Hardware-Specific Information and Troubleshooting

DRI User Guide : Hardware-Specific Information and Troubleshooting
Previous: General Trouble Shooting
Next: General Limitations and Known Bugs

10. Hardware-Specific Information and Troubleshooting

This section presents hardware-specific information for normal use and troubleshooting.

10.1. 3dfx Banshee, Voodoo3, Voodoo4 and Voodoo5 Series

10.1.1. Requirements

The 3dfx DRI driver requires special versions of the 3dfx Glide library. Different versions of Glide are needed for Banshee/Voodoo3 than for Voodoo4/5. The Glide libraries can be downloaded from the DRI website.

10.1.2. Configuration

Your XF86Config file's device section must specify the tdfx device. For example:

        Section "Device"
            Identifier  "Voodoo3"
            VendorName  "3dfx"
            Driver      "tdfx"
        EndSection

Or,

        Section "Device"
            Identifier  "Voodoo5"
            VendorName  "3dfx"
            Driver      "tdfx"
        EndSection

The Screen section should then reference the Voodoo device:

	Section "Screen"
	    Identifier  "Screen 1"
	    Device      "Voodoo3"
	    Monitor     "High Res Monitor"
	    DefaultDepth 16
	    Subsection "Display"
		Depth       16
		Modes       "1280x1024" "1024x768" "800x600" "640x480"
		ViewPort    0 0
	    EndSubsection
        EndSection

Or,

	Section "Screen"
	    Identifier  "Screen 1"
	    Device      "Voodoo5"
	    Monitor     "High Res Monitor"
	    DefaultDepth 24
	    Subsection "Display"
		Depth       16
		Modes       "1280x1024" "1024x768" "800x600" "640x480"
		ViewPort    0 0
	    EndSubsection
	    Subsection "Display"
		Depth       24
		Modes       "1280x1024" "1024x768" "800x600" "640x480"
		ViewPort    0 0
	    EndSubsection
        EndSection

The kernel module for 3dfx hardware is named tdfx.o and should be installed in /lib/modules/2.4.x/kernel/drivers/char/drm/. It will be automatically loaded by the Xserver if needed.

The DRI 3D driver for 3dfx hardware should be in /usr/X11R6/lib/modules/dri/tdfx_dri.so. This will be automatically loaded by libGL.so.

The Voodoo5 supports 3D rendering in 16 and 32 bpp modes. When running in 32bpp mode an 8-bit stencil buffer and 24-bit Z (depth) buffer are offered. When running in 16bpp mode only a 16-bit Z (depth) buffer is offered and stencil is implemented in software.

A software-based accumulation buffer is available in both 16 and 32bpp modes.

10.1.3. Troubleshooting

If you try to run an OpenGL application and see an error message similar to

      gd error (glide): gd error (glide): grSstSelect:  non-existent SSTgd error (glide): grSstSelect:  non-existent SSTS

it means that you have the wrong version of the Glide library for your hardware.

3D acceleration for Banshee and Voodoo3 is only supported in the 16 bit/pixel screen mode. Use xdpyinfo to verify that all your visuals are depth 16. Edit your XF86Config file if needed.
The /dev/3dfx device is not used for DRI; it's only for Glide on older 3dfx hardware.
Different versions of Glide are needed for Voodoo3 and Voodoo5. See the DRI website's resources page to download the right version of Glide.
Voodoo4/5 may be run at 24bpp (instead of 32bpp, the default) but 3D acceleration is not supported in that mode. 32bpp mode is fully 3D accelerated.

