True coherent memory on DX12

[kvark]

Currently, we expose all CPU-visible memory types in DX12 as coherent.
However, it's documented in D3D12's Map() and Unmap() that they may (or need) to be used to invalidate/flush the memory regions, and that they can be nested. Judging by that, I think this is closer to our non-coherent memory than coherent one.

We may need to strip the coherency property from all the memory types we expose in D3D12 (Note: we already do this for D3D11). Alternatively, the command buffers would track the buffer transitions to and from HOST, and we could flush/invalidate the relevant resources at Submit() time.

In any way, what we have today appears incorrect. I'm just surprised we haven't seen any issues associated with this.,, I discovered this when looking at gfx-rs/wgpu#1002 but in the end it was a red herring.

[kvark]

Got an interesting clarification from Microsoft that this documentation is outdated, and memory is truly coherent. No Map/Unmap required.

[ghost]

@kvark https://software.intel.com/content/www/us/en/develop/articles/tutorial-migrating-your-apps-to-directx-12-part-3.html

3.4.4 Readback of GPU Data
We don’t want to encourage persistent map of readback (write-back) memory. You can keep it mapped, but before reading data that was written by the GPU, you should do another Map() with a range to ensure the cache is coherent. This is free on systems which don’t need it, but ensures correctness on systems which do.

JFYI.