[Mesa-dev] [PATCH] radeonsi: don't use the real barrier instruction in tess ctrl shaders

Nicolai Hähnle nhaehnle at gmail.com
Thu Apr 7 00:30:37 UTC 2016


Reviewed-by: Nicolai Hähnle <nicolai.haehnle at amd.com>

On 06.04.2016 19:07, Marek Olšák wrote:
> From: Marek Olšák <marek.olsak at amd.com>
>
> ---
>   src/gallium/drivers/radeonsi/si_shader.c | 8 ++++++++
>   1 file changed, 8 insertions(+)
>
> diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
> index bf3f008..08da3e3 100644
> --- a/src/gallium/drivers/radeonsi/si_shader.c
> +++ b/src/gallium/drivers/radeonsi/si_shader.c
> @@ -4282,6 +4282,14 @@ static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
>   	struct si_shader_context *ctx = si_shader_context(bld_base);
>   	struct gallivm_state *gallivm = bld_base->base.gallivm;
>
> +	/* The real barrier instruction isn’t needed, because an entire patch
> +	 * always fits into a single wave.
> +	 */
> +	if (ctx->type == TGSI_PROCESSOR_TESS_CTRL) {
> +		emit_optimization_barrier(ctx);
> +		return;
> +	}
> +
>   	lp_build_intrinsic(gallivm->builder,
>   			   HAVE_LLVM >= 0x0309 ? "llvm.amdgcn.s.barrier"
>   					       : "llvm.AMDGPU.barrier.local",
>


More information about the mesa-dev mailing list