Minimize glActiveTexture calls #24492
Conversation
|
This is an interesting PR, thanks! Do you think it's possible to reproduce the redundant gl calls in a small live example? I've prepared a fiddle with a simple textured mesh. Do you think you can modify/enhance it so we have something to test? |
|
@Mugen87 The CSM example shows the issue pretty well - we're using CSM so that's where I initially saw it. Without this change it calls ActiveTexture 4 times (for the 4 shadow sections) when it changes material: And after applying this change it doesn't call it at all because the textures were already bound to that slot: Because this example only uses 2 materials, you don't see a ton of savings, but on our scenes with 80+ material changes the gl calls add up to a significant chunk of cpu time doing nothing. |
|
Nice! Is is worth having both Would it make sense to update all the current calls to |
We could definitely do that. The remaining usage of bindTexture is all around render target stuff (generating mipmaps, setting up framebuffer textures) that doesn't currently do any slot management. The remaining question is whether calling it with slot == undefined should set the slot to currentTextureSlot, or to the last slot. currentTextureSlot would minimize activeTexture calls and match the existing behavior the closest. Using the last slot could potentially keep from rebinding over textures, but that's more content dependent. |
|
So could it be this instead and make it work for both? function bindTexture( webglType, webglTexture, webglSlot ) { |
|
Much nicer! |
|
Thanks! |
* Don't set activetexture for textures that won't be bound * Fix texture state update of a previously bound texture * Combine bindTexture and bindTextureToSlot into a single function
Description
When running spector.js captures of a complex scene, I noticed a number of calls to glActiveTexture that were followed immediately by a separate call to glActiveTexture. This change checks to see if a texture will be bound (or uploaded) before calling ActiveTexture, improving performance by removing these useless gl calls.
This contribution is funded by Meta