bigint-syrk-blas: split Q window, if it does not fit into the --maxSharedMemory limit #205
Comments
4 tasks
vasdommes
added a commit
that referenced
this issue
Mar 7, 2024
…e --maxSharedMemory limit TODO: update bigint_syrk/Readme.md
vasdommes
added a commit
that referenced
this issue
Mar 8, 2024
…e --maxSharedMemory limit TODO: update also bigint_syrk/Readme.md Changed two end-to-end tests: set low --maxSharedMemory to enforce Q window splitting In unit tests, we set different shared memory limits - to calculate Q = P^T P without splitting, with splitting only P window, or with splitting both P and Q. Also supported both uplo=UPPER and LOWER for syrk
vasdommes
added a commit
that referenced
this issue
Mar 8, 2024
…e --maxSharedMemory limit TODO: update also bigint_syrk/Readme.md Changed two end-to-end tests: set low --maxSharedMemory to enforce Q window splitting In unit tests, we set different shared memory limits - to calculate Q = P^T P without splitting, with splitting only P window, or with splitting both P and Q. Also supported both uplo=UPPER and LOWER for syrk
vasdommes
added a commit
that referenced
this issue
Mar 8, 2024
Added new parameter for reduce_scatter() to cover all cases. Fix #205 bigint-syrk-blas: split Q window, if it does not fit into the --maxSharedMemory limit TODO: update also bigint_syrk/Readme.md Changed two end-to-end tests: set low --maxSharedMemory to enforce Q window splitting In unit tests, we set different shared memory limits - to calculate Q = P^T P without splitting, with splitting only P window, or with splitting both P and Q. Also supported both uplo=UPPER and LOWER for syrk. Fixed reduce_scatter(): Old version synchronized only upper half always, but for off-diagonal blocks Q_IJ, we need to synchronize all.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
See details in parent issue #207
If total size for P (input) and Q (output) windows exceeds
--maxSharedMemory, we should split P window, see #204.If it doesn't help (i.e. size of Q exceeds
--maxSharedMemory), we have to also split Q (choosing a minimal split factor to reduce performance overhead) and calculate subblocks ofQ = P^T Pindependently.The text was updated successfully, but these errors were encountered: