Information for Tseng Chipset Users The XFree86 Project, Inc. Dirk H. Hohndel, Koen Gadeyne and others. 03 Nov 1998 1. Supported chipsets The Tseng chipsets supported by XFree86 are ET3000, ET4000, ET4000/W32 and ET6000. Accelerated features of the ET4000/W32, W32i, W32p and ET6000 are supported by the SVGA driver. For details about the separate accelerated 8bpp (=256 color) ET4000/W32 and ET6000 server, refer to README.W32. Note that you should NOT be using XF86_W32 unless XF86_SVGA doesn't work on your hardware. No further development is being done on the W32 server; all new efforts go into the SVGA server. Some ET4000W32 ISA cards are known NOT to work with the SVGA server in this version (XFree86 3.3.1): they hang the machine... Use the W32 server XF86_W32 for these cards! 2. Terminology In the rest of this document, "8bpp" is short for "8 bits per pixel", which means a 256-color mode. Similarly, 15bpp refers to 32768 colors, 16bpp to 65536 colors , 24bpp to a "packed" 16 million color mode, and 32bpp to a "sparse" 16 million color mode (at 32bpp, only 24 of the 32 bits are actually used, hence the "sparse"). 15bpp is only used here to differentiate it from 16bpp, but they are both normally referred to as 16bpp. 15bpp is actually 16bpp with a 5-5-5 color weight (wasting one bit per pixel), while 16bpp is, well, 16bpp, with 5-6-5 color weight. 3. ET4000 driver features The SVGA driver for ET4000 chipsets supports all color depths (8, 15, 16, 24 and 24 bpp) on most ET4000 chips starting with the ET4000W32i. The ET4000W32 only supports 8bpp. Depending on the RAMDAC and the support code in the SVGA server, some cards may only support a few of these color depths, or even only 8bpp. On W32i and W32p chips all color depths are supported on the supported RAM- DACs (currently ICS5341, STG170x and Chrontel CH8398). These modes are also accelerated. Some W32p board implementations are limited to 1 MB of video memory in linear memory modes. This is a hardware limitation that cannot be solved in the driver. Since XFree86 requires linear memory for 16/24/32 bpp modes, the use- fulness of these cards for highcolor and truecolor applications is severely limited (those modes mostly use a lot of video memory). In addition, those cards also don't support acceleration in linear mode. This is a design choice in the driver code: if acceleration were to be sup- ported in linear mode, you'd only be able to use 768 kb of video memory, and the driver code would be twice as complex. Cards with a RAMDAC that is not yet supported will be limited in a similar manner as the older cards, i.e. to a maximum pixel clock of 86 MHz, whilst they actually might be able to go up to 135 MHz. As a result, 1280x1024 modes will only be possible when using interlacing, and non-interlaced modes are limited to about 1024x768 at 75 Hz refresh. For a non-interlaced 1280x1024x(256 colors) at say 135-MHz on a W32-type card, you need a w32p (with its 16-bit RAMDAC bus) with a multiplexing RAMDAC so that the w32p sees only (135/2 = 67.5) MHz, not 135 MHz. This requires special code only provided for cards using the ICS5341 GENDAC, the STG170x or the CH8398. This code seems to work fine for most people, except, with the ICS5341, for a small band of frequencies around 90MHz. Linear memory mode (especially important for some DGA clients, like xf86quake) is supported on all ET4000W32i and ET4000W32p cards, but not on the ET4000W32. See the section on linear memory for more information. There are some important issues related to linear memory. For the higher color depths (16, 24 and 32 bpp), linear memory mode is REQUIRED. It is enabled by default in these modes. There is no need to spec- ify that in the XF86Config file. Please read the section on linear memory below: it contains some vital information on how to avoid serious problems. To force "banked" mode in 8bpp modes (where linear memory mode is the default), put the following in the Device section of your XF86Config: Option "no_linear" Acceleration support is present, and enabled by default, for all W32 and ET6000 family chips. This is based on the new XFree86 acceleration interface (XAA). If you have problems with acceleration, acceleration can be disabled by putting the following in the Device section of your XF86Config: Option "noaccel" On some PCI systems (i.e. only on the ET6000 and the ET4000W32p), accelera- tion may cause occasional font corruption. This is probably caused by a badly written system BIOS that ignores the fact that the Tseng PCI devices have their "non-prefetchable" attribute set. On such a BIOS, a PCI feature called "write combining" (or "byte merging") is enabled for the Tseng video card, although it is not permitted. Some systems allow you to manually enable or disable the Write Combining feature in the BIOS setup (sometimes abbreviated to WC). Make sure WC is disabled for the VGA memory aperture. If you experience font corruption on your system and are unable to manually disable WC in your BIOS, font acceleration may be disabled using the follow- ing in the Device section of your XF86Config: Option "xaa_no_color_exp" Note that this will reduce the performance of the X server. 