Best practices for porting DirectML

[Axiverse]

I'm working on porting DirectML and am willing to contribute the work to the general library if you'd want it here. I have some questions on conventions for some of the struct classes. I'm sure more questions will come up as I continue the port, but these are the ones I have currently.

stackalloc

When calling some bindings, you copy into a stackalloc array before calling. I'm wondering why this is as SharpGenTools do create bindings which take in an array.

DML_BINDING_DESC

DirectML heavy uses DML_BINDING_DESC for all its operator descriptions and similar structures which use a type enum and a void ptr for the payload. Is there a recommended way of porting this to C#?

As they don't add any direct value, I've opted to collapse this out in favor of using the specializations which would then marshal directly into DML_BINDING_DESC. What do you think of this approach.

I have a draft pull request with some implementation and a partial port of the HelloDirectML. The part which has been ported compiles and runs without throwing any exceptions.

#256

[amerkoleci]

stackalloc

When calling some bindings, you copy into a stackalloc array before calling. I'm wondering why this is as SharpGenTools do create bindings which take in an array.
-------- I probably need to fork SharpGenTools to handle those goodies

DML_BINDING_DESC

DirectML heavy uses DML_BINDING_DESC for all its operator descriptions and similar structures which use a type enum and a void ptr for the payload. Is there a recommended way of porting this to C#?
-------- I will try to see how to handle it
As they don't add any direct value, I've opted to collapse this out in favor of using the specializations which would then marshal directly into DML_BINDING_DESC. What do you think of this approach.

I have a draft pull request with some implementation and a partial port of the HelloDirectML. The part which has been ported compiles and runs without throwing any exceptions.

#256

Thanks

[Axiverse]

I've updated the alloc structure to be closer to some other examples in the code. Everything handles it's own allocation and deallocation now. There's a number of allocations which need to happen for every call, but these descriptions shouldn't be used often in calls - only for setting up graphs.

[Axiverse]

On the stackalloc question, I'm just wondering why you do it and if it's necessary for me to do it as well.

[Axiverse]

I found some missing things. There's a few structures around DirectML 1.1 graphs which need to be wrapped. Otherwise I'd like to get some documentation in as well. Put a fix for SharpGenTools to (finally?) fix documentation inclusion. We will see. I might go ahead and start generating the documentation anyhow. 😃