Audio: Multiband DRC: Use optimized 4th order IIR filter version

[singalsu]

The simpler 4th order hard-coded IIR version saves MCPS in in band split filter bank and emphasis/de-emphasis IIR filters.

[singalsu]

WIP - I'll see if I can further improve the HiFi4 and HiFi5 IIR versions.

[singalsu]

WIP - I'll see if I can further improve the HiFi4 and HiFi5 IIR versions.

Just for the record, I tried this change for the inner loop but it's in practical application a bit slower (1.35 MCPS, while a separate test code indicated 296 cycles vs. earlier 310 cycles) than this proposed version:

		/* Load data */
		AE_LA32X2X2_IP(delay_y2y1, delay_x2x1, data_r_align, delay_r);

		/* Load coefficients */
		AE_LA32X2X2_IP(coef_a2a1, coef_b2b1, coef_align, coefp);
		AE_LA32X2X2_IP(coef_b0shift, gain, coef_align, coefp);

		acc = AE_MULF32RA_HH(AE_MOVAD32_H(coef_b0shift), in);  /* acc = b0 * in */
		AE_MULAAFD32RA_HH_LL(acc, coef_a2a1, delay_y2y1); /* + a2 * y2 + a1 * y1 */
		AE_MULAAFD32RA_HH_LL(acc, coef_b2b1, delay_x2x1); /* + b2 * x2 + b1 * x1 */
		AE_PKSR32(delay_y2y1, acc, 1);		     /* y2 = y1, y1 = acc(q1.31) */
		delay_x2x1 = AE_SEL32_LL(delay_x2x1, in);   /* x2 = x1, x1 = in */

		/* Store data */
		AE_SA32X2X2_IP(delay_y2y1, delay_x2x1, data_w_align, delay_w);

		/* Apply gain */
		acc = AE_MULF32R_LL(AE_MOVAD32_H(gain), delay_y2y1);	/* acc = gain * y1 */
		acc = AE_SLAI64S(acc, 17);		/* Convert to Q17.47 */

		/* Apply biquad output shift right parameter and then
		 * round and saturate to 32 bits Q1.31.
		 */
		acc = AE_SRAA64(acc, AE_MOVAD32_L(coef_b0shift));
		in = AE_ROUND32F48SSYM(acc);

So the current version with 7 words coefficients set becomes the proposal. Padding it to 8 words for two 128 bits loads didn't improve. In addition it needed a separate new function to copy and align the existing format coefficients.

[johnylin76]

So the current version with 7 words coefficients set becomes the proposal. Padding it to 8 words for two 128 bits loads didn't improve. In addition it needed a separate new function to copy and align the existing format coefficients.

IIUC the coefficient pattern of EQIIR isn't changed by this migration, the current config blobs will still be compatible. Is that correct?
If that is the case then LGTM

[kv2019i]

Minor: @singalsu "The omitting of outer loop for parallel biquads and check for null coefficients and use of fixed loop count of two" very complex language is, comments Yoda.