Information for S3 Chipset Users The XFree86 Project Inc. 27 February 1998 1. Supported hardware The current S3 Server supports the following S3 chipsets: 911, 924, 801/805, 928, 732 (Trio32), 764, 765, 775, 785 (Trio64*), 864, 868, 964, 968 and M65 (Aurora64V+). The S3 server will also recognise the 866, but it has not been tested with this chipset. If you have any problems or success with these, please report it to us. Nevertheless, this is not enough to support every board using one of these chipsets. The following list contains some data points on boards that are known to work. If your card is similar to one of the described ones, chances are good it might work for you, too. S3 801/805, AT&T 20C490 (or similar) RAMDAC o Orchid Fahrenheit 1280+ VLB o Actix GE32 8 and 15/16 bpp Note: Real AT&T20C490 RAMDACs should be automatically detected by the server. For others which are compatible, you need to provide a `Ramdac "att20c490"' entry in your XF86Config. Real AT&T 20C490 or 20C491 RAMDACs work with the "dac_8_bit" option. Some clones (like the Winbond 82C490) do not. The Orchid Fahrenheit 1280+ VLB may require `Option "nolinear"'. S3 805 VLB, S3 GENDAC (RAMDAC + clock synthesizer) o MIRO 10SD (available for VLB and PCI) It is not known whether all 10SDs use the S3 GENDAC. 8 and 15/16 bpp ClockChip "s3gendac" RamDac "s3gendac" S3 801/805, AT&T 20C490 RAMDAC, ICD2061A Clockchip o STB PowerGraph X.24 S3 (ISA) 8 and 15/16 bpp Note: Real AT&T20C490 RAMDACs should be automatically detected by the server. For others which are compatible, you need to provide a `Ramdac "att20c490"' entry in your XF86Config. ClockChip "icd2061a" RamDac "att20c490" Option "dac_8_bit S3 805, Diamond SS2410 RAMDAC, ICD2061A Clockchip o Diamond Stealth 24 VLB 8 and 15bpp(*) only. requires `Option "nolinear"' (*) The SS2410 RAMDAC is reportedly compatible with the AT&T20C490 in 15bpp mode. To make the server treat it as an AT&T20C490, you need to provide a `Ramdac "att20c490"' entry in your XF86Config. S3 801/805, Chrontel 8391 Clockchip/Ramdac o JAX 8241 o SPEA Mirage 8 and 15/16 bpp. The 8391 is compatible with the AT&T 20C490 RAMDAC ClockChip "ch8391" Ramdac "ch8391" Option "dac_8_bit" S3 928, AT&T 20C490 RAMDAC o Actix Ultra 8 and 15/16 bpp Note: Real AT&T20C490 RAMDACs should be automatically detected by the server. For others which are compatible, you need to provide a `Ramdac "att20c490"' entry in your XF86Config. Also, the server's RAMDAC probe reportedly causes problems with some of these boards, and a RamDac entry should be used to avoid the probe. Real AT&T 20C490 or 20C491 RAMDACs work with the "dac_8_bit" option. Some clones (like the Winbond 82C490) do not. S3 928, Sierra SC15025 RAMDAC, ICD2061A Clockchip o ELSA Winner 1000 ISA/EISA (``TwinBus'', not Winner1000ISA!!) o ELSA Winner 1000 VL 8, 15/16 and 24(32) bpp Supports 8bit/pixel RGB in 8bpp and gamma correction for 15/16 and 24bpp modes 24 bpp might get ``snowy'' if the clock is near the limit of 30MHz. This is not considered dangerous, but limits the usability of 24 bpp. D-step (or below) chips cannot be used with a line width of 1152; hence the most effective mode for a 1 MB board is about 1088x800x8 (similar to 2 MB, 1088x800x16). ClockChip "icd2061a" S3 928, Bt9485 RAMDAC, ICD2061A Clockchip o STB Pegasus VL 8, 15/16 and 24(32) bpp Supports RGB with sync-on-green if "sync_on_green" option is pro- vided and board jumper is set for BNC outputs. VLB linear addressing now occurs at 0x7FCxxxxx so that 64MB or more main memory can be supported without losing linear frame buffer access. ClockChip "icd2061a" Option "stb_pegasus" S3 928, Bt485 