[Mesa-dev] [PATCH 01/17] gallium: Add PIPE_CAP_MSAA_MODES

[Marek Olšák]

On Sat, Mar 19, 2016 at 5:09 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
> On Sat, Mar 19, 2016 at 12:02 PM, Bas Nieuwenhuizen
> <bas at basnieuwenhuizen.nl> wrote:
>> On Sat, Mar 19, 2016 at 4:25 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>>> On Sat, Mar 19, 2016 at 11:14 AM, Bas Nieuwenhuizen
>>> <bas at basnieuwenhuizen.nl> wrote:
>>>> That would limit us to supporting sample counts for which we have
>>>> texture formats.
>>>>
>>>> As far as I understand with radeonsi we can support 16 samples without
>>>> any attachments, but all formats are limited to <= 8 samples.
>>>
>>> So you're going to end up with a situation where GL_MAX_SAMPLES is
>>> less than GL_MAX_FRAMEBUFFER_SAMPLES? I don't know that that's a
>>> useful thing to have. This implementation still has the problem of
>>> only supporting POT MSAA levels (although tbh I'm not 100% sure
>>> there's hw out there that supports NPOT MSAA levels). If people really
>>> want this, I think the way to go would be to make
>>> is_format_supported() work with PIPE_FORMAT_NONE and do it that way.
>>>
>>> Also, are you *sure* that's the case on radeonsi? I find it very odd
>>> that the rasterizer would support a higher MSAA level than the highest
>>> attachment would...
>>
>> I am pretty confident that this is the case. I just tested 16 samples
>> (although this series seems to miss changing MaxFramebufferSamples),
>> and the driver disallows any texture format with > 8 samples [1].
>> Furthermore the proprietary driver on Windows seems to have
>> GL_MAX_SAMPLES=8 and GL_MAX_FRAMEBUFFER_SAMPLES=16 [2].
>
> OK. I still think it's crazy, but it is what it is :)

It's called EQAA (similar to CSAA). The hardware can do 16 unique
depth samples, but only 8 unique color samples can be stored. Other
than that, the rasterization hw supports 16x MSAA fully.

>
>>
>> Using PIPE_FORMAT_NONE to query the driver would probably be a bit
>> less error prone than the current code that sets the masks, so that
>> would be fine with me.
>
> Actually my earlier criticism about it only doing POT levels is a bit
> off -- after reading some more of the code, I just think that the
> settings to drivers were off - it should have been (1 << 8) | (1 <<
> 4), etc. This works up to 32x MSAA, which is not supported by anyone
> (for real, although NVIDIA blob drivers do fake it).

R300 can do 6x MSAA, but granted it won't support this extension.

>
> I do still prefer to avoid having separate places where this info is
> encoded... so I maintain my vote for using PIPE_FORMAT_NONE in
> is_format_supported.

Same here.

Marek