I need to use a lot mathematic functions for my DSP with fixed point. So I decided to use C64x DSP Library in order to get high computation performance. And what do I get? A set of useless functions!
For example, lets have a look to DSP_dotprod function. This function just computes a sum of products two 16-bit vectors (something like that: z = x[0]*y[0] + x[1]*y[1] + .. etc. But it has 32-bit accumulator only! It's easy to calculate that 32-bit accumulator might be overflow for input vectors with 5 elemets only! It seems ridiculous to use 32-bit sum register in that case! It must be at least 40-bit (long long). The real vectors usualy have the length of 160...320 elements.
Could anybody comment that situation?