some push/pop wmops added, also addedd FIX_1072_SPEEDUP_gainpanning and... (8d14bce5) · Commits · SA4 / Audio / IVAS BASOP

lib_com/ivas_dirac_com.c

+51 −0

Original line number	Diff line number	Diff line
		@@ -960,6 +960,7 @@ void computeDiffuseness_fixed(
		q_intensity = add( q_factor_intensity[0], min_q_shift2 );
		move16();

		push_wmops( "(IDR) LOOP1 DirACparams computeDiffuseness B <<-\|" );
		FOR( i = 0; i < DIRAC_NO_COL_AVG_DIFF; ++i )
		{
		/* Energy slow */
		@@ -967,6 +968,28 @@ void computeDiffuseness_fixed(

		q_tmp = add( q_factor_energy[i], min_q_shift1 );


		#ifdef FIX_1072_SPEEDUP_COMPUTEDIFUSENESSB
		Word16 shift_q = sub( q_tmp, q_ene );
		Word32 shiftEquiv;
		Word16 shift_qtotal;
		if( shift_q < 0 )
		{
		shiftEquiv = L_lshl( 0x80000000, shift_q );
		shift_qtotal = sub( min_q_shift1, 0 );
		}
		if( shift_q >= 0 )
		{
		shiftEquiv = L_add( 0x7FFFFFFF, 0 );
		shift_qtotal = sub( min_q_shift1, shift_q );
		}
		FOR( k = 0; k < num_freq_bands; k++ )
		{
		tmp = L_shl( p_tmp_c[k], shift_qtotal );
		energy_slow[k] = Madd_32_32_r( tmp, energy_slow[k], shiftEquiv );
		move32();
		}
		#else
		Word16 shift_q = sub( q_tmp, q_ene );
		IF( shift_q < 0 )
		{
		@@ -986,6 +1009,9 @@ void computeDiffuseness_fixed(
		move32();
		}
		}
		#endif



		q_ene = s_min( q_ene, q_tmp );

		@@ -993,6 +1019,28 @@ void computeDiffuseness_fixed(
		q_tmp = add( q_factor_intensity[i], min_q_shift2 );

		shift_q = sub( q_tmp, q_intensity );
		#ifdef FIX_1072_SPEEDUP_COMPUTEDIFUSENESSB
		if( shift_q >= 0 )
		{
		shiftEquiv = L_lshl( 0x7FFFFFFF, 0 );
		shift_qtotal = sub( min_q_shift2, shift_q );
		}
		if ( shift_q < 0 )
		{
		shiftEquiv = L_lshl( 0x80000000, shift_q );
		shift_qtotal = sub( min_q_shift2, 0 );
		}
		FOR( j = 0; j < DIRAC_NUM_DIMS; ++j )
		{
		p_tmp = buffer_intensity[j][i];
		FOR( k = 0; k < num_freq_bands; k++ )
		{
		tmp = L_shl( p_tmp[k], shift_qtotal );
		intensity_slow[j * num_freq_bands + k] = Madd_32_32_r( tmp, intensity_slow[j * num_freq_bands + k], shiftEquiv );
		move32();
		}
		}
		#else
		IF( shift_q > 0 )
		{
		FOR( j = 0; j < DIRAC_NUM_DIMS; ++j )
		@@ -1019,8 +1067,11 @@ void computeDiffuseness_fixed(
		}
		}
		}
		#endif

		q_intensity = s_min( q_intensity, q_tmp );
		}
		pop_wmops(); /push_wmops( "(IDR) LOOP1 DirACparams computeDiffuseness B <<-\|" );//

		min_q_shift1 = getScaleFactor32( intensity_slow, i_mult( DIRAC_NUM_DIMS, num_freq_bands ) );
		min_q_shift1 = sub( min_q_shift1, idiv1616( add( find_guarded_bits_fx( DIRAC_NUM_DIMS ), 1 ), 2 ) );

lib_com/options.h

+5 −1

Original line number	Diff line number	Diff line
		@@ -56,7 +56,7 @@

		#define SUPPORT_JBM_TRACEFILE /* Support for JBM tracefile, which is needed for 3GPP objective/subjective testing, but not relevant for real-world implementations */