RAMDAC, SC11412 Clockchip o SPEA Mercury 2MB VL 8, 15/16 and 24(32) bpp ClockChip "SC11412" Option "SPEA_Mercury" S3 928, Bt485 RAMDAC, ICD2061A Clockchip o #9 GXE Level 10, 11, 12 8, 15/16 and 24(32) bpp ClockChip "icd2061a" Option "number_nine" S3 928, Ti3020 RAMDAC, ICD2061A Clockchip o #9 GXE Level 14, 16 8, 15/16 and 24(32) bpp Supports RGB with sync-on-green ClockChip "icd2061a" Option "number_nine" S3 864, AT&T20C498, ICS2494 Clockchip o MIRO 20SD (BIOS 1.xx) The ICS2494 is a fixed frequency clockchip, you have to use X -probeonly (without a Clocks line in XF86Config) to get the cor- rect clock values. 8, 15/16 and 24(32) bpp S3 864, AT&T20C498 or STG1700 RAMDAC, ICD2061A or ICS9161 Clockchip o Elsa Winner1000PRO VLB o Elsa Winner1000PRO PCI o MIRO 20SD (BIOS 2.xx) o Actix GraphicsENGINE 64 VLB/2MB 8, 15/16 and 24(32) bpp ClockChip "icd2061a" S3 864, 20C498 or 21C498 RAMDAC, ICS2595 Clockchip o SPEA MirageP64 2MB DRAM (BIOS 3.xx) 8, 15/16 and 24(32) bpp Clockchip support is still sometimes flaky and on some machines problems with the first mode after startup of XF86_S3 or after switching back from VT have been seen; switching to next mode with CTRL+ALT+``KP+'' and back seems to solve this problem. Interlaced modes don't work correctly. Mirage P64 with BIOS 4.xx uses the S3 SDAC. ClockChip "ics2595" S3 864, S3 86C716 SDAC RAMDAC and Clockchip o Elsa Winner1000PRO o MIRO 20SD (BIOS 3.xx) o SPEA MirageP64 2MB DRAM (BIOS 4.xx) o Diamond Stealth 64 DRAM 8, 15/16 and 24 bpp S3 864, ICS5342 RAMDAC and Clockchip o Diamond Stealth 64 DRAM (only some cards) 8, 15/16 and 24 bpp ClockChip "ics5342" Ramdac "ics5342" S3 864, AT&T21C498-13 RAMDAC, ICD2061A Clockchip o #9 GXE64 - PCI 8, 15/16, 24(32) bpp ClockChip "icd2061a" Option "number_nine" S3 964, AT&T 20C505 RAMDAC, ICD2061A Clockchip o Miro Crystal 20SV 8, 15/16, 24(32) bpp ClockChip "icd2061a" Ramdac "att20c505" S3 964, Bt485 RAMDAC, ICD2061A Clockchip o Diamond Stealth 64 8, 15/16, 24(32) bpp ClockChip "icd2061a" S3 964, Bt9485 or AT&T 20C505 RAMDAC, ICS9161a Clockchip o SPEA Mercury 64 8, 15/16, 24(32) bpp ClockChip "ics9161a" Option "SPEA_Mercury" S3 964, Ti3020 RAMDAC, ICD2061A Clockchip o Elsa Winner2000PRO PCI 8, 15/16, 24(32) bpp ClockChip "icd2061a" S3 964, Ti3025 RAMDAC, Ti3025 Clockchip o Miro Crystal 40SV o #9 GXE64 Pro VLB o #9 GXE64 Pro PCI 8 bpp, 15, 16 and 24(32) bpp There are some known problems with the GXE64 Pro support, includ- ing some image shifting/wrapping at 15/16/24 bpp. We have found that #9 no longer support the GXE64 Pro at 1600x1200. They do however have a new (and more expensive) board called the GXE64Pro-1600 which uses a 220MHz RAMDAC instead of 135MHz part used on the other boards. S3 764 (Trio64) o SPEA Mirage P64 (BIOS 5.xx) o Diamond Stealth 64 DRAM o #9 GXE64 Trio64 8/15/16/24 bpp Note: The Trio64 has a builtin RAMDAC and clockchip, so the server should work with all Trio64 cards, and there is no need to specify the RAMDAC or clockchip in the XF86Config file. S3 732 (Trio32) o Diamond Stealth 64 DRAM SE 8/15/16/24 bpp Note: The Trio32 has a builtin RAMDAC and clockchip, so the server should work with all Trio32 cards, and there is no need to specify the RAMDAC or clockchip in the XF86Config file. S3 868, S3 86C716 SDAC RAMDAC and Clockchip o ELSA Winner 1000AVI o Diamond Stealth Video DRAM 8/15/16/24 bpp S3 868, AT&T 20C409 RAMDAC and Clockchip o ELSA Winner 1000AVI 8/15/16/24 bpp Note: pixelmultiplexing is not supported yet, therefore limited maximum dot clock for 8bpp (currently 67.5MHz, should be changed to 100MHz if pixmux isn't fixed prior to release) S3 968, Ti3026 RAMDAC, Ti3026 Clockchip