Hello ARM Community,
I have a pipeline of OpenCL kernels executions that are being applied one after another. The pattern is:
Output[i] = Kernel(Output[i-1])
There are about 150 enqueued kernel executions. The pipeline delivers correct results on Intel HD Graphics and NVidia Quadro. On Galaxy S7 (MALI T880) the results are correct only up to a certain kernel execution. The difference is not due to a floating point precision, but rather to a computation error. The interesting part is that until that kernel execution the results are absolutely identical with those of the other 2 GPUs. No shared or local memory is used. All allocated buffers are about 100 MB, where the maximum allocable size is 1/4 * 4GB. Synchronization can't be an issue neither, because clFinish(queue) is used before the result is being read.
Best Wishes,
James.