		/#define WMOPS/ /* Activate complexity and memory counters */
		//#define WMOPS /* Activate complexity and memory counters */
		#ifdef WMOPS
		/#define WMOPS_PER_FRAME/ /* Output per-frame complexity (writes one float value per frame to the file "wmops_analysis") */
		/#define WMOPS_DETAIL/ /* Output detailed complexity printout for every function. Increases runtime overhead */
		@@ -153,3 +153,7 @@
		#define FIX_881_HILBERT_FILTER /* VA: improve the precision of the Hilbert filter to remove 2kHz unwanted tone */
		#endif
		#define FIX_MINOR_SVD_WMOPS_MR1010X /* FhG: Minor WMOPS tuning, bit-exact to previous version, saves about 8.2 WMOPS for MR1010 */


		#define FIX_1072_SPEEDUP_gainpanning /* FhG: WMOPS tuning, in development*/
		#define FIX_1072_SPEEDUP_COMPUTEDIFUSENESSB /* "-" */
		No newline at end of file

lib_dec/ivas_dirac_dec.c

+18 −5

Original line number	Diff line number	Diff line
		@@ -2231,7 +2231,7 @@ void ivas_dirac_dec_render_sf_fx(
		move16();
		Word16 tmp1;

		push_wmops( "ivas_dirac_dec_render" );
		push_wmops( "ivas_dirac_dec_render (IDR)" );

		/* Initialize aux buffers */
		hDirAC = st_ivas->hDirAC;
		@@ -2341,6 +2341,7 @@ void ivas_dirac_dec_render_sf_fx(
		}
		ELSE IF( !( EQ_16( st_ivas->ivas_format, SBA_FORMAT ) \|\| EQ_16( st_ivas->ivas_format, SBA_ISM_FORMAT ) ) )
		{
		push_wmops( "(IDR) SBA_FORMAT \| SBA_ISM_FORMAT" );
		Word16 outchannels;
		idx_lfe = 0;
		move16();
		@@ -2409,6 +2410,7 @@ void ivas_dirac_dec_render_sf_fx(
		}
		}
		}
		pop_wmops(); /push_wmops( "(IDR) SBA_FORMAT \| SBA_ISM_FORMAT" );/
		}

		size = imult1616( hDirACRend->num_outputs_dir, hSpatParamRendCom->num_freq_bands );
		@@ -2706,6 +2708,7 @@ void ivas_dirac_dec_render_sf_fx(
		}
		}

		push_wmops( "(IDR) LOOP1" );
		FOR( slot_idx = 0; slot_idx < hSpatParamRendCom->subframe_nbslots[subframe_idx]; slot_idx++ )
		{
		index_slot = add( slot_idx_start, slot_idx );
		@@ -2923,6 +2926,8 @@ void ivas_dirac_dec_render_sf_fx(
		move16();
		BREAK;
		default:
		pop_wmops(); /* push_wmops( "ivas_dirac_dec_render (IDR)" );*/
		pop_wmops(); /push_wmops( "(IDR) LOOP1");//
		return;
		}
		q_proto_direct_buffer[slot_idx] = hDirACRend->h_output_synthesis_psd_state.proto_direct_buffer_f_q;
		@@ -2932,6 +2937,7 @@ void ivas_dirac_dec_render_sf_fx(
		/-----------------------------------------------------------------
		* Compute DirAC parameters at decoder side
		-----------------------------------------------------------------/
		push_wmops( "(IDR) LOOP1 DirACparams \|" );
		IF( EQ_16( hDirAC->hConfig->dec_param_estim, TRUE ) )
		{
		Copy( &hSpatParamRendCom->azimuth[md_idx][hDirAC->hConfig->enc_param_start_band], &azimuth[hDirAC->hConfig->enc_param_start_band], sub( hSpatParamRendCom->num_freq_bands, hDirAC->hConfig->enc_param_start_band ) );
		@@ -2980,8 +2986,11 @@ void ivas_dirac_dec_render_sf_fx(
		hDirACRend->q_buffer_energy[index - 1] = DirAC_mem.reference_power_q;
		move16();

		push_wmops( "(IDR) LOOP1 DirACparams computeDiffuseness <-\|" );
		computeDiffuseness_fixed( hDirACRend->buffer_intensity_real_fx, hDirACRend->buffer_energy_fx, num_freq_bands, hSpatParamRendCom->diffuseness_vector_fx[md_idx], hDirACRend->q_buffer_intensity_real, hDirACRend->q_buffer_energy, &hSpatParamRendCom->q_diffuseness_vector );
		pop_wmops(); /push_wmops( "(IDR) LOOP1 DirACparams computeDiffuseness <-\|" );//
		}
		pop_wmops(); /* push_wmops( "(IDR) LOOP1 DirACparams \|" );*/

