[Mesa-dev] [PATCH 00/18] Gallium blitter optimizations

[Brian Paul]

On 08/18/2017 10:25 AM, Marek Olšák wrote:
> On Fri, Aug 18, 2017 at 5:56 PM, Brian Paul <brianp at vmware.com> wrote:
>> On 08/17/2017 12:31 PM, Marek Olšák wrote:
>>>
>>> Hi,
>>>
>>> Major u_blitter changes:
>>>
>>> - All draw calls except cubemap blits will now use the draw_rectangle
>>>     callback. (cubemap blits are not used if the driver supports texture
>>>     views)
>>>
>>> - The VS POSITION output always covers the whole 2D clip space,
>>>     (-1,-1) -> (1,1). The final vertex positions and Z are determined by
>>>     the viewport transformation.
>>>
>>> - If VertexID is supported, the VS POSITION output is derived from
>>>     VertexID, and the vertex input with POSITION data is not present.
>>>
>>> - If the VS GENERIC output is not needed, it's not present.
>>>     (previously it was always there)
>>>
>>> - If there are no VS inputs, the blitter doesn't change vertex buffer
>>>     and vertex element states, leading to lower CPU overhead.
>>>     (measurable with trivial apps like glxgears)
>>>
>>>
>>> RadeonSI changes:
>>>
>>> - We only used the draw_rectangle callback for decompress and resolve
>>>     passes, and depth and color clears. Now we will always use it when
>>>     the blitter invokes it.
>>>
>>> The overall impact on RadeonSI is that only slow color clears and
>>> texture blits use a vertex shader that has 1 vertex input. Moving that
>>> input into user SGPRs would be possible, but that's not done here.
>>>
>>> Please review.
>>
>>
>> If you can wait a bit, I'd like to apply your patches to my tree and do some
>> testing to make sure our driver can handle these changes.
>
> Sure. The changes are here:
> https://urldefense.proofpoint.com/v2/url?u=https-3A__cgit.freedesktop.org_-7Emareko_mesa_log_-3Fh-3Dblitter-2Doptz&d=DwIFaQ&c=uilaK90D4TOVoH58JNXRgQ&r=Ie7_encNUsqxbSRbqbNgofw0ITcfE8JKfaUjIQhncGA&m=L5oxTYkKIi2pqgCQMRPNabejTUMjsMySB6IsbTx2FBg&s=R3XgONotN2Qd5Kj24pXcOsVaOIEOyzNKsFY8VpxtiPo&e=
>
> If somebody wants to test it on r600g, that would be appreciated too.

Seems OK with the VMware driver.

Tested-by: Brian Paul <brianp at vmware.com>

-Brian