CUDA and OpenMP

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello!
I’m a new user of CUDA!!

I wrote a linear solver for the CPU and I parallelized it using OpenMP.
I wrote a version of my linear solver for the GPU using CUDA.

I need to solve two systems, Ax1 = b1 and Ax2 = b2.
I have an 8-Cores machine and my idea was to use 1-Core as Host for my CUDA solver and the others 7 Cores for solving the others system.
In this way I can use 7 parallel threads to solve the first system and simultaneously solve the second system on my GPU.

Unfortunately my application stops responding when I launch the computation and it looks like the CUDA solver doesn’t return the control to his host thread.

Is my idea feasible? Can I run in parallel my two solvers, one on the CPU and the other on the GPU?

Thanks in advance!!
Best,

dario

2

Yes, it is feasible.
You just need to be careful about using the same thread to initialize and drive the GPU solver ( for example, you could use the master thread).

3

Hello!

Thank you very much!! You’re perfectly right!!

In order to measure the performance I first allocate the memory both in the CPU and GPU and then I launch the multi-threaded computation… In this way nobody assures me that the thread that initializes and allocates the GPU is the same thread that drives my CUDA solver.

I just changed the code and works sweetly!!
Thank you!!
Best,

dario