Originally Posted by Alexey Lukin
You guys go great lengths to prove that 32-bit floating point FIR filtering is inappropriate for SRC, but you fail to recognize a simple fact...
a typical optimizing compiler will do the summation in 64-bit double precision, either on FPU or on SSE2.
Please clarify the terms. Seeing that 64 bit = 32 bit double-precision. 64 bit double-precision = 128-bit. How many SHARCs you got in that tank?
Dr. Moorer was not against floating-point maths. He was against only using 32 bits. He wrote in his white paper on why 48 bits beats 32 that he'd recommend 64 bits (floating point) but that, at the time of writink, only MathLab was usink 64 bits - too "expensive" for practical audio applications. Also, the fixed-point maths was lettink him use idealized state variable filters - I believe, only for EDL strip equalization, rather than Melchoir's(?) SFC - than those which come automagically for floatink point, but which are not as precise.