o Elsa Winner 2000PRO/X-2 and /X-4 (Revsions <= F) o Elsa Winner 2000AVI-2 and -4 o Diamond Stealth 64 VIDEO VRAM 8/15/16/24 bpp S3 968, Ti3026 RAMDAC, ICS9161A Clockchip o Elsa Winner 2000PRO/X-2 and /X-4 (Revision G) 8/15/16/24 bpp Note: clock doubling doesn't work, yet, therefore the maximum usable dot clock is limited to about 120MHz. S3 964, IBM RGB 514/524/525/528 RAMDAC & Clockchip o Hercules Graphics Terminator 64 8/15/16/24 bpp s3RefClk 50 DACspeed 170 Option "slow_vram" S3 968, IBM RGB 514/524/525/528 RAMDAC & Clockchip o Genoa Genoa VideoBlitz III AV s3RefClk 50 DACspeed 170 o Hercules Graphics Terminator Pro 64 s3RefClk 16 DACspeed 220 This card may require the line: Invert_VCLK "*" 0 in each Display subsection. o STB Velocity 64 s3RefClk 24 DACspeed 220 o Number Nine FX Motion 771 s3RefClk 16 DACspeed 220 This card may require the line: Invert_VCLK "*" 0 in each Display subsection. o MIRO 80SV s3RefClk 16 DACspeed 250 8/15/16/24 bpp ELSA Winner 2000PRO/X-8 (S3 968, 8MB VRAM, 220MHz for 32bpp) The server has only been tested for "revision C" of this card (guess the serial number should start with C, but not sure since mine says Ser.No. A-0000.000.000;) which have an IBM RGB528A note the A; can't be probed though) depending on the mode line etc there may be some display distor- tions like: 1. many long horizontal lines/stripes 2. pixel jitter or short horizontal stripes like snow all over the screen 3. Like 2., but only when doing graphics ops (like opaque move of windows). 4. additional pixel at the left display edge and some missing pixels at the right edge. All of these problems can be fixed by small adjustments to the mode line (best to run `xvidtune' and make these adjustments interactively). E.g., for the first three problems, shift the display left or right a few steps. For the last problem, increasing HSyncEnd (making the hsync pulse longer) solves the problem. In some cases, a significant increase in the sync pulse width is needed, and rarely, it needs to be shortened (by decreasing HSyncEnd). In rare cases, InvertVCLK and/or EarlySC may need to be adjusted, followed by an adjustment of BlankDelay (see the bottom line of xvidtune). If you see any of these problems, please contact koenig@XFree86.org, and send details of: o Original mode showing the problem, o Tuned/fixed mode, including all flags from the bottom line of xvidtune, o Colour depth used for this tuned mode line, o Full server startup output. 2. 16bpp and 32bpp On 801/805 + AT&T490 Cards (like the Fahrenheit 1280+ VLB) only 15 and 16bpp are supported. 32bpp isn't available on this type of card. (There is a 24 bit mode under MS Windows, but it's not a 32bpp sparse mode but a real 3 bytes/pixel mode). 3. List of Supported Clock Chips ICD2061A ==> ClockChip "icd2061a" ICS9161A (ICD2061A compatible) ==> ClockChip "ics9161a" DCS2824-0 (Diamond, ICD2061A comp.) ==> ClockChip "dcs2824" S3 86c708 GENDAC ==> ClockChip "s3gendac" ICS5300 GENDAC (86c708 compatible) ==> ClockChip "ics5300" S3 86c716 SDAC ==> ClockChip "s3_sdac" ICS5342 GENDAC ==> ClockChip "ics5342" STG 1703 ==> ClockChip "stg1703" Sierra SC11412 ==> ClockChip "sc11412" ICS2595 ==> ClockChip "ics2595" TI3025 ==> ClockChip "ti3025" TI3026 ==> ClockChip "ti3026" IBM RGB 5xx ==> ClockChip "ibm_rgb5xx" Chrontel 8391 ==> ClockChip "ch8391" AT&T 20C409 ==> ClockChip "att20c409" AT&T 20C499 (untested) ==> ClockChip "att20c499" 4. List of Supported RAMDAC Chips If you have a RAMDAC that is not listed here, be VERY careful not to over- drive it using XF86_S3. Better contact the XFree86 team first to verify that running XF86_S3 will not damage your