Accelerator roadmap
This page is under construction. Following, additional functionality is planned for ILNumerics Accelerator:
- support more functions / function categories for low-level kernel building:
- low-level / elementary functions:
- complex & fcomplex element types
- boolean functions: any, all, lt / <, eq / ==, ...
- type conversion functions
- ensure exact same segment semantic as for ILMath / numpy handling of NaN, keeping of reduced dims, empty / scalar arrays, ...
higher-level functions:linear algebra (lapack), fftimplement toolbox functions
- low-level / elementary functions:
- enable more auto-optimizations:
- automatic (in)dependency detection in loops
- automatic loop transforms
better binary broadcasting segment switches- even more useful kernel code specializations
- enable / improve automatic mode
- improve cross device scheduling efficiency
- make recovering from buggy OpenCL GPU drivers more robust
- improve latency hiding
- file caching of kernels