[Mesa-dev] [PATCH v3 15/19] nir: use load_local_group_size

[Rob Clark]

On Sat, Mar 24, 2018 at 8:12 PM, Jason Ekstrand <jason at jlekstrand.net> wrote:
> On March 24, 2018 16:24:57 Rob Clark <robdclark at gmail.com> wrote:
>
> On Fri, Mar 23, 2018 at 4:59 PM, Jason Ekstrand <jason at jlekstrand.net>
> wrote:
> On Fri, Mar 23, 2018 at 1:35 PM, Karol Herbst <kherbst at redhat.com> wrote:
>
> On Fri, Mar 23, 2018 at 9:18 PM, Jason Ekstrand <jason at jlekstrand.net>
> wrote:
> On Fri, Mar 23, 2018 at 12:33 PM, Karol Herbst <kherbst at redhat.com>
> wrote:
>
> From: Rob Clark <robdclark at gmail.com>
>
> If local_size is not known at compile time, which is the case with
> clover, use the load_local_group_size intrinsic instead.
>
> Signed-off-by: Karol Herbst <kherbst at redhat.com>
> ---
> src/compiler/nir/nir_lower_system_values.c | 25
> +++++++++++++++++--------
> 1 file changed, 17 insertions(+), 8 deletions(-)
>
> diff --git a/src/compiler/nir/nir_lower_system_values.c
> b/src/compiler/nir/nir_lower_system_values.c
> index d507c28f421..ff4e09c8e61 100644
> --- a/src/compiler/nir/nir_lower_system_values.c
> +++ b/src/compiler/nir/nir_lower_system_values.c
> @@ -57,19 +57,28 @@ convert_block(nir_block *block, nir_builder *b)
> *    "The value of gl_GlobalInvocationID is equal to
> *    gl_WorkGroupID * gl_WorkGroupSize +
> gl_LocalInvocationID"
> */
> +         nir_ssa_def *local_size_def;
>
> -         nir_const_value local_size;
> -         memset(&local_size, 0, sizeof(local_size));
> -         local_size.u64[0] = b->shader->info.cs.local_size[0];
> -         local_size.u64[1] = b->shader->info.cs.local_size[1];
> -         local_size.u64[2] = b->shader->info.cs.local_size[2];
> +         /* if local_size[] is already known, use that, otherwise use
> +          * load_local_group_size intrinsic:
> +          */
> +         if (b->shader->info.cs.local_size[0]) {
> +            nir_const_value local_size;
> +            memset(&local_size, 0, sizeof(local_size));
> +            local_size.u64[0] = b->shader->info.cs.local_size[0];
> +            local_size.u64[1] = b->shader->info.cs.local_size[1];
> +            local_size.u64[2] = b->shader->info.cs.local_size[2];
> +
> +            local_size_def = nir_build_imm(b, 3, bit_size,
> local_size);
>
> +         } else {
> +            local_size_def = nir_load_local_group_size(b, bit_size);
> +         }
>
>
> I commented on an earlier patch about how the approach to building the
> 32/64-bit immediates is wrong.
>
> oh right, I totally forgot about that.
>
> Setting that aside, this patch looks fine to me in principal.  There's a
> part of me that doesn't like using cs.local_size[0] being the trigger
> but I
> think it's probably ok.  Maybe we should assert that cs_local_size is
> either
> all zero (second case) or all not zero (first case) just to be safe.
>
> I think the main problem here is, that even with OpenCL kernels you
> can specify it, but then overwrite it at runtime again. So yes I
> agree, that we need something better here.
>
>
> Oh, that's tricky then.  We could make nir_lower_system_values take a flag
> or OpenCL callers could just whack it all to 0 after spirv_to_nir.c.  Or you
> could do recompiles or something.
>
> I'm half-way towards thinking we should have MESA_SHADER_KERNEL vs
> MESA_SHADER_COMPUTE to deal with cl/cuda/etc vs glsl compute shaders..
>
> So far, I've been trying to avoid that, but maybe it would be a better
> solution..
>
> I think that would be a bit short-sighted.  There are cases where that might
> be a good idea but I doubt this is one of them. It may will be that someone
> will decide that runtime local sizes would make a great Vulkan extension and
> then we'd be switching on API for no good reason.
>

hmm, the extension case is a good counter-argument (and probably also
applies in the other cases I was thinking of but am too jet-lagged to
remember)..

I guess until someone comes up with a better idea, sticking to
info.cs.local_size[0]==0 => use intrinsic seems sane.. a
local_size[0]==0 is impossible.  Extra asserts for local_size[1..2]==0
is a good idea.

BR,
-R