Computational optimization by reducing tf-resolution (b991f4de) · Commits · IVAS Codec Public Collaboration / IVAS Codec

lib_com/options.h

+1 −0

Original line number	Diff line number	Diff line
		@@ -159,6 +159,7 @@
		#endif

		#define NOKIA_ADAPTIVE_BINAURAL_PROTOS
		#define NOKIA_ADAPTIVE_BINAURAL_PROTOS_OPT /* enable adaptive binaural prototype complexity optimizations */

		/* ################## End DEVELOPMENT switches ######################### */
		/* clang-format on */

lib_rend/ivas_dirac_dec_binaural_functions.c

+127 −2

Original line number	Diff line number	Diff line
		@@ -40,6 +40,8 @@
		#include "ivas_cnst.h"
		#include "ivas_rom_binauralRenderer.h"
		#include "ivas_rom_rend.h"
		#include "ivas_rom_com.h"

		#ifdef DEBUGGING
		#include "debug.h"
		#endif
		@@ -55,7 +57,13 @@
		#define IVAS_TDET_DUCK_MULT_FAC_PARA_BIN_LOW_BR ( 3.0f )

		#ifdef NOKIA_ADAPTIVE_BINAURAL_PROTOS
		#ifdef NOKIA_ADAPTIVE_BINAURAL_PROTOS_OPT
		/* powf(0.95f, 4.0f) for sub-frame smoothing instead of CLDFB slot */
		#define ADAPT_HTPROTO_IIR_FAC 0.81450625f
		#else
		#define ADAPT_HTPROTO_IIR_FAC 0.95f
		#endif

		#define ADAPT_HTPROTO_ILD_LIM_DB0 1.0f
		#define ADAPT_HTPROTO_ILD_LIM_DB1 4.0f
		#define ADAPT_HTPROTO_ROT_LIM_0 0.4f
		@@ -77,8 +85,12 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices( Decoder_Struc
		static void ivas_dirac_dec_binaural_process_output( Decoder_Struct *st_ivas, float output_f[][L_FRAME48k], float inRe[][CLDFB_NO_COL_MAX][CLDFB_NO_CHANNELS_MAX], float inIm[][CLDFB_NO_COL_MAX][CLDFB_NO_CHANNELS_MAX], const int16_t max_band_decorr, const uint8_t numInputChannels, const uint8_t firstSlot, const uint8_t slotEnd );

		#ifdef NOKIA_ADAPTIVE_BINAURAL_PROTOS
		#ifdef NOKIA_ADAPTIVE_BINAURAL_PROTOS_OPT
		static void adaptTransportSignalsHeadtracked( HEAD_TRACK_DATA_HANDLE hHeadTrackData, float inIm[][CLDFB_NO_COL_MAX][CLDFB_NO_CHANNELS_MAX], float inRe[][CLDFB_NO_COL_MAX][CLDFB_NO_CHANNELS_MAX], const uint8_t firstSlot, const uint8_t slotEnd, float Rmat[3][3] );
		#else
		static void adaptTransportSignalsHeadtracked( HEAD_TRACK_DATA_HANDLE hHeadTrackData, float inIm[][CLDFB_NO_COL_MAX][CLDFB_NO_CHANNELS_MAX], float inRe[][CLDFB_NO_COL_MAX][CLDFB_NO_CHANNELS_MAX], const uint8_t firstSlot, const uint8_t slotEnd, const uint8_t nBins, float Rmat[3][3] );
		#endif
		#endif

		static void ivas_dirac_dec_binaural_check_and_switch_transports_headtracked( HEAD_TRACK_DATA_HANDLE hHeadTrackData, float inIm[][CLDFB_NO_COL_MAX][CLDFB_NO_CHANNELS_MAX], float inRe[][CLDFB_NO_COL_MAX][CLDFB_NO_CHANNELS_MAX], const uint8_t firstSlot, const uint8_t slotEnd, const uint8_t nBins, float Rmat[3][3] );

