[Mesa-dev] [PATCH 3/3] mesa: don't check mapped buffers in every draw call if drivers allow it

[Dave Airlie]

On 16 May 2017 at 08:21, Thomas Helland <thomashelland90 at gmail.com> wrote:
> 2017-05-15 21:08 GMT+00:00 Marek Olšák <maraeo at gmail.com>:
>> On Mon, May 15, 2017 at 10:00 PM, Thomas Helland
>> <thomashelland90 at gmail.com> wrote:
>>> Would this be a win for a game limited by the winsys/buffer handling?
>>> I played some Hitman the other day, and noticed my gpu-utilization
>>> was hovering at about 80%. Some profiling later and I concluded
>>> that one thread was pegged at 100% doing buffer handling work.
>>> amdgpu_cs_add_buffer, amdgpu_lookup_buffer, the cso_hash,
>>> and a couple other functions related to the winsys where the hottest.
>>> Should I expect that to be mitigated by your threaded_gallium work,
>>> or by this patch series? If so, I will not delve into that much further =)
>>
>> amdgpu_cs_add_buffer and amdgpu_lookup_buffer overhead will definitely
>> be hidden by threaded gallium. cso_hash will not.
>>
>> I don't think what you're seeing is related, but I'd be interested in
>> a screenshot of sysprof showing what you described.
>>
>> Marek
>
> The below dropbox link has two screenshots showing the situation.
> The data was collected by attaching perf to the process,
> and the report was made with "perf report --hierarchy".

did you use perf record --call-graph dwarf?

Helps as well if you have no frame ptrs.

Dave.