I am working on making a DX12, Vulkan framework run on CPU and GPU in parallel.

Decided to finish the Vulkan implementation before DX12. (Eat the veggies before having the steak XDD)

I have a few questions about the usage of ID3D12CommandAllocator:

• Different sized command lists should use different allocators so the allocators dont grow to worst size
  • Does this mean that I need to know the size of the command list before calling CreateCommandList and pass the appropriate allocator?
• Try to keep number of allocators to a minimum
  • What are the pitfalls if I create a command allocator per list? This way each allocator will never grow too large for the list. In addition, there will be no need for synchronization.

Most of the examples I have seen just use a pool of allocators and do fence based synchronization. I can modify that to also consider command list size but before that any advice on this will really help me to understand the internal workings of the ID3D12CommandAllocator in a better way.

The idea behind reusing allocators is to reuse block of memory needed by the command list generation, possibly. There is not a single strategy here, especially because you still don't know internally how the allocators behaves and where may be the sweet spots.

I can only at best give you ONE idea on how to reuse them  Because an allocator can only bound to one active command list at a time, you can create a command allocator per thread producing command list (imagine a job queue system, a job produce a command list ). You can then have a pool of allocator to alternate between frames and give you enough time for the gpu to execute them before you again need to reset a set of allocators. 

On 8/9/2017 at 6:20 PM, mark_braga said:

Different sized command lists should use different allocators so the allocators dont grow to worst size

• Does this mean that I need to know the size of the command list before calling CreateCommandList and pass the appropriate allocator?

 

Yes to the first part, not exactly (but close) to the bulleted question.  If you can swing it, you should consider using the same or similar allocators for the same command lists.  For example, let's say you're doing deferred rendering and are drawing some number of items.  Your pass to fill the gbuffer may take (for example's sake) 200 calls to various functions on the command list, which are written to the allocator.  Your shadowmap pass has to be run for each shadow-casting light, so you'll have somewhere in the neighborhood of 200 * numLights entries in the command list, and therefore that allocator.  If in a subsequent frame, you use the allocator you initially used for the gbuffer pass for the shadow map pass, that allocator will grow to the worst-case scenario size, i.e., the size required by the shadowmap pass.  If instead you keep that allocator designated for gbuffer stuff only, it'll only be sized enough for your worst-case gbuffer pass, so could save on your overall memory budget.

To answer whether you need to know the size of the command list before calling CreateCommandList, not precisely, but it's helpful to have some idea of how the command list will be used so you have a ballpark idea of how it will affect the allocator bound to it.

On 8/9/2017 at 6:20 PM, mark_braga said:

Try to keep number of allocators to a minimum

• What are the pitfalls if I create a command allocator per list? This way each allocator will never grow too large for the list. In addition, there will be no need for synchronization.

 

Keep them to a minimum, but don't dig yourself into a hole.  You can reuse an existing command list to write into a separate allocator than the one you initially created it with while the initial allocator's content is being consumed by the GPU.  You still must set and track your fence values to be sure that the GPU is done with the allocator you're about to reset and reuse.  You should create enough to cover your maximum latency, but if it comes down to it, you still must be able to block and wait for the GPU to finish it's work if all the command allocators are occupied.  This is like the pool of allocators @galop1n mentioned in his post.