solve merge issue with main (58c015d2) · Commits · SA4 / Audio / IVAS BASOP

.gitlab-ci.yml

+2 −2

Original line number	Diff line number	Diff line
		@@ -506,7 +506,7 @@ stages:
		.ivas-pytest-sanitizers-anchor: &ivas-pytest-sanitizers-anchor
		stage: test
		needs: ["build-codec-linux-make"]
		timeout: "300 minutes"
		timeout: "420 minutes"
		rules:
		- if: $CI_PIPELINE_SOURCE == 'push'
		when: never
		@@ -522,7 +522,7 @@ stages:
		- *build-reference-and-dut-binaries
		- make clean
		- make -j CLANG=$CLANG_NUM
		- if [[ $CLANG_NUM == 3 ]]; then export UBSAN_OPTIONS="suppressions=scripts/ubsan.supp,report_error_type=1"; fi
		- if [[ $CLANG_NUM == 3 ]]; then export UBSAN_OPTIONS="suppressions=scripts/ubsan_basop.supp,report_error_type=1"; fi
		- testcase_timeout=$TESTCASE_TIMEOUT_LTV_SANITIZERS
		- python3 -m pytest $TEST_SUITE -v --tb=no --update_ref 1 --html=report.html --self-contained-html --junit-xml=report-junit.xml --testcase_timeout $testcase_timeout --ref_encoder_path $DUT_ENCODER_PATH --ref_decoder_path $DUT_DECODER_PATH
		artifacts:

lib_com/cldfb.c

+1 −1

Original line number	Diff line number	Diff line
		@@ -1213,7 +1213,7 @@ void cldfbSynthesis_ivas_fx(
		/cplxMult(&iBuffer[2i], &iBuffer[2i+1],-imagBuffer[k][2i], imagBuffer[k][M1-1-2i], rot_vctr_re[i], rot_vctr_im[i]);/
		iBuffer_fx[2 * i] = Msub_32_32( Mpy_32_32( ( L_negate( imagBuffer_fx[k][2 * i] ) ), rot_vctr_re_fx[i] ), imagBuffer_fx[k][( M1 - 1 ) - ( i * 2 )], rot_vctr_im_fx[i] ); // Qx
		move32();
		iBuffer_fx[2 * i + 1] = Madd_32_32( Mpy_32_32( ( L_negate( imagBuffer_fx[k][2 * i] ) ), rot_vctr_im_fx[i] ), imagBuffer_fx[k][( M1 - 1 ) - ( i * 2 )], rot_vctr_re_fx[i] ); // Qx
		iBuffer_fx[2 * i + 1] = Msub_32_32( Mpy_32_32( imagBuffer_fx[k][( M1 - 1 ) - ( i * 2 )], rot_vctr_re_fx[i] ), imagBuffer_fx[k][2 * i], rot_vctr_im_fx[i] ); // Qx
		move32();
		}

lib_com/ivas_stereo_ica_com_fx.c

+24 −38

Original line number	Diff line number	Diff line
		@@ -131,24 +131,23 @@ static void interpTargetChannel_fx(
		Word32 spread_factor2_fx;
		Word64 tempD1_fx, tempD2_fx;

		d = negate( sub( currShift, prevShift ) );
		IF( d >= 0 )
		d = sub( prevShift, currShift );

		IF( d == 0 )
		{
		/* this can happen in DFT->TD switching */
		return;
		}

		signShift = 1;
		move16();
		}
		ELSE

		if ( d < 0 )
		{
		signShift = -1;
		move16();
		}

		IF( d == 0 )
		{
		/* this can happen in DFT->TD switching */
		return;
		}

		N = L_shift_adapt;
		move16();
		Word32 *table_pointer = NULL;
		@@ -207,7 +206,7 @@ static void interpTargetChannel_fx(

		FOR( j = lim1; j <= lim2; j++ )
		{
		ptr2_fx[i] = L_add( Mpy_32_32( win_fx[j * INTERP_FACTOR1 - i], ptr1_fx[j] ), ptr2_fx[i] ); // qsynth
		ptr2_fx[i] = Madd_32_32( ptr2_fx[i], win_fx[j * INTERP_FACTOR1 - i], ptr1_fx[j] ); // qsynth
		move32();
		}
		}
		@@ -225,44 +224,31 @@ static void interpTargetChannel_fx(
		tempD1_fx = W_deposit32_l( table_D1_pointer[abs( d )] ); // Q35
		tempD2_fx = W_mult0_32_32( 3, table_D1_pointer[abs( d )] ); // Q35