4. ET6000 driver features In addition to the features in the ET4000 driver, the SVGA ET6000 server sup- ports all possible color depths in the SVGA server: 8bpp, 16bpp (both at 5-5-5 and 5-6-5 color resolutions), 24bpp and 32 bpp. Linear memory mode (as opposed to the VGA default, banked memory layout) is supported. It is required and enabled by default for the 16/24/32 bpp modes. For 8bpp, the default is linear mode for PCI cards and banked mode for ISA/VLB cards. To force linear memory at 8bpp, put the following in the SVGA section of your XF86Config: Option "linear" Acceleration is supported and is enabled by default, and accelerates all color depths on the ET6000. Acceleration can be disabled by adding the fol- lowing in the Device section of your XF86Config: Option "noaccel" The hardware cursor is supported in all color depths. Due to a hardware limi- tation in the ET6000, only a limited set of colors is supported (2 signifi- cant bits per color component). This may cause some (small) cursor color errors. If absolute cursor color accuracy is required, the hardware cursor should not be enabled. However, in most applications, this will not be a problem. The hardware cursor can be enabled using Option "hw_cursor" There is a problem with the hardware cursor at high dotclocks (above approx. 110MHz) at which point the cursor does strange things when partly off the left-hand side of the screen. On older ET6000 chip revisions, DoubleScan modes currently don't work with the hardware cursor: only the top half of the cursor is visible. If you want to use DoubleScan modes (320x200 is a popular one), then do not enable the hardware cursor. Most recent ET6000 cards and the ET6100 do not exhibit this problem. On some PCI systems, acceleration may cause occasional font corruption. As described above, this is caused by a bug in your system BIOS or a wrong set- ting of the write combining feature in that BIOS. If you are unable to fix the BIOS or force the option off, font acceleration may be disabled using the following in the Device section of your XF86Config: Option "xaa_no_color_exp" When using accelerated high color-depths (24bpp and 32bpp), high-resolution modes (starting somewhere around 800x600) may cause temporary "garbage" lines to the right of the screen while the accelerator is busy. The garbage should not be persistent: it should go away as soon as the server is left alone. This is a memory bandwidth problem, and thus cannot be resolved (except by not allowing such modes at all, which is what is done in the current driver). Ignoring it is one option (it isn't destructive). Disabling acceleration in the Device section of the XF86Config file is another option: since the accel- erator is not being used, there is ample bandwidth to avoid such problems. 5. Clock selection problems with some ET4000 boards XFree86 has some problems getting the clock selection right with some ET4000 boards when the server is started from a high-resolution text mode. The clock selection is always correct when the server is started from a standard 80x25 text mode. This problem is indicated when the reported clocks are different when the server is started from the high-resolution text mode from what they are when it is started from the 80x25 text mode. To allow the server to work cor- rectly from the high-resolution text mode, there are some Option flags that may be set in XF86Config. To find out which flags to set, start the server with the -probeonly flag from an 80x25 text mode and look at the information printed by the server. If the line: VGAXXX: ET4000: Initial hibit state: low is printed, put the following in the SVGA, VGA16 and VGA2 sections of your XF86Config: Option "hibit_low" If the line: VGAXXX: ET4000: Initial hibit state: high is printed, put the following in the SVGA, VGA16 and VGA2 sections of your XF86Config: Option "hibit_high" 6. Text mode restore problems In XFree86 1.3, an option flag ``force_bits'' was provided as an experiment to attempt to alleviate text-restoration problems that some people experi- enced. We have now made the behavior of this option the default, hence the flag has been removed. Hopefully the past text-restoration problems are alleviated in XFree86 2.0. 7. Basic configuration It is recommended that you generate an XF86Config file using the XF86Setup' or xf86config' program, which should produce a working high-resolution 8bpp configuration. You may want to include mode timings in the Monitor section that better fit your monitor (e.g 1152x864 modes). The driver options are described in detail in the next section; here the basic options are hinted at. If graphics redrawing goes wrong on accelerated chips (ET4000W32 and ET6000), first try the "noaccel" option, which disables all accelerated functions. 8. general options in the XF86Config file The following options are of particular interest to the Tseng driver. Each of them must be specified in the svga' driver section of the XF86Config file, within the Screen subsections of the depths to which they are applicable (you can enable options for all depths by specifying them in the Device section). Option "noaccel" (ET4000W32p, et6000) This option will disable the use of any accelerated functions. This is likely to help with some problems related to DRAM timing, high dot clocks, and bugs in accelerated functions, at the cost of performance (which will still be rea- sonable on a local or PCI bus). This option applies only to those chips where acceleration is supported. Option "fast_dram" "slow_dram" These options set the DRAM speed of certain cards where it applies. The "slow_dram" option is always enabled on ET4000, and ET4000W32. If enabled, it slows down DRAM timing, which may avoid some memory-related problems. If your card starts up with a black screen (and possibly a system hang), this option might be needed. The "fast_dram" option will cause the driver to speed up DRAM timings, which may also avoid screen-related problems (streaking, stripes, garbage, ...). It may also increase those very same effects. All in all, these are potentially dangerous options: they could crash your machine as soon as you start the server. Use them with caution. option "w32_interleave_off" "w32_interleave_on" (W32i, W32p) Force memory interleaving off or on. W32i and W32p chips can increase memory bandwidth when they have 2MB or more video mem- ory. Normally the VGA BIOS sets the W32i or W32p chip to the correct mode. If you suspect problems with memory sizing or interleaving, fooling around with these options may improve the situation. It may also make things worse. These options are not normally needed: the server will use the correct value automati- cally. Setting this option the wrong way will result in a com- pletely distorted display. option "pci_burst_off" "pci_burst_on" (W32p) This option disables or enables PCI bursts on the W32p chip if it's a PCI card. Normally, a good BIOS will set the motherboard and the VGA card to the same setting, but if both don't match, you may experience garbage on the screen (e.g. mouse droppings). These options allow you to match the W32p burst setting to the motherboard setting. videoram 1024 (or another value) (all chips) This option will override the detected amount of video memory, and pretend the given amount of memory is present on the card. This is useful on cards with 2Mbyte of memory whose DRAM configu- ration is not compatible with the way the driver enables the upper megabyte of memory, or if the memory detection goes wrong. It must be specified in the Device section. Clockchip "et6000" (et6000) This enables programmable clocks, but obviously only on the et6000. It must be specified in the Device section. Normally the server will automatically use this feature when it detects an ET6000. Use it only when you suspect auto-detection is not work- ing. Note that some frequencies may be unstable (resulting in a `Wavy' screen). Only tried and tested frequencies (like the default clocks) are guaranteed to be stable. If this happens, try a slightly different frequency in the modeline (like 0.5 MHz more or less). The monitor should still be capable of syncing to this frequency, but the clockchip may already be outside an unstable region. Option "linear" (ET4000W32i, ET4000W32p, ET6000) This enables linear addressing, which is the mapping of the entire framebuffer to a high address beyond system memory, making the full length of the framebuffer directly accessible. In this way, slow SVGA bank switching (where only a small fraction of the framebuffer is visible at one time) is not necessary. It enhances performance at 256 colors, and is currently required for 16bpp, 24bpp, and 32bpp. MemBase 0xE0000000. (or a different address) (ET4000W32, ET6000) This sets the physical memory base address of the linear frame- buffer. It must be specified in the Device section. It may be required for non-PCI linear addressing configurations, and might be useful for PCI-based systems where auto-detection fails. How- ever, almost all PCI systems will not need this. Read the section on linear memory base address issues below! Read the section on linear memory base address issues below! (Message repeated for a very good reason) Use this option ONLY if you have trouble with the default MemBase used by the server, or if the server explicitly states that you must provide one. Option "pci_retry" (ET4000W32p on PCI bus, ET6000) This enables the PCI bus retry function, which is a performance enhancing mode for local bus or PCI bus-based systems, where the VGA controller will put the bus in a hold state (sort of like wait-states) when the server tries to start a new accelerated operation but the accelerator is still busy with the previous operation. This is the fastest way to drive a VGA card (no busy-waiting loops needed), but it also stresses some hardware that is timing- dependent (tape drives, sound cards, etc). See also the trouble shooting section. 9. linear memory base address (MemBase) issues First a WARNING: defining a bad MemBase may cause serious injury or death (to your operating system, of course). Especially defining the MemBase to be inside the range of system memory is a ticket to hell. 9.1 What you should know BEFORE trying another MemBase Rule #1: first, let the server find a memory base by itself, without specify- ing it. Make sure you "sync" all files to disk and close all critical appli- cations. Make sure nothing bad will happen to your filesystems if you have to jump for the power switch soon. The most critical cards are the ET4000W32p rev a and rev b on VESA local bus (VLB). The server will autodetect a linear base address that doesn't work on all systems. The least critical cards are PCI-bus cards. The PCI BIOS normally takes care of assigning a good MemBase, and you should never have to deal with all the mumbo-jumbo below. If the server gets it wrong, you may end up with a severe system crash (e.g. if it maps the video memory right on top of your system memory). If this hap- pens, RESET IMMEDIATELY. Do not try to shut down cleanly, because the X- server, thinking it writes to the VGA memory, will write to system memory instead, and you'll be writing corrupted data to disk. If you did a sync prior to starting the server, there will be no harm done (only a filesystem check which should end up clean). DO NOT attempt to redirect the server out- put to a file on the system you're testing on (that will write data after you synced). These are worst-case scenarios, and it is very unlikely this will happen to you. The text above is to make sure you are properly prepared, so that noth- ing serious happens. When the server can't find a working linear memory base, it's time to experi- ment. The rest of this section deals with that. 9.2 Choosing a MemBase Choosing a suitable MemBase can be quite tricky. If you have no way of deter- mining the