How do you recommend deallocating device memory when using VexCL

[agerlach]

@ddemidov , you must be getting sick seeing all my questions here and on the odeint project. I am about 1 month away from finishing my dissertation and I truly appreciate all your help and patients.

As I previously mentioned, I am using odeint + vexcl to do parameters studies on ensembles of ODE's. The main portion of my code is in Matlab and I call odeint + vexcl through a mex interface. When I do a single call everything works perfectly, but I do encounter problems when I call the mex in a loop.

After around 8-9 mex calls, I find that no devices are available when creating a vex::Context. I check this with:

vex::Context ctx( vex::Filter::Type(CL_DEVICE_TYPE_GPU) );
if (!ctx) throw std::runtime_error("No devices available.");

By running

nvidia-smi

in the terminal I see that with each mex call I lose ~ 100MB in device RAM. Once I am out of RAM I encounter the failure mentioned. Interestingly, once I recompile the mex, the device RAM is freed.

So, how do you recommend freeing device memory allocated by

typedef vex::vector< float >    vector_type;
typedef vex::multivector< float, 5 > state_type;
...
vector_type input( ctx.queue() , input_host );
state_type X(ctx.queue(), n);

[ddemidov]

The memory associated with vex containers should in principle be deallocated when the containers are destroyed. Do you have to initialize vex::Context on every iteration? Could you initialize that only once? I have not worked with a mex interface, because I prefer to create a shared library and interact with it through a loadlibrary/calllib calls.

In case you absolutely have to repeatedly initialize vex::Context, you can try to call vex::purge_kernel_caches() before destroying the context. That could possibly help with the issue.

[agerlach]

Sorry for taking so long to respond to this. I got put on another project for the last week. vex::perge_kernel_caches() works.

I did not know about loadlibrary/calllib in Matlab until now. I'll look into that in the future. Thanks!

[ddemidov]

Ok, that means the memory was spent on creating new OpenCL contexts and caching compiled kernels. If your function looks like this:

vex::Context ctx(...);

// do the work

vex::purge_kernel_caches();

then you could try to make the instance of vex::Context static:

// Create context once:
static vex::Context ctx(...);

// do the work

// Keep cache of compiled kernels across function calls:
// vex::purge_kernel_caches();

This way you won't spend time on context initialization and recompilation of the same kernels over and over again.