Go Back   nV News Forums > Hardware Forums > Benchmarking And Overclocking

Newegg Daily Deals

Reply
 
Thread Tools
Old 05-26-03, 12:41 AM   #85
Behemoth
radeon 9800 pro
 
Behemoth's Avatar
 
Join Date: Apr 2003
Location: Darkness Falls
Posts: 841
Default

Quote:
Originally posted by Sazar
yes.. but in order to put in on the suspect list you need to have some evidence...

I STILL can't understand the possible logic you are using to suggest that futuremark is cheating... what are they cheating in and how can they possibly achieve this ?

ok, if there aint any evidence to eliminate a particular possibilty i would keep it on suspect list at least for a while just because it is possible hehe
now nvidia have been accusing futuremark how unfair 3DMark is, think along this way i dont think its too hard to see why futuremark wouldnt do cheating things in order to give unfair advantage to who they want
but i dont want to go over this thing again, i didnt see why its impossible so its possible
Behemoth is offline   Reply With Quote
Old 05-26-03, 03:06 AM   #86
StealthHawk
Guest
 
Posts: n/a
Default

Quote:
Originally posted by Behemoth
this is one way of shuffling instructions that manually reduces workload:

original program:

do a;
do b;
if (c == true) {do d;}
else { do e;}
do f;
do g;
do h;
do i;
do j;

do a;
if (c == true) {do d;}
else {do e;
do b;
do f;
do g;
do i;
do j;
}
do h;
Are these two things actually functionally and mathematically equivalent? Because they do not do the same thing. Although with your explantion they do have the same output. But to my understanding, have the same output DOES NOT make two things functionally or mathematically equivalent.

Here's what ATI said about their instruction shuffling:
1) it produces the same output as Futuremark's original code.
2) it is functionally equivalent to Futuremark's code.
3) it is mathematically equivalent to Futuremark's code.

To me, that implies that they are not skipping any steps.

What you are proposing is that they are putting a spin on the truth. If that is your opinion, then that is your opinion.

OR, are you actually saying that your code examples are both functionally and mathematically equivalent. If you are saying that, then please provide the basis for what definitions of functional equivalence and mathematical equivalence are.

I mean, for example, if you know the output of something is 4 you could caluclate it one of several ways.

1) 2+2 = 4
2) 2^2 = 4
3) cube root(64) = 4

The output of all 3 methods are the same, but I would not consider any of them to be functionally or mathematically equivalent. Maybe I don't know the definition.

edit: I would also like to throw this quesiton into the air. Suppose your definition of functionally and mathematically equivalent is correct, and that as long as output is always correct, there is equivalence.

You gave some examples of instructions that are bypassed, but because they do not effect the IQ, the output is still the same. That's all nice and good in hypotheticals, but are REAL shader programs like that? Can you really throw out instructions and have them not affect the resulting image?
  Reply With Quote
Old 05-26-03, 03:09 AM   #87
StealthHawk
Guest
 
Posts: n/a
Default

Quote:
Originally posted by Behemoth
i believe futuremark is possible to cheat. but why would futuremark want to cheat? please forgive me i dont want to go over it once again.
thanks for your kind info on the couple of other things. i am sure they may change my view on nvidia.
Hmm, Sazar posts a link to Sweeney's thoughts on optimization, and you say you will read it.

Are you saying when I posted links to dicussions on the same subject you don't read it
  Reply With Quote
Old 05-26-03, 03:12 AM   #88
Behemoth
radeon 9800 pro
 
Behemoth's Avatar
 
Join Date: Apr 2003
Location: Darkness Falls
Posts: 841
Default

Quote:
Originally posted by StealthHawk
Hmm, Sazar posts a link to Sweeney's thoughts on optimization, and you say you will read it.

Are you saying when I posted links to dicussions on the same subject you don't read it
sorry StealthHawk, i forgot to thank you for your links, dont be mad, i havent read Sazar stuff yet
i will read them all
Behemoth is offline   Reply With Quote
Old 05-26-03, 03:27 AM   #89
StealthHawk
Guest
 
Posts: n/a
Default

Please read this post by CatalystMaker here

Something very important that he says.
Quote:
The second comment I would like to relate that to is from Tim Sweeney, the lead developer for the UT games. (also found on Beyond3d)

".. any code optimization performed on a function that does not change the resulting value of the function for any argument, is uncontroversially considered a valid optimization. "

That guys is exactly what we did. We optimized the code without changing the end resutl.
Notice that he is quoting Tim Sweeney. Notice that Sweeney says that any function that has the same result as the original for any argument is a valid optimization. Then notice that CatalystMaker says that is exactly what they did.

Now look at your example. It does not have the same output for every argument. If c = false then you are skipping instructions.
  Reply With Quote
Old 05-26-03, 03:27 AM   #90
Behemoth
radeon 9800 pro
 
Behemoth's Avatar
 
Join Date: Apr 2003
Location: Darkness Falls
Posts: 841
Default

Quote:
Originally posted by StealthHawk
Are these two things actually functionally and mathematically equivalent? Because they do not do the same thing. Although with your explantion they do have the same output. But to my understanding, have the same output DOES NOT make two things functionally or mathematically equivalent.

Here's what ATI said about their instruction shuffling:
1) it produces the same output as Futuremark's original code.
2) it is functionally equivalent to Futuremark's code.
3) it is mathematically equivalent to Futuremark's code.

To me, that implies that they are not skipping any steps.

What you are proposing is that they are putting a spin on the truth. If that is your opinion, then that is your opinion.

OR, are you actually saying that your code examples are both functionally and mathematically equivalent. If you are saying that, then please provide the basis for what definitions of functional equivalence and mathematical equivalence are.

I mean, for example, if you know the output of something is 4 you could caluclate it one of several ways.