		@@ -575,7 +587,11 @@ static void ivas_dirac_dec_binaural_internal(
		if ( nchan_transport == 2 )
		{
		#ifdef NOKIA_ADAPTIVE_BINAURAL_PROTOS
		#ifdef NOKIA_ADAPTIVE_BINAURAL_PROTOS_OPT
		adaptTransportSignalsHeadtracked( st_ivas->hHeadTrackData, Cldfb_RealBuffer_in, Cldfb_ImagBuffer_in, firstSlot, slotEnd, Rmat );
		#else
		adaptTransportSignalsHeadtracked( st_ivas->hHeadTrackData, Cldfb_RealBuffer_in, Cldfb_ImagBuffer_in, firstSlot, slotEnd, nBins, Rmat );
		#endif
		#endif

		ivas_dirac_dec_binaural_check_and_switch_transports_headtracked( st_ivas->hHeadTrackData, Cldfb_ImagBuffer_in, Cldfb_RealBuffer_in, firstSlot, slotEnd, nBins, Rmat );
		@@ -1355,12 +1371,19 @@ static void adaptTransportSignalsHeadtracked(
		float inRe[][CLDFB_NO_COL_MAX][CLDFB_NO_CHANNELS_MAX],
		const uint8_t firstSlot,
		const uint8_t slotEnd,
		#ifndef NOKIA_ADAPTIVE_BINAURAL_PROTOS_OPT
		const uint8_t nBins,
		#endif
		float Rmat[3][3] )
		{
		int16_t slot, ch, bin, louderCh;
		#ifdef NOKIA_ADAPTIVE_BINAURAL_PROTOS_OPT
		float ILD, mono_factor_ILD, mono_factor_rotation, mono_factor, y_val, ene_proc, ene_target;
		uint8_t n_slots_per_sf, sf_idx, n_sf;
		#else
		float re[2], im[2], ILD, mono_factor_ILD, mono_factor_rotation, mono_factor, y_val;
		float proc_re[2], proc_im[2], sum_re, sum_im, ene_proc, ene_target, mf;
		#endif

		/* Determine head-orientation-based mono factor.
		Rmat[1][1] entry informs how close the ears are aligned according to transport signals. */
		@@ -1369,9 +1392,110 @@ static void adaptTransportSignalsHeadtracked(
		mono_factor_rotation = fmaxf( 0.0f, fminf( 1.0f, mono_factor_rotation ) );

		/* Adapt transport signals in frequency bands */
		#ifdef NOKIA_ADAPTIVE_BINAURAL_PROTOS_OPT
		/* optimization grouping CLDFB bins into MASA bands (they are readily available in ROM and suitable for the task) AND group CLDFB slots into sub-frames */
		n_slots_per_sf = CLDFB_NO_COL_MAX / MAX_PARAM_SPATIAL_SUBFRAMES;
		n_sf = ( slotEnd - firstSlot ) / n_slots_per_sf;

		for ( sf_idx = 0; sf_idx < n_sf; sf_idx++ )
		{
		float eqVal;
		uint8_t start_slot, stop_slot;
		int16_t band_idx, bin_lo, bin_hi;

		start_slot = firstSlot + sf_idx * n_slots_per_sf;
		stop_slot = start_slot + n_slots_per_sf;

		for ( band_idx = 0; band_idx < MASA_FREQUENCY_BANDS; band_idx++ )
		{
		float ch_nrg[2]; /* storage for input signal channel energies */
		bin_lo = MASA_band_grouping_24[band_idx];
		bin_hi = MASA_band_grouping_24[band_idx + 1];

		for ( ch = 0; ch < 2; ch++ )
		{
		ch_nrg[ch] = 0.0f;
		for ( slot = start_slot; slot < stop_slot; slot++ )
		{
		for ( bin = bin_lo; bin < bin_hi; bin++ )
		{
		ch_nrg[ch] += ( inRe[ch][slot][bin] * inRe[ch][slot][bin] ) + ( inIm[ch][slot][bin] * inIm[ch][slot][bin] );
		}
		}
		hHeadTrackData->chEneIIR[ch][band_idx] *= ADAPT_HTPROTO_IIR_FAC;
		hHeadTrackData->chEneIIR[ch][band_idx] += ( 1.0f - ADAPT_HTPROTO_IIR_FAC ) * ch_nrg[ch];
		}

		/* Determine ILD */
		ILD = fabsf( 10.0f * log10f( fmaxf( 1e-12f, hHeadTrackData->chEneIIR[0][band_idx] ) / fmaxf( 1e-12f, hHeadTrackData->chEneIIR[1][band_idx] ) ) );
		if ( hHeadTrackData->chEneIIR[1][band_idx] > hHeadTrackData->chEneIIR[0][band_idx] )
		{
		louderCh = 1;
		}
		else
		{
		louderCh = 0;
		}

