Just a though;
The way I've understanded, the compression bases heavily on the
quantization. It's selected with two methods;
- The compressed bitstream with quantized values should be more
than the selected bitrate.
- The max value of quantized values in cant be over (8191 + 14)
The evaluation is done this order.
I cant see WHY the later have to be done by iterations.
Based on the original in quantize() & simplified;
...
step = pow(2.0, quantizerStepSize * 0.25);
dest[counter] = nint(pow(fabs(source[counter]) / step, 0.75) - 0.0946);
Why iterate the buffer with different quantizerStepSize -> step?
Just calculate source max value and use inverse function to calculate
quantizerStepSize.
The bitstream size checking is done with binary search, with function
that takes TOP / BOTTOM values and halfs the range for after calculating
bitstream. Better way to implement this is to have a estimated value
(based on the quant_init or even better, use the result from previous)
and start adding a step for given direction. When the sign changes,
half the step. The current method takes around 9 iterations of huffman
compressions, using the previous value as start and first step 8.0,
the iterations drop to 5 - 4. (I tried this one, at least it work ;-)
This idea can be used for better quality; quantizerStepSize is a double
value, but with the current implementation the iterations are done with
integer steps..
..Or have I misunderstood something critical?
- Juha Laukala
