[Mesa-dev] [PATCH 1/2] radeonsi: Lock a mutex when checking scratch relocations.

[Michel Dänzer]

On 22.04.2016 17:19, Marek Olšák wrote:
> On Fri, Apr 22, 2016 at 6:41 AM, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
>>
>> (Curiously, there is an old "this is probably not needed anymore" comment on
>> the PS_PARTIAL_FLUSH in si_context_gfx_flush, but this may be wrong: since
>> shaders can write memory, and the kernel may want to swap buffers around, we
>> have to be extra careful about this...)
> 
> PS_PARTIAL_FLUSH is not emitted in this case, because SURFACE_SYNC
> does the same thing.
> 
> Also, the kernel emits EVENT_WRITE_EOP(CACHE_FLUSH_AND_INV_TS_EVENT |
> TC_ACTION_EN | TCL1_ACTION_EN) as part of the fence sequence, but it
> doesn't wait. SI also flushes SMEM L1 and ICACHE before
> EVENT_WRITE_EOP. The GPU scheduler waits for the fence when there is a
> dependency (we can't have inter-process dependencies though). Memory
> migrations always wait for idle.

Note that TTM BO migration always waits for idle due to a limitation of
the current implementation, which may be lifted at some point. Not by
design.


-- 
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer