Opinion about memory allocation methods

[duckdoom5]

I'm building a rendering framework for my game engine and I'm wondering about what's the best route for me to go with.

In your opinion, should I use the allocator interface and use vma/dx ma or is it better to do more manual memory allocation. What are the pros and cons?

I'm mostly concerned about the ability to create as many textures and resources up front and then only assign the memory when the resource is actually used. I'm also planning to add ray tracing support eventually

[dzhdanNV]

Without dipping into the details, I would say:

• VMA/D3D12MA is the best if memory management is needed. Say, resources are planned to be created and destroyed on the fly. MA documentation is great and... comprehensive. MA support in NRI is limited. Extending this can be done per request
• HelperInterface::AllocateAndBindMemory is still useful to organize a bunch of resources into one heap, aka a single logical memory block (potentially even with aliasing)
• Core functionality allows you to create your own MA...

P.S. Really, ResourceAllocator should offer more MA functionality somehow...

[dzhdanNV]

I wish VMA and D3D12MA, which are completely different, get merged into a single API-agnostic project MA. Yes, it must be graphics specific, of course, but it should not use any GAPI calls explicitly. In this case, it could be just used as a 3rd party in NRI-based projects without a need to be integrated into NRI itself. It's my wish, it won't happen, I guess.

[duckdoom5]

Thanks for your answer, very helpful!

I'll go with the ResourceAllocator then and request features if I find anything missing

[dzhdanNV]

Really, ResourceAllocator should offer more MA functionality somehow...

ResourceAllocator interface has been merged into Core in a better form. Now:

• a Memory can be allocated via AllocateMemory with AllocateMemoryDesc::useVMA = true. It allows to sub-allocate (alias, place) resources in a memory block allocated by AMD D3D12/Vulkan Memory Allocator. Also it allows to "create an own allocator" inside the allocator :)
• 2 ways to allocate and bind resources to memory: Vulkan style (classic, compatible with old-school Vulkan) and D3D12 style (a different one):
  • VK style: https://github.com/NVIDIA-RTX/NRI/blob/main/Include/NRI.h#L111
  • D3D12 style: https://github.com/NVIDIA-RTX/NRI/blob/main/Include/NRI.h#L126
• newly added CreatePlacedX functions may accept a Nri[Device/DeviceUpload/HostUpload/HostReadback]Heap macro instead of memory, offset pair to create a placed resource in corresponding implicit heaps via D3D12/VK MA
• hooked up D3D12 tight alignment !

[duckdoom5]

Awesome, looking good 🚀