[Mesa-dev] [PATCH v2 03/13] nir: Add a writemask to store intrinsics.

[Jason Ekstrand]

On Sat, Dec 12, 2015 at 2:00 PM, Jason Ekstrand <jason at jlekstrand.net> wrote:
> On Sat, Dec 12, 2015 at 12:22 PM, Kenneth Graunke <kenneth at whitecape.org> wrote:
>> On Saturday, December 12, 2015 08:29:37 AM Jason Ekstrand wrote:
>>> On Dec 11, 2015 1:24 PM, "Kenneth Graunke" <kenneth at whitecape.org> wrote:
>> [snip]
>>> > @@ -302,12 +302,12 @@ LOAD(shared, 1, 1, NIR_INTRINSIC_CAN_ELIMINATE)
>>> >  #define STORE(name, srcs, indices, flags) \
>>> >     INTRINSIC(store_##name, srcs, ARR(0, 1, 1, 1), false, 0, 0, indices,
>>> flags)
>>> >
>>> > -/* src[] = { value, offset }. const_index[] = { base } */
>>> > -STORE(output, 2, 1, 0)
>>> > -/* src[] = { value, vertex, offset }. const_index[] = { base } */
>>> > -STORE(per_vertex_output, 3, 1, 0)
>>> > +/* src[] = { value, offset }. const_index[] = { base, write_mask } */
>>> > +STORE(output, 2, 2, 0)
>>> > +/* src[] = { value, vertex, offset }. const_index[] = { base, write_mask
>>> } */
>>> > +STORE(per_vertex_output, 3, 2, 0)
>>> >  /* src[] = { value, block_index, offset }. const_index[] = { write_mask
>>> } */
>>> > -STORE(ssbo, 3, 1, 0)
>>> > +STORE(ssbo, 3, 2, 0)
>>>
>>> I don't think you meant to change SSBOs.
>>
>> Oops.  Yeah, that was a mistake when resolving merge conflicts.
>> Changed for v3.
>>
>>> >  /* src[] = { value, offset }. const_index[] = { base, write_mask } */
>>> >  STORE(shared, 2, 1, 0)
>>
>> While we're here, this comment seems wrong, doesn't it?  There's only
>> one index, but the comment lists two.  (It's wrong before my patch...)
>
> Yes, we should fix that.  It's only one index.  That was my mistake
> while rebasing on shared variables.

Scratch taht.  This is correct.  When doing shared variables, I
decided to go with base+offset since they are also opaque.  That was
intentional.