Hi, I have a Tegra PX2 and:
“Linux ettin-a 4.4.15-rt25-tegra #1 SMP PREEMPT RT Sat Feb 18 09:04:00 PST 2017 aarch64 aarch64 aarch64 GNU/Linux”
and when I try to make the CUDA 8.0 sample cuSolverSp_LinearSolver, I get an error due to the fact that Makefile has a check:
This sample is not supported on aarch64
ifeq ((TARGET_ARCH),aarch64)
(info >>> WARNING - cuSolverSp_LinearSolver is not supported on aarch64 - waiving sample <<<)
SAMPLE_ENABLED := 0
endif
However if I set SAMPLE_ENABLED, I am able to compile cuSolverSp_LinearSolver and even run the sample for the Cholesky solving which is what interests me.
- Why is this limitation for ARM 64b? Can I ignore it for Cholesky solver?
- Why is the internal to Parker GPU so much slower than the external GPU?
CPU and internal to Parker GPU (./cuSolverSp_LinearSolver -device=1):
timing chol: CPU = 0.197825 sec , GPU = 2.152372 sec
CPU and external to Parker GPU (./cuSolverSp_LinearSolver):
timing chol: CPU = 0.197933 sec , GPU = 0.147527 sec
Thank you and regards,
Pepe