some more mods to FIX1072_SPEEDUP_ivas_dirac_dec_output_synthesis_process_slot (abc297aa) · Commits · SA4 / Audio / IVAS BASOP

lib_rend/ivas_dirac_output_synthesis_dec.c

+41 −15

Original line number	Diff line number	Diff line
		@@ -821,8 +821,6 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		Word16 k;
		IF( ch_idx != 0 )
		{
		push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop A <<<<<-\|" );
		;
		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;
		@@ -912,7 +910,7 @@ 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*/
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop A <<<<<-\|" );/

		push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop B <<<<<-\|" );
		#ifdef FIX1072_SPEEDUP_ivas_dirac_dec_output_synthesis_process_slot
		FOR( ; k < num_freq_bands; k++ )
		@@ -926,16 +924,20 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		}
		ELSE
		{
		Word16 diff_c_exp;
		Word16 diff_aab_exp;
		IF( reference_power[k + ( ch_idx + 1 ) * num_freq_bands] == 0 )
		{
		mpy_a_a_b = Mpy_32_32( a, a ); // Q = (h_dirac_output_synthesis_state->q_direct_responses + (15 - b_exp) - 15) + (h_dirac_output_synthesis_state->q_direct_responses) - 31
		mpy_diff_aab = Mpy_32_32( L_sub( L_shl( 1, q_diffuseness ), diffuseness[k] ), mpy_a_a_b ); // Q = 2*(h_dirac_output_synthesis_state->q_direct_responses) - b_exp - 31 + q_diffuseness -31
		mpy_diff_c = Mpy_32_32( diffuseness[k], c ); // Q = q_diffuseness - 4
		//q_diff_aab = add( h_dirac_output_synthesis_state->direct_responses_q, add( sub( h_dirac_output_synthesis_state->direct_responses_q, 31 ), sub( q_diffuseness, 31 ) ) );
		q_diff_aab = sub( add( h_dirac_output_synthesis_state->direct_responses_q, add( h_dirac_output_synthesis_state->direct_responses_q, q_diffuseness ) ), 62 );
		q_diff_c = sub( q_diffuseness, 4 );
		//q_diff_aab = sub( add( h_dirac_output_synthesis_state->direct_responses_q, add( h_dirac_output_synthesis_state->direct_responses_q, q_diffuseness ) ), 62 );
		diff_aab_exp = sub( 31 + 62, add( h_dirac_output_synthesis_state->direct_responses_q, add( h_dirac_output_synthesis_state->direct_responses_q, q_diffuseness ) ) );
		//q_diff_c = sub( q_diffuseness, 4 );
		diff_c_exp = sub( 31 + 4, q_diffuseness );

		sqr_inp = BASOP_Util_Add_Mant32Exp( mpy_diff_c, sub( 31, q_diff_c ), mpy_diff_aab, sub( 31, q_diff_aab ), &sqr_exp ); /q(31-sqr_exp)/
		sqr_inp = BASOP_Util_Add_Mant32Exp( mpy_diff_c, diff_c_exp, mpy_diff_aab, diff_aab_exp, &sqr_exp ); /q(31-sqr_exp)/
		}
		ELSE
		{
		@@ -945,16 +947,44 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		mpy_diff_aab = Mpy_32_32( L_sub( L_shl( 1, q_diffuseness ), diffuseness[k] ), mpy_a_a_b ); // Q = 2*(h_dirac_output_synthesis_state->q_direct_responses) - b_exp - 31 + q_diffuseness -31
		mpy_diff_c = Mpy_32_32( diffuseness[k], c ); // Q = q_diffuseness - 4
		//q_diff_aab = add( add(h_dirac_output_synthesis_state->direct_responses_q , sub( sub( 15, b_exp ), 15 )), add( sub( h_dirac_output_synthesis_state->direct_responses_q, 31 ), sub( q_diffuseness, 31 ) ) );
		q_diff_aab = add( sub( h_dirac_output_synthesis_state->direct_responses_q, b_exp ), ( sub( add( h_dirac_output_synthesis_state->direct_responses_q, q_diffuseness ), 62 ) ) );
		q_diff_c = sub( q_diffuseness, 4 );
		//q_diff_aab = add( sub( h_dirac_output_synthesis_state->direct_responses_q, b_exp ), ( sub( add( h_dirac_output_synthesis_state->direct_responses_q, q_diffuseness ), 62 ) ) );
		diff_aab_exp = sub( sub( add( sub( 31 + 62, h_dirac_output_synthesis_state->direct_responses_q ), b_exp ), h_dirac_output_synthesis_state->direct_responses_q ), q_diffuseness );
		//q_diff_c = sub( q_diffuseness, 4 );
		diff_c_exp = sub( 31 + 4, q_diffuseness );