board. RAMDACs that are grouped together below are treated as compatible with each other as far as the server is concerned. For example, the server will report "bt485" when you actually specify RAMDAC "bt9485", or "s3_gendac" when you specify RAMDAC "ics5300". ATT20C409 ==> RAMDAC "att20c409" ATT20C490 ==> RAMDAC "att20c490" ATT20C491 ==> RAMDAC "att20c491" CH8391 ==> RAMDAC "ch8391" ATT20C498 ==> RAMDAC "att20c498" ATT21C498 ==> RAMDAC "att21c498" ATT22C498 ==> RAMDAC "att22c498" ATT20C505 ==> RAMDAC "att20c505" BT485 ==> RAMDAC "bt485" BT9485 ==> RAMDAC "bt9485" IBMRGB514 ==> RAMDAC "ibm_rgb514" IBMRGB525 ==> RAMDAC "ibm_rgb525" IBMRGB524 ==> RAMDAC "ibm_rgb524" IBMRGB526 ==> RAMDAC "ibm_rgb526" IBMRGB528 ==> RAMDAC "ibm_rgb528" S3_GENDAC ==> RAMDAC "s3gendac" ICS5300 ==> RAMDAC "ics5300" S3_SDAC ==> RAMDAC "s3_sdac" ICS5342 ==> RAMDAC "ics5342" S3_TRIO32 ==> RAMDAC "s3_trio32" S3_TRIO64 ==> RAMDAC "s3_trio64" S3_TRIO64 ==> RAMDAC "s3_trio" SC11482 ==> RAMDAC "sc11482" SC11483 ==> RAMDAC "sc11483" SC11484 ==> RAMDAC "sc11484" SC11485 ==> RAMDAC "sc11485" SC11487 ==> RAMDAC "sc11487" SC11489 ==> RAMDAC "sc11489" SC15025 ==> RAMDAC "sc15025" STG1700 ==> RAMDAC "stg1700" STG1703 ==> RAMDAC "stg1703" TI3020 ==> RAMDAC "ti3020" TI3025 ==> RAMDAC "ti3025" TI3026 ==> RAMDAC "ti3026" None of the above ==> RAMDAC "normal" If you feel adventurous you could also open up your computer and have a peek at your RAMDAC. The RAMDAC is usually one of the larger chips (second or third largest chip that is NOT an EPROM) on the board. The markings on it are usually - For example: @@ @@ AT&T ATT20C490-11 9339S ES 9869874 This is an AT&T 20C490 with a speed grade of 110 MHz. This would then mean that you put a `DacSpeed 110' line in your XF86Config file. Be advised that some RAMDACs have different modes that have different limits. The manufac- turer will always mark the chip naming the higher limits, so you should be careful. The S3 server knows how to handle the limits for most of the RAM- DACs it supports providing the DacSpeed is specified correctly. Chips labeled -80 or -8 should use `DacSpeed 80' in the device section. S3 86C716-ME SDAC ==> DacSpeed 110 SC15025-8 ==> DacSpeed 80 ATT20C490-80 ==> DacSpeed 80 IBM 8190429 ==> DacSpeed 170 IBM 03H5428 ==> DacSpeed 170 IBM 03H6447 ==> DacSpeed 170 IBM 03H6448 ==> DacSpeed 220 IBM 03H5319 ==> DacSpeed 220 IBM 63G9902 ==> DacSpeed 250 IBM 37RGB514CF17 ==> DacSpeed 170 IBM 37RGB524CF22 ==> DacSpeed 220 ^^ 5. Additional Notes Note that the Sierra SC1148{5,7,9} will not be distinguished from the Sierra SC1148{2,3,4} by the probe. The only difference between the two series as far as the server is concerned is that the {2,3,4} is capable of 15bpp, while the {5,7,9} is capable of 16bpp. So if you have a SC1148{5,7,9} and want to use 16bpp instead of 15bpp, you will have to specify a RAMDAC "sc11485" line as shown above. Some RAMDACs (like the Ti3025) require some mode timing consideration for their hardware cursor to work correctly. The Ti3025 requires that the mode have a back porch of at least 80 pixel-clock cycles. A symptom of this not being correct is the HW cursor being chopped off when positioned close to the right edge of the screen. 