[PATCH libdrm] amdgpu: add amdgpu_bo_handle_type_kms_noimport

Christian König christian.koenig at amd.com
Tue Jul 17 07:33:36 UTC 2018


Am 17.07.2018 um 09:26 schrieb Michel Dänzer:
> On 2018-07-17 08:50 AM, Christian König wrote:
>> Am 16.07.2018 um 18:05 schrieb Michel Dänzer:
>>> On 2018-07-13 08:47 PM, Marek Olšák wrote:
>>> [SNIP]
>>> Other opinions?
>> I understand the reason why Marek wants to do this, but I agree that
>> this is a little bit dangerous if used incorrectly.
>>
>> On the other hand I don't see any other way to sanely handle it either.
> Sanely handle what exactly? :) I still haven't seen any description of
> an actual problem, other than "the handle is stored in the hash table".

Well the problem is that it's not "the handle" but rather "all handles" 
which are now stored in the hash table.

To begin with that is quite a bunch of wasted memory, not talking about 
the extra CPU cycles.

>> MM interop with GFX usually imports BOs on each frame, so that can hurt
>> here.
> That's always the same set of BOs in the steady state, right? So it's
> easy to make the repeated lookups fast, by moving them to the start of
> their hash buckets.

Yeah, that can help with the CPU cycles but it is still not ideal.

I think that Mareks change is justified, but we should add a comment 
explaining the restrictions.

Christian.


More information about the amd-gfx mailing list