Wide range of values for observed for reference_power resulting in precision loss and high MLDs in ivas_dirac_dec_render processing

Test case: ltv-MASA 2TC at 96 kbps, 48kHz in, 48kHz out, MONO out

In the above test case, at around 7712 frame, the values inside ivas_dirac_dec_render processing for reference_power and reference_power_smooth buffers are having quite a large range with values ranging from 2.xx to 10^11 order. Since these buffers are in Word32, the Q factor was coming out as Q(-11) which was unable to handle these smaller values. Word64 implementation also doesn’t seem to be feasible because there are some multiplication and division operations happening for its calculation and usage down the line could not be accommodated in Word64. For further processing, reference_power_smooth_buffers are used to calculate gains which are modifying cldfb related buffers inside ivas_dirac_dec_output_synthesis_process_subframe_psd_ls_fx. So, due to this precision loss, the real and imaginary buffers are getting calculated incorrectly further leading to deviation in output and high MLDs.

Float code dumps:

Name	Value	Type
reference_power_smooth,60	0x00fb0ed8 {1.12123781e+11, 2.13012419e+10, 3.20389120e+09, 4.32607437e+09, 1.43548698e+09, 1.38198003e+09, ...}	float[60]
[0]	1.12E+11	float
[1]	2.13E+10	float
[2]	3.20E+09	float
[3]	4.33E+09	float
[4]	1.44E+09	float
[5]	1.38E+09	float
[6]	3.44E+09	float
[7]	835644864	float
[8]	78562392	float
[9]	126561176	float
[10]	40910900	float
[11]	25921034	float
[12]	33086312	float
[13]	29657100	float
[14]	18425232	float
[15]	12699304	float
[16]	8861048	float
[17]	13667408	float
[18]	12157381	float
[19]	3789297.25	float
[20]	9753415	float
[21]	8467907	float
[22]	3889820	float
[23]	8308914.5	float
[24]	16013249	float
[25]	3653945.5	float
[26]	1804171.13	float
[27]	604622.125	float
[28]	924181.375	float
[29]	704615.875	float
[30]	2040965.25	float
[31]	1561034.75	float
[32]	4265408	float
[33]	5864695	float
[34]	2893515.25	float
[35]	1526915.75	float
[36]	3238846.5	float
[37]	1193969.5	float
[38]	1449454	float
[39]	2964750.5	float
[40]	2883687.75	float
[41]	1898733.63	float
[42]	2171816.25	float
[43]	2113912	float
[44]	1890157	float
[45]	486523.563	float
[46]	282687.844	float
[47]	903418.125	float
[48]	241957	float
[49]	362024.063	float
[50]	17602.918	float
[51]	5.04815102	float
[52]	5.02293968	float
[53]	3.54574561	float
[54]	4.74774218	float
[55]	4.09441185	float
[56]	3.72185445	float
[57]	3.90770555	float
[58]	4.39547396	float
[59]	3.69693184	float

Fixed code dumps:

Name	Value	Type
reference_power_smooth_fx,60	0x013c10c8 {54692936, 10400828, 1563414, 2112119, 701430, 674762, 1680343, 407737, 38359, 61674, 20040, ...}	int[60]
[0]	54692936	int
[1]	10400828	int
[2]	1563414	int
[3]	2112119	int
[4]	701430	int
[5]	674762	int
[6]	1680343	int
[7]	407737	int
[8]	38359	int
[9]	61674	int
[10]	20040	int
[11]	12791	int
[12]	16132	int
[13]	14600	int
[14]	9001	int
[15]	6307	int
[16]	4406	int
[17]	6835	int
[18]	5965	int
[19]	1837	int
[20]	4867	int
[21]	4131	int
[22]	1881	int
[23]	4057	int
[24]	7805	int
[25]	1740	int
[26]	861	int
[27]	307	int
[28]	468	int
[29]	351	int
[30]	1014	int
[31]	752	int
[32]	2103	int
[33]	2877	int
[34]	1424	int
[35]	747	int
[36]	1616	int
[37]	606	int
[38]	703	int
[39]	1451	int
[40]	1411	int
[41]	943	int
[42]	1054	int
[43]	1051	int
[44]	935	int
[45]	240	int
[46]	142	int
[47]	448	int
[48]	108	int
[49]	179	int
[50]	9	int
[51]	0	int
[52]	0	int
[53]	0	int
[54]	0	int
[55]	0	int
[56]	0	int
[57]	0	int
[58]	2	int
[59]	2	int