MemBase your card uses, trying to put it a few Mb above the system memory is a good first guess. E.g. if you have 16 Mb of RAM, defining MemBase 0x01000000 (=16M) or 0x01400000 (=20M) may work. However, this may only work on non-PCI systems, as PCI systems mostly map all hardware above the 2GB mark. But then again, on PCI systems the server is almost always able to detect the correct linear memory base address. The only exception are those systems with more than one PCI VGA card. On most VESA local bus (VLB) boards, there is an additional problem with address decoding. Most motherboards only decode the first 32, 64 or 128 MB of address space (a good pointer is to check the amount of DRAM that can be installed on the board: it will at least decode as much address space as it supports DRAM). On such boards, you MUST specify a MemBase inside that range, or the actual address may wrap back onto system memory: if your system only decodes 128MB of addresses, and you set the MemBase to 128 MB, it will actually be decoded as being on address 0, which is probably exactly where your kernel memory is located. That is why the general guideline of putting the MemBase just above the system memory is a sound one: it stands most chance of actually being inside the decoded address range of the board. Unless your motherboard's entire memory space is filled with RAM. 9.3 An alternative approach If you don't know how much memory address space your motherboard decodes (and who does?), try using a "non-trivial" address, like 0x1FC00000, which has enough bits set to "1" to work on any motherboard, even if a few are not decoded. Keep in mind that using for example 0x10000000 may end up right on top of your system memory if the motherboard doesn't decode all upper address bits. You will only do that once. 9.4 When all else fails... Some other VLB boards can only map the linear framebuffer above the 1GB mark (0x80000000 and up), so you must use a MemBase that is higher or equal to 0x80000000. Some other VLB boards can only map the linear framebuffer BELOW the 16 MB mark. So you may want to try booting your system with up to 12 MB of memory (some operating systems allow you to supply a boot-time parameter that limits the memory to a certain amount, so you don't have to open your computer to try this), and set the MemBase to 0x00C00000 (=12M). Unfortunately, there is no easy way to tell what system you have (these details are mostly not in the motherboard manuals). Trial and error is the only road to success here. The server code will provide a default that works on most boards... but yours won't be one of those, of course. 9.5 Restrictions There are some limits as to where the linear memory base may be put. On any ET4000W32, it must have a 4MB granularity (i.e. it can be put at 16M or at 20M, but not at 18M). On ET6000, it needs a 16M granularity (note: the ET6000 driver should be able to determine the linear memory base automatically, so you should never need to define MemBase in the first place). On ET4000W32i, things are worse: the linear address base is hardwired on the card, and there is no reliable way to read it back from the card. You need to know the address in some way, and specify it. The current code does an intel- ligent guess at it, but this is no guarantee. On ISA cards, things are much more simple: ISA only uses 24 address lines, and hence the linear memory MUST lie within the 16 MB boundary. Together with the 4MB granularity of the linear memory base address on ET4000 cards, this means that you cannot have more than 12 MB of system memory in the machine if you want to use linear memory. Hence, the only realistic MemBase for ISA cards is 0x00C00000. This is also what the server will automatically choose if it detects an ISA W32 card. WARNING: you must not have over 12 MB of system memory in this case. Or if you have it, you must disable access to all memory above the 12 MB mark. Some operating systems allow you to specify at startup how much memory it is allowed to use, so you don't have to unplug some memory each time you want to use linear memory. 9.6 Some boards simply cannot work in linear mode Yes, and in that case, you're out of luck. There can be at least two reasons for this. The first is the most common: the board manufacturer has left out the neces- sary connections and hardware to be able to use linear addressing. This means that no coding effort on this planet can help you with your problem: it is physically impossible to use linear addressing. The second reason is that the current XFree86 Tseng linear addressing code is incompatible with the way your board is designed. The XFree86 Tseng code assumes a 1:1 mapping of the address lines from the bus (either ISA, VLB or PCI) to the address lines on the Tseng VGA chip. As unlikely as it may sound, this may NOT be the case! Some very rare boards do not have such a 1:1 mapping (e.g. two address lines swapped). It is possible to support this type of hardware, but at this moment, this has not been implemented yet. Other boards use external address decoding hardware that combines a number of address lines on the bus to a (smaller) number of address lines to the VGA chip. One such board for example uses three NOR gates (one 74F02 chip) to combine the 6 upper address lines to three address pins on the W32i chip. Obviously, this represents a 2:1 mapping, and not a 1:1 mapping. Therefor, this board is not "compatible" with the way XFree86 implements linear mode. 9.7 How can I see if the linear address is wrong? Simple: nothing works, or your machine locks solid, or it crashes, or a zil- lion of other things. However, sometimes it is not always as obvious. Sometimes nothing bad hap- pens: you just get a black screen, or a screen with rubbish on it, but noth- ing is drawn on it. Sometimes you get a core dump when the first application starts. If acceleration is enabled in those cases, you will almost always see multi- ple "WAIT_ACL: timeout" messages in the server output. That is because the accelerator registers are also mapped in the linear memory, and if linear memory doesn't work, then also the accelerator doesn't work. NOTE however that a WAIT_ACL message doesn't necessarily mean the linear mem- ory address is bad. There are a number of other reasons for this message as well. But if you never saw these messages at 8bpp banked, then there's a good chance you have a linear memory problem ("banked" is the opposite of "lin- ear", and is the default mode when "option linear" is not in the XF86Config file). 