Ut-oh.  You have run into an application where even your streamlined C code is too slow and you need to break the code down to assembly.  If you want to get decent optimization with the VLIW (Very Long Instruction Word) for your DSP, you have to go to hand assembly.  Don't let me mislead, this is within the experiences but it is a slow process.  It requires a detailed knowledge of instruction times, bus characteristics which control numbers of clock cycles for stores and fetches across different busses and DMA channels.

Experience includes assembler in C6000 series TI DSPs, 6502, 8051/52, Toshiba micros, Intel PXA 255.