		sqr_inp = BASOP_Util_Add_Mant32Exp( mpy_diff_c, sub( 31, q_diff_c ), mpy_diff_aab, sub( 31, q_diff_aab ), &sqr_exp ); /q(31-sqr_exp)/
		sqr_inp = BASOP_Util_Add_Mant32Exp( mpy_diff_c, diff_c_exp, mpy_diff_aab, diff_aab_exp, &sqr_exp ); /q(31-sqr_exp)/

		}
		}
		sqr = Sqrt32( sqr_inp, &sqr_exp ); /Q(31-sqr_exp)/
		sqr = L_shr( sqr, 2 ); /Q(31-sqr_exp)/
		}

		IF( h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + k] != 0 )
		{
		IF( LT_16( sub( 31, sqr_exp ), h_dirac_output_synthesis_state->q_cy_cross_dir_smooth ) )
		{
		h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + k] = L_shr( h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + k], sub( h_dirac_output_synthesis_state->q_cy_cross_dir_smooth, sub( 31, sqr_exp ) ) ); /h_dirac_output_synthesis_state->q_cy_cross_dir_smooth->Q( 31- sqr_exp )/
		move32();
		Q_temp_cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + k] = sub( 31, sqr_exp );
		move16();
		}
		ELSE
		{
		sqr = L_shr( sqr, sub( sub( 31, sqr_exp ), h_dirac_output_synthesis_state->q_cy_cross_dir_smooth ) ); /Q(31- sqr_exp)->h_dirac_output_synthesis_state->q_cy_cross_dir_smooth/
		Q_temp_cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + k] = h_dirac_output_synthesis_state->q_cy_cross_dir_smooth;
		move16();
		}
		h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + k] = L_add( h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + k], sqr ); /Q_temp_cy_cross_dir_smooth_fx[ch_idx num_freq_bands + k]*/
		move32();
		}
		ELSE
		{
		h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + k] = L_add( h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + k], sqr ); /Q(31- sqr_exp)/
		move32();
		Q_temp_cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + k] = sub( 31, sqr_exp );
		move16();
		}
		#else
		FOR( ; k < num_freq_bands; k++ )
		{
		@@ -1011,7 +1041,6 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		sqr = Sqrt32( sqr_inp, &sqr_exp ); /Q(31-sqr_exp)/
		sqr = L_shr( sqr, 2 ); /Q(31-sqr_exp)/
		}
		#endif

		IF( h_dirac_output_synthesis_state->cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + k] != 0 )
		{
		@@ -1038,12 +1067,13 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		Q_temp_cy_cross_dir_smooth_fx[ch_idx * num_freq_bands + k] = sub( 31, sqr_exp );
		move16();
		}
		#endif


		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop B <<<<<-\|" );/
		}
		ELSE
		{
		push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop C <<<<<-\|" );
		Word32 sqr_inp, mpy_diff, sqr;
		Word16 sqr_exp;
		/Diffuseness modellling nrg compensation/
		@@ -1081,8 +1111,6 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		move16();
		}
		}
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop C <<<<<<-\|" );/
		push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop D <<<<<<-\|" );
		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
		@@ -1116,8 +1144,6 @@ void ivas_dirac_dec_output_synthesis_process_slot_fx(
		move16();
		}
		}

		pop_wmops();/push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop D <<<<<-\|" );//
		}
		}
		pop_wmops(); /push_wmops( "(IDR) LOOP1 PSDs PATH3 B3.3 bigloop <<<<-\|" );//