10.1.4. Performance and Features

Normally, buffer swapping in double-buffered applications is synchronized to your monitor's refresh rate. This may be overridden by setting the FX_GLIDE_SWAPINTERVAL environment variable. The value of this variable indicates the maximum number of swap buffer commands can be buffered. Zero allows maximum frame rate.
On Voodoo4/5, rendering with 16-bits/texel textures is faster than using 32-bit per texel textures. The internalFormat parameter to glTexImage2D can be used to control texel size. Quake3 and other games let you control this as well.
The glTexEnv mode GL_BLEND is not directly supported by the Voodoo3 hardware. It can be accomplished with a multi-pass algorithm but it's not implemented at this time. Applications which use that mode, such as the Performer Town demo, may become sluggish when falling back to software rendering to render in that mode.
The Voodoo3/Banshee driver reverts to software rendering under the following conditions:
- Setting GL_LIGHT_MODEL_COLOR_CONTROL to GL_SEPARATE_SPECULAR_COLOR.
- Enabling line stippling or polygon stippling.
- Enabling point smoothing or polygon smoothing.
- Enabling line smoothing when line width is not 1.0. That is, antialiased lines are done in hardware only when the line width is 1.0.
- Using 1-D or 3-D texture maps.
- Using the GL_BLEND texture environment.
- Using stencil operations.
- Using the accumulation buffer.
- Using glBlendEquation(GL_LOGIC_OP).
- Using glDrawBuffer(GL_FRONT_AND_BACK).
- Using glPolygonMode(face, GL_POINT) or glPolygonMode(face, GL_LINE).
- Using point size attenuation (i.e. GL_DISTANCE_ATTENUATION_EXT).
- Using glColorMask(r, g, b, a) when r!=g or g!=b.
The Voodoo5 driver reverts to software rendering under the same conditions Voodoo3 with three exceptions. First, stencil operations are implemented in hardware when the screen is configured for 32 bits/pixel. Second, the GL_BLEND texture env mode is fully supported in hardware. Third, glColorMask is fully supported in hardware when the screen is configured for 32 bits/pixel.
As of January, 2001 the second VSA-100 chip on the Voodoo5 is not yet operational. Therefore, the board isn't being used to its full capacity. The second VSA-100 chip will allow Scan-Line Interleave (SLI) mode for full-screen applications and games, potentially doubling the system's fill rate. When the second VSA-100 chip is activated glGetString(GL_RENDERER) will report Voodoo5 instead of Voodoo4.
The lowest mipmap level is sometimes miscolored in trilinear- sampled polygons.
The GL_EXT_texture_env_combine extension is supported on the Voodoo4 and Voodoo5.

10.1.5. Known Problems

The lowest mipmap level is sometimes miscolored in trilinear- sampled polygons (Voodoo3/Banshee).
Fog doesn't work with orthographic projections.
The accuracy of blending operations on Voodoo4/5 isn't always very good. If you run Glean, you'll find some test failures.
The Glide library cannot be used directly; it's only meant to be used via the tdfx DRI driver.
SSystem has problems because of poorly set near and far clipping planes. The office.unc Performer model also suffers from this problem.

10.2. Intel i810

10.2.1. Requirements

A kernel with AGP GART support (such as Linux 2.4.x) is needed.

10.2.2. Configuration

Your XF86Config file's device section must specify the i810 device, and specify a usable amount of video ram to reserve.

        Section "Device"
            Identifier  "i810"
            VendorName  "Intel"
            Driver      "i810"
	    Option	"AGPMode" "1"
	    VideoRam    10000
        EndSection

The Screen section should then reference the i810 device:

	Section "Screen"
	    Identifier  "Screen 1"
	    Device      "i810"
	    Monitor     "High Res Monitor"
	    DefaultDepth 16
	    Subsection "Display"
		Depth       16
		Modes       "1280x1024" "1024x768" "800x600" "640x480"
		ViewPort    0 0
	    EndSubsection
        EndSection

The kernel module for the i810 is named i810.o and should be installed in /lib/modules/2.4.x/kernel/drivers/char/drm/. It will be automatically loaded by the Xserver if needed.

The DRI 3D driver for the i810 should be in /usr/X11R6/lib/modules/dri/i810_dri.so. This will be automatically loaded by libGL.so.

10.2.3. Troubleshooting

3D acceleration for the i810 is only available in the 16 bit/pixel screen mode at this time. 32bpp acceleration is not supported by this hardware. Use xdpyinfo to verify that all your visuals are depth 16. Edit your XF86Config file if needed.
The i810 uses system ram for video and 3d graphics. The X server will ordinarily reserve 4mb of ram for graphics, which is too little for an effective 3d setup. To tell the driver to use a larger amount, specify a VideoRam option in the Device section of your XF86Config file. A number between 10000 and 16384 seems adequate for most requirements. If too little memory is available for DMA buffers, back and depth buffers and textures, direct rendering will be disabled.

10.2.4. Performance and Features

Basically all of the i810 features which can be exposed through OpenGL 1.2 are implemented. However, the following OpenGL features are implemented in software and will be slow:

Stencil buffer and accumulation buffer operations
Blend subtract, min/max and logic op blend modes
glColorMask when any mask is set to false
GL_SEPARATE_SPECULAR_COLOR lighting mode
glDrawBuffer(GL_FRONT_AND_BACK)
Using 1D or 3D textures
Using texture borders

10.3. Matrox G200 and G400

10.3.1. Requirements

A kernel with AGP GART support (such as Linux 2.4.x) is needed.

