[Mesa-dev] [PATCH v2 2/2] r600: set rounding mode for texture array layer selection

[Roland Scheidegger]

Am 29.06.2018 um 16:35 schrieb Gert Wollny:
> The evaluation of the array layer index is "floor(z+0.5)", and the default
> rounding mode doesn't correctly evaluate this. Therefore, set the rounding
> mode to "trunc" and z-filter mode to "point".
> For other textures make sure the the default rounding mode and z-filter are
> used.
> 
> Fixes single-sample tests out of:
>   dEQP-GLES3.functional.texture.shadow.2d_array.*
>   dEQP-GLES3.functional.texture.vertex.2d_array.*
>   dEQP-GLES3.functional.texture.filtering.2d_array.*
> (With the single sample tests the rounding accuracy is tested too)
> 
> v2: * reword comments and commit message
>     * clear S_03C008_TRUNC_COORD for all non-array types
> 
> Signed-off-by: Gert Wollny <gert.wollny at collabora.com>
> ---
>  src/gallium/drivers/r600/evergreen_state.c | 22 ++++++++++++++++++++++
>  1 file changed, 22 insertions(+)
> 
> diff --git a/src/gallium/drivers/r600/evergreen_state.c b/src/gallium/drivers/r600/evergreen_state.c
> index a484f0078a..b8ed4fea33 100644
> --- a/src/gallium/drivers/r600/evergreen_state.c
> +++ b/src/gallium/drivers/r600/evergreen_state.c
> @@ -2413,6 +2413,28 @@ static void evergreen_emit_sampler_states(struct r600_context *rctx,
>  		rstate = texinfo->states.states[i];
>  		assert(rstate);
>  
> +		/* For texture arrays the formula select the layer is (floor(z + 0.5)) and
to select

> +		 * apparently tha hardware doesn't trigger this when the texture is in ARRAY mode
> +		 * Neither does the default z-rounding mode provide the required 0.5 shift
> +		 * nor does it round with sufficinet accuracy. Consequently set the coordinate
sufficient. Albeit I'd still argue the rounding is done with quite
sufficient accuracy to be spec compliant.

> +		 * interpolation and truncate mode here to get "floor" for positive coordinates.
> +		 * Adding the 0.5 offset is done in the shader.
> +		 * Also  make sure that for other texture types the default is used.
You could also mention this alters all coordinates.

So, I'm still really not fond of the idea, since there's no proof it
causes any issues rather than venture into unknown territory, but whatever.

Acked-by: Roland Scheidegger <sroland at vmware.com>


> +		 */
> +		struct r600_pipe_sampler_view	*rview = texinfo->views.views[i];
> +		if (rview) {
> +			rstate->tex_sampler_words[0] &= C_03C000_Z_FILTER;
> +			enum pipe_texture_target target = rview->base.texture->target;
> +			if (target == PIPE_TEXTURE_2D_ARRAY ||
> +				target == PIPE_TEXTURE_CUBE_ARRAY ||
> +				target == PIPE_TEXTURE_1D_ARRAY) {
> +				rstate->tex_sampler_words[0] |= S_03C000_Z_FILTER(V_03C000_SQ_TEX_Z_FILTER_POINT);
> +				rstate->tex_sampler_words[2] |= S_03C008_TRUNCATE_COORD(1);
> +			} else {
> +				rstate->tex_sampler_words[2] &= C_03C008_TRUNCATE_COORD;
> +			}
> +		}
> +
>  		radeon_emit(cs, PKT3(PKT3_SET_SAMPLER, 3, 0) | pkt_flags);
>  		radeon_emit(cs, (resource_id_base + i) * 3);
>  		radeon_emit_array(cs, rstate->tex_sampler_words, 3);
>