		/-----------------------------------------------------------------
		* frequency domain decorrelation
		@@ -3083,6 +3092,7 @@ void ivas_dirac_dec_render_sf_fx(
		}

		/Compute PSDs/
		push_wmops( "(IDR) LOOP1 PSDs \|" );
		h_dirac_output_synthesis_params = &( hDirACRend->h_output_synthesis_psd_params );
		h_dirac_output_synthesis_state = &( hDirACRend->h_output_synthesis_psd_state );
		num_channels_dir = hDirACRend->num_outputs_dir;
		@@ -3165,6 +3175,7 @@ void ivas_dirac_dec_render_sf_fx(
		}
		ELSE
		{
		push_wmops( "(IDR) LOOP1 PSDs PATH3 <-\|" );
		ivas_dirac_dec_output_synthesis_process_slot_fx( reference_power_fx,
		DirAC_mem.reference_power_q,
		p_onset_filter_fx,
		@@ -3182,6 +3193,7 @@ void ivas_dirac_dec_render_sf_fx(
		md_idx,
		hodirac_flag,
		hDirAC->hConfig->dec_param_estim );
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 <-\|" );//
		}

		IF( hDirAC->hConfig->dec_param_estim )
		@@ -3252,7 +3264,9 @@ void ivas_dirac_dec_render_sf_fx(
		v_add_fixed( reference_power_fx, reference_power_smooth_fx, reference_power_smooth_fx, hSpatParamRendCom->num_freq_bands, 1 );
		q_reference_power_smooth = sub( q_reference_power_smooth, 1 );
		}
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs \|" );/
		}
		pop_wmops(); /push_wmops( "(IDR) LOOP1" );/

		minimum_s( q_proto_direct_buffer, hSpatParamRendCom->subframe_nbslots[subframe_idx], &hDirACRend->h_output_synthesis_psd_state.proto_direct_buffer_f_q );
		IF( EQ_16( hDirACRend->synthesisConf, DIRAC_SYNTHESIS_GAIN_SHD ) )
		@@ -3581,7 +3595,6 @@ void ivas_dirac_dec_render_sf_fx(
		/-----------------------------------------------------------------
		* CLDFB synthesis (and binaural rendering)
		-----------------------------------------------------------------/

		index_slot = slot_idx_start_cldfb_synth;
		move16();

		@@ -3963,6 +3976,7 @@ void ivas_dirac_dec_render_sf_fx(
		}
		}


		hSpatParamRendCom->slots_rendered = add( hSpatParamRendCom->slots_rendered, hSpatParamRendCom->subframe_nbslots[subframe_idx] );
		move16();
		hSpatParamRendCom->subframes_rendered = add( hSpatParamRendCom->subframes_rendered, 1 );
		@@ -4077,8 +4091,7 @@ void ivas_dirac_dec_render_sf_fx(
		}
		}
		}

		pop_wmops();
		pop_wmops(); /push_wmops( "ivas_dirac_dec_render (IDR)" );/

		return;
		}

lib_rend/ivas_dirac_output_synthesis_dec.c

+86 −3

Original line number	Diff line number	Diff line
		@@ -710,11 +710,13 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		}
		}

