[Mesa-dev] [PATCH 2/3] i965/fs: Make emit_uniformize a no-op for immediates

Iago Toral itoral at igalia.com
Mon Oct 19 01:59:00 PDT 2015


On Sun, 2015-10-18 at 21:31 -0700, Kristian Høgsberg Kristensen wrote:
> An immdiate is already uniform so just return it up front. Without this,
> brw_fs_surface_builder ends up passing immediate surface indices through
> SHADER_OPCODE_BROADCAST. This writes to a stride 0 dst, which we can't
> constant propagate out of, and further, we don't constant propagate into
> the typed/untype read/write opcodes at all.  The end result is that all
> typed/untyped read/write/atomics end up as indirect sends.
> 
> Code generation should always produce either an immediate or an actual
> indirect surface index, so we can fix this by just special casing
> immediates in emit_uniformize.
> 
> Signed-off-by: Kristian Høgsberg Kristensen <krh at bitplanet.net>
> ---
>  src/mesa/drivers/dri/i965/brw_fs_builder.h | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_builder.h b/src/mesa/drivers/dri/i965/brw_fs_builder.h
> index df10a9d..98ce71e 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_builder.h
> +++ b/src/mesa/drivers/dri/i965/brw_fs_builder.h
> @@ -394,6 +394,9 @@ namespace brw {
>           const dst_reg chan_index = component(vgrf(BRW_REGISTER_TYPE_UD), 0);
>           const dst_reg dst = component(vgrf(src.type), 0);
>  
> +         if (src.file == IMM)
> +            return src;
> +

We can move this check to the top of emit_uniformize(), right?

>           ubld.emit(SHADER_OPCODE_FIND_LIVE_CHANNEL, chan_index);
>           ubld.emit(SHADER_OPCODE_BROADCAST, dst, src, chan_index);
>  

Iago



More information about the mesa-dev mailing list