[Mesa-dev] [PATCH 3/3] i965/fs/nir: Don’t let nir lower nir_intrinsic_load_subgroup_all_mask

[Jason Ekstrand]

On Tue, Oct 31, 2017 at 10:55 AM, Neil Roberts <nroberts at igalia.com> wrote:

> Instead of letting nir lower nir_intrinsic_load_subgroup_all_mask this
> is now generated directly. This is more efficient because it can be
> calculated in the compiler based on the dispatch width.
>
> Sadly it’s still not totally ideal because the constant doesn’t seem
> to get propagated and there is still a redundant MOV.
> ---
>  src/intel/compiler/brw_compiler.c | 2 +-
>  src/intel/compiler/brw_fs_nir.cpp | 7 ++++++-
>  2 files changed, 7 insertions(+), 2 deletions(-)
>
> diff --git a/src/intel/compiler/brw_compiler.c b/src/intel/compiler/brw_
> compiler.c
> index 8df0d2e..f02fceb 100644
> --- a/src/intel/compiler/brw_compiler.c
> +++ b/src/intel/compiler/brw_compiler.c
> @@ -57,7 +57,7 @@ static const struct nir_shader_compiler_options
> scalar_nir_options = {
>     .lower_unpack_snorm_4x8 = true,
>     .lower_unpack_unorm_2x16 = true,
>     .lower_unpack_unorm_4x8 = true,
> -   .lower_subgroup_all_mask = true,
> +   .lower_subgroup_all_mask = false,
>     .lower_subgroup_masks = true,
>     .max_subgroup_size = 32,
>     .max_unroll_iterations = 32,
> diff --git a/src/intel/compiler/brw_fs_nir.cpp
> b/src/intel/compiler/brw_fs_nir.cpp
> index 9202b0f..b73edc9 100644
> --- a/src/intel/compiler/brw_fs_nir.cpp
> +++ b/src/intel/compiler/brw_fs_nir.cpp
> @@ -4185,7 +4185,12 @@ fs_visitor::nir_emit_intrinsic(const fs_builder
> &bld, nir_intrinsic_instr *instr
>        break;
>     }
>
> -   case nir_intrinsic_load_subgroup_all_mask:
> +   case nir_intrinsic_load_subgroup_all_mask: {
> +      uint32_t mask = ~UINT32_C(0) >> (32 - dispatch_width);
> +      bld.MOV(retype(dest, BRW_REGISTER_TYPE_Q), brw_imm_d(mask));
>

In SIMD32, you're going to get unintentional sign-extension here.  I think
you want UQ and ud.

--Jason


> +      break;
> +   }
> +
>     case nir_intrinsic_load_subgroup_eq_mask:
>     case nir_intrinsic_load_subgroup_ge_mask:
>     case nir_intrinsic_load_subgroup_gt_mask:
> --
> 2.9.5
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20171031/cd5e5c54/attachment.html>