import pyopencl as cl

kernel_source = """
__kernel void evaluate(){

  /* Variable declarations */

  size_t elementIndex = get_global_id(1);

  size_t i, j;

  double testValue[3];
  double trialValue[1];

  double shapeIntegral[3][1];

  for (i = 0; i < 3; ++i)
    for (j = 0; j < 1; ++j)
      shapeIntegral[i][j] = 0.0;



  testValue[0] = 1.0; testValue[1] = 1.0; testValue[2] = 1.0;
  trialValue[0] = 1.0;

    for (i = 0; i < 3; ++i)
      for (j = 0; j < 1; ++j){
        shapeIntegral[i][j] += testValue[i] * trialValue[j];
        }
  if (elementIndex == 0)
        printf("Value: %e\\n",shapeIntegral[0][0]);
  if (elementIndex == 0)
        printf("Value: %e\\n",shapeIntegral[1][0]);

}
"""

ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx, ctx.devices[0]);
prg = cl.Program(ctx, kernel_source)
prg.build()
kernel = prg.evaluate
kernel(queue, (1, 2), (1, 1))