| commit | 12b771f3265881c668d95be23abd223df78f5875 | [log] [tgz] |
|---|---|---|
| author | Ben Vanik <ben.vanik@gmail.com> | Wed Mar 08 12:27:45 2023 -0800 |
| committer | GitHub <noreply@github.com> | Wed Mar 08 12:27:45 2023 -0800 |
| tree | 6fdeccbb527db78fc653d2b5c174db6e4dbd37d8 | |
| parent | f67abe1dfb02d50e4938fd3c2aa3cc0fceb4b32b [diff] |
Making iree-benchmark-trace work with stateful traces. (#12552) This reworks it from precaching calls and running over those to instead reprocessing the entire trace YAML each iteration but only timing the calls. This allows for pretty much all trace features (with a few tweaks). `<stdin>` modules are supported by passing `--capture_stdin` to the tool so that it captures it once and reuses it for all reads. Inputs are reparsed each iteration so that traces which take ownership of the buffers or mutate them in place don't interfere. By default devices and modules are reused across iterations but that can be disabled with `--reuse_devices=false` and `--reuse_modules=false`. This approach does result in slower overall runs depending on trace content as libyaml is stupid slow but the timings aren't impacted thanks to tight scoping. In the future we could have a cached representation of the traces that keep the buffers around and avoid the need for document reparsing but I'd rather do that once the workflows around this have stabilized. I fully expect this tool to be rewritten at some point but at least now we can run stateful traces and benchmark things like the caching allocator in real(ish) execution sequences. Traces look like this, with a special "Timing Active" plot showing when we're actually timing (during calls):  Zooming in each call event can be seen getting timed: 
IREE (Intermediate Representation Execution Environment, pronounced as “eerie”) is an MLIR-based end-to-end compiler and runtime that lowers Machine Learning (ML) models to a unified IR that scales up to meet the needs of the datacenter and down to satisfy the constraints and special considerations of mobile and edge deployments.
See our website for project details, user guides, and instructions on building from source.
IREE is still in its early phase. We have settled down on the overarching infrastructure and are actively improving various software components as well as project logistics. It is still quite far from ready for everyday use and is made available without any support at the moment. With that said, we welcome any kind of feedback on any communication channels!
See our website for more information.
IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.
