Marek Looking To Tackle Large RadeonSI Performance Bottleneck

[phoronix]

Phoronix: Marek Looking To Tackle Large RadeonSI Performance Bottleneck

Prolific Mesa developer Marek Olšák is looking to tackle what he thinks is the "biggest performance bottleneck at the moment" for the RadeonSI Gallium3D driver...

http://www.phoronix.com/scan.php?page=news_item&px=RadeonSI-BO-Move-Bottleneck

[oooverclocker]

Nice! Loosening these remaining brakes is exactly what we need.

[yoshi314]

Anyone has a human readable explanation of that technical description?

[Sethox]

Since it's pretty much technical anyways, it's pointless to bring up a more "human readable explanation". If you really want to know what it means, well the internet is your resource to find out the details.
What I take from this, is that people are bisecting the problem and trying to find a solution, emphasis on trying to find a solution because of the Quote: "but the situation is still far from ideal".

[boxie]

Anyone has a human readable explanation of that technical description?

My take on it: Lots of unnecessary transfers between Main Memory <->CPU <-> GPU <-> VGA RAM - i.e. bandwidth overutilisation problems.

if they can cut down on the amount of stuff sent/evicted they can use more of that bandwidth for other required things

[mitch074]

If this is also the cause for the huge slowdown in downtown Pragues in DE:MD, then I'm all for it.

[follower]

Dear Michael! Why are you so obsessed with the reviews and news of AMD video cards? Almost half of all your articles about AMD products! But they occupy only 25% of the market at best. Thanks anyway for your work. From Russia with respect 8)

[schmidtbag]

Not knowing enough about the technical details of this, I can't tell if addressing this problem would contradict Mesa's descriptor+uniform behavior, or, if it was more of a regression that just needs some tweaking.

Anyway, more performance is always welcome. Though, it seems to me the #1 performance issue seems to be some sort of bottleneck (likely CPU). For a long while, many benchmarks have shown GPUs of varying performance levels and/or architectures topping out around the same frame rate.

[artivision]

Cache, memory and buffer (management), are always the top problems.

[geearf]

Is this for all radeonsi cards? The ML seems to be about amdgpu though.