PDA

View Full Version : Improved Instruction Scheduler in Det50


StealthHawk
10-04-03, 08:20 PM
So...has anyone actually been able to prove how much, if any, Det50's shader compiler has improved compared to Det40? It shouldn't really be too hard when using this tool (http://www.nvnews.net/vbulletin/showthread.php?s=&threadid=18912) in conjunction with different drivers.

Ruined
10-05-03, 02:40 PM
According to 3dcenter, from 45.23 to 52.10 there is an average 31.5% increase in speed for partial precision, and 25 percent increase in shader speed for full precision in Shadermark 2.0. IIRC Shadermark 2.0 came out after 52.10 so I don't think its an app specific optimization. They also stated that 52.10 returned IQ to the level of the 40 series, fixed from 51.75.

gokickrocks
10-05-03, 02:53 PM
Originally posted by Ruined
They also stated that 52.10 returned IQ to the level of the 40 series, fixed from 51.75.

they have updated their article and it says otherwise (in regards to AF anyways)

volt
10-05-03, 03:33 PM
The numbers are nice, but how about IQ? :p

theultimo
10-05-03, 03:51 PM
Well, from what i read in the review, the 52.10's are faster, but basically kill AF and Trilinnear filtering.

Carfax
10-05-03, 04:40 PM
Originally posted by gokickrocks
they have updated their article and it says otherwise (in regards to AF anyways)

Well, the AF can still work properly through the application mode I believe. Hopefully, this is just a bug and the final revision will resolve it..

Matt-AWD
10-05-03, 04:55 PM
Originally posted by Carfax
Well, the AF can still work properly through the application mode I believe. Hopefully, this is just a bug and the final revision will resolve it..

It's always a "bug" :rolleyes:

Kihon
10-05-03, 05:05 PM
From my understanding, AF works properly as long as you set drivers to application - same as ATI.

However, their numbers were based on setting AF in drivers, hence their results for UT2003 atleast are based on incorrect AF settings. The solution seems to be rerun the tests with AF set by application rather than drivers.

or am I completely misreading it?

serAph
10-06-03, 10:55 AM
could someone fill me in on what an instruction scheduler is and its function in the driver/hardware/whatever?

thx

volt
10-06-03, 11:11 AM
...our driver team has developed a next-generation automatic shader optimizer that vastly improves GeForce FX pixel shader performance across the board...

Derek Perez
Director of Public Relations
NVIDIA Corp.

Is that it ? :)

ragejg
10-06-03, 11:16 AM
yep, that's it. :)

gokickrocks
10-06-03, 11:23 AM
Originally posted by serAph
could someone fill me in on what an instruction scheduler is and its function in the driver/hardware/whatever?

thx

it schedules instructions (literally) by putting them in an order that would be more efficient for execution

serAph
10-06-03, 11:47 AM
Originally posted by gokickrocks
it schedules instructions (literally) by putting them in an order that would be more efficient for execution

at what level? And does it more "efficiently" ogranize them by priority or something?

fivefeet8
10-06-03, 11:57 AM
Could this be why there is a 2 to 3x performance gain in DM-ICetomb and DM-Rustatorium maps in UT2k3 from the 45.23 dets to the 51.75's? Could be.

serAph
10-06-03, 12:02 PM
sorry to post ANOTHER question, but:
Does modifying or "optimizing" this feature result in compromised IQ of any form?

mrsabidji
10-06-03, 01:25 PM
Originally posted by serAph
at what level? And does it more "efficiently" ogranize them by priority or something?

From what I read, it reorganizes shader instructions taking advantage of the ability of the FXs to handle long shader instructions and tries to reduce the number of registers used, since this is what seems to kill nv3x performance most of the time (I'm no coder, so feel free to correct me if I'm wrong).

Originally posted by serAph
sorry to post ANOTHER question, but:
Does modifying or "optimizing" this feature result in compromised IQ of any form?

Theoretically, it shouldn't.

mrsabidji

serAph
10-06-03, 01:42 PM
ayte cool. Thx :D

StealthHawk
10-06-03, 02:42 PM
Originally posted by fivefeet8
Could this be why there is a 2 to 3x performance gain in DM-ICetomb and DM-Rustatorium maps in UT2k3 from the 45.23 dets to the 51.75's? Could be.

UT2003 uses shaders which bog down gfFX cards :confused: I don't think that is correct. I don't even know if gfFX cards use PS1.4 in UT2003. Someone can test it with PS1.1 and see if it gets any faster(doubt it).

StealthHawk
10-06-03, 02:44 PM
Originally posted by serAph
sorry to post ANOTHER question, but:
Does modifying or "optimizing" this feature result in compromised IQ of any form?

Obviously that depends on where the extra performance from Det50 is coming from. Is there an improved instruction scheduler? How improved is it?

Do the drivers have any other automatic mechanisms in place that increase performance?

Or is performance coming from application specific optimizations? Or a combination of the above? Is IQ being compromised in any of these hypothetical stages? All questions that need answering.