		push_wmops( "(IDR) LOOP1 PSDs PATH3 B <<-\|" );
		test();
		IF( dec_param_estim == FALSE && hodirac_flag )
		{
		IF( EQ_16( hDirACRend->synthesisConf, DIRAC_SYNTHESIS_GAIN_SHD ) )
		{
		push_wmops( "(IDR) LOOP1 PSDs PATH3 B1<<<-\|" );
		v_multc_fixed( hSpatParamRendCom->energy_ratio1_fx[md_idx], -MAX_32 /-1 Q31/, aux_buf, num_freq_bands ); /* 30 + 31 - 31 -> 30 */
		v_addc_fixed( aux_buf, ONE_IN_Q30 /1 Q30/, aux_buf, num_freq_bands ); /30/
		Copy32( hSpatParamRendCom->energy_ratio1_fx[md_idx],
		@@ -737,19 +739,24 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		move16();
		h_dirac_output_synthesis_state->direct_power_factor_q = 30;
		move16();
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B1<<<-\|" );//
		}
		ELSE
		{
		push_wmops( "(IDR) LOOP1 PSDs PATH3 B2<<<-\|" );
		ivas_dirac_dec_compute_gain_factors_fx( num_freq_bands,
		hSpatParamRendCom->diffuseness_vector_fx[md_idx],
		h_dirac_output_synthesis_state->direct_power_factor_fx,
		h_dirac_output_synthesis_state->diffuse_power_factor_fx,
		&h_dirac_output_synthesis_state->direct_power_factor_q,
		&h_dirac_output_synthesis_state->diffuse_power_factor_q );
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B2<<<-\|" );/
		}
		}
		ELSE IF( EQ_16( dec_param_estim, TRUE ) )
		{
		push_wmops( "(IDR) LOOP1 PSDs PATH3 B3<<<-\|" );
		push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.1<<<<-\|" );
		/* compute direct responses */
		ivas_dirac_dec_compute_directional_responses_fx( hSpatParamRendCom,
		hDirACRend,
		@@ -764,7 +771,7 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		sh_rot_max_order,
		p_Rmat,
		hodirac_flag );

		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.1<<<<-\|" );/
		{
		IF( h_dirac_output_synthesis_state->direct_responses_square_fx )
		{
		@@ -811,12 +818,14 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		Q_temp_cy_cross_dir_smooth_fx[kk] = h_dirac_output_synthesis_state->q_cy_cross_dir_smooth;
		move16();
		}

		push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop <<<<-\|" );
		FOR( ch_idx = 0; ch_idx < s_min( 4, nchan_transport ); ch_idx++ )
		{
		Word16 k;
		IF( ch_idx != 0 )
		{
		push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop IF <<<<<-\|" );
		;
		Word32 a, c;
		Word16 b, b_exp, sqr_exp, q_diff_aab, q_diff_c;
		Word32 mpy_a_a_b, mpy_diff_c, mpy_diff_aab;
		@@ -906,6 +915,8 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		}
		}
		c = Madd_32_16( ONE_IN_Q27 /1 Q27/, L_sub( h_dirac_output_synthesis_params->diffuse_compensation_factor_fx, ONE_IN_Q27 /1 Q27/ ), 5461 ); /Diffuseness modellling nrg compensation/ /* 1.0 / 6.0 = 5461 in Q15/ /Q27*/

		push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop IF nfreqbds <<<<<<-\|" );
		FOR( ; k < num_freq_bands; k++ )
		{
		a = h_dirac_output_synthesis_state->direct_responses_fx[ch_idx * num_freq_bands + k]; // Q = h_dirac_output_synthesis_state->q_direct_responses
		@@ -987,9 +998,12 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		move16();
		}
		}
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop IF nfreqbds <<<<<<-\|" );/
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop IF <<<<<-\|" );/
		}
		ELSE
		{
		push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop ELSE <<<<<-\|" );
		Word32 sqr_inp, mpy_diff, sqr;
		Word16 sqr_exp;
		/Diffuseness modellling nrg compensation/
		@@ -1027,6 +1041,7 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		move16();
		}
		}
		push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop ELSE nfreqbds <<<<<<-\|" );
		FOR( ; k < num_freq_bands; k++ )
		{
		mpy_diff = Mpy_32_32( diffuseness[k], L_sub( h_dirac_output_synthesis_params->diffuse_compensation_factor_decorr_fx, ONE_IN_Q29 /1 Q29/ ) ); // Q = q_diffuseness - 1
		@@ -1060,8 +1075,12 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		move16();
		}
		}
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop ELSE nfreqbds <<<<<<-\|" );/
		pop_wmops();/push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop ELSE <<<<<-\|" );//
		}
		}
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop <<<<-\|" );//

		Word16 temp = MAX_16; /q0/
		move16();
		tmp16 = imult1616( num_freq_bands, num_channels_dir );
		@@ -1078,9 +1097,21 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		}
		free( Q_temp_cy_cross_dir_smooth_fx );
		/Directional gain (panning)/
		push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 gainpanning <<<<-\|" );