6. Reference clock value for IBM RGB 5xx RAMDACs Cards with IBM RGB5xx RAMDACs use several different input frequencies for the clock synthesizer which can't be probed without some knowledge of the text mode clocks (which may be a wrong assumption if you're using non-standard text modes). Here is the procedure you should use to find out the input fre- quency: First run X -probeonly >& outfile and check the output for the probed clock chip which might look like this: (--) S3: Using IBM RGB52x programmable clock (MCLK 66.000 MHz) (--) S3: with refclock 16.000 MHz (probed 15.952 & 16.041) ^^^^^^ ^^^^^^^^^^^^^^^^^^^^^^ there will be a "good guessed" value which will be used and two probed values in brackets based on the 25MHz and 28MHz text clocks. This probing can only work if you run a normal 80x25 or 80x28 text mode! The refclock values known so far are: STB Velocity 64 24 Mhz Genoa VideoBlitz II AV 50 MHz Hercules S3 964 50 MHz Hercules S3 968 16 MHz #9 Motion 771 16 MHz depending on the quartz on your card and maybe other features like an addi- tional clock chip on the Genoa card (which as a 14.3MHz quartz). If you claim that your card has a 16MHz clock, but it really uses 50MHz, all pixel clocks will be tripled and a 640x480 mode with 25MHz will use a 75MHz pixel clock, so be very careful. If you found the right refclock, you should set it in the config file (device section) e.g. with s3RefClk 16 or s3RefClk 50 so that that this value will be used even if you use another text mode and probing fails! 7. Hints for LCD configuration (S3 Aurora64V+) If LCD is active the CRT will always output 1024x768 (or whatever is the _physical_ LCD size) and smaller modes are zoomed to fit on the LCD unless you specify Option "lcd_center" in the device section. The pixel clock for this physical size (e.g. 1024x768) mode... o ...can explicitly set in the config file (device section) with e.g. `Set_LCDClk 70' (resulting 70 MHz pixel clock being used for all modes when LCD is on) o ...is taken from the _first_ mode in the modes line iff this mode's dis- play size is the same as the physical LCD size o ...the default LCD pixel clock of BIOS initialisation setup is used. This value is output at server startup in the line `LCD size ...' unless you're specifying a value using `Set_LCDClk ...' If LCD is _not_ active, the normal mode lines and pixel clocks are used for the VGA output. Whenever you switch output sources with Fn-F5, the Xserver won't get informed and pixel clock and other settings are wrong. Because of this you have to switch modes _after_ switch output sources! Then the server will check which outputs are active and select the correct clocks etc. So the recommended key sequence to switch output is Fn-F5 Ctrl-Alt-Plus Ctrl-Alt-Minus and everything should be ok.. on the Toshiba keypad you can first hold down Ctrl-Alt, then press `Fn' addi- tionally before pressing Plus/Minus too to avoid to explicitly enable/disable the numeric keypad for mode switching. 8. How to avoid ``snowing'' display while performing graphics operations For cards with the S3 Vision864 chip, there is an automatic correction which depends on the pixel clock and the memory clock MCLK at which the S3 chip operates. For most clock chips this value can't be read (only the S3 SDAC allows reading the MCLK value so far), so this value has to be estimated and specified by the user (the default is 60 [MHz]). With the new `s3MCLK' entry for your XF86Config file, now you can specify e.g. s3MCLK 55 for a 55 MHz MCLK which will reduce snowing. Smaller MCLK values will reduce performance a bit so you shouldn't use a too low value (55 or 50 should be a good guess in most cases). Below is a small shell script which might be useful to determine the approxi- mate value for MCLK (about +/- 1-2 MHz error). Before running this script you have to add the line s3MNadjust -31 255 to the device section in your XF86Config file and restart X Windows. With this option (which is for testing and debugging only) you'll get lots of dis- astrous display flickering and snowing, so it should be removed again immedi- ately after running the test script below. Running this script will use xbench and/or x11perf to run a test to determine the MLCK value, which is printed in MHz. Up to 4 tests are run, so you'll get up to 4 estimates (where the first might be the most accurate one). #!