Check sum with 2% error instead of abs error. #45
Conversation
This seems to help validate single precision runs where the sum is as close as we can expect for these large arrays.
|
This is extremely sensitive to not just array size, but also number of threads (for OpenMP). For example, running with the default array size with just 2 threads using the Intel compiler produces an error a little over 2%. For a single thread running with 1GB arrays (2^28) elements, I'm getting an error of 44%, which is worryingly large. That said, I'm not sure there's a better solution. We may get lower errors by tweaking the initial/scalar constants to keep the dot result from getting too large (where the coarseness of the floating point representation becomes an issue)? |
|
Oh wow, those are big errors. I wonder if we should normalise the dot product by the array size just in the checking script. Maybe that would net out a factor of the array size? |
|
Closing as this doesn't sufficiently solve Issue #20 |
This seems to help validate single precision runs where the sum
is as close as we can expect for these large arrays.
This fixes #20.