[Mesa-dev] [PATCH] r600g: Support TGSI_SEMANTIC_HELPER_INVOCATION

[Ilia Mirkin]

On Mon, Nov 16, 2015 at 8:31 AM, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
> Hi Glenn,
>
> On 14.11.2015 00:11, Glenn Kennard wrote:
>>
>> On Fri, 13 Nov 2015 18:57:28 +0100, Nicolai Hähnle <nhaehnle at gmail.com>
>> wrote:
>>
>>> On 13.11.2015 00:14, Glenn Kennard wrote:
>>>>
>>>> Signed-off-by: Glenn Kennard <glenn.kennard at gmail.com>
>>>> ---
>>>> Maybe there is a better way to check if a thread is a helper invocation?
>>>
>>>
>>> Is ctx->face_gpr guaranteed to be initialized when
>>> load_helper_invocation is called?
>>>
>>
>> allocate_system_value_inputs() sets that if needed, and is called before
>> parsing any opcodes.
>
>
> Sorry, you're right, I missed the second change to the inputs array there.
>
>
>>> Aside, I'm not sure I understand correctly what this is supposed to
>>> do. The values you're querying are related to multi-sampling, but my
>>> understanding has always been that helper invocations can also happen
>>> without multi-sampling: you always want to process 2x2 quads of pixels
>>> at a time to be able to compute derivatives for texture sampling. When
>>> the boundary of primitive intersects such a quad, you get helper
>>> invocations outside the primitive.
>>>
>>
>> Non-MSAA buffers act just like 1 sample buffers with regards to the
>> coverage mask supplied by the hardware, so helper invocations which have
>> no coverage get a 0 for the mask value, and normal fragments get 1.
>> Works with the piglit test case posted at least...
>
>
> Here's why I'm still skeptical: According to the GLSL spec, the fragment
> shader is only run once per pixel by default, even when MSAA is enabled.
> _However_, if a shader statically accesses the SampleID, _then_ it must be
> run once per fragment. The way I understand it, your change forces the
> fragment shader to access SampleID, even when people ostensibly use
> HelperInvocation in the hope of optimizing something.

GPU's don't operate based on GLSL specs. Per-sample shading is enabled
separately.

>
> In the usual MSAA operation of only running the fragment shader once per
> pixel, HelperInvocation should be the same as SampleMask != 0, right? It
> seems like the right thing to do is to _not_ allocate the
> TGSI_SEMANTIC_SAMPLEID when TGSI_SEMANTIC_HELPER_INVOCATION is used, and
> then use different code paths in load_helper_invocation based on which of
> the source registers are actually there.

This, however is a good point -- for MSAA presumably the sample id
will always work out to 0, but the bottom bit of the sample mask may
not be set. In the non-SSAA case, should probably check if the whole
mask != 0.

  -ilia