Removing typechange tasks does not produce incorrect results

[abouteiller]

Original report by Nuria Losada (Bitbucket: nuriallv, GitHub: nuriallv).

---

When removing the type change tasks on those kernels using them, I would have expected results to be incorrect in the using master parsec as data won’t be correctly propagated to successors tasks (whether incorrectly unpacked or reusing the predecessor repo for multiple outs). However, results seems to be good enough. I’m attaching the modified JDF and these are the tests I’ve run which I believe are using those kernels (with the exception of zgetrf_qrf).

#dplasma_zgeqrf_mpi
mpirun -np 4 ./testing_zgeqrf -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgeqrf -P 2 -Q 4 -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgeqrf -P 4 -Q 2 -M 487 -N 283 -K 97 -t 7 -x -v=5

#dplasma_zgeqrf_dtd_mpi
mpirun -np 4 ./testing_zgeqrf_dtd -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgeqrf_dtd -P 2 -Q 4 -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgeqrf_dtd -P 4 -Q 2 -M 487 -N 283 -K 97 -t 7 -x -v=5

#dplasma_zgelqf_mpi
mpirun -np 4 ./testing_zgelqf -M 287 -N 383 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgelqf -P 2 -Q 4 -M 287 -N 383 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgelqf -P 4 -Q 2 -M 287 -N 383 -K 97 -t 7 -x -v=5

#dplasma_zunmqr_mpi
mpirun -np 4 ./testing_zunmqr -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zunmqr -P 2 -Q 4 -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zunmqr -P 4 -Q 2 -M 487 -N 283 -K 97 -t 7 -x -v=5

#dplasma_zunmlq_mpi
mpirun -np 4 ./testing_zunmlq -M 287 -N 383 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zunmlq -P 2 -Q 4 -M 287 -N 383 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zunmlq -P 4 -Q 2 -M 287 -N 383 -K 97 -t 7 -x -v=5

#dplasma_zgeqrf_hqr_mpi
mpirun -np 4 ./testing_zgeqrf_hqr -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgeqrf_hqr -P 2 -Q 4 -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgeqrf_hqr -P 4 -Q 2 -M 487 -N 283 -K 97 -t 7 -x -v=5

#dplasma_zgelqf_hqr_mpi
mpirun -np 4 ./testing_zgelqf_hqr -M 287 -N 383 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgelqf_hqr -P 2 -Q 4 -M 287 -N 383 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgelqf_hqr -P 4 -Q 2 -M 287 -N 383 -K 97 -t 7 -x -v=5

#dplasma_zunmqr_hqr_mpi
mpirun -np 4 ./testing_zunmqr_hqr -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zunmqr_hqr -P 2 -Q 4 -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zunmqr_hqr -P 4 -Q 2 -M 487 -N 283 -K 97 -t 7 -x -v=5

#dplasma_zunmlq_hqr_mpi
mpirun -np 4 ./testing_zunmlq_hqr -M 287 -N 383 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zunmlq_hqr -P 2 -Q 4 -M 287 -N 383 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zunmlq_hqr -P 4 -Q 2 -M 287 -N 383 -K 97 -t 7 -x -v=5

#dplasma_zgeqrf_systolic_mpi
mpirun -np 4 ./testing_zgeqrf_systolic -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgeqrf_systolic -P 2 -Q 4 -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgeqrf_systolic -P 4 -Q 2 -M 487 -N 283 -K 97 -t 7 -x -v=5

#dplasma_zgelqf_systolic_mpi
mpirun -np 4 ./testing_zgelqf_systolic -M 287 -N 383 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgelqf_systolic -P 2 -Q 4 -M 287 -N 383 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zgelqf_systolic -P 4 -Q 2 -M 287 -N 383 -K 97 -t 7 -x -v=5

#dplasma_zunmqr_systolic_mpi
mpirun -np 4 ./testing_zunmqr_systolic -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zunmqr_systolic -P 2 -Q 4 -M 487 -N 283 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zunmqr_systolic -P 4 -Q 2 -M 487 -N 283 -K 97 -t 7 -x -v=5

#dplasma_zunmlq_systolic_mpi
mpirun -np 4 ./testing_zunmlq_systolic -M 287 -N 383 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zunmlq_systolic -P 2 -Q 4 -M 287 -N 383 -K 97 -t 7 -x -v=5
mpirun -np 8 ./testing_zunmlq_systolic -P 4 -Q 2 -M 287 -N 383 -K 97 -t 7 -x -v=5

‌

[abouteiller]

did you try with larger matrices? That may indicate that the tests are not large enough to actually find bugs.