The sum of the data size partitioned into each accelerator and the expected stack size cannot exceed the accelerators local memory size (246 KB).
The data size for each accelerator can be calculated as follows:
sum = 0; for_each(IO_BUFFER) { sum += data_transfer_size; } sum += expected_stack_size; if(sum > (246 * 1024)) return error;
The data_transfer_size for each buffer is computed as follows (element_type is determined by the field labeled Element type: in the buffer dialog):