[PATCH 11/16] gpu: nova-core: add falcon register definitions and base code

Alexandre Courbot acourbot at nvidia.com
Thu May 1 14:07:25 UTC 2025


On Thu May 1, 2025 at 9:22 AM JST, Joel Fernandes wrote:
> Hi Alex,
>
> On 4/30/2025 8:09 PM, Alexandre Courbot wrote:
>>>> I am just not seeing the benefits of not using dyn for
>>>> this use case and only drawbacks. IMHO, we should try to not be doing the
>>>> compiler's job.
>>>>
>>>> Maybe the only benefit is you don't need an Arc or Kbox wrapper?
>>> That's not a huge concern for me, it's only one single allocation per Engine,
>>> correct?
>> Correct. Note that for other engines we will be able to store the HALs as
>> static singletons instead of building them on the heap like I am
>> currently doing. The reason for doing this on falcon is that the
>> dual-dimension of the instances makes it more complex to build and look
>> them up.
>> 
>> ... or maybe I could just use a macro? Let me try that and see whether
>> it works.
>
> Do you mean a macro for create_falcon_hal which adds an entry to this?
>
>  let hal = match chipset {
>     Chipset::GA102 | Chipset::GA103 | Chipset::GA104 | Chipset::GA106
> |Chipset::GA107 => { .. }
>
>
> Actually it would be nice if a single macro defined both a chipset and created
> the hal together in the above list, that way the definition of a "chipset" is in
> a singe place. Kind of like what I did in the vbios patch for various BiosImage.
> But not sure how easy it is to do for Falcon.
>
> Or perhaps you meant a macro that statically allocates the Engine + HAL
> combination, and avoids need for Arc/KBox and their corresponding allocations?

I was thinking of a macro to create all the Chipset * Engine static
instances of HALs, and generate the body of a lookup function to return
the right one for a given Chipset at runtime.

But trying to write it, I realized it wasn't as easy as I thought since
generics cannot be used as macro parameters - i.e. if you have 
<E: Engine> as a generic and pass `E` to the macro, it will see... `E`
and not whatever was bound to `E` when the code is monomorphized (macros
are expanded before generics it seems).

A solution to that would involve new traits and a bunch of boilerplate,
which I have decided is not worth the trouble to save one 8-bytes object
on the heap per falcon instance. :) I'll keep things as they currently
are for now.


More information about the dri-devel mailing list