Android Vulkan Performance

[vinsentli]

I tested on Qualcomm Adreno (TM) 750, 730, and other chips, and found that Vulkan consumes more CPU and memory compared to OpenGL ES. This seems inconsistent with our expectations.
Do you see the same in your test data?

When I used Android Studio to analyze the performance, I discovered that ResourcesBinder::updateBindings() consumes the vast majority of the time in RenderCommandEncoder::draw().
Finally, the CPU spends most of its time in vkCreateDescriptorPool(), vkAllocateDescriptorSet(), vkUpdateDescriptorSets().

According to the documentation https://github.com/rudybear/vulkan_best_practice_for_mobile_developers/blob/master/samples/performance/descriptor_management/descriptor_management_tutorial.md
DescriptorSets should be reused, rather than reallocated every frame.

Do you also think so?

If this is indeed the problem, do you have any optimization suggestions? I can proceed to implement them based on your plan.

[corporateshark]

@vinsentli Thanks for bringing this up!

Yes, descriptors can definitely benefit from a good caching strategy and reuse. It’s on our roadmap, but we don’t have any specific timeline yet. If you’re able to demonstrate measurable performance improvements with your approach, we’d be happy to take a look at a pull request.

Btw, did you consider using BindGroups in IGL? They are much faster and don't have any caching problems.

[vinsentli]

I've noticed BindGroups. I've used BindGroups for textures, but haven't succeeded yet for UBOs. The reason is that we have a relatively high number of draw calls, and the UBO structure is quite complex. I have managed to reduce calls to vkAllocateDescriptorSets() by reusing VkDescriptorSet objects. However, it has been challenging to reduce the number of vkUpdateDescriptorSets() calls.

[corporateshark]

@vinsentli I see. Please go ahead with a PR.

[vinsentli]

PR:#336