[Mesa-dev] [PATCH] gallium: add pipe_grid_info::partial_block

[Ilia Mirkin]

On Tue, Jan 8, 2019 at 7:26 PM Marek Olšák <maraeo at gmail.com> wrote:
>
> On Tue, Jan 8, 2019 at 7:18 PM Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>>
>> On Tue, Jan 8, 2019 at 6:21 PM Marek Olšák <maraeo at gmail.com> wrote:
>> >
>> > On Tue, Jan 8, 2019 at 5:25 PM Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>> >>
>> >> Why does this need to be in p_state? And who is responsible for
>> >> setting it (and how will it be set)?
>> >
>> >
>> > Oh right, there is a way to get it out of p_state.h if needed.
>> >
>> > It should be set to 0 by default.
>> >
>> > If your thread block is 8x8x1, but you need to launch 10x8x1 threads, set partial_block = {2, 0, 0}. It will launch the following thread blocks:
>> > 8x8x1
>> > 2x8x1
>> >
>> > It's the same as launching 16x8x1 threads and doing this at the beginning of the compute shader:
>> >   if (globalThreadID.x >= 10) return;
>>
>> But that all sounds like something a state tracker wouldn't care
>> about, right? In e.g. GLSL you can specify the block to be 10x8x1 and
>> let the backend work it all out. Should st/mesa care about this (or
>> clover or whatever)?
>
>
> The block size should be a multiple of 64 on radeonsi to utilize all SIMD lanes. If you want to launch 8192+1 threads with the block size of 64, you need to launch 1 partial block with the block size of 1 at the end. OpenGL can't do this.

Ohhhhhhhhhh. So the partial-ness applies to the last-executed block.
If you have a local_size=(2,2,1), and you want your global grid to be,
say, (5,4,1), with unextended GL you might run it as groups = (3,2,1)
which would end up invoking a bunch of bits you don't want, and so
this partial_size is a way to say that you don't want the last "line"
to be executed at all.

That makes sense, and seems like a reasonable thing to have in
pipe_grid_info. The documentation did not make that clear the first
time I read it, but now I'm having trouble suggesting improvements to
it. So I think it's fine.

The p_state.h bits are Acked-by: Ilia Mirkin <imirkin at alum.mit.edu> .
Can't speak as to the radeonsi bits.

Cheers,

  -ilia