[Mesa-dev] [RFC PATCH] gallium: add a common uploader to pipe_context

[Brian Paul]

On 02/06/2017 12:11 PM, Marek Olšák wrote:
> On Mon, Feb 6, 2017 at 5:15 PM, Brian Paul <brianp at vmware.com> wrote:
>> On 02/03/2017 02:41 PM, Marek Olšák wrote:
>>>
>>> On Fri, Feb 3, 2017 at 9:45 PM, Brian Paul <brianp at vmware.com> wrote:
>>>>
>>>> On 02/01/2017 02:23 PM, Brian Paul wrote:
>>>>>
>>>>>
>>>>> On 01/27/2017 04:00 AM, Marek Olšák wrote:
>>>>>>
>>>>>>
>>>>>> On Fri, Jan 27, 2017 at 10:05 AM, Nicolai Hähnle <nhaehnle at gmail.com>
>>>>>> wrote:
>>>>>>>
>>>>>>>
>>>>>>> On 27.01.2017 00:51, Marek Olšák wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> From: Marek Olšák <marek.olsak at amd.com>
>>>>>>>>
>>>>>>>> For lower memory usage and more efficient updates of the buffer
>>>>>>>> residency
>>>>>>>> list. (e.g. if drivers keep seeing the same buffer for many
>>>>>>>> consecutive
>>>>>>>> "add" calls, the calls can be turned into no-ops trivially)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> This makes sense to me, but how are you planning to deal with the bind
>>>>>>> flags? They are currently set differently for different upload mgrs.
>>>>>>> We
>>>>>>> should probably do away with them entirely anyway.
>>>>>>
>>>>>>
>>>>>>
>>>>>> Drivers can set the bind flags they need. Some drivers will set all 3
>>>>>> bind flags. Other drivers don't have to set any.
>>>>>
>>>>>
>>>>>
>>>>> I need to look into this part more closely.  I think we may have trouble
>>>>> mixing constants with index/vertex data in our VMware driver...
>>>>
>>>>
>>>>
>>>> Marek,
>>>>
>>>> Your patch series, as-is, did indeed cause trouble with our VMware
>>>> driver.
>>>> We need to keep constants in a separate buffer.
>>>>
>>>> The good news is I don't think this is a huge problem and I've updated (a
>>>> subset of) your patches to accommodate both your needs and ours.
>>>>
>>>> The basic idea is to add a pipe_context::get_stream_uploader() hook that
>>>> allows drivers to use just one or separate uploaders for
>>>> vertex/index/constant data.  Plus, I added a
>>>> pipe_context::unmap_stream_uploaders() helper, but this isn't strictly
>>>> necessary.
>>>>
>>>> WIP patch attached (only lightly tested).  Let me know what you think.
>>>
>>>
>>> Can we simply add these 2 fields into pipe_context instead of the
>>> callback?
>>>
>>> pipe_context::stream_uploader // vertex + index
>>> pipe_context::const_uploader
>>
>>
>> Yeah, that might work too.  Though, now I have to test the case of vertex
>> data and index data being in the same VBO.  We may need three uploader
>> pointers...
>
> What's the issue with svga that it needs different buffers for each
> type? OpenGL allows the same buffer to be used for vertex, index,
> const, even the same buffer range.

Our virtual device protocol uses DX10 conventions and in DX10, constant 
buffers are distinct from vertex/index buffers.  That is, you can't set 
both D3D10_BIND_CONSTANT_BUFFER and D3D10_BIND_VERTEX_BUFFER for one 
buffer.  Luckily, we haven't come across any GL apps that need that (and 
it wouldn't be too hard to work around it if needed).

I was pretty sure that vertex+index data in one buffer was OK but I 
wanted to check.  Since I don't think we have any piglit tests (or apps) 
that exercise that specific case, I just wrote one and it seems to work 
fine.

So, pipe_context::stream_uploader + pipe_context::const_uploader should 
be OK for us.  Do you want to implement that?

-Brian