		/* Determine ILD-based mono factor */
		mono_factor_ILD = ( ILD - ADAPT_HTPROTO_ILD_LIM_DB0 ) / ( ADAPT_HTPROTO_ILD_LIM_DB1 - ADAPT_HTPROTO_ILD_LIM_DB0 );
		mono_factor_ILD = fmaxf( 0.0f, fminf( 1.0f, mono_factor_ILD ) );

		/* Combine mono factors */
		mono_factor = mono_factor_ILD * mono_factor_rotation;

		/* Mix original audio and sum signal according to determined mono factor */
		for ( ch = 0; ch < 2; ch++ )
		{
		if ( ch != louderCh )
		{
		float band_nrg = 0.0f;

		for ( slot = start_slot; slot < stop_slot; slot++ )
		{
		for ( bin = bin_lo; bin < bin_hi; bin++ )
		{
		/* mono sum signal with the computed weight + rest from the original channel */
		inRe[ch][slot][bin] = mono_factor * ( inRe[0][slot][bin] + inRe[1][slot][bin] ) + ( 1.0f - mono_factor ) * inRe[ch][slot][bin];
		inIm[ch][slot][bin] = mono_factor * ( inIm[0][slot][bin] + inIm[1][slot][bin] ) + ( 1.0f - mono_factor ) * inIm[ch][slot][bin];
		band_nrg += ( inRe[ch][slot][bin] * inRe[ch][slot][bin] ) + ( inIm[ch][slot][bin] * inIm[ch][slot][bin] );
		}
		}
		hHeadTrackData->procChEneIIR[ch][band_idx] *= ADAPT_HTPROTO_IIR_FAC;
		hHeadTrackData->procChEneIIR[ch][band_idx] += ( 1.0f - ADAPT_HTPROTO_IIR_FAC ) * band_nrg;
		}
		else
		{
		/* processed signal is input. use the original channel, so no need to compute new signals or signal energy */
		hHeadTrackData->procChEneIIR[ch][band_idx] *= ADAPT_HTPROTO_IIR_FAC;
		hHeadTrackData->procChEneIIR[ch][band_idx] += ( 1.0f - ADAPT_HTPROTO_IIR_FAC ) * ch_nrg[ch];
		}
		}

		/* Equalize */
		ene_target = hHeadTrackData->chEneIIR[0][band_idx] + hHeadTrackData->chEneIIR[1][band_idx];
		ene_proc = hHeadTrackData->procChEneIIR[0][band_idx] + hHeadTrackData->procChEneIIR[1][band_idx];
		eqVal = fminf( 4.0f, sqrtf( ene_target / fmaxf( 1e-12f, ene_proc ) ) );

		for ( slot = start_slot; slot < stop_slot; slot++ )
		{
		for ( ch = 0; ch < 2; ch++ )
		{
		for ( bin = bin_lo; bin < bin_hi; bin++ )
		{
		inRe[ch][slot][bin] *= eqVal;
		inIm[ch][slot][bin] *= eqVal;
		}
		}
		}
		}
		}
		#else
		/* original contribution */
		for ( slot = firstSlot; slot < slotEnd; slot++ )
		{
		float eqVal[60];

		for ( bin = 0; bin < nBins; bin++ )
		{
		/* Determine channel energies */
		@@ -1421,6 +1545,7 @@ static void adaptTransportSignalsHeadtracked(
		}
		}
		}
		#endif

		return;
		}

lib_rend/ivas_stat_rend.h

+5 −0

Original line number	Diff line number	Diff line
		@@ -246,8 +246,13 @@ typedef struct ivas_binaural_head_track_struct
		float lrSwitchInterpVal;

		#ifdef NOKIA_ADAPTIVE_BINAURAL_PROTOS
		#ifdef NOKIA_ADAPTIVE_BINAURAL_PROTOS_OPT
		float chEneIIR[2][MASA_FREQUENCY_BANDS]; /* independent of the format. MASA bands are suitable for the task and readily available in ROM. */
		float procChEneIIR[2][MASA_FREQUENCY_BANDS];
		#else
		float chEneIIR[2][CLDFB_NO_CHANNELS_MAX];
		float procChEneIIR[2][CLDFB_NO_CHANNELS_MAX];
		#endif
		#endif

		int16_t shd_rot_max_order;