10.3.2. Configuration

Your XF86Config file's device section must specify the mga device:

        Section "Device"
            Identifier  "MGA"
            VendorName  "Matrox"
            Driver      "mga"
	    Option	"AGPMode" "1"
	    VideoRam    32768
        EndSection

The Screen section should then reference the MGA device:

	Section "Screen"
	    Identifier  "Screen 1"
	    Device      "MGA"
	    Monitor     "High Res Monitor"
	    DefaultDepth 16
	    Subsection "Display"
		Depth       16
		Modes       "1280x1024" "1024x768" "800x600" "640x480"
		ViewPort    0 0
	    EndSubsection
        EndSection

To use a 32bpp screen mode, use this Screen section instead:

	Section "Screen"
	    Identifier  "Screen 1"
	    Device      "MGA"
	    Monitor     "High Res Monitor"
	    DefaultDepth 24
            DefaultFbBpp 32
	    Subsection "Display"
		Depth       24
		Modes       "1280x1024" "1024x768" "800x600" "640x480"
		ViewPort    0 0
	    EndSubsection
        EndSection

The kernel module for the G200/G400 is named mga.o and should be installed in /lib/modules/2.4.x/kernel/drivers/char/drm/. It will be automatically loaded by the Xserver if needed.

The DRI 3D driver for the G200/G400 should be in /usr/X11R6/lib/modules/dri/mga_dri.so. This will be automatically loaded by libGL.so.

10.3.3. Performance and Features

Software rendering will be used under any of the following conditions:

Using glDrawBuffer(GL_FRONT_AND_BACK).
Using point, line, or triangle smoothing.
Using glLogicOp.
Using glPolygonStipple or glLineStipple.
Using 1D or 3D textures.
Using texture borders.
Using glDepthFunc(GL_NEVER).
Using the accumulation buffer.

The AGP mode may be set to 1, 2, or 4. One is used by default. Higher AGP speeds may result in unreliable performance depending on your motherboard.

Compaq has funded the implementation of AGP accelerated ReadPixels and DrawPixels in this driver. With this implementation, on a G400 drawing directly from AGP memory (exported to the client), throughput of up to 1 GB/sec has been measured.

Additionally Compaq's funding has produced several new extensions in Mesa, including one (packed_depth_stencil_MESA) which enables Read/DrawPixels functionality to operate directly on the packed 24/8 depth/stencil buffers of this hardware.

In order to access this functionality, the application must ensure that all pixel processing operations are disabled. There are in addition a fairly complex set of rules regarding which packing/unpacking modes must be used, and which data formats are supported, and alignment constraints. See the files in lib/GL/mesa/src/drv/mga/DOCS for a summary of these. The extension definitions are included in the Mesa 3.4 source distribution.

10.3.4. IRQ Assignment

There have been problems in the past with the MGA driver being very sluggish when the DRI is enabled (to the point of being unusable.) This is caused by the graphics card not having an interrupt assigned to it. The current DRI trunk will attempt to detect this condition and bail out gracefully.

The solution to the above problem is to assign an interrupt to your graphics card. This is something you must turn on in your system BIOS configuration. Please consult your system BIOS manual for instructions on how to enable an interrupt for your graphics card.

10.3.5. MGA HAL lib

MGAHALlib.a is a binary library Matrox has provided for use under Linux to expose functionality for which they can not provide documentation. (For example TV-Out requires MacroVision be enabled on the output.) This binary library also sets the pixel/memory clocks to the optimal settings for your Matrox card.

Currently the MGAHAL library is required for the G450 to work. You can download this from the driver section on Matrox's website: www.matrox.com/mga

Here modifications to the DRI build instructions which make the mga ddx driver use the MGAHAL library:

	  1.Put the following define in your host.def file
		  #define UseMatroxHal YES
	  2. Place mgaHALlib.a in the following directory
		  xc/programs/Xserver/hw/xfree86/drivers/mga/HALlib/

You can use DualHead on the G400/G450 DH cards by creating two device sections which both point to the same BusID. For most AGP devices the BusID will be "PCI:1:0:0". Configure your screen section as you would normally configure XFree86 4.x Multihead. It should be noted that currently the second head does not support direct rendering.

10.3.6. Known Problems

None.

10.4. ATI Rage 128

10.4.1. Requirements

A kernel with AGP GART support (such as Linux 2.4.x) is needed.

