> Ok, I must have confused the transforms :)
The main confusion is with the application note numbers.
At http://pentium.intel.com/drg/mmx/appnotes/ we can find:
ap533.htm with MPEG1 Audio Kernels with the synthesis sub-band filter
algorithm used
in mp3 decoding, that are very similar to the analysis sub-band filter
used in the encoding. The filter includes the IMDCT. The work to be done
is to employ the same optizations to the forward MDCT and analysis
sub-band filter bank.
Also, in ap555.htm we find an MMX optimized implementation of 16 bit
FFT. The title says that it is real FFT, but it is a complex FFT. The
job here is to find out if 16 bit FFT produces the same bit stream as
the float pointing calculations. I bet this is almost the case since the
FFT is used only to direct the psycho acoustic model. So small
diferences
would not decrease to much the quality.
ap557.htm has a dot product (like a convolution) and mvmult.htm has an
efficient vector/matrix multiply that could be used to speed up some
parts of the code.
And finally, for the speed paranoics, ap527.htm has the MMX
optimizations to separate the bits in the bitstream. Similar techiques
could be employed to optimize the huffman coding and bit assembly.
I think we should think about employing fixed point or even integer
arithmetic in some part of the code, since with MMX (or SIMD in general
we can take a 2 to 8 gain in speed).
MM
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
