thinking about validating the input/output buffer

[psteinb]

in order to make gearshifft more versatile, we could think about comparing the input and the output buffer after a iFFT(FFT(input) roundtrip. This way gearshifft can not only be used for benchmarking but also for hardware checks on the memory subsystem. This is can be very useful.

[tdd11235813]

this is already done in benchmark_executor.hpp and benchmark_data.hpp.
One could argument, that an error of a single transform might be not visible if it becomes "corrected" on the inverse transform. For real hardware checks a second evaluation path executed on a different memory system must be considered, i.e. result of forward and inverse transform on device A must be compared with results gathered on device B.

Since your suggestion is already implemented in gearshifft, we could rephrase it to a feature request where gearshifft should provide an interface to get the results of each transform to enable comparison.

[psteinb]

ok, that sounds reasonable. for a first try, I would however not open the API more. I think an automated way to cross check the transform results would be enough for now after all repetitions are done. the error bound should be configurable though (through a compile flag IMHO is enough)

[tdd11235813]

ah I knew I forgot a code line: the error bound and the number of runs can be changed in application.hpp. At the moment these are template values, so required at compile-time.

I will come back to your feature request at a later time. I also noted to extend the README on how error checking works in gearshifft. The idea to use a gearshifft wiki becomes stronger ...