Timothy B. Terriberry authored 17 years ago and Jean-Marc Valin committed 17 years ago

This eliminates an extra O(nm) lookups on decode, and reduces the rate control
 from O(nm^2) to O(nm), in addition to eliminating O(m) lookups on both encode
 and decode.
Although the interface is slightly more complex, the internal code is also
 simpler.

and mainly in the mid-range.

in energy quantisation when possible.

a hint that there has to be at least one iteration.

mix_pitch_and_residual()

find_spectral_pitch instead of using celt_rcp(celt_sqrt(x))

we're not going to use.

normalise_bands(). Also, another bunch of restrict qualifiers -- mainly
to tell the compiler there's no aliasing between the array bounds and the
data.

a division

first three bands (instead of total MDCT window energy)

the multiplications

remove unnecessary copying and some conditional branches.

X and Y.

to assume denominator is positive.

some calls to ec_enc_uint() by calls to ec_enc_bits() (saves divisions)

a use for that anymore now that the overlap can be less than one frame.

the more complex "n-best search" and replaced it with a greedy search

require a branch.

optimisations.