10. Mode issues The accelerated driver on ET4000W32/W32i/W32p and ET6000 needs at least 1K bytes of scratch space in video memory. Consequently, if you want accelera- tion, a 1024x1024 virtual resolution should not be used with a 1Mbyte card. This also means that a 1024x768 mode at 24bpp on a 2.25 MB ET6000 card cannot be accelerated, since you've used up all the memory for the display. The same thing goes for the ET6000 hardware cursor: it also requires 1kb of free video memory. If that memory is not available, the hardware cursor can- not be used. The use of a higher dot clock frequencies has a negative effect on the per- formance of graphics operations on non-et6000 cards (the effect is much less, or even non-existing, on ET6000 cards), especially BitBlt, when little video memory bandwidth is left for drawing. Memory bandwidth is the speed at which data can be pumped into the memory while the RAMDAC is pulling it out to dis- play it on the screen. Higher dot-clocks (mostly related to higher resolutions) consume more band- width, so that less of it is left for drawing into the framebuffer. With a working accelerator, things become increasingly crammed, because modern accelerators can consume huge amounts of bandwidth (but they also give you high speeds in return). High color depths also need extra bandwidth. If you are short on memory bandwidth (see the separate section on this) and experience blitting slowness or screen "glitches", try using the lowest dot clock that is acceptable; for example, on a 14" or 15" screen 800x600 with high refresh (50 MHz dot clock) is not so bad, with a large virtual screen. Tseng chips are mostly known for their (very) good memory bandwidth, so you should only start to see problems in the higher regions. It does not make much sense performance-wise to use the highest clock (85 MHz) for 1024x768 at 76 Hz on a 1 MB ET4000W32; the card will very slow, because there is almost no bandwidth left for drawing. A 75 MHz dot clock results in 70 Hz which should be acceptable. If you have a monitor that sup- ports 1024x768 at 76 Hz with a 85 MHz dot clock, an 1MB card is a poor match anyway. The ET4000W32i and ET4000W32p have a special feature that almost doubles mem- ory bandwidth (+70%) using "interleaving" between the two banks. Upgrading to 2MB is a real bonus on these cards. This is not true for W32 cards or for ET6000 cards. 11. Acceleration issues The XFree Acceleration Architecture makes extensive use of the unused video memory on the VGA card. If there is not enough free video memory, some accel- eration features will be disabled or crippled, resulting in less performance. To avoid this from happening, try to keep an absolute minimum of 16 kb of free memory, in addition to the 1kb already reserved by the accelerator. In practice, this small amount of memory should not be a problem. Most cards nowadays have 2 MB of video memory, and running 1280x1024 still leaves plenty of memory unused. Even a 1600x1200 desktop will leave over 170kb unused, which will then be used by the accelerator to enhance performance. Most 1MB cards cannot display modes larger than 1024x768 with a decent refresh rate, leaving 256kb unused. The order in which free memory is used to accelerate certain features is as follows. If no video memory is unused (i.e. all of it is used for display memory), no acceleration can be used at all -- not even a hardware cursor on the ET6000. If the hardware cursor is enabled (ET6000 only) and there's at least 1kb of free video memory, 1kb is used for that. If there is at least 1kb of free memory remaining after this, most accelera- tion features are enabled as well, reserving an extra 1kb of video memory. If there's still some free memory, some extra acceleration features are enabled. These require more free video memory, depending on the virtual screen width and the color depth (bpp). The server will print out how much memory it used if it could. If there's still some free video memory, it is used as a pixmap cache. This way, small patterns and images can be kept in the video memory so that they don't need to be transferred into the video memory each time they're needed. This is beneficial because transferring an image over the bus to the video memory takes a lot more time than letting the accelerator blit it from the pixmap cache to the display memory. 