Performance optimizations

[rudedogg]

I've identified a few things that could help improve the performance of the library. Currently, there's a bit of lag in the SwiftUI demo view when moving the noise property sliders around and using a large texture size.

Benchmarks

• Need to get some baseline numbers for how long generating noise images takes. The numbers should guide us through the decisions about performance below.
• Benchmarks of GameplayKit's GKNoiseSource would be nice to have and compare against

Storage Modes

See https://developer.apple.com/documentation/metal/setting_resource_storage_modes for storage mode information.

• Optimize resource storage modes for macOS
  • https://developer.apple.com/documentation/metal/setting_resource_storage_modes/choosing_a_resource_storage_mode_in_macos
• Optimize resource storage modes for iOS
  • https://developer.apple.com/documentation/metal/setting_resource_storage_modes/choosing_a_resource_storage_mode_in_ios_and_tvos

Shader Optimizations

• Benchmark whether avoiding branches helps. All of the shaders make pretty heavy use of branches, so this could be a huge win
  • Advanced Metal Shader Optimization - WWDC 2016
  • Metal Performance Optimization Techniques - WWDC2015 Looks like it covers using the debugger to get a time spent - % for each statement in the shader code

Thread Groups

• Calculate an optimal thread group size
  • https://developer.apple.com/documentation/metal/creating_threads_and_threadgroups
  • https://developer.apple.com/documentation/metal/calculating_threadgroup_and_grid_sizes

Asynchronous Command Buffer

• No idea if possible when we are grabbing the texture, but should check if we can use waitUntilScheduled() instead of waitUntilCompleted()

Miscellaneous Performance Links

• https://developer.apple.com/documentation/metal/render_pipelines/optimizing_performance_with_gpu_counters
• https://developer.apple.com/documentation/metal/render_pipelines/optimizing_performance_with_pipeline_statistics