From: moskewicz AT MEM DOT po DOT com
Date: Sat, 15 Jun 1996 04:00:16 -0400
Message-Id: <9606150800.AA15815@mercury>
To: Sengan DOT Short AT durham DOT ac DOT uk
Cc: djgpp AT delorie DOT com
Subject: Re: Are near pointers worth it?

[Quotes edited for brevity.]
[That is, I accidently leaned on the delete line key.]
>>>> and VESA 2 linear frame buffers. If I were to use near pointers
would
>>>> it make it any faster?
>> 
>>>I think not. I got about a 5% speedup of my putpixel() function from
>>>using near pointers, and that is the absolute worst case becuase it
>>>involves a segment load per pixel. If you are just setting the
selector
>> 
>> I would agree totally, using near pointers is not worth it. Remember,
>> GCC
>> doesn't touch %fs and %gs, so just put your selector there, and you
only
>> load it once. Just watch out for the _far* family, as they will
clobber
>> %fs
>
>Surely the point is that the access to video memory is so slow that
large
>variations in execution speed in the processor are masked by the wait
states
>enforced by the graphics cards. If (ever) some graphics card is
introduced
>which has the same access time as normal memory is introduced, it seems
to me
>that difference between the different ways of accessing memory will be
clearer.

Surely not!

Profiling results for mode 4101h, 640x480 256 color.
Running in 32 bit protected mode with linear framebuffer

 1.0428s for    30000 lines   =>   28768.23 lines/s
 1.4810s for      300 clears  =>     202.57 clears/s,   59.35 Mb/s
 1.0628s for      150 bitBlts =>     141.13 bitBlt/s,   41.35 Mb/s

Baseline values:

REP STOSD in system memory:   31.28 Mb/s
REP MOVSD in system memory:   29.87 Mb/s

No cache thrashing.

Unless I am throughly confused, these numbers indicate that video
memory access on my ATI mach64 2MB VRAM PCI card is faster than 
system DRAM access. This profile was produced by profile, from 
SciTech. It comes with a big list of various card/cpu profiles
in a text file, ready to use as a reference...
I just ignore people who say "My code draws 10^9 sprites per year,
is that fast?" without mentioning their cpu or video card... :)

>best for its solution: I might want a fast program accessing hardware
really
>fast, and not want memory protection for one application, and yet want
memory
>protection for another.

Nah, near pointers are just no good, ever! :) Seriously, though, they
have
serious problems, and little benefit (but of course, I won't back that
up).

>Sengan

----
moskewicz AT mem DOT po DOT com