[Starpu-devel] Execute on 2 OpenCL devices

[Julian Holzmüller <h-julian@gmx.at>]

Hi, i have a problem with my implementation of jacobi equation.

I want to execute my program parallel on a CPU and GPU wich are 2 OpenCL devices.

Unfortunately it is only executed on one OpenCL device. The other Device is only waiting...

It looks like this (in this case only on CPU-Device): http://imageshack.us/a/img21/5926/fjrz.png

Here is my function where i submit:

int execute_on(uint32_t where, float *matrix, float *matrixTemp)

{

struct starpu_codelet cl;

starpu_data_handle_t matrixHandle, matrixTempHandle;

starpu_matrix_data_register(&matrixHandle, 0, (uintptr_t)matrix, N, N, N, sizeof(float));

starpu_matrix_data_register(&matrixTempHandle, 0, (uintptr_t)matrixTemp, N, N, N, sizeof(float));

starpu_codelet_init(&cl);

cl.where = where;

        cl.cpu_funcs[0] = cpu_codelet;

        cl.opencl_funcs[0] = opencl_codelet;

        cl.nbuffers = 2;

cl.modes[0] = STARPU_RW,

cl.modes[1] = STARPU_RW,

        cl.model = NULL;

struct starpu_task *task = starpu_task_create();

task->cl = &cl;

task->callback_func = NULL;

        task->handles[0] = matrixHandle;

        task->handles[1] = matrixTempHandle;

int ret = starpu_task_submit(task);

starpu_task_wait_for_all();

starpu_data_unregister(matrixHandle);

starpu_data_unregister(matrixTempHandle);

return 0;

}

my funktion call:

ret = starpu_opencl_load_opencl_from_file("jacobi.cl", &opencl_code, NULL);

STARPU_CHECK_RETURN_VALUE(ret, "starpu_opencl_load_opencl_from_file");

ret = execute_on(STARPU_OPENCL, matrix, matrixTemp);

i execute it with:

STARPU_OPENCL_ON_CPUS=1 STARPU_DISABLE_ASYNCHRONOUS_COPY=1 STARPU_DISABLE_ASYNCHRONOUS_OPENCL_COPY=1 STARPU_SCHED=dmda STARPU_GENERATE_TRACE=1 ./jacobi

What doing i wrong? How can i manage that the program will run parallel on both devices? The performance is not important at this point.