[Mesa-dev] [PATCH 2/2] radeonsi: move PSIZE and CLIPDIST unique IO indices after GENERIC

[Dieter Nützel]

It improves Uningine Heaven performance here (RX 580, 8 GB),too.

without Tessellation

Before:
FPS:	  79.0
Score:	  1991
Min FPS:  20.8
Max FPS: 189.8

After:
FPS:	  79.1
Score:	  1993
Min FPS:  19.6
Max FPS: 185.9


with Tessellation

Before:
FPS:	  67.7
Score:	  1705
Min FPS:   8.8
Max FPS: 182.6

After:
FPS:	  68.3
Score:	  1720
Min FPS:  15.9
Max FPS: 179.6


System

Platform:	Linux 4.20.0-amd-staging-4.11-1.g7262353-default+ x86_64
CPU model:	Intel(R) Xeon(R) CPU X3470 @ 2.93GHz (2925MHz) x8
GPU model:	Unknown GPU (256MB) x1

Marek is this the 'tessellation regression' you tried to solve?

For this series:
Tested-by: Dieter Nützel <Dieter at nuetzel-hh.de>

Dieter

Am 03.06.2017 18:04, schrieb Marek Olšák:
> From: Marek Olšák <marek.olsak at amd.com>
> 
> Heaven LDS usage for LS+HS is below. The masks are "outputs_written"
> for LS and HS. Note that 32K is the maximum size.
> 
> Before:
>   heaven_x64: ls=1f1 tcs=1f1, lds=32K
>   heaven_x64: ls=31 tcs=31, lds=24K
>   heaven_x64: ls=71 tcs=71, lds=28K
> 
> After:
>   heaven_x64: ls=3f tcs=3f, lds=24K
>   heaven_x64: ls=7 tcs=7, lds=13K
>   heaven_x64: ls=f tcs=f, lds=17K
> 
> All other apps have a similar decrease in LDS usage, because
> the "outputs_written" masks are similar. Also, most apps don't write
> POSITION in these shader stages, so there is room for improvement.
> (tight per-component input/output packing might help even more)
> 
> It's unknown whether this improves performance.
> ---
>  src/gallium/drivers/radeonsi/si_shader.c        | 18 
> +++++++++++-------
>  src/gallium/drivers/radeonsi/si_state_shaders.c |  4 +++-
>  2 files changed, 14 insertions(+), 8 deletions(-)
> 
> diff --git a/src/gallium/drivers/radeonsi/si_shader.c
> b/src/gallium/drivers/radeonsi/si_shader.c
> index ddfaa3b..3a86c0b 100644
> --- a/src/gallium/drivers/radeonsi/si_shader.c
> +++ b/src/gallium/drivers/radeonsi/si_shader.c
> @@ -129,32 +129,36 @@ unsigned
> si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned
> in
>  /**
>   * Returns a unique index for a semantic name and index. The index 
> must be
>   * less than 64, so that a 64-bit bitmask of used inputs or outputs 
> can be
>   * calculated.
>   */
>  unsigned si_shader_io_get_unique_index(unsigned semantic_name, 
> unsigned index)
>  {
>  	switch (semantic_name) {
>  	case TGSI_SEMANTIC_POSITION:
>  		return 0;
> -	case TGSI_SEMANTIC_PSIZE:
> -		return 1;
> -	case TGSI_SEMANTIC_CLIPDIST:
> -		assert(index <= 1);
> -		return 2 + index;
>  	case TGSI_SEMANTIC_GENERIC:
> +		/* Since some shader stages use the the highest used IO index
> +		 * to determine the size to allocate for inputs/outputs
> +		 * (in LDS, tess and GS rings). GENERIC should be placed right
> +		 * after POSITION to make that size as small as possible.
> +		 */
>  		if (index < SI_MAX_IO_GENERIC)
> -			return 4 + index;
> +			return 1 + index;
> 
>  		assert(!"invalid generic index");
>  		return 0;
> -
> +	case TGSI_SEMANTIC_PSIZE:
> +		return SI_MAX_IO_GENERIC + 1;
> +	case TGSI_SEMANTIC_CLIPDIST:
> +		assert(index <= 1);
> +		return SI_MAX_IO_GENERIC + 2 + index;
>  	case TGSI_SEMANTIC_FOG:
>  		return SI_MAX_IO_GENERIC + 4;
>  	case TGSI_SEMANTIC_LAYER:
>  		return SI_MAX_IO_GENERIC + 5;
>  	case TGSI_SEMANTIC_VIEWPORT_INDEX:
>  		return SI_MAX_IO_GENERIC + 6;
>  	case TGSI_SEMANTIC_PRIMID:
>  		return SI_MAX_IO_GENERIC + 7;
>  	case TGSI_SEMANTIC_COLOR: /* these alias */
>  	case TGSI_SEMANTIC_BCOLOR:
> diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c
> b/src/gallium/drivers/radeonsi/si_state_shaders.c
> index 8ac4309..f36997b 100644
> --- a/src/gallium/drivers/radeonsi/si_state_shaders.c
> +++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
> @@ -1226,21 +1226,23 @@ static void
> si_shader_selector_key_hw_vs(struct si_context *sctx,
>  		ps_disabled = sctx->queued.named.rasterizer->rasterizer_discard ||
>  			      (!ps_colormask &&
>  			       !ps_modifies_zs &&
>  			       !ps->info.writes_memory);
>  	}
> 
>  	/* Find out which VS outputs aren't used by the PS. */
>  	uint64_t outputs_written = vs->outputs_written;
>  	uint64_t inputs_read = 0;
> 
> -	outputs_written &= ~0x3; /* ignore POSITION, PSIZE */
> +	/* ignore POSITION, PSIZE */
> +	outputs_written &= ~((1ull <<
> si_shader_io_get_unique_index(TGSI_SEMANTIC_POSITION, 0) |
> +			     (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_PSIZE, 
> 0))));
> 
>  	if (!ps_disabled) {
>  		inputs_read = ps->inputs_read;
>  	}
> 
>  	uint64_t linked = outputs_written & inputs_read;
> 
>  	key->opt.hw_vs.kill_outputs = ~linked & outputs_written;
>  }