Skip to content

Latest commit

 

History

History
125 lines (100 loc) · 4.83 KB

CHANGELOG.md

File metadata and controls

125 lines (100 loc) · 4.83 KB
Raw

UpTune, Copyright Shaojie Xiang (sx233@cornell.edu), 2018-2019

See the LICENSE file for copyright information.

TODO

  • support rpc session for multi-platform communication
  • support automatic deployment to local/aws/gcp
  • support autotuning with inter-variable constraint

2020-11-18

  • Use an async task scheduler

2020-11-12

  • Updated naming conventions in source code
  • Added plotting scripts to capture relationship between compile_time and
  • Switched to ray's wait API for task overtime
  • Remove OpenTuner global database
  • Updated the new constraint and search interface (@ut.rule)
  • Support Async execution of different executors (the democratized tuners shouls also be controlled by the contralized cointroller)
  • New scripts for Intel-AOCL synthesis report parsing
  • Add new APIs to query global_id and local_id

2019-12-25

  • added permutation support & rename
  • updated api for easy integration
  • support extern tuner as plugin-in into autotvm (builder + runner required. also for samplers and optimizer + cost model)

2019-12-11

  • support deamon auto-tuning server with ray-like API
  • replace returned value as symbolic variable

2019-11-09

  • added placeholder replacement feature for tuning non-python file

2019-11-05

  • added baysien net structure recovery for causal learning
  • added c++ template for building lower level intrinscs
  • added support for importing history data

2019-10-10

  • fixed the sqlalchemy cfg generation issue
  • added the runtime limit increase reminder
  • added the cfg display support with --cfg arg

2019-09-11

  • added constraint features
  • added support for tuning across gpus
  • updated target decorator (support ut.tune inside)

2019-09-01

  • updated api for automatic tuning type inference
  • updated module import mechanism for constraint feature

2019-07-21

  • unified interface for controller initilization
  • added adaptive work dierctory selection on shared FS
  • added Amazon AWS support for cluster auto-tuning

2019-05-30

  • replaced message passing interface with file access
  • added __tmp__ to store template files

2019-05-28

  • fixed ZMQ msg encoding issue (in intrusive mode)
  • updated XGBoost objective to squared error
  • incorporated single stage tuning into current flow
  • enabled automatic clean-up func before running

2019-05-27

  • resolved control conflicts of intrusive / template tuning
  • supported single/double quotes in regex value detection
  • migration to Python 3.7

2019-05-21

  • added support of decouple tuning
  • support real-time tuning result plotting

2019-03-21

  • added ResNet-18 Example for HeteroCL

2019-03-11

  • introduced intrusive APIs for specifying tuning variable
  • added OS environment specifier for different tuning mode
  • added ZeroMQ framework for proposal distribution over cluster

2019-03-02

  • added README and basic use cases
  • fixed Jinja tojson render error for Boolean vars with filter patch

2019-02-26

  • added the objective declaration support in annotation
  • added report-api mapping mode and auto-test shell script for LAMBDA
  • disabled RUNNING status in opentuner's desired result
  • updated cache method of ML instance for re-training

2019-02-21

  • updated command line interface to suppotr mixed command and meta-parameter
  • updated learning models abstraction for re-training and parameterization
  • added multi-stage auto-tuning support based on fine-grained control logics
  • added example of LAMBDA on Quartus LUT flow (for hybrid/online/vinilla mode)
  • added tojson filter to Jinja template render for string type

2019-02-18

  • added HeteroCL schedule tuning example for BNN acceleration
  • added CLI clean mode (clean imtermediate files with "uptune clean")
  • update the regex pattern for accurate matching
  • add support to limit process runtime (process killed if timeout)

2019-02-17

  • enabled Command Line Interface ("python -m uptune.on cmd" or "uptune cmd")
  • added Jinja Template Engine for code generation on ray run() function
  • added Template Parsing and Generation for auto-tuning and proposal assignment
  • added fine-grained auto-tuning control APIs in api.py
  • test and Error cases + Environment set-up scripts
  • added tuning data type template (TuneFloat + TuneResult)
  • enabled Metaclass Mode for class registration

2019-02-09

  • added tuning data type template (TuneInt + TuneEnum)
  • added learning model support as modular plugins for search space pruning

2018-12-17

  • added support for calling function spawned in ThreadPool with time limit (defined by user)
  • added EZTuner GCC benchamark(gccflags/eztune.py) over OpenTuner baseline(gccflags.py)
  • added Support for recording timing out results in global DB (for ML training)