		IF( EQ_16( signShift, 1 ) )
		{
		tempF1_fx = ONE_IN_Q12; // Q12
		move32();
		}
		ELSE
		{
		tempF1_fx = -ONE_IN_Q12; // Q12
		move32();
		}
		tempF1_fx = L_sub( imult3216( factor_fx, d ), tempF1_fx ); // Q12
		FOR( k = 0; k < sub( N, 1 ); k++ )
		{
		Word32 local = L_sub( W_extract_l( W_shr( W_mult0_32_32( tempF1_fx, spread_factor2_fx ), 31 ) ), ONE_IN_Q12 ); // Q12
		Word32 sign_local;
		IF( local > 0 )

		if ( EQ_16( signShift, 1 ) )
		{
		sign_local = 1;
		tempF1_fx = ONE_IN_Q12; // Q12
		move32();
		}
		ELSE

		tempF1_fx = L_sub( imult3216( factor_fx, d ), tempF1_fx ); // Q12
		FOR( k = 0; k < N - 1; k++ )
		{
		sign_local = -1;
		move32();
		}
		Word32 local_int = W_extract_l( W_shr( W_abs( local ), 12 ) ); // Q0
		Word32 local = Madd_32_32( -ONE_IN_Q12, tempF1_fx, spread_factor2_fx ); // Q12
		Word32 local_int = L_shr( local, 12 ); // Q0
		Word32 res_a1, res_a2, res_a3;
		Word32 res_b1, res_b2, res_b3;
		Word32 res_c1, res_c2, res_c3;
		Word32 res_d1, res_d2, res_d3;
		Word64 local_int_scaled;
		Word32 local_int_scaled;
		Word64 res_a, res_b, res_c, res_d;
		Word64 tempa, tempb;
		Word64 mult_a_D1, mult_b_D2;
		local_int = W_extract_l( W_mult0_32_32( sign_local, local_int ) ); // Q0
		local_int_scaled = W_deposit32_l( L_shl( local_int, 12 ) ); // Q12
		local_int_scaled = L_shl( local_int, 12 ); // Q12
		lim1 = extract_l( local_int ); // Q0
		IF( W_sub( local_int_scaled, local ) > 0 ) // Q21
		if ( L_sub( local_int_scaled, local ) > 0 ) // Q12
		{
		lim1 = sub( lim1, 1 ); // Q0
		}
		@@ -387,7 +373,7 @@ static void targetCh_AlignStereoDFT_fx(
		}
		FOR( i = 0; i < L_shift_adapt; i++ )
		{
		target_fx[i] = L_add( Mpy_32_32( alpha_fx, fadeInBuff_fx[i] ), Mpy_32_32( L_sub( ONE_IN_Q31, alpha_fx ), fadeOutBuff_fx[i] ) ); // qsynth
		target_fx[i] = Madd_32_32( Mpy_32_32( alpha_fx, fadeInBuff_fx[i] ), L_sub( ONE_IN_Q31, alpha_fx ), fadeOutBuff_fx[i] ); // qsynth
		move32();

		alpha_fx = L_add_sat( alpha_fx, winSlope_fx ); // Q31

lib_com/ivas_tools.c

+81 −43

Original line number	Diff line number	Diff line
		@@ -942,6 +942,12 @@ Word16 matrix_product_mant_exp_fx(
		Word16 *Zp_fx_e = out_e;
		Word16 row, col;
		Word16 x_idx, y_idx;
		Word64 temp;
		Word16 temp_e;
		Word16 prod_e = add( X_fx_e, Y_fx_e );

		Word16 max_exp = -31;
		move16();

		/* Processing */
		test();
		@@ -957,17 +963,28 @@ Word16 matrix_product_mant_exp_fx(
		{
		FOR( i = 0; i < colsX; ++i )
		{
		( *Zp_fx ) = 0;
		move32();
		( *Zp_fx_e ) = 0;
		move16();
		temp = 0;
		move64();

