Ceygen Change Log¶

This file mentions changes between Ceygen versions that are important for its users. Most recent versions and changes are mentioned on top.

Changes in 0.4 since 0.3¶

eigen_version() function introduced to get Eigen version.
llt module introduced with Cholesky matrix decomposition.
dtype enhanced to provide C char, short, int, long and float types in addition to C double type. nonint_dtype introduced for non-integer numeric types. If you get no suitable method found or Invalid use of fused types, type cannot be specialized Cython errors, specify the specialization explicitly: ceygen.elemwise.add_vv[double](np.array(...), np.array(...)). This unfortunately slows down compilation and makes resulting modules bigger, but doesn’t affect performance and makes Ceygen more generic.
power_vs() and power_ms() functions were added to the reductions module.

reductions module was added with vector, matrix, row-wise and column-wise sums.
Simple benchmarks for many functions have been added, define BENCHMARK or BENCHMARK_NUMPY environment variable during test execution to run them; define SAVE environment variable to save timings into .pickle files that can be visualized by support/visualize_stats.py.
Added code paths optimized for C-contiguous and F-contiguous matrices and vectors using fancy C++ dispatching code. Rougly 40% speed gains in core.dot_mm() (for common matrix sizes), 300% gains for core.dot_mv() and core.dot_vm() starting with 16*16, 30% gains for vector-vector operations and slight gains at other places.
Internal Ceygen .pxd files (e.g. eigen_cython.pxd) are no longer installed.
-fopenmp is now added by default to build_ext cflags and ldflags to enable parellelising core.dot_mm() in Eigen; speedups are noticeable for matrices 64*64 and bigger. Can be easily disabled.
dtype.vector() and dtype.matrix() convenience functions added; their usage in other modules leads to speedups because it circumvents Cython shortcoming.
core.set_is_malloc_allowed() added to aid in debugging and tests.