Date: Sun, 25 Aug 1996 13:10:02 -0500 (CDT) From: Rob Snow <rsnow@lgc.com> To: freebsd-hackers@FreeBSD.ORG Subject: [Fwd: Fastvid,NT,Pentium Pro Performance: Expanation] (fwd) Message-ID: <Pine.SGI.3.93.960825130712.16821A-110000@dympna>
next in thread | raw e-mail | index | archive | help
This message is in MIME format. The first part should be readable text,
while the remaining parts are likely unreadable without MIME-aware tools.
Send mail to mime@docserver.cac.washington.edu for more info.
--------------1CFBAE3959E2B60015FB7483
Content-Type: TEXT/PLAIN; CHARSET=us-ascii
Content-ID: <Pine.SGI.3.93.960825130712.16821C@dympna>
Here is some more information on the post I made a couple of weeks ago
about performance features for PPros systems and the PCI bus. This
doesn't make a ton of sense to me. However, running the DOS FASTVID
program on my Natoma increases video output from ~25MPixel/sec to
~75MPixel/sec. I thought this might be some interesting info for you
hacker dudes.
BTW, it's from comp.sys.intel
-Rob
--------------1CFBAE3959E2B60015FB7483
Content-Type: MESSAGE/NEWS
Content-ID: <Pine.SGI.3.93.960825130712.16821D@dympna>
Content-Description:
Path: dildog.lgc.com!news.sesqui.net!newsfeed.rice.edu!uw-beaver!news.u.washington.edu!news.uoregon.edu!hunter.premier.net!news-res.gsl.net!news.gsl.net!news.mathworks.com!newsfeed.internetmci.com!news.sprintlink.net!news-stk-200.sprintlink.net!news.sprintlink.net!news-chi-13.sprintlink.net!itnews.sc.intel.com!news.fm.intel.com!ornews.intel.com!news.jf.intel.com!usenet
From: Jeff Sponaugle <Jeff_Sponaugle@ccm.jf.intel.com>
Newsgroups: comp.sys.intel,comp.sys.ibm.pc.hardware.chips,alt.comp.periphs.mainboard.asus,alt.comp.periphs.mainboard.tyan,comp.os.linux.misc
Subject: Fastvid,NT,Pentium Pro Performance: Expanation
Date: Thu, 08 Aug 1996 09:50:10 -0700
Organization: Intel Corporation
Message-ID: <320A1AC2.3588@ccm.jf.intel.com>
Reply-To: Jeff_Sponaugle@ccm.jf.intel.com
NNTP-Posting-Host: jsponau.jf.intel.com
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
X-Mailer: Mozilla 3.0b4 (Win95; I)
Xref: dildog.lgc.com comp.sys.intel:96071 comp.sys.ibm.pc.hardware.chips:93295 alt.comp.periphs.mainboard.asus:23893 alt.comp.periphs.mainboard.tyan:3373 comp.os.linux.misc:130758
Subject: Fastvid,NT,Pentium Pro Performance: Explanation
Over the past few weeks, there has been a great deal of discussion
about the speed of a Pentium Pro system when compared to an
equivalent Pentium systems. In particular, the discussion has focused
around the program "fastvid", and certain "PCI features". After reading
many such threads, I felt it would be helpful to clarify these issues
and put to rest some concerns.
First, a disclaimer: I'm an engineer at Intel, however I do not speak
for Intel, nor do I represent Intel on this forum in any way. The
opinions expressed here are mine and mine alone.
I believe the first clarification should be what fastvid does.
Fastvid (the dos version), in its current form, does three things.
The first is to enable write posting. This is enabled by toggling
bit #1 in PCI config register offset 0x53. That bit enables "P6.0 to
PCI Write Posting" IMPORTANT: This bit is specific to the Intel
82450Kx/Gx Orion PCIset. This bit is slightly different in the 440FX
PCIset) This feature was not enabled in some of the first Pentium Pro
systems shipped due to a bug with the 82450Kx/Gx (stepping a3 I believe).
Enabling this feature in a system with this older rev of the chipset can
cause data loss and random lockups. (of course, many people have enabled
this without problem, and the performance of PCI writes is greatly
improved...)
If you have an older 82450 system, use this feature with caution.
If you have a new 82450, or a Natoma, the feature should already
be enabled.
The second and third thing fastvid does has nothing to do with the
chipset, but only with the processor itself. The Pentium Pro
Processor has a set of MSR (Machine specific registers) that
control the kind of caching performed on different ranges of
physical system memory. (e.g. applies to the real physical
address, not the virtual address) These MSRs are called MTTRs
(Memory type range register). There are three different kinds of
MTTRs. Ones for fixed memory ranges, Ones for variable ranges, and a
single registers for all other ranges not included in other MTTRs.
(i.e.. default)
This is where fastvid comes in. By default, the caching applied to
the physical address space used by the video adapter is set to none.
There are two MTTRs that must be adjusted in order to change this.
Most video adapters have two distinct physical address ranges, one in
the more standard location (0xa0000), and another flat linear (large)
buffer at some other location (0xfc000000 for the Matrox Millennium).
Fastvid changes both the fixed and variable MTTRs for these ranges
to enable a caching protocol called "Write Combining". (see pages
11-12 thru 11-26 in The Pentium Pro Family Developer's Manual, Volume
3: Operating System Writer's Manual)
"Write Combining" is defined as: {per Intel Docs} " System memory
locations are not cached and coherency is not enforced by the
processor's bus coherency protocol. Speculative reads are allowed.
Writes may be delayed and combined in the write buffer to reduce
memory accesses. This type of cache-control is appropriate for frame
buffers, where the order of the writes is unimportant as long as the
writes update memory so they can be seen on the graphics display."
This kind of caching is IDEAL for frame buffers. The speed up is
great due to the reduction in small PCI write cycles. THE IMPORTANT
THING TO NOTE HERE IS THAT "Write Combining" IS NOT A FEATURE OF THE
CHIPSET, BUT OF THE PENTIUM PRO PROCESSOR ITSELF.
When you run fastvid on a newer Natoma system (which has write
posting, as well as several other performance features enabled
already), the only enhancement you get is the addition of the
Write Combining feature.
Also NOTE: Running fastvid on a newer Orion or Natoma system (
that already has write posting enabled), will not directly speed up
other PCI devices. The MTTRs are specific to the physical addresses
of the Video device, and have no direct effect outside of that range.
The only noticeable effect on other devices might be the change in bus
traffic, thus providing more bandwidth to other devices.
Lastly, for those of you running NT (which I hope for Pro users you
are!), I do have an alpha NT driver that programs the "Write Combining"
features, as well as can toggle to write posting bit in the Orion
chipset. I'll post later this week with information about where to
get a copy. (If you run linux, and are interesteing in a driver to
turn this feature on, let me know)
I hope this clarifies what fastvid does. Many people still ask
why Intel has not put code to do this into the bios. The answer should
be evident from the above explanation. First, the MTTRs have to be
mapped to a particular physical address range. Not all video cards
map the linear physical frame buffers to the same place. Secondly
even if they did, some cards might use certain memory ordering
restrictions, which would be broken by the weak ordering of the
"Write Combining" method. IN MY OPINION, it is the responsibility
of the Video BIOS and or Video Driver to detect the presence of MTTRS
(from the CPUID instruction), and to enable the features that could
speed up screen access.
Jeff Sponuagle (jeff_sponaugle@ccm.jf.intel.com)
--------------1CFBAE3959E2B60015FB7483--
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.SGI.3.93.960825130712.16821A-110000>