Enabling features

[ib00]

One thing that is currently not addressed is enabling features when GPU is initialized and created.

For example, if I want to use raytracing, mesh shaders and cooperative matrix (ie, tensor cores) I would need to enable those
features when device is created. Are there any plans to specify high level features (raytracing, meshlets, cluster RT, etc.)?

This of course complicates the device creation tremendously, but maybe if the number of these features are limited, it could be
doable. Perhaps a limited number of high level features could be specified by the user at gpu init time.

Thoughts?

[leotmp]

Is there a reason for not automatically enabling those features when they're supported? I was thinking of adding new features in the same way that the raytracing feature works, i.e. it's enabled automatically if supported and then you can check if it's supported via features_available().

[ib00]

I am not a Vulkan expert and I don't know if there's a cost associated with enabling all available features.

It seems perfectly reasonable to enable all high level features. We would then need another query function
that returns all available/enabled features for the GPU device.

[leotmp]

That already exists in the API, it's features_available()(it's used in the raytracing example)

[leotmp]

I may have misunderstood, did you also mean enabling features that aren't currently supported by no_gfx? In that case we could also consider providing an escape hatch and let users add Vulkan features directly, or let them create the vulkan device themselves and initialize the no_gfx context with that. We could probably just start with adding features as mentioned though. If you need anything in particular let me know. I know I'll certainly need the external memory extension to be able to use Open Image Denoise.

[ib00]

Enabling features and creating a device is definitely one of the most annoying features of Vulkan and GPU APIs in general.

Not really sure what a sane way of handling this is. In the past, I have created some number of features that I was interested in (eg, ray tracing, mesh shaders, cooperative matrix, support for atomics, etc.). Alternatively, a string list of features/extensions could be passed in as well. Neither options is satisfying. Even for raytracing, there are so many different things that can be enabled (clustered RT, micro displacements, etc.). And I don't think there exists a generic solution,

It would be nice to be able to enable raytracing (and cluster RT), mesh shaders, cooperative matrix, descriptor heaps, etc.
if supported by physical device / driver.

[leotmp]

What do you think of this?

// Optional API-specific arguments to existing no_gfx procedures.
// *_set_* type procedures adds an argument for one or more specific procedures (specified in a comment)
// and they all get restored after the target call. Arguments are thread-local.
// For example:
// vk_set_instance_extensions sets the instance extensions for the next gpu.init() call, the latter
// of which will reset the instance extensions argument to an empty slice.

vk_set_instance_extensions: proc(extensions: []cstring) : _vk_set_instance_extensions      // init
vk_set_opt_instance_extensions: proc(extensions: []cstring) : _vk_set_opt_instance_extensions  // init
vk_set_device_extensions: proc(extensions: []cstring) : _vk_set_device_extensions          // init
vk_set_opt_device_extensions: proc(extensions: []cstring) : _vk_set_opt_device_extensions  // init
// pNext will be set automatically. Any structure type in 'structs' that is already used
// will overwrite no_gfx's struct, so remember to include all features used by no_gfx!
vk_set_device_pnext_chain: proc(structs: []any) : _vk_set_device_pnext_chain               // init

You can then check for extension support with the existing get_vulkan_device() and get_vulkan_physical_device()