Numerical Stability in consecutive matrix multiplication #3304
Description
Is there anybody that can shed some light on the implementation of consecutive matrix multiplication?
Since i was implementing a forward-backward algorithm with many matrix multiplied together, which could lead to a numerical problem. A common way of doing this is to just transfer the conventional vector dot operation to a "LogOfSumExp" process with max-subtraction trick involved , in which situation the elements in matrix are regarded as energy(log probability) instead of the original probability mass.
My question is that whether i can take advantage of some optimized cblas functions to make a new version of matrix multiplication utility function corresponding to those already in math_function.cpp ?
and if i just handcraft all this functions element by element , Do i have to suffer a severe efficiency loss ?