sokol_gfx.h: SG_PIXELFORMAT_R16F (compute) support as a feature?

[exavi]

Hi there!

Would you consider allowing SG_PIXELFORMAT_R16F as a valid compute destination for supported backends? (I think it might be another case of it being restricted by WebGPU too)

It might seem like an obscure format but even a basic SSAO implementation could use it (my current use case 😅) otherwise at the moment one would have to use R32F (which works but it's a bit of a waste).

If you would consider it I'd be happy to start a PR and add support for the backends I can test (Metal, OpenGL, Vulkan) this time looking at how you set up the INT10_N2 feature flag, to add the feature flag too... to reduce work on your end. 😄

[floooh]

This is actually a bit easier, since pixel format capabilities are already optional by default.

Look for the functions:

• _sg_gl_init_caps_glcore and _sg_gl_init_caps_gles3, both call into _sg_gl_init_pixelformats_compute when GPU-compute is supported
• _sg_mtl_init_caps
• _sg_wgpu_init_caps
• _sg_vk_init_caps
• _sg_d3d11_init_caps

Currently all backends except D3D11 use the same conservative capabilities looking like this, which is basically what WebGPU can do:

    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RGBA8]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RGBA8SN]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RGBA8UI]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RGBA8SI]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RGBA16UI]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RGBA16SI]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RGBA16F]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_R32UI]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_R32SI]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_R32F]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RG32UI]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RG32SI]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RG32F]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RGBA32UI]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RGBA32SI]);
    _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_RGBA32F]);

D3D11 is already a bit smarter, it checks the actual pixel caps provided by D3D:

https://github.com/floooh/sokol/blob/da2079a382a90b97dcf0cf4496e273068a725313/sokol_gfx.h#L13128

For the other backends we'd need to research what's allowed.

E.g. in Metal I usually check this table:

https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf

For Vulkan and desktop GL there's no such convenient lookup table though.

[exavi]

sorry I'm not sure I'm following, I enabled _sg_pixelformat_compute_all(&_sg.formats[SG_PIXELFORMAT_R16F]); in sg*_init_caps() locally on my end, but want to know if that would be something you'd accept in mainline sokol_gfx.h :)

or do you mean that there's already a way to enable them via some API ?

[floooh]

Which backends are you testing with? E.g. D3D11 should already work.

On the other backends we'd need to do some research whether R16F for compute shader access is an optional feature behind an extension or guaranteed to be supported.

Except for D3D11, all backends currently only support the 'safe subset' which works for WebGPU without extensions.

[exavi]

I've tested OpenGL & Metal locally already with an SSAO implementation that I can flip between R32F/R16F and can test Vulkan (only on Linux, if that matters, just need to push it to test it on another machine)

[floooh]

At least on Metal R16F is universally supported for all shader access types, so we can just add the _sg_pixelformat_compute_all there...

For GL and WebGPU I'll need to do some research (e.g. there may be extensions we could check or require).

...let me just take care of the implementation, less back and forth ;) (I'll see if I can take care of it within the next few days).

[floooh]

Note to self:

WebGPU texture formats tier1 and tier2 extensions are relevant:

https://www.w3.org/TR/webgpu/#texture-formats-tier1

https://www.w3.org/TR/webgpu/#texture-formats-tier2

(TL;DR: tier1 enabled read-only / write-only... e.g. in sokol-gfx it would basically be:

• only tier1 available: read=false, write=true
• tier2 available: read=true, write=true

In Vulkan, R16F support is behind the shaderStorageImageExtendedFormats feature flag, e.g. search here: https://registry.khronos.org/VulkanSC/specs/1.0-extensions/man/html/VkPhysicalDeviceFeatures.html

...apparently in desktop GL it's always supported, see format table here:

https://registry.khronos.org/OpenGL-Refpages/gl4/html/glBindImageTexture.xhtml

...but not in GLES 3.1/3.2:

https://registry.khronos.org/OpenGL-Refpages/es3/html/glBindImageTexture.xhtml

[floooh]

FWIW, all taken care of in #1476, will merge this tomorrow.

[exavi]

thanks!