[Mesa-dev] [PATCH] nv50: report only 16 texure_samplers

[Christoph Bumiller]

On 12.10.2013 18:52, Roland Scheidegger wrote:
> Am 12.10.2013 17:19, schrieb Christoph Bumiller:
>> On 12.10.2013 02:47, Emil Velikov wrote:
>>> On 12/10/13 01:25, Roland Scheidegger wrote:
>>>> Am 12.10.2013 02:02, schrieb Brian Paul:
>>>>> On 10/11/2013 10:44 AM, Emil Velikov wrote:
>>>>>> Current mesa code(cso and drivers) expect and use only up-to 16
>>>>>> texture samplers.
>>>>>>
>>>>>> Verbatum copy from the nvc0 driver.
>>>>>>
>>>>>> Cc "9.1" <mesa-stable at lists.freedesktop.org>
>>>>>> Cc "9.2" <mesa-stable at lists.freedesktop.org>
>>>>>> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=70212
>>>>>> Reported-by: Aaron Watry <awatry at gmail.com>
>>>>>> Signed-off-by: Emil Velikov <emil.l.velikov at gmail.com>
>>>>>> ---
>>>>>>   src/gallium/drivers/nouveau/nv50/nv50_screen.c | 4 ++++
>>>>>>   1 file changed, 4 insertions(+)
>>>>>>
>>>>>> diff --git a/src/gallium/drivers/nouveau/nv50/nv50_screen.c
>>>>>> b/src/gallium/drivers/nouveau/nv50/nv50_screen.c
>>>>>> index f454ec7..3f81cc4 100644
>>>>>> --- a/src/gallium/drivers/nouveau/nv50/nv50_screen.c
>>>>>> +++ b/src/gallium/drivers/nouveau/nv50/nv50_screen.c
>>>>>> @@ -249,7 +249,11 @@ nv50_screen_get_shader_param(struct pipe_screen
>>>>>> *pscreen, unsigned shader,
>>>>>>      case PIPE_SHADER_CAP_INTEGERS:
>>>>>>         return 1;
>>>>>>      case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
>>>>>> +      return 16; /* would be 32 in linked (OpenGL-style) mode */
>>>>>> +      /*
>>>>>> +   case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLER_VIEWS:
>>>>>>         return 32;
>>>>>> +      */
>>>>>>      default:
>>>>>>         NOUVEAU_ERR("unknown PIPE_SHADER_CAP %d\n", param);
>>>>>>         return 0;
>>>>>>
>>>>> Since PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLER_VIEWS doesn't really exist,
>>>>> I'd rather see it removed entirely.
>>>> Actually it doesn't seem to exist at all?
>>>>
>>> Indeed and afaics it never did :)
>>>
>>> As the commit says it's a verbatim copy of nvc0, which also started with
>>> 32 TEXTURE_SAMPLERS.
>>>
>>>>> If you wanted to future-proof the code you could do
>>>>>    return MIN2(32, PIPE_MAX_SAMPLERS);
>>>>>
>>>>> in case we bump PIPE_MAX_SAMPLERS to 32 one of these days.
>>>>>
>>>>>
>>>>> In any case, Reviewed-by: Brian Paul <brianp at vmware.com>
>>>>>
>>>> Well I think there is quite some hw out there which can only do 16
>>>> samplers but more sampler views as this is what d3d10/11 wants (16
>>>> samplers max per stage, but 128 sampler views). So making it queryable
>>>> may have some benefits, but OpenGL can't really make any use of it in
>>>> any case.
>>>>
>>> I'm not entirely sure what is the case in here, as I'm a bit short of
>>> knowledge about the hardware, especially with the lack of documentation.
>> That comment's there as a reminder that gallium should have that cap.
>>
>> On nv50 have 2 big tables in VRAM, of texture view (or "sampler view" in
>> gallium, "shader resource view" in d3d) descriptors (TIC) and sampler
>> descriptors (TSC).
>> These are mapped to texture and sampler "units" via a binding table
>> (that happens as a result of calling glBindTexture).
>> In the shader you have to select the units, you have 4 bits for the
>> sampler and 5 bits for the texture unit index.
> Hmm if you have only 5 bits for the texture unit index how do you access
> 128 different sampler views as required by d3d10?
>

Good question ... I should have checked rather than just going by my
memory. Sorry.

So, the story goes like this: actually the shader instruction has 5 bits
for samplers and 8 bits for textures, but the binding table only has 16
entries for samplers (don't ask me why the shader has 5 bits then), so
the 32 that was there initially was probably not an actual undocumented
hardware limit but rather "mesa loops through all of these every time so
don't advertise more than necessary" (where necessary is what people
expect because of the NV blob).

>> You can set the hardware to linked mode (LINKED_TSC) so that sampler
>> unit index automatically corresponds to the texture view unit index.
>> Now you can't select them independently but you access 32 bindings of
>> them tied together (which is how OpenGL works).
>> But gallium requires them to be selectable independently we can't use that.
>>
>> With Kepler they removed the binding table and you can select the
>> descriptors directly via a 32-bit register so you can access (1 << 20)
>> texture views and (1 << 12) samplers, there's no problem there.
> For bindless textures I assume (a neat extension).
>
> Roland
>
>
>>> FWIW I've intentionally added/copied the SAMPLER_VIEWS, as I feel it's
>>> beneficial in the long term. That is after going through the build
>>> system(s), my plan is to jump into the nv50 driver + vdpau st and some
>>> of the missing extensions and other things along the way :)
>>>
>>> Cheers
>>> Emil
>>>
>>>> Roland
>>>> _______________________________________________
>>>> mesa-dev mailing list
>>>> mesa-dev at lists.freedesktop.org
>>>> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>>>>
>>> _______________________________________________
>>> mesa-dev mailing list
>>> mesa-dev at lists.freedesktop.org
>>> http://lists.freedesktop.org/mailman/listinfo/mesa-dev