Issue in dft stereo at 32 kbps for residual spectral lines decoding
with 7b78b916 and this bitstream :
[ltv_STEREO_32k.bit.32000.zip](/uploads/3a8916cb3ae06b6781cbb3d652545456/ltv_STEREO_32k.bit.32000.zip)
and the command line : STEREO 32 bit syn
As show un the figure below, a major problem can be seen in the synthesis.
After some debugging, it seems the problem come from that part of the code (in IVAS_Stereo_dft_dec_fx.c):
```
IF( I != ECSQ_GLOBAL_GAIN_INDEX_ALL_ZERO )
{
ECSQ_init_instance_fx( &ecsq_inst, 0 /*dummy index*/, &range_uni_dec_state );
rc_uni_dec_init_fx( &range_uni_dec_state, bit_stream_side + *nb_bits, max_bits - *nb_bits );
hStereoDft->res_global_gain_fx = ECSQ_dequantize_gain_fx( I );
move32();
#ifdef DEBUGGING
dbgwrite_txt( (const float *) ( &hStereoDft->res_global_gain_fx ), 1, "fixed_res_global_gain.txt", NULL );
#endif
ecsq_inst.config_index = sub( shl( hStereoDft->res_cod_mode[k_offset], 1 ), 1 );
move16();
>>>>> ECSQ_decode( &ecsq_inst, hStereoDft->res_cod_line_max, dec );
>>>>> n_bits = rc_uni_dec_virtual_finish_fx( &range_uni_dec_state );
set_zero_fx( res_buf_fx, STEREO_DFT_N_8k );
FOR( Word16 c = 0; c < hStereoDft->res_cod_line_max; c++ )
{
dec[c] = shl_sat( dec[c], 8 );
move16();
}
ECSQ_dequantize_vector_fx( dec, hStereoDft->res_global_gain_fx, hStereoDft->res_cod_line_max, res_buf_fx );
#ifdef DEBUGGING
dbgwrite_txt( (const float *) res_buf_fx, hStereoDft->res_cod_line_max, "fixed_res_buf.txt", NULL );
#endif
}
```
ECSQ_decode( ..., ..., dec ); -> The last indice of dec is different between float and fix.
n_bits = rc_uni_dec_virtual_finish_fx( ... ); -> here it is not the same function as float, but using the "float" doesn't solve the problem as the n_bits output is the same (and wrong compared to float)
Given n_bits differs from float, it leads to a wrong bitstream decoding for the core decoder, which is likely the cause of the bad synthesis.
This is happening at frame 2769. @multrus, @norvell, I'm not sure who knows more about ECSQ coding, please feel free to tag someone else
issue