[Mesa-dev] [PATCH 3/3] radeonsi: improve the computation and comment of scratch_waves

[Nicolai Hähnle]

For the series:

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

On 08.06.2016 17:36, Marek Olšák wrote:
> From: Marek Olšák <marek.olsak at amd.com>
>
> 2% isn't much. If you think the number should be decreased, please speak up.
> ---
>   src/gallium/drivers/radeonsi/si_pipe.c | 22 ++++++++++++++++++----
>   1 file changed, 18 insertions(+), 4 deletions(-)
>
> diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
> index 5d7d2f3..0c601da 100644
> --- a/src/gallium/drivers/radeonsi/si_pipe.c
> +++ b/src/gallium/drivers/radeonsi/si_pipe.c
> @@ -236,11 +236,25 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen,
>   				     R600_COHERENCY_SHADER);
>   	}
>
> -	/* XXX: This is the maximum value allowed.  I'm not sure how to compute
> -	 * this for non-cs shaders.  Using the wrong value here can result in
> -	 * GPU lockups, but the maximum value seems to always work.
> +	uint64_t max_threads_per_block;
> +	screen->get_compute_param(screen, PIPE_SHADER_IR_TGSI,
> +				  PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK,
> +				  &max_threads_per_block);
> +
> +	/* The maximum number of scratch waves. Scratch space isn't divided
> +	 * evenly between CUs. The number is only a function of the number of CUs.
> +	 * We can decrease the constant to decrease the scratch buffer size.
> +	 *
> +	 * sctx->scratch_waves must be >= the maximum posible size of
> +	 * 1 threadgroup, so that the hw doesn't hang from being unable
> +	 * to start any.
> +	 *
> +	 * The recommended value is 4 per CU at most. Higher numbers don't
> +	 * bring much benefit, but they still occupy chip resources (think
> +	 * async compute). I've seen ~2% performance difference between 4 and 32.
>   	 */
> -	sctx->scratch_waves = 32 * sscreen->b.info.num_good_compute_units;
> +	sctx->scratch_waves = MAX2(32 * sscreen->b.info.num_good_compute_units,
> +				   max_threads_per_block / 64);
>
>   	/* Initialize LLVM TargetMachine */
>   	r600_target = radeon_llvm_get_r600_target(triple);
>