Usage best practices

This page contains a list of best practices for getting the most out of IREE, spanning model authoring, ahead-of-time compilation, and runtime use. Treat these as a collection of ideas to consider or areas to start benchmarking when working on your own applications.

Introduction

Common themes include:

• Give the compiler as much information as possible
• Give the compiler opportunities to batch work together or defer computation
• Keep compute devices saturated with work through pipelining
• Use dense math where possible, particularly for inner loop bodies
• Limit synchronization points between devices like CPUs and GPUs
• Profile early and often, using the right tools for each level of granularity

Practices for model authoring

Track state within your model when possible

If your model is stateful prefer to store that state directly within your program rather than externalizing it through arguments and return values. By keeping state inside your program the compiler is better able to reason about it and function calls will have lower overhead.

If you do externalize state, try to pack that state into a limited number of arguments.

See the variables and state sample for further guidance on tracking and using state.

Limit uses of dynamic shapes

While IREE aims to support general dynamic shapes use, it is better able to optimize parts of programs where shapes are static. Slow varying dimensions like batch index or timestamp are safer uses of dynamic shapes than faster varying dimensions like the x/y/channel dimensions of images.

See the dynamic shapes sample for further guidance on using dynamic shapes.

Practices for compilation settings

TODO: which compiler targets to use (try both CUDA and Vulkan?)

TODO: use the most specific LLVM target triple you can?

Practices for runtime use

TODO: sample code, profile numbers

Tuning runtime settings

When running on the CPU, the task system flags specified in iree/task/api.c give control over how worker threads will be created. For example, the --task_topology_group_count=3 flag can be set to explicitly run on three workers rather than rely on heuristic selection that defaults to one worker per detected physical core.

If running on a single thread or system with no threading support the local-sync HAL driver can be used instead of the multithreaded local-task HAL driver to reduce dependencies and code size. When running with the local-sync driver all execution happens inline on the thread invoking the IREE runtime and will block until it has completed.

Do the minimum amount of work: cache queries and reuse buffers

When using IREE's runtime libraries, try to front-load queries, particularly queries using strings that look up into maps like iree_runtime_session_call_by_name, so that hot sections of code are doing the minimum amount of work: routing inputs through buffers, scheduling runtime calls, and routing outputs through other buffers.