10.4.2. Configuration

Your XF86Config file's device section must specify the ati device:

        Section "Device"
            Identifier  "Rage128"
            VendorName  "ATI"
            Driver      "ati"
	    Option	"AGPMode" "1"
	    Option	"UseCCEFor2D" "false"
        EndSection

The Screen section should then reference the Rage 128 device:

	Section "Screen"
	    Identifier  "Screen 1"
	    Device      "Rage128"
	    Monitor     "High Res Monitor"
	    DefaultDepth 16
	    Subsection "Display"
		Depth       16
		Modes       "1280x1024" "1024x768" "800x600" "640x480"
		ViewPort    0 0
	    EndSubsection
	    Subsection "Display"
		Depth       32
		Modes       "1280x1024" "1024x768" "800x600" "640x480"
		ViewPort    0 0
	    EndSubsection
        EndSection

The kernel module for the Rage 128 is named r128.o and should be installed in /lib/modules/2.4.x/kernel/drivers/char/drm/. It will be automatically loaded by the Xserver if needed.

The DRI 3D driver for the Rage 128 should be in /usr/X11R6/lib/modules/dri/r128_dri.so. This will be automatically loaded by libGL.so.

You may also set your screen depth to 32 for 32bpp mode.

10.4.3. Performance and Features

While PCI Rage 128 based cards are supported, they do not yet support PCI GART, so they will not perform as well as their AGP counterparts.

For AGP cards, the AGP mode may be set to 1, 2, or 4. One is used by default. Higher AGP speeds may result in unreliable performance depending on your motherboard.

Note that even at 32bpp there is no alpha channel.

The following OpenGL features are implemented in software and will be slow:

accumulation buffer operations
stencil, when using a 16bpp screen
Blend subtract, min/max and logic op blend modes
GL_SEPARATE_SPECULAR_COLOR lighting mode
glDrawBuffer(GL_FRONT_AND_BACK)
Using 1D or 3D textures
Using texture borders

10.4.4. Known Problems

If you experience stability problems you may try setting the UseCCEFor2D option to true. This will effectively disable 2D hardware acceleration. Performance will be degraded, of course.

10.5. ATI Radeon

10.5.1. Requirements

A kernel with AGP GART support (such as Linux 2.4.x) is needed.

10.5.2. Configuration

Your XF86Config file's device section must specify the ati device:

        Section "Device"
            Identifier  "Radeon"
            VendorName  "ATI"
            Driver      "ati"
	    Option	"AGPMode" "1"
        EndSection

The Screen section should then reference the Radeon device:

	Section "Screen"
	    Identifier  "Screen 1"
	    Device      "Radeon"
	    Monitor     "High Res Monitor"
	    DefaultDepth 16
	    Subsection "Display"
		Depth       16
		Modes       "1280x1024" "1024x768" "800x600" "640x480"
		ViewPort    0 0
	    EndSubsection
	    Subsection "Display"
		Depth       32
		Modes       "1280x1024" "1024x768" "800x600" "640x480"
		ViewPort    0 0
	    EndSubsection
        EndSection

The kernel module for the Radeon is named radeon.o and should be installed in /lib/modules/2.4.x/kernel/drivers/char/drm/. It will be automatically loaded by the Xserver if needed.

The DRI 3D driver for the Radeon should be in /usr/X11R6/lib/modules/dri/radeon_dri.so. This will be automatically loaded by libGL.so.

You may also set your screen depth to 32 for 32bpp mode.

10.5.3. Performance and Features

While this driver supports many of the features of ATI Radeon cards, we do not yet fully support the card's TCL features. This work is progressing, but is not yet ready.

The AGP mode may be set to 1, 2, or 4. One is used by default. Higher AGP speeds may result in unreliable performance depending on your motherboard.

The following OpenGL features are implemented in software and will be slow:

Blend subtract, blend min/max and blend logicops
Stencil and accumulation operations
1D and 3D textures
Texture borders

The GL_EXT_texture_env_combine, GL_EXT_texture_env_add and GL_EXT_texture_env_dot3 extensions are supported (or will be soon supported in the new driver based on Mesa 3.5).

We hope to implement support for the following features in the future:

Vertex transformation, clipping and lighting (TCL)
Hardware stencil buffer
Cube map textures
3D textures
Three texture units

10.5.4. Known Problems

Certain (early?) revisions of the AMD Irongate chipset have AGPGART problems which effect Radeon, and other graphics cards. The card may work unreliably, or not work at all. If the DRM kernel module is not loaded, the 2D Xserver may work. There's hope that this can be fixed in the future.

10.6. 3DLabs Oxygen GMX 2000

The driver for this hardware was experimental and is no longer being developed or supported.

DRI User Guide : Hardware-Specific Information and Troubleshooting
Previous: General Trouble Shooting
Next: General Limitations and Known Bugs