The Apache SystemML 1.0.0 release was approved on December 12, 2017. The release includes enhancements, features, and additions as listed below.
- JMLC/MLContext parameter passing w/ and w/o meta data.
- Improved CPU convolution function performance, including sparsity.
- Enhanced/additional NN layers, e.g. conv2d transpose and depthwise convolution
- NN architecture summary table.
New Scripts / Algorithms
- Enhanced PCA.
Features & new Functions
order with multiple order-by columns.
- Improved rewrites, e.g. merging of statement block sequences.
- IPA effectiveness.
- Automatic vectorization of indexing pairs.
- Instruction generation for memory efficiency.
- Instruction code organization.
- [Experimental] Code generation, code gen optimizer, and multi-threaded codegen operators.
Robustness & Performance
- ParFor/HOP Memory budgets for Spark cluster configurations.
- JMLC prepared scripts (thread affinity for outputs and configs, script cloning, configuration management).
- Sparse-dense binary cell wise operations.
- GPU conv2d and memory estimates.
- Additional kernels,e.g. right indexing.
- Single precision backend.
- SystemML Lite artifact, a minimumm-size uber JAR for embeddability (w/o Hadoop or Spark dependencies).
- Compression on by default.
- Exploitation of native BLAS libraries.
- JDK 8 (dropped JDK 7).
- jCUDA for Windows & Linux (x86_64, ppc64le) included.
- Fine-grained runtime statistics.
- Refactored configurations parameters to have sysml prefix
- Performance test suite.
- Refreshed/new examples and notebooks (DML examples), tutorial.
- Features, Improvements, Bug fixes, etc.