Releases: intel/yask
Version 2.4
Kernel: Add ability to use temporal wave-front blocking across MPI ranks.
Compiler: Add ability to specify equation grouping using regular expressions.
Now requires g++ 4.9 or later.
Version 2.3
Added scratch-grids feature.
Version 2.2
API functions now throw C++ and Python exceptions instead of just terminating the program.
Version 2.1.5
Improves performance when domain sizes are not multiples of vector-cluster sizes.
Several other small fixes and improvements since v2.1.0.
Version 2.1
Added some code to src/stencils/Iso3dfdStencil.hpp that illustrates applying a simple Cerjan ABC via three 1D arrays. It adds about 11% runtime to this "iso3dfd' acoustic stencil using a 1024^3 problem size on KNL and only about 0.0001% more memory.
Version 2.0
Significant new features including arbitrary grid dimensions and APIs for C++ and Python.
Version 1.0
v1.0.0 Update README.md
Binary reorg
- Scripts and binaries have been moved into the 'bin' directory, and several have been renamed to start with 'yask' instead of 'stencil'. Similar changes will occur in the 'src' dir later.
- Turned off streaming stores by default on all architectures.
- Updated the docs to reflect the above changes.
- Run script (now called 'yask.sh') now creates a log file in the 'logs' dir. Useful for keeping a record of configs and perf.
Xeon tuning update
Update tuning on iso3dfd and awp stencils for Xeon platforms (BDW, etc.).
Sub-blocking
- Major new features include addition of sub-block hierarchy between blocks and vector-clusters.
- Added ability to provide default cmd-line options in the Makefile. This is much more expressive and flexible than the old way of setting a few macros.
- Updated the example recipes and results for KNL, BDW, and HSW architectures.
- Make it easier to copy settings from the stencil-tuner back into the Makefile.