12. ET6000 memory size facts and fiction The ET6000 uses a special kind of video memory called MDRAM (multi-bank DRAM). It may have a non-power-of-two amount of MDRAM: 2.25 or even 4.50 MB. Especially 2.25 MB MDRAM is popular, since this can support 1024x768 at 24bpp without needing 4MB of RAM. There are a few less intuitive problems with this. First of all, All memory above the 4 MB limit is a waste of money, because the ET6000 cannot use this memory for anything at all. There are boards with 4.5 MB around, but that extra 0.5 MB is a waste. The ET6000 can only refresh 4 MB of (M)DRAM (refresh register). It can only access 64 banks of 64KB in VGA mode (bank select register). All accelerated commands use a 22-bit address (=4MB) inside the video memory. You get the idea... There is no way for the ET6000 to use anything above the 4Mb limit. And Secondly (more importantly): you may not have 2.25 MB at all! There have been several reports about ET6000 cards that were sold with (supposedly) 2.25 MB of MDRAM, but which turned out to be standard 2MB MDRAM cards. People have been having trouble with these all along, since sometimes the X-server used to detect this as 2.25 MB (or even 2.5 MB) due to internal chip design and also due to faulty BIOSs. This memory detection problem has been fixed now, and the server should detect the correct amount of memory. Do NOT define the amount of memory in the XF86Config yourself, unless you are absolutely sure about the amount. There is a simple way to determine the amount of MDRAM on your card beyond doubt. Look at the video card. There is one large chip with 204 pins on it, which is the ET6000. One socketed rectangular chip, mostly with a sticker on it,is the BIOS. The remaining big chips are (mostly) 2 or 4 other large square chips on it with the following markings: MDRAM MD9xy ("xy" is a two-digit number) SJ-5-100 (this may differ, but it will have the same layout) and a nice logo next to all that with 4 diamonds and the name "MoSys" under- neath. The "xy" number tells you how much MEGABITS there are in that one chip. The amount of RAM on the card is then: ("xy" * number_of_MDRAM_chips) / 8 Mbytes On my board, there are two MD908 chips, which means I have (08 * 2) / 8 = 2 MB of MDRAM. Boards with two MD909 chips have 2.25 MB, etc. Current MDRAM chips are MD904, MD906, MD908, MD909, MD910, MD916, MD918 and MD920. 13. ET6000 memory bandwidth hype and the impact on video modes Tseng has always had wet dreams about memory bandwidth, and their press announcements about the ET6000 memory bandwidth are no exception. They claim the ET6000 using MDRAM is capable of reaching an incredible 1.2 Gbytes/sec of bandwidth. That would surpass just about everything on the mar- ket (even SGI). And that would be true, _if_ they actually used the fastest available MDRAMs on their boards, which they don't. The stunning 1.2 GByte mark is only reached when using 4 MDRAM chips at their max clock rate of 166 MHz. But due to design limitations, the first-generation ET6000 can only drive the memo- ries at 92 MHz (that will change when the ET6100 and ET6300 hit the streets). This means the max. theoretical bandwidth available on current ET6000 boards is "only" 360 MB/sec on boards with 2 MDRAM chips, and 720 MB/sec on boards with 4 MDRAM chips. And this assumes a best-case situation (=extremely long bursts -- the MDRAMs use a shared address/data bus, much like the PCI bus does). In the real world, unaligned accesses both from the PCI bus and the accelerator will reduce the effective available bandwidth. The current ET6000 boards peak out at about 225 MB/sec, with 2 or 4 MDRAMs. Whatever you may have read in press releases, the ET6000 has a 32-bit memory bus (not 128 bits; that's only the accelerator data path within the chip, if anything). That means that, with their 16-bit busses, 2 MDRAM chips already use the full bus capacity. Having 4 memory chips on an ET6000 board will not give you extra memory bandwidth. Memory bandwidth limits the maximum resolution you can use at a given color depth. The ET6000 RAMDAC can cope with 135 MHz in any situation. But the RAM cannot. At 32bpp (sparse 16M color mode), using a 135 MHz pixel clock would require a memory bandwidth of 135*4 = 540 MB/sec, which the current ET6000 boards simply cannot cope with. And then you still need some spare bandwidth for the PCI bus and the accelerator. That is why some modes will be refused, depending on your MDRAM memory lay- out, even if the amount of memory would permit such a mode. See also the trouble shooting section to see what can happen if too little memory band- width is available. 14. Linear addressing and 16bpp/24bpp/32bpp modes Currently the 16-bit (32768 or 65536 colors), 24-bit (16M colors, packed pixel), and 32-bit (16M colors, sparse) pixel support in the SVGA server requires linear addressing. This restriction may be removed in a future ver- sion, but with nearly all new cards using the PCI bus (where linear address- ing poses no problem), removing the linear addressing requirement presently has a lower priority than other features. Option "linear" can be specified in a depth-specific screen section to enable linear addressing; a MemBase set- ting (in the device section) is probably also required on non-PCI based sys- tems, and optionally on PCI systems that have trouble finding out for them- selves where the MemBase is. Non-PCI cards are not (or not well) supported in linear memory mode at this moment. Some of them don't support it at all, and some of the ones that do have so many address decoding bugs that it isn't feasible to provide a work- ing solution. For the most part, many of the accelerated features in the 8bpp server have been implemented to support 16, 24, and 32 bpp modes for the W32 and the ET6000. So although there are now up to 4 times as many bits to display, the X server shouldn't feel overly sluggish. Note also that the 24bpp and 32bpp modes are only supported on a limited set of cards, and with at least 2Mb of memory. An ET6000 with 2.25 MB MDRAM is cheap-and-sound, since it can support exactly 1024x768 at 24bpp (using all available video memory). On all other video cards, you