[RFC] Adding options for overriding base quantization matrices & adaptive quantizer parameters

[erikandre]

Hi jpegli team!

I've been working on tuning jpegli for some non-standard use cases, which can greatly benefit from being able to override some of the currently hard-coded constants defined in quant.cc. This includes:

• Setting a different base quant matrix for YCbCr, the global scale and choosing a scaling mode (linear/non-linear)
• Controlling the parameters for the adaptive quantizer, at least to set zero_bias_mul and zero_bias_offset, replacing the default value of 0.5f

I've got draft version of these changes, but wanted to check there's interest to include these in the official API before starting to work on polished push requests.

[erikandre]

Setting a different base quant matrix for YCbCr, the global scale and choosing a scaling mode (linear/non-linear)

Turns out this can be achieved using the existing jpegli_add_quantization_table() API as long as we account for the transformation of the table passed to the function. I assume that the (q + 50) / 100 transformation is done to stay compatible with the equivalent function in libjpeg?

Controlling the parameters for the adaptive quantizer, at least to set zero_bias_mul and zero_bias_offset, replacing the default value of 0.5f

I'll submit a PR for this change in the next couple of days. I understand that this is a pretty low level configuration that won't make sense for most users of jpegli, but hopefully it can still be accepted if the API is done in a good way

[szabadka]

Setting a different base quant matrix for YCbCr, the global scale and choosing a scaling mode (linear/non-linear)

Turns out this can be achieved using the existing jpegli_add_quantization_table() API as long as we account for the transformation of the table passed to the function. I assume that the (q + 50) / 100 transformation is done to stay compatible with the equivalent function in libjpeg?

Yes, we could not do anything differently in these API functions, since we wanted to be a drop-in replacement for libjpeg-turbo.

Controlling the parameters for the adaptive quantizer, at least to set zero_bias_mul and zero_bias_offset, replacing the default value of 0.5f

I'll submit a PR for this change in the next couple of days. I understand that this is a pretty low level configuration that won't make sense for most users of jpegli, but hopefully it can still be accepted if the API is done in a good way

Maybe it's not too late to start doing something like mozjpeg (and also libjxl) does. We could have some generic setter functions for int/float parameters and enum values as keys.

See e.g.
https://github.com/libjxl/libjxl/blob/main/lib/include/jxl/encode.h#L1306
https://github.com/mozilla/mozjpeg/blob/master/jcext.c#L115