New TTM Allocator For AMDGPU Graphics Memory Landing With Linux 5.11

[PCJohn]

By the way, what is the time of page allocation? The rumours are: it is slow.

The page size is 64KiB on most gpus. Time depends on amount of allocated memory. And it is also very dependent if you are using standard memory or you allocate it as sparse.
Radeon RX 460, win7:
standard memory: 2MiB (32*64KiB) block: 3.13us cost per page, 64MiB (1024*64KiB): 139ns cost per page
sparse memory: 2MiB (32*64KiB) block: 19.8us, 64MiB (1024*64KiB): 3.00us

Nvidia's turing is not much better (measured on Quadro RTX 3000, win10):
standard memory: 2MiB (32*64KiB) block: 3.13us, 64MiB (1024*64KiB): 220ns
sparse memory: 2MiB (32*64KiB) block: 13.0us, 64MiB (1024*64KiB): 2.42us

Intel UHD 630, win10:
standard memory: 2MiB (32*64KiB) block: 741ns, 64MiB (1024*64KiB): 23.5ns
sparse memory: 2MiB (32*64KiB) block: 15.0us, 64MiB (1024*64KiB): 2.48us

So, we have the winner ;-) : standard memory on integrated Intel. Who would believe that Intel graphics might beat AMD and Nvidia :-) (joke).
Looking forward whatever this new AMD allocation bring.
Sorry, no Linux tests currently. My new Dell laptop still does not work well with Ubuntu 20.04 :-( .
John

[xxmitsu]

What's to elaborate? Vulkan is supported on the PS4 and PS5 platform, its not the native rendered which is some in-house spin that Sony came up with but it apparently is allowed. Also Vulkan is used on Switch console. DX12 is proprietary to xbox and windows.

Are you sure ? I'm not sure that's the case. As far as I've read they have different low-level API, similar to Vulkan. Named GNM and GNMX of which documentation is not publicly available, but instead only available through an NDA.

How do you know the PS4/5 consoles support Vulkan. Is it something official or just speculation ?

[theriddick]

How do you know the PS4/5 consoles support Vulkan. Is it something official or just speculation ?

I've just read before that its a OPTION for developers. Guess only people with SDK kits would know for sure.

[duby229]

Is TTM still worth it compared to GEM?

My understanding is that GEM is used to setup memory and TTM is used to perform memory management. They work together, one has a certain job and the other has a different job.

EDIT: From 2013
https://www.phoronix.com/forums/foru...-of-gem-vs-ttm

It's not that complex. GEM has two parts:

1. the API
2. the backend memory manager implementation

In radeon (and nouveau and just about all the other kms drivers), we use the GEM API, but use TTM as the backend memory manager since the GEM backend does not adequately handle things like vram. You still need a backend regardless of what API you expose to userspace.

[geearf]

My understanding is that GEM is used to setup memory and TTM is used to perform memory management. They work together, one has a certain job and the other has a different job.

EDIT: From 2013
https://www.phoronix.com/forums/foru...-of-gem-vs-ttm

Ooooh, thank you!
I thought GEM was like a newer version of TTM by Intel, I guess I was very wrong!