1) 2+2 = 4
2) 2^2 = 4
3) cube root(64) = 4

The output of all 3 methods are the same, but I would not consider any of them to be functionally or mathematically equivalent. Maybe I don't know the definition.
i knew you would question it, i was trying to say both programs were identical only in a particular case.
if i could just re-order some of the order-non-critical instructions to gain 8% performance boost that works in every case, i just got a better implementation of the same program, but now ati wouldnt use a better implementation in next driver because people think its a cheat!?! this is where i cant believe it, ati is unbelievable, ati should have used a better implementation and prove it to the public and say it is a better revised PS2 shader.
but removing it only makes me think it only works on GT4 sky and water, yes they are mathematically, functionally and visually identical to futuremark's, but only on GT4 sky and water.
Behemoth is offline   Reply With Quote
Old 05-26-03, 03:30 AM   #91
Behemoth
radeon 9800 pro
 
Behemoth's Avatar
 
Join Date: Apr 2003
Location: Darkness Falls
Posts: 841
Default

Quote:
Originally posted by StealthHawk
Please read this post by CatalystMaker here

Something very important that he says.

Notice that he is quoting Tim Sweeney. Notice that Sweeney says that any function that has the same result as the original for any argument is a valid optimization. Then notice that CatalystMaker says that is exactly what they did.

Now look at your example. It does not have the same output for every argument. If c = false then you are skipping instructions.
wow you are correct, the program model of mine is what i think ati is doing, the shaders only work on GT4 sky and water
Behemoth is offline   Reply With Quote
Old 05-26-03, 03:30 AM   #92
StealthHawk
Guest
 
Posts: n/a
Default

Quote:
Originally posted by Behemoth
i knew you would question it, i was trying to say both programs were identical only in a particular case.
Exactly!

They are only the same in a particular case. Now read my post above yours.

CatalystMaker affirms that ATI's optimizations are functionally equivalent, ie, that they are the same in all cases! Unless you think he is lying.

Also, you should note that I edited the post you responded to before you posted a reply. Although I basically went over the same thing(ie, that there are only equivalent in specific situations, not in all situations.).
  Reply With Quote

Old 05-26-03, 03:33 AM   #93
StealthHawk
Guest
 
Posts: n/a
Default

Quote:
Originally posted by Behemoth
wow you are correct, the program model of mine is what i think ati is doing, the shaders only work on GT4 sky and water
Well, yes. Because they re-wrote the shader, by shuffling the instructions.

Even though in 3dmark03 the shader program has a specific correct output every time, that does not change the fact that they are using a general algorithm, which hypothetically coudl have different inputs, and thus different outputs.
  Reply With Quote
Old 05-26-03, 03:38 AM   #94
Behemoth
radeon 9800 pro
 
Behemoth's Avatar
 
Join Date: Apr 2003
Location: Darkness Falls
Posts: 841
Default

Quote:
Originally posted by StealthHawk
Well, yes. Because they re-wrote the shader, by shuffling the instructions.

Even though in 3dmark03 the shader program has a specific correct output every time, that does not change the fact that they are using a general algorithm, which hypothetically coudl have different inputs, and thus different outputs.
hmm maybe its just me, i cant believe they dont use it just because people think its a cheat when its actually not a cheat. this is just stupid, imo.
edit: and obviously ati have to detect and alter the codes in order to facilitate the optimization, if the catalyst maker wanted it so bad to back out the optimization like he said in your link, its even more unbelievable, since all he has to do is swap the better ordering shaders in next driver, no more driver detection, no more code alteration, it should work in all cases, hence its very hard not to think they are mathematically functionally visually identical on GT4 sky and water only.

Last edited by Behemoth; 05-26-03 at 03:53 AM.
Behemoth is offline   Reply With Quote
Old 05-26-03, 04:25 AM   #95
Behemoth
radeon 9800 pro
 
Behemoth's Avatar
 
Join Date: Apr 2003
Location: Darkness Falls
Posts: 841
Default

lets investigate further on your example:

1) 2+2=4
2) 2^2=4


assume operation ^ is PS2 shader
assume operation + is IHV siturational shader
assume both integers on the left side are inputs.
assume the integer on right side is output.

futuremark is testing PS2 shader performance of a video card, since the driver maker knows in this particular case, with input 2,2, both shaders are mathematically functional outputly identical, if the driver maker substitute a IHV siturational shader +, do you think it is a cheat?

Last edited by Behemoth; 05-26-03 at 04:30 AM.
Behemoth is offline   Reply With Quote
Old 05-26-03, 04:55 AM   #96
Behemoth
radeon 9800 pro
 
Behemoth's Avatar
 
Join Date: Apr 2003
Location: Darkness Falls
Posts: 841
Default

here comes another contender:

3) 2?2=3.7(oops)

this stupid video card driver maker think 3.7 and 4 is around the same, hence he swaps in his IHV specific shader "?" too, just for speed, now even a dog will say this is cheating
Behemoth is offline   Reply With Quote
Reply


Thread Tools

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off
Forum Jump


Similar Threads
Thread Thread Starter Forum Replies Last Post
"GPU has fallen off the bus" error on 650M unless a CUDA program run first amonakov NVIDIA Linux 0 06-18-12 06:34 PM
Getting the proprietary nvidia driver to run with Debian 3.0 r0 (woody) Katchina404 NVIDIA Linux 9 01-12-03 08:49 AM
On The 333 Boards Does Anyone Run At 166fsb john19055 CPUs, Motherboards And Memory 22 07-30-02 07:39 PM

All times are GMT -5. The time now is 03:29 PM.


Powered by vBulletin® Version 3.7.1
Copyright ©2000 - 2014, Jelsoft Enterprises Ltd.
Copyright 1998 - 2014, nV News.