[Mesa-dev] [PATCH 5/5] radeonsi: add a workaround for a compute VGPR-usage LLVM bug

[Nicolai Hähnle]

On 12.07.2016 22:52, Marek Olšák wrote:
> From: Marek Olšák <marek.olsak at amd.com>
>
> This patch may be dropped depending on feedback.
>
> Cc: 12.0 <mesa-stable at lists.freedesktop.org>
> ---
>   src/gallium/drivers/radeonsi/si_shader.c | 33 ++++++++++++++++++++++++++++++++
>   1 file changed, 33 insertions(+)
>
> diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
> index 79c3496..81d071e 100644
> --- a/src/gallium/drivers/radeonsi/si_shader.c
> +++ b/src/gallium/drivers/radeonsi/si_shader.c
> @@ -6705,6 +6705,39 @@ int si_compile_tgsi_shader(struct si_screen *sscreen,
>
>   	radeon_llvm_dispose(&ctx.radeon_bld);
>
> +	/* Validate SGPR and VGPR usage for compute to detect compiler bugs. */

There should probably be some comment about which versions of LLVM are 
affected.

Apart from this and the comment on patch #1, the series is

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

> +	if (sel->type == PIPE_SHADER_COMPUTE) {
> +		unsigned *props = sel->info.properties;
> +		unsigned wave_size = 64;
> +		unsigned max_vgprs = 256;
> +		unsigned max_sgprs = sscreen->b.chip_class >= VI ? 800 : 512;
> +		unsigned max_sgprs_per_wave = 128;
> +		unsigned min_waves_per_cu =
> +			DIV_ROUND_UP(props[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
> +				     props[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
> +				     props[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH],
> +				     wave_size);
> +		unsigned min_waves_per_simd = DIV_ROUND_UP(min_waves_per_cu, 4);
> +
> +		max_vgprs = max_vgprs / min_waves_per_simd;
> +		max_sgprs = MIN2(max_sgprs / min_waves_per_simd, max_sgprs_per_wave);
> +
> +		if (shader->config.num_sgprs > max_sgprs ||
> +		    shader->config.num_vgprs > max_vgprs) {
> +			fprintf(stderr, "LLVM failed to compile a shader correctly: "
> +				"SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
> +				shader->config.num_sgprs, shader->config.num_vgprs,
> +				max_sgprs, max_vgprs);
> +
> +			/* Just terminate the process, because dependent
> +			 * shaders can hang due to bad input data, but use
> +			 * the env var to allow shader-db to work.
> +			 */
> +			if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
> +				exit(0);
> +		}
> +	}
> +
>   	/* Add the scratch offset to input SGPRs. */
>   	if (shader->config.scratch_bytes_per_wave)
>   		shader->info.num_input_sgprs += 1; /* scratch byte offset */
>