Cache bounding box values between register reads in DX11/GL #9648
Conversation
stenzek
force-pushed
the
dx11-gl-bbox-caching
branch
from
aec023f
to
72ade61
Compare
| @@ -36,43 +50,101 @@ void BoundingBox::Shutdown() | |||
| glDeleteBuffers(1, &s_bbox_buffer_id); | |||
| } | |||
|
|
|||
| void BoundingBox::Set(int index, int value) | |||
| void BoundingBox::Flush() | |||
There was a problem hiding this comment.
I don't know enough about BoundingBox stuff to know if this is worth it but longer term it makes me wonder if we could split some of this out to a AbstractBoundingBox class that does some of this common logic and calls into the backend specific functionality. This is near identical to the D3D11::Flush() . I haven't looked at the other backends to know if they are similar but you mentioned there is caching there too..
|
This makes bounding box effects much faster in Paper Mario TTYD on D3D11 with an NVIDIA GPU. OpenGL got a marginal performance gain as well. However, despite what the PR description seems to imply, D3D12 is very slow. |
stenzek
force-pushed
the
dx11-gl-bbox-caching
branch
from
72ade61
to
46128ba
Compare
There was a problem hiding this comment.
One thing I'm noticing (and it seems to extend to the VK and DX12 cache implementations too) is that bounding box behaviour is ill-defined during save-state load.
At the very least, perhaps the caches should be flushed after a save-state load to prevent dirty cached bbox values from being loaded. But even that could produce invalid results if there is a save state between the finish of a draw and the bbox values being read.
I'm not sure if this is worth worrying about. This shouldn't effect regular users, just people doing debugging with save-states, which is already a hairy proposition with lots of edge cases
stenzek
force-pushed
the
dx11-gl-bbox-caching
branch
from
46128ba
to
79eebd7
Compare
stenzek
force-pushed
the
dx11-gl-bbox-caching
branch
from
79eebd7
to
7fd0a52
Compare
|
FifoCI detected that this change impacts graphical rendering. Here are the behavior differences detected by the system:
automated-fifoci-reporter |
|
^ I think this fifoci diff is a false herring, bounding box shouldn't affect anything here. |
This was present in the Vulkan and DX12 renderers, but not DX11/GL.
Quick test shows 60fps->90fps at 1x, and 55fps->80fps at 3x on an AMD GPU.
Bounding box is broken on OpenGL/AMD due to what seems to be a driver bug - the atomic only works on one of the fields, so I can't test it.