PDA

View Full Version : Didn't know Supersampling had such a performance hit


Saintster
02-23-06, 01:36 AM
I installed Far Cry and had my settings on 8aa 16 af and supersampling. MY fps were from 25 to 50 I kicked the super to multi and my fps shot to 95-110 resolution of 1280x1024. I didn't know supersampling had that much of an affect.

ynnek
02-23-06, 01:47 AM
and knowing is half the battle!

Go JOE!

Saintster
02-23-06, 03:07 AM
so true!!:D

jolle
02-23-06, 05:28 AM
pure 2x2 (4x) supersampling is VERY heavy, it pretty much renders the image 2x height 2x width and downsamples that I think.. so every pixel is the result of 4 subpixels.
If you mean Transparancy SSAA, that only applies SS to alphablended textures, which Farcry has ALOT of, so the hit there scales with the amount of such textures.

Saintster
02-23-06, 11:27 AM
pure 2x2 (4x) supersampling is VERY heavy, it pretty much renders the image 2x height 2x width and downsamples that I think.. so every pixel is the result of 4 subpixels.
If you mean Transparancy SSAA, that only applies SS to alphablended textures, which Farcry has ALOT of, so the hit there scales with the amount of such textures.

Wow...thanks for the info. It will be nice down the road when the hardware and software advance to take the hit out of 2x2 supersampling it's a great feauture. So multisampling is most likely just 1x the height and width? Thanks again for the explanation!:thumbsup:

jolle
02-23-06, 02:01 PM
Multisampling only samples around polygon edges on the screen, and does some sort of avraging on the color values to smooth them out, the performance hit of doing so isnt very big since the % of pixels on polygon edges/outlines isnt very big, and it doesnt scale or work similar to SSAA.

But they dont do anything on textures with transparancys, stuff like most of the vegitation in games, grates, fences etc.. since their edges arent polygons, only transparent textures.
Super Sampling, which I think is a older and more "brute force" way to approach edge aliasing, works on the entire image, so it will help this type of texture aswell as polygon edges, and the sort of aliasing we are starting to see within Normal mapped surfaces, or Vdisp/offset/parallax mapped surfaces that have a sortof virtual edges within them.

Transparancy SSAA is a mix of both works, it only does SSAA on the textures that needs it, and traditional MSAA on geometry, the hit taken should scale with the amount of Alphablended textures in the scene.

john19055
02-25-06, 10:31 AM
It looks good ,but it can be a performance killer on the newer games.

rhink
02-25-06, 12:33 PM
Main difference in super and multi performance is supersampling is a 4x fillrate hit. Multisampling doesn't 'scale' the image quite the same as supersampling does; it still takes the same number of samples per pixel. It has zero fillrate hit, only a memory bandwidth hit. And due to smart memory compression techniques, much of the memory bandwidth hit can be eliminated. Thus, multisampling is far, far more efficient. Major drawback, as has been mentioned, is not sampling edges of geometry simulated by textures with transparency.

That can be fixed by selectively using supersampling on transparent textures (transparency aa).

For everything else you get the best of both worlds by using high quality texture filtering, together with multisampling, and you get an image that's fairly close to the supersampled image, for a fraction of the performance hit.

The reason why supersampling is not, and will not be practical for today's games is pixel shaders. Since 2x2 doubles your resolution, then scales it back down, you're quadrupling the number of pixels that need to be rendered. If you're running shaders on those pixels, the performance hit can be immense. (maybe in an architecture where you have many many many more pixel pipelines than ROPs, the hit won't be so noticable- but I'm not sure it's smart to optimize your architecture for supersampling in such a way, when multisampling can be done so much more cheaply). In situations where you're bound by the pixel shader, the performance hit from multisampling can even approach zero, however (the performance cost of the rendering may in fact hide the performance cost of multisampling).

One day we may get to the point where hardware is so immensely powerful, that it just doesn't matter, and we can take the performance hit.

But software has a way of sucking up hardware resources as much as possible. I think we'll get a bigger realism boost for the foreseeable future by spending the extra processing time on more detailed shaders and higher resolutions, instead of spending those resources on supersampling.