Hello all,
I am using the ezDSP C5515 kit and I am developing a speech processing algorithm. At some point I need to do some operations on Q13.18 and I am looking to use the mul32() function supplied by the TI DSPLIB.
From what I have seen from the application in the examples folder the code works for numbers in Q0.31 format. However, when the input has N fractional bits the output has 2*(N-16)+1 fractional bits which puts a very severe limit on precision. Through trial and error I found a work-around and where you need to shift the input to the left by {(N+1)-[2*(N-16)+1]}/2 in order to get an output with the desired N fractional bits. However, this puts a problem on the maximum value allowed in the input so as the shift and the multiplication will not overflow.
When I opened the code of mul32.asm I found out that it is based on cross-multiplication in order to be able to calculate a 32-bit fractional multiplication without a 64-bit register. I noticed also the example 5.9 in spru376a which has the same code. As I understand the principle is to multiply high_byte(X)*high_byte(Y)+high_byte(X)*low_byte(Y)+low_byte(X)*high_byte(Y)+low_byte(X)*low_byte(Y). However both in mul32.asm and in example 5.9 the last argument (low*low) is missing. Why is that? I see that high*high requires a shift. What shift is required for low*low? Will adding this argument make the method work for formats other than Q0.31?
I would appreciate some comments on this!
All the best,
Elias
