0 Members and 1 Guest are viewing this topic.
PCIeSpeedTest provides a synthetic benchmark to measure CPU->GPU and GPU->CPU transfer speed using ATI CAL. This is done by allocating a local resource on the GPU using calResAllocLocal2D() and allocating a remote resource on the CPU using calResAllocLocal2D(). The remote resource allocation is done in uncached memory space to optimize transfer performance. calMemCopy() is used to issue the data transfer.Allocation sizes start at 16 bytes and is incremented by 2 for every data point. Allocation continues until the physical memory size is reached on the device or the allocation routine returns unsuccessfully. It is expected that you will not be able to allocate the entire memory space as various system allocations exist on the device that are not visible to the CAL user.To amortize away moment to moment system variations, PCIeSpeedTest queues 100 calMemCopy()s back to back before checking on the completion event.PCIeSpeedTest_random is also provided to help generate a complete performance view by generating random data size points. It will round robin through all of the available compatible GPUs in the system. It will also use variable binning to spread the data points across the entire data size range. By default, PCIeSpeedTest_random will generate 100 data points for each GPU.