[Mesa-dev] [PATCH] R600/SI: radeonsi: Program RASTER_CONFIG for harvested GPUs v3

[Michel Dänzer]

On 10.09.2014 05:08, Tom Stellard wrote:
> v2:
>    - Write RASTER_CONFIG for all SEs.
>
> v3:
>    - Set GRBM_GFX_INDEX.INSTANCE_BROADCAST_WRITES bit.
>    - Set GRBM_GFX_INFEX.SH_BROADCAST_WRITES bit when done setting
>      PA_SC_RASTER_CONFIG.
>    - Get num_se and num_sh_per_se from kernel.
>
> https://bugs.freedesktop.org/show_bug.cgi?id=60879
>
> CC: "10.3 10.2" <mesa-stable at lists.freedesktop.org>

I think you should drop the 'R600/SI: ' prefix from the commit summary, 
as this change isn't in the LLVM backend.


> diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
> index 1d6ae86..32b93d3 100644
> --- a/src/gallium/drivers/radeonsi/si_state.c
> +++ b/src/gallium/drivers/radeonsi/si_state.c
> @@ -3060,24 +3071,97 @@ void si_init_config(struct si_context *sctx)
>   			break;
>   		}
>   	} else {
> +		unsigned raster_config = 0;
> +		unsigned sh_per_se = MAX2(sctx->screen->b.info.max_sh_per_se, 1);
> +		unsigned num_se = MAX2(sctx->screen->b.info.max_sh_per_se, 1);

I think num_se should use sctx->screen->b.info.max_se, not .max_sh_per_se?

Maybe it's unnecessary anyway though:


> +		for (i = 0; i < num_se; i++) {
> +			si_pm4_set_reg(pm4, GRBM_GFX_INDEX,
> +				SE_INDEX(i) |
> +				SH_BROADCAST_WRITES |
> +				INSTANCE_BROADCAST_WRITES);
> +			si_pm4_set_reg(pm4, R_028350_PA_SC_RASTER_CONFIG, raster_config);
> +		}

Since this uses the same raster_config value for all SEs, couldn't it 
just use a single write with SE_BROADCAST_WRITES enabled in GRBM_GFX_INDEX?


-- 
Earthling Michel Dänzer            |                  http://www.amd.com
Libre software enthusiast          |                Mesa and X developer