[NNNN] DynamicBufferObjects (FromAddress) proposal

[mapodaca-nv]

Dynamic Buffer Objects introduces the ability to create buffer objects directly from GPUVAs.

[llvm-beanz]

I have two high-level thoughts on this, which are probably worth exploring.

Should this just be pointers?

If we just had pointers in HLSL would we be considering this at all? If this feature is just working around not having pointers, I'd rather we figure out how to add pointers or references.

Memory safety

I don't love that we have no bounds checking on these. It should be pretty straightforward to add a size value to the new functions and support bounds checking. We've previously discussed not adding raw pointers to HLSL and instead adding bounds-checked "range" constructs specifically to avoid introducing memory safety issues.

Similar to conversations that we've had in the past about bounds checking on resources, I think that if we do add strictly unsafe memory operations we will need to think about how to make it opt-in, and something captured in the shader's feature flags so that an application that accepts untrusted inputs (like a web browser) can set a flag on PSO creation to disallow the feature.

[mapodaca-nv]

Should this just be pointers?

If we just had pointers in HLSL would we be considering this at all? If this feature is just working around not having pointers, I'd rather we figure out how to add pointers or references.

I see your point but there are some advantages to this model. Once the objects are created from the GPUVA, they can be passed to existing, legacy shader code without requiring any changes. And we maintain the strong typing, alignments, and indexing syntax. Even once we have full support for raw pointers, casting those pointers to an object type like this can provide some benefits.

Memory safety

I don't love that we have no bounds checking on these. It should be pretty straightforward to add a size value to the new functions and support bounds checking. We've previously discussed not adding raw pointers to HLSL and instead adding bounds-checked "range" constructs specifically to avoid introducing memory safety issues.

FWIW, this concern seems to conflict with the first desire to add raw pointers. Furthermore, adding a length parameter does nothing more than force the implementation to implement manual bounds checking (something the shader code can implement itself, if desired) on a value that has no more guarantees of being valid than the GPUVA. i.e. If we cannot trust the any part of the shader code then we must not trust all the shader code.

Similar to conversations that we've had in the past about bounds checking on resources, I think that if we do add strictly unsafe memory operations we will need to think about how to make it opt-in, and something captured in the shader's feature flags so that an application that accepts untrusted inputs (like a web browser) can set a flag on PSO creation to disallow the feature.

There is a flag to detect if this feature is being used.

[damyanp]

This is nearly ready to get a number assigned and merged, with any further follow ups tracked by issues.

There's one proposed change to set the metadata up correctly.

We also need to get a sponsor assigned. I took the liberty of setting @llvm-beanz as the sponsor in the proposed metadata change, so we should wait for his final approval before merging this.

[damyanp]

Suggested change

	# Dynamic Buffer Objects
	* Proposal: [NNNN](NNNN-dynamic-buffer-objects.md)
	* Author(s): [Mike Apodaca](https://github.com/mapodaca-nv)
	* Sponsor: TBD
	* Status: **Under Consideration**
	---
	title: "NNNN - Dynamic Buffer Objects"
	params:
	authors:
	- mapodaca-nv: Mike Apodaca
	sponsors:
	- llvm-beanz: Chris Bieneman
	status: Under Consideration
	---