BenchPress
A short summary of all changes and upgrades:
Models
- Implement Pytorch version of BERT
- Added iterative compilation training in BERT model
- Implement online data generator for training and sampling; Masks kernels on stream.
- Implement active data generator for sampling only; Creates generations of good candidates based on distance from target features.
- Add support for TPUs.
- Add support for data parallelism in multiple GPUs.
- Added compilation rate, sample features fields in samples DB.
Datasets
- Implement BigQuery module for querying large datasets, in C, C++, OpenCL, Java, Python and GO.
- Fetched a hybrid dataset (BigQuery + recursive miner) of 40,000 subject openCL files.
- Enabled official OpenCL 2.2 API and now CL types are supported in training and generation.
- Decreased overall rejection rate of corpuses; achieved 48% rejection rate in 40k kernel corpus.
- Sample set feature is added; original corpus is pickled and can be masked with new specs.
- Plot overall input features of encoded corpus.
System
-
Removed Bazel as a build system and migrated to CMake.
-
Upgraded full build to LLVM-9. Now both LLVM-6 and 9 are equally supported.
-
Add memory management modules for RAM and GPU memory.
-
Implement flask dashboard
-
Implement interactive plotly plots.
-
Bug fixes.