Audio: Optimize IIR performance

[singalsu]

This patch optimizes the buffer copying and output scaling to
other format than int32_t. The main saving is from not using
read/write frag buffer access functions for every sample.

The saving of processing cycles consumption varies per platform
but a second order IIR stereo EQ on TGL shows 43% improvement for
average copy() duration.

Signed-off-by: Seppo Ingalsuo seppo.ingalsuo@linux.intel.com

[singalsu]

Note: There's still plenty of read/write frag buffer access remaining elsewhere. I will address them later in another PR.

[lgirdwood]

These should all be static inline in the header.

[singalsu]

Sure, need to think how to make in a nice way the generic vs. hifi3 headers. Here I was hoping compiler does the inline but it's of course not guaranteed.

[lgirdwood]

not when this is an archive. you can do

/* the header file */
#if USE_THE_STATIC_INLINE_VERSION
static inline int16_t iir_df2t_s16(struct iir_state_df2t *iir, int16_t x)
{
	return sat_int16(Q_SHIFT_RND(iir_df2t(iir, ((int32_t)x) << 16), 31, 15));
}
#else
/* just declare func */
int16_t iir_df2t_s16(struct iir_state_df2t *iir, int16_t x);
#endif  

[singalsu]

@lgirdwood I'm testing approach to have in iir_df2t.h next code chunk.

/* Inline functions with or without HiFi3 intrinsics */
#if IIR_HIFI3
#include "iir_df2t_hifi3.h"
#else
#include "iir_df2t_generic.h"
#endif

The inline functions e.g. in iir_df2t_hifi3.h are like:

static inline int16_t iir_df2t_s16(struct iir_state_df2t *iir, int16_t x)
{
	ae_f32x2 y = iir_df2t(iir, ((int32_t)x) << 16);

	return AE_ROUND16X4F32SSYM(y, y);
}

Is such OK?

[lgirdwood]

should be , but what is the iir_df2t definition ?

[singalsu]

It's https://github.com/thesofproject/sof/blob/main/src/math/iir_df2t_generic.c or https://github.com/thesofproject/sof/blob/main/src/math/iir_df2t_hifi3.c depending on build.

[keyonjie]

Looks nice improvement to me, thanks a lot @singalsu

[lyakh]

A nice optimisation! Not very obvious either - this mostly just seems to remove checking for buffer wrapping on each sample access! But yes, making some of those one-line wrapper functions inline would help a bit more!

[singalsu]

Here's a mistake. This instruction does not saturate to 24 bits. Need in addition to shift left with saturation by 8, shift right by 8. I think my cmocka test for 24 bit treated S24_LE as S32_LE so it can't detect overflow in 24 bits.

[kv2019i]

Nice!