need at least 4MB of video memory to do this. With only 2MB of (M)DRAM, 960x720 is the best you can hope for. In the XF86Config "Screen" section, a "Display" subsection must be defined for each depth that you want to run, with separate Modes and virtual screen size. Example (2Mb of video memory): Section "screen" SubSection "Display" Depth 8 Virtual 1280 1024 ViewPort 0 0 Modes "640x480" "800x600" "1024x768" EndSubSection SubSection "Display" Depth 16 Virtual 1024 992 ViewPort 0 0 Modes "640x480" "800x600" "1024x768" EndSubSection SubSection "Display" Depth 24 Virtual 960 720 ViewPort 0 0 Modes "640x480" "800x600" EndSubSection SubSection "Display" Depth 32 Virtual 832 600 ViewPort 0 0 Modes "640x480" "800x600" EndSubSection EndSection 15. Trouble shooting with the SVGA Tseng driver First of all, make sure that the default modes selected from your XF86Config are supported by your monitor, i.e. make sure the horizontal sync limit is correct. It is best to start with standard 640x480x256 with a 25.175 MHz clock (by specifying a single horizontal sync of 31.5) to make sure the driver works on your configuration. The default mode used will always be the first mode listed in the modes line, with the highest dot clock listed for that resolution in the timing section. Some general hints: o Put Option "slow_dram" in the Device Section. o Put Option "pci_burst_off" in the Device Section. o Put Option "w32_interleave_off" in the Device Section. o Take out the Hercules monochrome adapter, if you have one. Many config- urations of the ET4000/W32 series do not allow one in the system. o Get a motherboard with its local bus running at 33 MHz. Many, if not all, ET4000/W32 boards will surely behave in a funny way on a 50-MHz bus. You may have to use a wait state or two, but first try without any. o Cold-boot your machine. Do not run anything that messes with the video hardware, including other X servers, before running XF86_SVGA. o In case of an ET6000 card, try specifying chipset "et6000" in the Device Section. The card normally auto-probes from the PCI bus, but on some systems, another on-board VGA card, although disabled, may cause the ET6000 server to want to use the other card. Note that some VESA standard mode timings may give problems on some monitors (try increasing the horizontal sync pulse, i.e. the difference between the middle two horizontal timing values, or try multiples of 16 or 32 for all of the horizontal timing parameters). There is a video signal, but the screen doesn't sync. You are using a mode that your monitor cannot handle. If it is a non-standard mode, maybe you need to tweak the timings a bit. If it is a standard mode and frequency that your monitor should be able to handle, try to find different timings for a similar mode and frequency combination. Horizontal jitter at high dot clocks. This problem shows up especially when drawing operations such as scrolling or blitting are in progress. There is currently no easy fix for this, You can try the "fast_dram" option, or use a lower dot clock. If that is not sufficient, the "noaccel" option will almost always help (it leaves more bandwidth for the RAMDAC). In most cases, this is caused by the video memory not being able to provide pixel data to the RAMDAC fast enough, so it gets fed with garbage. `Wavy' screen. Horizontal waving or jittering of the whole screen, continuously (independent from drawing operations). You are probably using a dot clock that is too high; it is also possible that there is interference with a close MCLK. Try a lower dot clock (sometimes even dropping it by 0.5 MHz may work). You can also try to tweak the mode timings; try increasing the second horizontal value somewhat. Here's a 65 MHz dot clock 1024x768 mode (about 60 Hz) that might help: "1024x768" 65 1024 1116 1228 1328 768 783 789 818 Crash or hang after start-up (probably with a black screen). Try the "noaccel" option. Check that the BIOS settings are OK; in particular, disable caching of 0xa0000-0xaffff. Disabling hidden DRAM refresh may also help. On Linux systems, if "APM" (power management) support is enabled in the kernel, the server may start up in power-save mode or with a black screen. Rebuild your kernel with APM support disabled. Crash, hang, or trash on the screen after a graphics operation. This may be related to a bug in one of the accelerated functions, or a problem with the BitBLT engine. Try the "noaccel" option. Also check the BIOS settings. `ACL: TIMEOUT' messages from the server. Same as for the above entry. However, on some systems, the prob- lem will not go away no matter what you do. It may be related to the operating system you use (it has only been seen on Linux sys- tems, and even then it depends on the kernel versions). Some- times, choosing another MemBase may help. Occasional erroneous pixels in text, pixel dust when moving window-frame Probably related to MCLK setting that is too high (can happen with linear addressing even though banked mode runs OK). Most (if not all) ET6000 cards are sold with the MCLK slightly over clocked for performance (the current norm is 90 or 92 MHz), which may cause these problems. There is currently no fix for this. If the pixel dust is only temporary (it disappears as soon as noth- ing moves on the screen anymore), then memory bandwidth is proba- bly the cause. The only solution is to disable acceleration, or, if that doesn't help, using a lower pixel clock. Textmode is not properly restored This has been reported on some configurations. Sometimes a Chipset line will fix this. Normally you should be able to restore the textmode font using a utility that sets it (setfont, runx, restorefont on Linux). Mostly