Buffer overflows and barrier divergence in OpenCL-D TQ and TQH

[mn416]

First, thanks for this benchmark suite. I found it very straightforward to use.

A few issues:

Oclgrind reports a buffer overflow in the OpenCL-D TQ benchmark. The value *next in the TaskQueue_gpu kernel is used to index queue without knowing that it is in bounds, both before the while loop is entered, and at the end of the while loop.

Oclgrind also reports a barrier synchronisation issue in the while loop: shared local data t is read and then modified without a barrier in between, forming a race condition between work items in a workgroup.

I believe there are similar issues in the TQH benchmark.

Oclgrind also reported synchronisation issues in other benchmarks, which may be worth investigating.