DX12 Performance Discussion And Analysis Thread

http://imgur.com/Xr0Jhjo

AMD presentation June 2016, Async Compute retired in GCN 1.0 in April.

 

GCN 1.0 = GCN gen 1, GCN 1.1 = GCN gen 2,... If we still insist on different nomenclature then AMD uses.

 

The linked slide doesn't say anything about retirement. Did you intend to link a different slide?

 

Don't mess with colors in posts please, especially not extremes. There is a light and dark theme.

 

Hm, when I published some Time Spy numbers in our mag a couple of months ago, there was no increase with AC on a GCN 1.0 as well. And interestingly enough, neither with GCN 1.1. Both (HD7970 & R7 260X) were tested with 16.7.3 drivers. Of course neither Maxwell nor Kepler scored an increase.

 

Anyone tested AC on GCN 1.0 under Vulkan?

 

I do and it works, but Vulkan use a different librairies and maybe that why.

 

I see somes gains on GCN 1.1(aka GCN 2) but no one test on old GCN 1.0 (aka GCN 1).

 

Have you still the R7 260X ? Can you test Async Compute with that tool ? https://mega.nz/#!IJNQ0S4a!oLn5-FzqBqZ2potDqYRtY6nahF3bcvAPZQ26PUIMFvA

If anyone can send me the results with GCN 1.1(aka GCN 2) like R9 290, R9 290X etc. I would be glad

 

We have it in the lab, sure. But pre-xmas I'm awfully busy atm. Will see what I can do.
I remember that tool from one of the threads here. Maybe I even have older results collecting dust on my hard drive.

 

"Official" technical response: http://gpuopen.com/concurrent-execution-asynchronous-queues/#comment-6937

 

And he saw it was very good.. and he approved it completely.

 

If someone ask, Async Compute still disabled in new driver (16.12.2)

 

More related to DX11 and up, but well was not sure where to post it:

GPUOpen AGS 5.0 - Shader compiler controls : http://gpuopen.com/ags-5-0-shader-compiler-controls/

 

I also find it odd that some people are still using GCN 1.0, 1.1, 1.2, 1.3, etc unofficial names. AMD might have used GCN 1.1 unofficially at some point, but I don't remember them ever using GCN 1.2 or 1.3.

Official names are: GCN1, GCN2, GCN3 and GCN4. Vega is most likely GCN5. Please don't call Vega GCN 1.4, that would just confuse people

Example: https://gpuopen.com/wp-content/uploads/2016/08/AMD_GCN3_Instruction_Set_Architecture_rev1.1.pdf
Short = GCN3
Long = Graphics Core Next, Generation 3

 

To nitpick, Vega will use what will probably be called GCN5, but it won't be GCN5, just like with Polaris where they decided they need (again) name for the whole architecture (Polaris is architecture which uses GCN4 and other elements like UVD-something and VCE-something, while before they just called the whole thing GCNx)

 

PIX for Windows (Beta) is finally out!

https://blogs.msdn.microsoft.com/pix/2017/01/17/introducing-pix-on-windows-beta/

 

Wow that could really be paramount in benchmarking DX12 games. The memory allocation could be an interesting one given Raja's remarks, if it also studies actual usage.

 

What are the advantages of these versus profiling tools in licensed engines? You get memory allocation profiling tools in Unity/Unreal for example.

 

I'm referring to benchmarking by analysts, not game developers.