/bin/sh exec 2> /dev/null scale=2 calc() { m=`awk 'BEGIN{printf "%.'$scale'f\n",'"( $1 + $2 ) / $3; exit}" ` [ -z "$m" ] && m=` echo "scale=$scale; ( $1 + $2 ) / $3" | bc ` [ -z "$m" ] && m=` echo "$scale $1 $2 + $3 / pq" | dc ` echo $m } run_xbench() { r=` ( echo 1; echo 2; echo 3; echo 4 ) | xbench -only $1 | grep rate ` [ -z "$r" ] && return cp="$2 $3" set $r calc $3 $cp } run_x11perf() { r=` x11perf $1 | grep trep | tr '(/)' ' ' ` [ -z "$r" ] && return cp="$2 $3" set $r calc `calc 1000 0 $4` $cp } run_x11perf "-rect500 -rop GXxor" 3.86 5.53 # 0 1 # 4.11 5.52 # run_xbench invrects500 4.63 5.48 # 0 1 # 4.69 5.48 # run_x11perf "-rect500 -rop GXcopy" -16.42 13.90 # 0 1 # -14.99 13.88 # run_xbench fillrects500 -7.81 13.57 # 0 1 # -8.53 13.58 # exit 9. New S3 SVGA driver There is a new experimental S3 driver for non-ViRGE S3 chipsets in the XF86_SVGA server. This is definitely an ALPHA quality driver and hasn't been well tested, and has some known problems. Because of this, the configuration programs will install XF86_S3 by default rather than this one. But if you're adventurous or had some problems with XF86_S3, you might want to give it a try. The driver includes generic S3 support which should work on all non-ViRGE S3 chips (in theory, that is). It also has improved support for chips that sup- port S3's new style memory mapped I/O. These chips include the 868, 968 and recent Trio64 variants (not the plain old Trio64s). Chips that are capable of using the new style MMIO will use it automatically. The option "NO_MMIO" can be used to turn this off. Performance for chips using the new style MMIO is expected to be better than XF86_S3, especially on a PCI bus. Performance without MMIO, however, is expected to be roughly comparable to XF86_S3 (faster in some areas, slower in others). All color depths achievable with XF86_S3 should be possible with these drivers. Additionally, packed 24 bpp "sort of" works for the 868 and 968. Your results may vary. Nearly all the options and features supported by XF86_S3 are supported by this driver. Additionally, the standard XAA/SVGA server options such as NO_ACCEL, SW_CURSOR, and NO_PIXMAP_CACHE are also supported. XF86_S3 fea- tures which are NOT supported in this driver are DPMS support and gamma cor- rection. The driver supports the PCI_RETRY option when using MMIO and a PCI card. This option can give large performance boosts for some operations, but has a tendency to hog the bus. Because of this, the option is not set by default. Most hardware combinations may not have any problems using this option, but sound card glitches during intensive graphics operations have been reported on some. One shortcoming worth noting is that this driver does not yet contain the work-around for some S3 PCI BIOSs that report their memory usage incorrectly. This can result in conflicting address spaces. If this is the case on your hardware you should run XF86_S3 once and write down the address that your card is relocated to (as printed out in the server output). Then you can force the server to use this address with the MemBase field in the XF86Config (see the man page on XF86Config). Generated from XFree86: xc/programs/Xserver/hw/xfree86/doc/sgml/S3.sgml,v 3.37.2.4 1998/02/27 02:34:38 dawes Exp $ $XConsortium: S3.sgml /main/14 1996/02/21 17:45:58 kaleb $