Change #278721

Comments

avcodec/ac3dec: avoid coefficient truncation in float decoder
The float decoder dequantized spectral coefficients into an int32_t
buffer using mantissa >> exponent. This discarded the fractional part
before conversion to float and substantially raised the dither noise
floor for low-level coefficients.

Store decoded coefficients as INTFLOAT and apply 2^-exponent directly
in the floating-point path. Keep the fixed-point decoder unchanged, and
perform coupling, rematrixing and final scaling in the native type.

This also removes the now-unused fmtconvert dependency. Add a synthetic
FATE test that measures the corrected dither level. Existing PCM
references remain within one sample except fate-ac3-2.0, whose measured
maximum difference is two samples.

Fixes #23684.

Changed files

• configure
• libavcodec/ac3dec.c
• libavcodec/ac3dec.h
• tests/fate/ac3.mak