I've thought about this a little bit, and I think we'd rather just
decrease the bitsize of the intrinsic rather than add a whole new one.
The separate intrinsic isn't really buying you anything, I don't think
it's going to make anything simpler.
On Thu, Jul 6, 2017 at 4:48 PM, Matt Turner <matts...@gmail.com> wrote:
> Some hardware, like i965, doesn't support group sizes greater than 32.
> In that case, we can use the ballot32 intrinsic instead, which will
> simplify our code generation.
> ---
> src/compiler/nir/nir.h | 2 ++
> src/compiler/nir/nir_intrinsics.h | 3 +++
> src/compiler/nir/nir_opt_intrinsics.c | 17 +++++++++++++++++
> src/intel/compiler/brw_compiler.c | 1 +
> 4 files changed, 23 insertions(+)
>
> diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
> index 3591048574..a3d8df0b19 100644
> --- a/src/compiler/nir/nir.h
> +++ b/src/compiler/nir/nir.h
> @@ -1844,6 +1844,8 @@ typedef struct nir_shader_compiler_options {
> */
> bool use_interpolated_input_intrinsics;
>
> + unsigned max_subgroup_size;
> +
> unsigned max_unroll_iterations;
> } nir_shader_compiler_options;
>
> diff --git a/src/compiler/nir/nir_intrinsics.h
> b/src/compiler/nir/nir_intrinsics.h
> index 96ecfbc338..61ad9f6643 100644
> --- a/src/compiler/nir/nir_intrinsics.h
> +++ b/src/compiler/nir/nir_intrinsics.h
> @@ -106,6 +106,9 @@ INTRINSIC(ballot, 1, ARR(1), true, 1, 0, 0, xx, xx, xx,
> NIR_INTRINSIC_CAN_ELIMIN
> INTRINSIC(read_invocation, 2, ARR(0, 1), true, 0, 0, 0, xx, xx, xx,
> NIR_INTRINSIC_CAN_ELIMINATE)
> INTRINSIC(read_first_invocation, 1, ARR(0), true, 0, 0, 0, xx, xx, xx,
> NIR_INTRINSIC_CAN_ELIMINATE)
>
> +/* Equivalent to ballot, but only returns the low 32-bits */
> +INTRINSIC(ballot32, 1, ARR(1), true, 1, 0, 0, xx, xx, xx,
> NIR_INTRINSIC_CAN_ELIMINATE)
> +
> /*
> * Memory barrier with semantics analogous to the compute shader
> * groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
> diff --git a/src/compiler/nir/nir_opt_intrinsics.c
> b/src/compiler/nir/nir_opt_intrinsics.c
> index 0cd75d8b28..023e169a99 100644
> --- a/src/compiler/nir/nir_opt_intrinsics.c
> +++ b/src/compiler/nir/nir_opt_intrinsics.c
> @@ -69,6 +69,23 @@ opt_intrinsics_impl(nir_function_impl *impl)
> replacement = nir_imm_int(&b, NIR_TRUE);
> break;
> }
> + case nir_intrinsic_ballot: {
> + assert(b.shader->options->max_subgroup_size != 0);
> + if (b.shader->options->max_subgroup_size > 32)
> + continue;
> +
> + nir_intrinsic_instr *ballot =
> + nir_intrinsic_instr_create(b.shader, nir_intrinsic_ballot32);
> + nir_ssa_dest_init(&ballot->instr, &ballot->dest, 1, 32, NULL);
> + ballot->src[0] = intrin->src[0];
> +
> + nir_builder_instr_insert(&b, &ballot->instr);
> +
> + replacement = nir_pack_64_2x32_split(&b,
> + &ballot->dest.ssa,
> + nir_imm_int(&b, 0));
> + break;
> + }
> default:
> break;
> }
> diff --git a/src/intel/compiler/brw_compiler.c
> b/src/intel/compiler/brw_compiler.c
> index 9ad5d8e4ee..6970bcb40f 100644
> --- a/src/intel/compiler/brw_compiler.c
> +++ b/src/intel/compiler/brw_compiler.c
> @@ -57,6 +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,
> + .max_subgroup_size = 64, /* FIXME */
> .max_unroll_iterations = 32,
> };
>
> --
> 2.13.0
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
