Open dionhaefner opened 4 years ago
I guess you can call bohrium._bh.get_data_pointer(bh_ary, copy2host=False, allocate=True)
, which will return a point (as a Python integter) to the GPU memory. If the data is on main memory, it will be copied to the GPU before returning.
Is there a way to force Bohrium to copy an array to GPU?
I want to benchmark Bohrium against some other libraries, so I do something like this:
But on GPU, the data is not copied during
prepare_inputs
, only when the data is accessed duringrun_benchmark
. The timings I get for Bohrium are thus higher than for other libraries, which copy the data right away.I found that the following works:
But it feels a bit stupid.