[Mesa-dev] [PATCH] radv: add support for anisotropic filtering on VI+

[Fredrik Höglund]

On Thursday 27 October 2016, Fredrik Höglund wrote:
> On Thursday 27 October 2016, Roland Scheidegger wrote:
> > On 10/26/2016 06:02 PM, Fredrik Höglund wrote:
> > > Ported from radeonsi.
> > > ---
> > >  src/amd/vulkan/radv_device.c | 30 +++++++++++++++++++++++++++---
> > >  1 file changed, 27 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
> > > index 512d366..9184308 100644
> > > --- a/src/amd/vulkan/radv_device.c
> > > +++ b/src/amd/vulkan/radv_device.c
> > > @@ -1742,26 +1742,50 @@ radv_tex_bordercolor(VkBorderColor bcolor)
> > >  	return 0;
> > >  }
> > >
> > > +static unsigned
> > > +radv_tex_aniso_filter(unsigned filter)
> > > +{
> > > +	if (filter < 2)
> > > +		return 0;
> > > +	if (filter < 4)
> > > +		return 1;
> > > +	if (filter < 8)
> > > +		return 2;
> > > +	if (filter < 16)
> > > +		return 3;
> > > +	return 4;
> > > +}
> > > +
> > >  static void
> > >  radv_init_sampler(struct radv_device *device,
> > >  		  struct radv_sampler *sampler,
> > >  		  const VkSamplerCreateInfo *pCreateInfo)
> > >  {
> > > -	uint32_t max_aniso = 0;
> > > -	uint32_t max_aniso_ratio = 0;//TODO
> > > +	uint32_t max_aniso = pCreateInfo->anisotropyEnable && pCreateInfo->maxAnisotropy > 1.0 ?
> > > +					(uint32_t) pCreateInfo->maxAnisotropy : 0;
> > > +	uint32_t max_aniso_ratio = radv_tex_aniso_filter(max_aniso);
> > >  	bool is_vi;
> > >  	is_vi = (device->instance->physicalDevice.rad_info.chip_class >= VI);
> > >
> > > +	if (!is_vi && max_aniso > 0) {
> > > +		radv_finishme("Anisotropic filtering must be disabled manually "
> > > +		              "by the shader on SI-CI when BASE_LEVEL == LAST_LEVEL\n");
> > > +		max_ansi = max_aniso_ratio = 0;
> > Does that really compile? I don't see any such variable.
> 
> No, it does not.  It turns out I fixed that locally, but in the wrong commit.
> 
> > (I'm actually curious what happens if you don't disable aniso in this 
> > case, does the math end up all crazy and it samples completely wrong 
> > texels? Seems rather lame...)
> 
> I'm curious about that myself.  But I don't own any of the GPU's that
> need the workaround, so I can't tell you.

I discussed this issue with Marek on IRC, and apparently nothing should
break if the shader doesn't disable it.  It's just desirable to do so.

I'll update the patch to allow anisotropic filtering to be enabled on
those GPU's.

> > Roland
> > 
> > 
> > > +	}
> > > +
> > >  	sampler->state[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo->addressModeU)) |
> > >  			     S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo->addressModeV)) |
> > >  			     S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo->addressModeW)) |
> > >  			     S_008F30_MAX_ANISO_RATIO(max_aniso_ratio) |
> > >  			     S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo->compareOp)) |
> > >  			     S_008F30_FORCE_UNNORMALIZED(pCreateInfo->unnormalizedCoordinates ? 1 : 0) |
> > > +			     S_008F30_ANISO_THRESHOLD(max_aniso_ratio >> 1) |
> > > +			     S_008F30_ANISO_BIAS(max_aniso_ratio) |
> > >  			     S_008F30_DISABLE_CUBE_WRAP(0) |
> > >  			     S_008F30_COMPAT_MODE(is_vi));
> > >  	sampler->state[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo->minLod, 0, 15), 8)) |
> > > -			     S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo->maxLod, 0, 15), 8)));
> > > +			     S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo->maxLod, 0, 15), 8)) |
> > > +			     S_008F34_PERF_MIP(max_aniso_ratio ? max_aniso_ratio + 6 : 0));
> > >  	sampler->state[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo->mipLodBias, -16, 16), 8)) |
> > >  			     S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo->magFilter, max_aniso)) |
> > >  			     S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo->minFilter, max_aniso)) |
> > >
> > 
> > 
> 
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>