Single precision preference about brian2cuda HOT 10 CLOSED

Just a quick comment about that: we actually have a preference in Brian2 exactly for this (core.default_float_dtype), but this has been broken for quite a while now because we did not consequently use this preference and instead just assumed double in quite a few places. Therefore, this preference will currently raise an error when you try to set it to single (see brian-team/brian2#417).

from brian2cuda.

Currently we are using double type everywhere except for random number generation with curand, where we generate floats... Any reason why this was implemented this way? If not, we should probably generate doubles by default?

from brian2cuda.

As long as brian-team/brian2#417 is not fixed, we could use our own preference and define our own c_data_type function and transform numpy.float64 and numpy.float32 either both to float or both to double, depending on the preference.

Additionally the SynapticPathway declaration on objects.cu should use the same preference (currently hard coded to double).

from brian2cuda.

We should use the brian-team/brian2@float32_support branch for our final benchmarks (potentially need to rebase on brian2 master), implement single precision from brian2CUDA side such that it works if the brian2 preference works and add a warning/error until brian2 has merged the float32_support branch.

from brian2cuda.

ensure that our templates always use type from user preference (including random number generator, sizeof() etc.)

from brian2cuda.

find a reasonable solution to get float32 branch from brian2 that allows syncing to the brian2 master updates easily

from brian2cuda.

lets discuss the failing tests situation when everything is implemented and (seems to be) working

from brian2cuda.

compare also visually with examples cpp vs cuda

from brian2cuda.

Just saw this comment on the timestep function implementation. Potential rounding error problems for single precision floats. We should think about that.

from brian2cuda.

seems like in brian-team/brian2#981 Marcel described by the remark that timestep and related quantities always should be doubles even in single precision mode, this seems very reasonable and to be followed

from brian2cuda.