black or blue screen when using accelerated driver features If you are seeing a mostly black or blue screen, with only a few icons (pixmaps) displayed, this section applies to you. There can be several causes for this. One is if the amount of memory is not detected (or specified) correctly. If the server's autodetection mechanism detects too much memory, accelerated features will not work. Define the amount of memory in the XF86Config file. This seems to happen sometimes on some 2.25 MB ET6000 cards, where the server detects 2.5 MB instead (add videoram "2304" in this particular case). If that doesn't help, disabling acceleration (option "noaccel") is the only solution. Problems with DMA hardware (floppy, tape) On some systems, the accelerated server will interfere with other hardware that uses ISA DMA. Most notably is the PC floppy con- troller and sound cards. The floppy interface cannot cope with inordinately long bus-holds, which may occur during large accel- erated operations. The Linux-ftape module for example (a floppy- tape driver) will generate lots of "write error" messages when running a backup or restore operation while the X-server is in use. These errors should not be fatal, but that all depends on how well the operating system handles these conditions. Linux seems to cope. There are two possible solutions: disable acceleration using the "noaccel" option, or disable PCI-retry (which is causing the large bus delays) by removing the "pci_retry" option. This will cause a very small slowdown of accelerated operations. The "pci_retry" option applies not only to the PCI bus systems, but has a similar effect on other busses. "Cannot read colourmap from VGA. Will restore with default" If this error occurs, the server was unable to properly initial- ize the RAMDAC, and tries to restore a default color map. On some unsupported RAMDACs, this will have the adverse effect of remov- ing all color altogether, leaving you with a bunch of weird col- ors, or with a completely black screen. If that happens, add the ramdac "normal" statement to the Device section in your XF86Con- fig file. In most cases, this will solve the color problem. Why does the server report my ModeLine with only half the pixel clock? For ET4000W32p cards at 8bpp, some modes using a clock over 75 MHz (e.g. a 1152x910 mode with 95 MHz pixel clock) will produce the following message in the Xserver output: (--) SVGA: Mode "1152x910" will use pixel multiplexing And later, when the accepted modelines are reported: (**) SVGA: Mode "1152x910": mode clock = 47.500 This is normal, because with pixel multiplexing, only half the clock is needed as two pixels are sent to the RAMDAC per clock pulse. For other screen drawing related problems, try the "noaccel" option. If you are having driver-related problems that are not addressed by this doc- ument, or if you have found bugs in accelerated functions, you can try con- tacting the XFree86 team. In fact, reports (success or failure) are very welcome, especially on config- urations that have not been tested. You can do this via the BetaReport form (mail it to report@XFree86.org). You may want to keep an eye on forthcoming beta releases at www.xfree86.org. 16. Acknowledgments Most of these stem from the old XF86_W32 server. That code was used exten- sively for getting the SVGA server to work on all the Tseng cards, so they are still somewhat valid. Glenn G. Lai wrote the original XF86_W32 server. It was modified by Dirk Hohndel and Koen Gadeyne to support some more hardware. Jerry J. Shekhel (jerry@msi.com) gave me (GGL) the 1-M Mirage ET4000/W32 VLB board on which the initial development (X_W32) was done. X11R6 and The XFree86 Project provide the base code for XF86_W32. Hercules Computer Technology Inc. lent me (GGL) a 2-M Hercules Dynamite Pro VLB board for the development that led to XF86_W32. They donated a Dynamite Power PCI to The XFree86 Project, that was used by DHH to extend the server. Tseng Labs kindly donated (KMG) an ET6000-based board (a Jazz Multimedia G- Force 128), which spurred the development of the ET6000 code. They also pro- vided an ET6100 evaluation board. Heiko Eissfeldt provided an ET4000W32p_rev_b board which allowed us to get better support for those rev_a and rev_b boards. Gyorgy Krajcsovits donated an ET4000W32p + CH8398 board. A Really Good Move! Numerous testers have given me feedback for X_W32 and later XF86_W32. I apologize for my failure to keep track of the people who tested X_W32, but the names of the people involved with the XF86_W32 testing are listed below: Linux: bf11620@coewl.cen.uiuc.edu (Byron Thomas Faber) dlj0@chern.math.lehigh.edu (David Johnson) peterc@a3.ph.man.ac.uk (Peter Chang) dmm0t@rincewind.mech.virginia.edu (David Meyer) nrh@philabs.Philips.COM (Nikolaus R. Haus) jdooley@dbp.caltech.edu (James Dooley) thumper@hitchcock.eng.uiowa.edu (Timothy Paul Schlie) klatta@pkdla5.syntex.com (Ken Latta) robinson@cnj.digex.net (Andrew Robinson) reggie@phys.washington.edu (Reginald S. Perry) sjm@cs.tut.fi (M{kinen Sami J) engel@yacc.central.de (C. Engelmann) use cengelm@gwdg.de postgate@cafe.net (Richard Postgate) are1@cec.wustl.edu (Andy Ellsworth) bill@celtech.com (Bill Foster) FreeBSD: ljo@ljo-slip.DIALIN.CWRU.Edu (L Jonas Olsson) Several people have developed code for the SVGA Tseng driver (this list is incomplete): o Glenn G. Lai o Dirk H. Hohndel o Koen Gadeyne o OEyvind Aabling o Dejan Ilic o Mark Vojkovich o Harald Nordgard Hansen o David Bateman o Gyorgy Krajcsovits o Kurt Olsen Generated from XFree86: xc/programs/Xserver/hw/xfree86/doc/sgml/tseng.sgml,v 3.15.2.18 1998/11/13 13:00:48 dawes Exp $ $XConsortium: tseng.sgml /main/6 1996/10/27 11:06:09 kaleb $