MSAA + HDR benchmarks ?

[KimB]

Z Fill rate with MSAA test confirms that the sample output is the same as all R300+ chips; only Xenos from ATI does more.

Er, wait wait wait. Why does this:

Normal (No AA) 	7059.0 	8000.2
2x 	5759.4 	7479.3
4x 	5690.0 	2999.8

...imply that the hardware only does 2x AA samples per clock? Isn't this directly stating that the hardware does 4x AA samples per clock, given the negligible performance drop from 2x to 4x for color fillrate?

Of course, I'm not sure how to interpret the z-only fill, but the color-only seems to imply a change.

Edit: Actually, I wonder if the low z-only fill is related to the performance improvements that the memory controller tweak has produced for 4x AA in games like Doom3?

 

[Xmas]

Er, wait wait wait. Why does this:

Normal (No AA) 	7059.0 	8000.2
2x 	5759.4 	7479.3
4x 	5690.0 	2999.8

...imply that the hardware only does 2x AA samples per clock? Isn't this directly stating that the hardware does 4x AA samples per clock, given the negligible performance drop from 2x to 4x for color fillrate?

Color is per pixel, depth is per sample.

 

[KimB]

I'm not sure how that changes things here.

 

[Mintmaster]

I would never play below 1280x1024 regardless of the FSAA or lacker there of I can use.

Then you're a resolution whore that doesn't give a crap about image quality

Really, all you guys saying you prefer one step higher resolution with no AA have got to be smoking crack. I think I'll take a high resolution picture from a digital camera, and downsize it two ways: One, I'll take single samples and fill 2x2 squares, and with the other I'll properly downsample to create 3x3 squares. Even though the latter is coarser, it'll look much better. If you step back, this'll be the equivalent of 16x12 w/o AA vs. 10x7 w/AA. That's two resolution steps - 12x10 would be even worse.

Maybe I'll run my first poll.

 

[Mintmaster]

I'm not sure how that changes things here.

I think what XMas meant was colour is samples per clock, and Z is pixels per clock. Not too sure myself, actually.

 

[Dave Baumann]

On a pure fill-rate test like that colour writes are bandwidth limited, Z aren't as much (note: this is the XL, which doesn't have the gobs of bandwidth the XT does; its also greater than half the theoretical samples).

 

[KimB]

On a pure fill-rate test like that colour writes are bandwidth limited, Z aren't as much (note: this is the XL, which doesn't have the gobs of bandwidth the XT does; its also greater than half the theoretical samples).

Except that the XL only has a theoretical max fillrate of 8000MPix/sec, so shouldn't that limit its 4x FSAA fillrate to 4000MPix/sec?

 

[KimB]

I think what XMas meant was colour is samples per clock,

Except if it was samples per clock, it'd be increasing with higher levels of FSAA, as otherwise it'd be no better than supersampling.

 

[OpenGL guy]

On a pure fill-rate test like that colour writes are bandwidth limited, Z aren't as much (note: this is the XL, which doesn't have the gobs of bandwidth the XT does; its also greater than half the theoretical samples).

To elaborate a little, Z is limited by how many Z comparisons you can do per clock, color is not.

 

[Tridam]

This fillrate test is useless as it doesn't say anything about MSAA performance. What's the point about using MSAA without writing Z values ?

Color + Z fillrate should be used instead of color only.

 

[OpenGL guy]

This fillrate test is useless as it doesn't say anything about MSAA performance. What's the point about using MSAA without writing Z values ?

Color + Z fillrate should be used instead of color only.

I wouldn't say the results are useless, it's a test of color compression.

 

[KimB]

Not necessarily. Early-z optimizations will allow better performance than the pure fillrate test would imply.

So it appears that the X1k ROPs are in fact able to output more than just two color samples per clock, but are limited by how quickly the z-compares are done. So it's not worlds better than a normal 2-sample per pipe per clock design, but still better depending upon the efficiency of early-z optimizations.

This will be particularly true for games that make an initial z-only pass.

 

[OpenGL guy]

Not necessarily. Early-z optimizations will allow better performance than the pure fillrate test would imply.

Nope. As soon as Z is enabled, you will be limited by the Z comparisons.

 

[Tridam]

I wouldn't say the results are useless, it's a test of color compression.

Right. So let's say that adding Z+Color results should tell more about MSAA performance.

BTW, the Z fillrate value for MSAA 4x looks too low. There is no reason for the Z fillrate to decrease by a factor higher than 2 between MSAA 2x and 4x. I get 3888 MPix/s with MSAA 4x and 7457 MPix/s with MSAA 2x.

 

[Xmas]

To elaborate a little, Z is limited by how many Z comparisons you can do per clock, color is not.

That's what I meant. NVidia uses a different implementation where color writes are limited to two samples per clock. Blending however is "full speed" for both IHVs when MSAA is enabled, IIRC, if there are no edge pixels. Though "full speed" for NV40 means only 8 blends per clock.

An implementation could take advantage of early accept for the Z-test, so it could write more Z-samples than it can test per clock.

 

[OpenGL guy]

Right. So let's say that adding Z+Color results should tell more about MSAA performance.

Not if you are trying to measure color-only fillrate

 

[Jawed]

Well I'm suitably confused now.

Is the test Dave is using actually testing what he intends it to test?

Jawed

 

[bdmosky]

Then you're a resolution whore that doesn't give a crap about image quality

Actually my main problem is the fact that I do have an LCD with a native resolution of 1280x1024 and I hate how lower resolutions look when they are interpolated on my screen regardless of the FSAA level.

 

[Tridam]

Not if you are trying to measure color-only fillrate

The color-only fillrate tells only a part of MSAA performance (color compression efficiency) and z-only fillrate tells another part (Z-only pass performance). But they don't tell everything so in most cases Color+Z fillrate will give more informations.

 

[Tridam]

Well I'm suitably confused now.

Is the test Dave is using actually testing what he intends it to test?

Jawed

Well Dave doesn't talk about color-only results so it's difficult to know what he intends to test but it should be color compression. Anyway, what I think is that it's misleading to talk about MSAA performance without giving the color+z results.