[Mesa-dev] [PATCH 1/2] radeonsi: don't enable scratch just for SGPR spills

Nicolai Hähnle nhaehnle at gmail.com
Tue Jun 7 23:18:14 UTC 2016


On 08.06.2016 01:05, Marek Olšák wrote:
> From: Marek Olšák <marek.olsak at amd.com>
>
> Diff from shader-db:
>    Scratch: 3221504 -> 17408 (-99.46 %) bytes per wave

Wow.

This really needs to be fixed in LLVM. In the meantime, this series is

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

> ---
>   src/gallium/drivers/radeonsi/si_shader.c | 17 +++++++++++++++--
>   1 file changed, 15 insertions(+), 2 deletions(-)
>
> diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
> index 6e8eefb..d392807 100644
> --- a/src/gallium/drivers/radeonsi/si_shader.c
> +++ b/src/gallium/drivers/radeonsi/si_shader.c
> @@ -5825,6 +5825,18 @@ void si_shader_binary_read_config(struct radeon_shader_binary *binary,
>   	unsigned i;
>   	const unsigned char *config =
>   		radeon_shader_binary_config_start(binary, symbol_offset);
> +	bool really_needs_scratch = false;
> +
> +	/* LLVM adds SGPR spills to the scratch size.
> +	 * Find out if we really need the scratch buffer.
> +	 */
> +	for (i = 0; i < binary->reloc_count; i++) {
> +		const struct radeon_shader_reloc *reloc = &binary->relocs[i];
> +
> +		if (!strcmp(scratch_rsrc_dword0_symbol, reloc->name) ||
> +		    !strcmp(scratch_rsrc_dword1_symbol, reloc->name))
> +			really_needs_scratch = true;
> +	}
>
>   	/* XXX: We may be able to emit some of these values directly rather than
>   	 * extracting fields to be emitted later.
> @@ -5859,8 +5871,9 @@ void si_shader_binary_read_config(struct radeon_shader_binary *binary,
>   		case R_0286E8_SPI_TMPRING_SIZE:
>   		case R_00B860_COMPUTE_TMPRING_SIZE:
>   			/* WAVESIZE is in units of 256 dwords. */
> -			conf->scratch_bytes_per_wave =
> -				G_00B860_WAVESIZE(value) * 256 * 4 * 1;
> +			if (really_needs_scratch)
> +				conf->scratch_bytes_per_wave =
> +					G_00B860_WAVESIZE(value) * 256 * 4;
>   			break;
>   		default:
>   			{
>


More information about the mesa-dev mailing list