-
2020Q3 Core
Due by September 30, 2020 Last updated about 9 hours agoDiscussion issue: #2290 Tracking the things we want done by the end…
Discussion issue: #2290 Tracking the things we want done by the end of Q3 in the IREE core compiler and runtime. Issues here are representative of what we think we can do + 20% with the expectation being that we won't complete everything.
-
2020Q4 Core
Due by December 31, 2020 Last updated 2 days agoTracking the things we want done by the end of Q4 in the IREE core …
Tracking the things we want done by the end of Q4 in the IREE core compiler and runtime. Issues here are representative of what we think we can do + 20% with the expectation being that we won't complete everything (if COVID weren't a thing). At this point, add ±75% to all estimates.
-
Replace legacy SPIR-V codegen with structured ops
No due date Last updated 3 days agoMilestone to track regressions from deprecation on Vulkan/SPIR-V co…
Milestone to track regressions from deprecation on Vulkan/SPIR-V codegen path due to deprecation of legacy SPIR-V codegeneration (which used IndexComputation pass to lower from XLA to SPIR-V directly)
-
Infrastructure Fixit
No due date Last updated 26 days ago -
TensorFlow SavedModel to Run Tutorial
Past due by 4 months Last updated about 1 month agoAs discussed on discord, we should target an e2e tutorial showcasin…
As discussed on discord, we should target an e2e tutorial showcasing the non-source based flow for compiling a TensorFlow model, saving it, loading it, and running via the command line, Python (and Java?) bindings. This should specifically be focused on doing so from installable artifacts (i.e. via pip install) vs local builds.
-
Dynamic Shapes
No due date Last updated 3 months agoEnd-to-end support for partial and dynamic shapes. Currently blocked on pending MLIR/TF design work.
-
Tracy Instrumentation
No due date Last updated 4 months agoAdding Tracy profiler support to the runtime and build configurations.
-
Model: Hotword Model DNN
No due date Last updated 4 months ago -
Benchmarking Infrastructure
No due date Last updated 6 months agoIn order to start hill-climbing performance of both the compiler-ge…
In order to start hill-climbing performance of both the compiler-generated sequencer and executable IR and runtime we'll need to gather enough metrics and be able to act on them. The goal is to not necessarily try to compare with existing frameworks but instead allow us to compare our various backends and incremental performance improvements.
-
Cellular Batching
No due date Last updated 11 months agoA core feature of the IREE design is cellular batching to allow run…
A core feature of the IREE design is cellular batching to allow runtime batching at a fine granularity. There are several approaches that combine to layer on benefits of coalescing (to exploit temporal cache locality) and batching (to increase arithmetic density). The same infrastructure can also be used to efficiently schedule multiple work streams and avoid pipeline stalls.