide is now done per pixel, which is correct across the whole triangle. Two stages sharing coordinates, a projected second stage, and 3-component (volume and cube) coordinates are also fixed.
- Environment bump mapping uses XP's layout. 3DMark 2001's environment-bump test now draws.
- Clears and device validation work when no render target is bound.
**DirectDraw, flips and the display**
- Page flips go through the 3D command ring using XP's sequence, an idle wait before the swap. This fixed a ring hang in flip-heavy tests. A mirrored second output now flips along with the first.
- **No-vsync flips (VX855 / VX875):** a flip without vsync shows from the next scan line instead of waiting a whole frame.
- The pointer no longer freezes on screen after a program locks the desktop. Each program's locks are released when it exits.
- Windowed programs' stretched Present no longer faults. Clipped stretches are done by the 3D engine.
- The first full-screen DOS session of a boot is no longer black (the BIOS's clock select is restored).
- WaitForVerticalBlank reports an error when the wait times out. A primary destroyed while locked no longer leaves the pointer bracket open. A zero-height overlay is refused.
**Stability and recovery**
- After any 3D engine reset, the 2D side tests the engine before the desktop uses it. If the engine fails that test, the desktop draws in software for that mode instead of corrupting.
- Repeated 3D hangs: after the reset limit, 3D turns off for the session and the desktop keeps working. The display FIFO registers are restored after every reset.
- Locking fixes:
- the page lock runs under the engine lock;
- a front-buffer wait no longer disturbs a newer flip;
- an unreported mode change resets the flip state.
**OpenGL**
- `GL_CLAMP` with linear filtering stays on the engine for cube, rectangle and volume textures.
- Lighting fixes for two-sided lighting and normal rescaling.
- Locked vertex arrays are drawn as one batch. In HarborBench's OpenGL draw test, frame time fell from 88 ms to 32 ms.
- Front-buffer and render-to-texture fixes (allocation failures, freeing under the engine lock).
**2D**
- A polyline wholly outside the clip is skipped. If the engine is lost partway through a polyline, the CPU finishes it.
**Settings (new)**
- `ACCEL_D3D_GUARDBAND`: the guard band, VX855 / VX875 only. 1 on (the default), 0 off.
- `ACCEL_FLIP_TEAR`: no-vsync flips from the next scan line, VX855 / VX875 only. On by default.
- `ACCEL_3D_RINGFLIP`: flips through the command ring. On by default.
I already have more builds cooking, but feel free to test and report bugs.
Looks like you've been busy! I will be trying this new driver out on my Cx0 Monday and I'll report back.
This is my next github release. 0.723:
VIA Chrome9 (VX800 / VX855 / VX875 / VX900) Windows 98 driver: changes from .637 to .723
GAME FIXES
3DMark 2001 Dragothic: the water no longer shows flashing blue and orange triangles. The fault was in the chip's projected-texture mode, so the driver now sends this kind of texture in the normal mode. Build .723.
Hellbender: fog works in hardware mode (no more "Your 3D adapter will not allow you to see fog" box), and the missing pieces of the picture are back. Build .722.
3DMark 2001: scenes no longer look dark. Spot lights, specular highlights and lights numbered above 15 are now lit as Direct3D specifies. Build .646. No more "failed to validate material", build .649.
GTA3: objects sent with an infinite depth value are drawn instead of dropped. Build .642.
DIRECT3D
The driver's own hardware T&L is much faster: four vertices at a time with SSE2, with point lights, spot lights and specular lighting all on the fast path, and its set-up kept between draws when nothing changed.
Bump mapping (environment-mapped bump on stage 1). Build .662.
Flat shading and edge flags through the T&L. DirectX 5/6 programs' draws are batched like DirectX 7/8 ones. Build .663.
A third texture stage, projected textures on the second unit, and state blocks keep their lighting state. Builds .646-.668.
Triangles cut by the near plane get the right corners when the driver is short of scratch space. Build .718.
Primitives whose every vertex has rhw 0, NaN or infinity are drawn, as the software path does, instead of dropped. Build .722.
Guard-band clipping for the driver's T&L, VX855 / VX875 only for now. Build .698. Projected, cube and volume textures are clipped to the screen instead, build .721.