		Word16 temp_q = sub( add( h_dirac_output_synthesis_state->direct_power_factor_q, h_dirac_output_synthesis_state->direct_responses_q ), 31 );
		IF( LT_16( temp_q, h_dirac_output_synthesis_state->q_cy_cross_dir_smooth ) )
		{
		#ifdef FIX_1072_SPEEDUP_gainpanning /is there any difference in any bitstream?/
		Word16 temp_q1 = sub( temp_q, h_dirac_output_synthesis_state->q_cy_cross_dir_smooth );
		FOR( Word16 kk = 0; kk < tmp16; kk++ )
		{
		h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[kk] = L_shl( h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[kk], temp_q1 ); /h_dirac_output_synthesis_state->q_cy_cross_dir_smooth ->temp_q/
		move32();
		}
		h_dirac_output_synthesis_state->q_cy_cross_dir_smooth = temp_q;
		move16();
		#else
		FOR( Word16 kk = 0; kk < tmp16; kk++ )
		{
		h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[kk] = L_shl( h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[kk], sub( temp_q, h_dirac_output_synthesis_state->q_cy_cross_dir_smooth ) ); /h_dirac_output_synthesis_state->q_cy_cross_dir_smooth ->temp_q/
		@@ -1088,7 +1119,53 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		}
		h_dirac_output_synthesis_state->q_cy_cross_dir_smooth = temp_q;
		move16();
		#endif

		}
		#ifdef FIX_1072_SPEEDUP_gainpanning
		Word16 temp_q1 = sub( h_dirac_output_synthesis_state->q_cy_cross_dir_smooth, temp_q );
		FOR( ch_idx = s_min( 4, nchan_transport ); ch_idx < num_channels_dir; ch_idx++ )
		{
		IF( NE_16( temp_q, h_dirac_output_synthesis_state->q_cy_cross_dir_smooth ) )
		{
		Word16 i;
		Word32 aux;
		IF(temp_q1 < 0)
		{
		Word32 temp_q1_equiv = L_lshl( 0x80000000, temp_q1 );
		FOR( i = 0; i < num_freq_bands; i++ )
		{
		aux = Mpy_32_32( h_dirac_output_synthesis_state->direct_power_factor_fx[i], h_dirac_output_synthesis_state->direct_responses_fx[ch_idx * num_freq_bands + i] );
		h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + i] = Madd_32_32( h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + i] , aux, temp_q1_equiv );
		move32();
		}
		}
		ELSE
		{
		FOR( i = 0; i < num_freq_bands; i++ )
		{
		aux = Mpy_32_32( h_dirac_output_synthesis_state->direct_power_factor_fx[i], h_dirac_output_synthesis_state->direct_responses_fx[ch_idx * num_freq_bands + i] );
		aux = L_shl( aux, temp_q1 );
		h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + i] = L_add( h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + i], aux );
		move32();
		}
		}

		}
		ELSE
		{
		Word16 i;
		FOR( i = 0; i < num_freq_bands; i++ )
		{
		h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + i] = Madd_32_32( h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + i], h_dirac_output_synthesis_state->direct_power_factor_fx[i], h_dirac_output_synthesis_state->direct_responses_fx[ch_idx * num_freq_bands + i] );
		move32();
		}
		}


		}

		#else
		FOR( ch_idx = s_min( 4, nchan_transport ); ch_idx < num_channels_dir; ch_idx++ )
		{
		v_mult_fixed( h_dirac_output_synthesis_state->direct_power_factor_fx,
		@@ -1107,6 +1184,9 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		num_freq_bands, 0 ); /Q(h_dirac_output_synthesis_state->q_cy_cross_dir_smooth)/
		}

		#endif
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 gainpanning <<<<-\|" );/

		/Diffuse gain/
		FOR( ch_idx = s_min( 4, nchan_transport ); ch_idx < num_channels_diff; ch_idx++ )
		{
		@@ -1124,7 +1204,8 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		&h_dirac_output_synthesis_state->cy_auto_diff_smooth_fx[ch_idx * num_freq_bands_diff],
		num_freq_bands_diff, 0 ); /h_dirac_output_synthesis_state->q_cy_auto_diff_smooth/
		}

		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B3 <<-\|" );//
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B <<-\|" );//
		return;
		}
		ELSE
		@@ -1143,7 +1224,9 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		h_dirac_output_synthesis_state->direct_power_factor_q = 31;
		move16();
		}
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B3<<<-\|" );//
		}
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B <<-\|" );//

		diff_start_band = 0;
		move16();