Mesa 22.2 RadeonSI Adds Option To Disable AMD Infinity Cache Usage, Other Changes

[phoronix]

Phoronix: Mesa 22.2 RadeonSI Adds Option To Disable AMD Infinity Cache Usage, Other Changes

AMD's Radeon open-source Linux graphics driver developers remain very busy preparing for next-generation RDNA3 GPU support...

https://www.phoronix.com/scan.php?page=news_item&px=Mesa-22.2-Option-Disable-MALL

[schmidtbag]

It's a little funny to have a headline discussing a new feature that only hurts performance.

[kiffmet]

schmidtbag It's great to determine how much of a speed bump the cache provides. Also might be useful to quickly see if the behavior of a bug changes when disabling the cache, which makes it faster/easier to localize the source.

[schmidtbag]

schmidtbag It's great to determine how much of a speed bump the cache provides. Also might be useful to quickly see if the behavior of a bug changes when disabling the cache, which makes it faster/easier to localize the source.

Haha I think you mean "bump in speed the cache provides" since a speed bump is typically what you use to decrease speed. Joking aside, I agree it would be interesting to see how much of an impact it makes.

[bridgman]

It's a little funny to have a headline discussing a new feature that only hurts performance.

I think the intent is to avoid trampling cache contents with data that will only be accessed once... in order to improve overall performance.

[CommunityMember]

It's a little funny to have a headline discussing a new feature that only hurts performance.

It is not at all unusual that for certain specific use cases that a capability actually hurts performance. Creating a formal method to disable those capabilities may have been the result of a request by some org who has run into those use cases.

[illwieckz]

AMD heavily talked up this all-new cache level with Radeon RX 6000 series GPUs. It's interesting now that they are adding a RadeonSI driver option to disable it. Granted, this should just be for debugging/profiling purposes or those curious to see how much of an impact it makes in different games. This can be activated using the AMD_DEBUG=mall_noalloc debug option when running with Mesa 22.2-devel and an AMDGPU Linux kernel driver build supporting AMDGPU_VM_PAGE_NOALLOC.

More precisely:

This can be activated using the AMD_DEBUG=mall_noalloc

Are you meaning the cache can be activated this way or are you meaning the disablement can be activated this way, meaning the cache can be disabled this way?

[jrch2k8]

More precisely:



Are you meaning the cache can be activated this way or are you meaning the disablement can be activated this way, meaning the cache can be disabled this way?

reading it logically AMD_DEBUG=mall_noalloc, should be the latter, this disable cache allocations

[Deathsimple]

Well, we have a rather funny misunderstanding here. This is not about disabling the cache, but rather fine tune which data enters the cache.

See normally all data which is written to VRAM enters the infinity cache so that when it is read again you can get it directly from the cache instead of waiting until VRAM came along with the data.

The problem is the size of the cache is limited, so it something doesn't make sense that data which you know won't be read for a while enters the cache and eventually kicks out data which will be used much more sooner.

Because of this you can now set a flag which says hey this specific memory region doesn't enter the cache on writes. Completely disabling it was just for testing if the flag works or not.

Cheers,
Christian.