Faster triangle culling and batch handling. Builds .709-.715.
DIRECTDRAW
Page flips go through the 3D command ring, with or without vertical sync. On the VX855 / VX875 a flip without vsync is shown from the next scan line.
Stretched blits and windowed Present, including a partly covered window, are drawn by the 3D engine. Builds .667-.671.
Fixes when you Alt+Tab out of a game or the mode changes, for the mouse pointer after a game exits, and for WaitForVerticalBlank timing out. Builds .705-.706.
OPENGL
Faster lighting on the SSE path, and less state sent to the chip between draws. Builds .639, .709-.713.
Fixes for the front buffer in windowed programs (Quake II's menu) and for render-to-texture. Builds .705-.706.
DESKTOP, DOS AND STABILITY
The 2D engine is checked again after any 3D reset, so the desktop stays right after a game exits. Build .683.
Full-screen DOS on laptop panels (Samsung NC20 and similar), build .673. The 3D engine is stopped before any BIOS mode set, including entering a full-screen DOS box, build .646.
Fixed two possible lock-ups between 3D programs and the desktop. Builds .706, .713.
Mode sets with a refresh rate use the BIOS timing correctly. Build .713.
A program could lose an open file after another Direct3D program ended (seen as test logs cut short). Build .707.
Polylines partly outside a window are drawn correctly, including when the engine is reset mid-line. Build .705.
TEST TOOLS (separate download, run RUNALL.BAT)
New picture tests for projected textures and the water scene, checked against a reference the test draws itself.
New tests: VXRHW (unusable rhw values), VXDX7 --enum (each device's capabilities), VXREUSE (handles across programs).
VXD3DMON (the profiler) no longer freezes the VX800. Build .713.
Some test results from your latest release:
- Hellbender renders properly now and does not show the missing fog message anymore
- Jedi Knight 2 still has the same glitches, but this is the only OpenGL game with issues I tested so far.
- Some more games: Expendable, MDK2, Indiana Jones and the Emperors Tomb all seem to be fine.
I discovered a new issue I didn't notice before: Video playback (like .avi videos) is broken when using the newer Windows Media Player 6 that came with Windows 98. It only shows a black window where the video should be. When using the original media player (C:\WINDOWS\mplayer.exe), the same video file renders fine. I can get a picture with media player 6 when I move a window on top of the video window and then move it aside again. Then, for ~1 second, I get a video picture before it goes black again.
Since the original mplayer uses the Video for Windows API while the newer one should use DirectShow, my suspicion would be that there is a problem in the latter one when using your driver.
I also did some benchmark runs with 3DMark2000 and 3DMark2001 on both Win98 with your driver and on Windows XP with the vendor driver.
Hardware TnL performance is slower than software TnL on both systems, the vendor driver also exposes pixel shaders and runs the "Nature" test in 3DM2001.
I also noticed that there is a ~60fps cap in your drivers (or vsync) so the numbers above might not be completely accurate. 3DMark99 also shows constant 0fps although is runs perfectly fine.
Set of new log files attached in case it helps.
Which Jedi 2, there are two them, can you believe that? I’m assuming your meaning the one from 1997? It seemed to work except for the rendered video, so that is a bug we’ll work through.
360alaska2 wrote on Today, 17:11:Which Jedi 2, there are two them, can you believe that? I’m assuming your meaning the one from 1997? It seemed to work except for the rendered video, so that is a bug we’ll work through.
I'm talking about this one https://en.wikipedia.org/wiki/Star_Wars_Jedi_ … I:_Jedi_Outcast
I would add that the T&L is under construction and the pixel shader support is incoming. More to come on that end.
Thank you for testing and documenting bugs, is the code 12 on install still an issue for you?
I’ve ordered an OLPC 1.5 in an attempt to check laptop support.
360alaska2 wrote on Today, 18:01:is the code 12 on install still an issue for you?
Yes, it is, no change here. But this might also be a BIOS/ACPI bug since I noticed that even the standard VGA driver installation ends up with this error.
I was kind of concerned about that, I know enough to know that the bios typically gives the resources out, can you select non-PNP os and try to force the resources. Can you give exactly what it was before and I already see that you put what it was after.