		FOR( k = 0; k < rowsX; ++k )
		{
		x_idx = add( k, imult1616( i, rowsX ) );
		y_idx = add( k, imult1616( j, rowsY ) );
		( Zp_fx ) = BASOP_Util_Add_Mant32Exp( Zp_fx, Zp_fx_e, Mpy_32_32( X_fx[x_idx], Y_fx[y_idx] ), add( X_fx_e, Y_fx_e ), Zp_fx_e ); /Q31 - Zp_fx_e*/
		move32();
		x_idx = k + i * rowsX;
		y_idx = k + j * rowsY;
		temp = W_mac_32_32( temp, X_fx[x_idx], Y_fx[y_idx] ); // X_fx_e + Y_fx_e
		}
		/* Maximize accumulated value to 32-bit */
		temp_e = W_norm( temp );
		temp = W_shl( temp, temp_e );
		if ( 0 == temp )
		{
		temp_e = prod_e;
		move16();
		}
		*Zp_fx_e = sub( prod_e, temp_e );
		move16();
		( *Zp_fx ) = W_extract_h( temp );
		move32();
		max_exp = s_max( max_exp, *Zp_fx_e ); // Find the max exp
		Zp_fx++;
		Zp_fx_e++;
		}
		@@ -987,17 +1004,27 @@ Word16 matrix_product_mant_exp_fx(
		{
		FOR( i = 0; i < rowsX; ++i )
		{
		( *Zp_fx ) = 0;
		move32();
		( *Zp_fx_e ) = 0;
		move16();
		temp = 0;
		move64();
		FOR( k = 0; k < colsX; ++k )
		{
		x_idx = add( i, imult1616( k, rowsX ) );
		y_idx = add( j, imult1616( k, rowsY ) );
		( Zp_fx ) = BASOP_Util_Add_Mant32Exp( Zp_fx, Zp_fx_e, Mpy_32_32( X_fx[x_idx], Y_fx[y_idx] ), add( X_fx_e, Y_fx_e ), Zp_fx_e ); /Q31 - Zp_fx_e*/
		move32();
		x_idx = i + k * rowsX;
		y_idx = j + k * rowsY;
		temp = W_mac_32_32( temp, X_fx[x_idx], Y_fx[y_idx] ); // X_fx_e + Y_fx_e
		}
		/* Maximize accumulated value to 32-bit */
		temp_e = W_norm( temp );
		temp = W_shl( temp, temp_e );
		if ( 0 == temp )
		{
		temp_e = prod_e;
		move16();
		}
		*Zp_fx_e = sub( prod_e, temp_e );
		move16();
		( *Zp_fx ) = W_extract_h( temp );
		move32();
		max_exp = s_max( max_exp, *Zp_fx_e ); // Find the max exp
		Zp_fx++;
		Zp_fx_e++;
		}
		@@ -1017,18 +1044,27 @@ Word16 matrix_product_mant_exp_fx(
		{
		FOR( i = 0; i < colsX; ++i )
		{
		( *Zp_fx ) = 0;
		move32();
		( *Zp_fx_e ) = 0;
		move16();
		temp = 0;
		move64();
		FOR( k = 0; k < colsX; ++k )
		{
		x_idx = add( k, imult1616( i, rowsX ) );
		y_idx = add( j, imult1616( k, rowsY ) );
		( Zp_fx ) = BASOP_Util_Add_Mant32Exp( Zp_fx, Zp_fx_e, Mpy_32_32( X_fx[x_idx], Y_fx[y_idx] ), add( X_fx_e, Y_fx_e ), Zp_fx_e ); /Q31 - Zp_fx_e*/
		move32();
		x_idx = k + i * rowsX;
		y_idx = j + k * rowsY;
		temp = W_mac_32_32( temp, X_fx[x_idx], Y_fx[y_idx] ); // X_fx_e + Y_fx_e
		}

		/* Maximize accumulated value to 32-bit */
		temp_e = W_norm( temp );
		temp = W_shl( temp, temp_e );
		if ( 0 == temp )
		{
		temp_e = prod_e;
		move16();
		}
		*Zp_fx_e = sub( prod_e, temp_e );
		move16();
		( *Zp_fx ) = W_extract_h( temp );
		move32();
		max_exp = s_max( max_exp, *Zp_fx_e ); // Find the max exp
		Zp_fx++;
		Zp_fx_e++;
		}
		@@ -1050,17 +1086,26 @@ Word16 matrix_product_mant_exp_fx(
		{
		FOR( i = 0; i < rowsX; ++i )
		{
		( *Zp_fx ) = 0;
		move32();
		( *Zp_fx_e ) = 0;
		move16();
		temp = 0;
		move64();
		FOR( k = 0; k < colsX; ++k )
		{
		x_idx = add( i, imult1616( k, rowsX ) );
		y_idx = add( k, imult1616( j, rowsY ) );
		( Zp_fx ) = BASOP_Util_Add_Mant32Exp( Zp_fx, Zp_fx_e, Mpy_32_32( X_fx[x_idx], Y_fx[y_idx] ), add( X_fx_e, Y_fx_e ), Zp_fx_e ); /Q31 - Zp_fx_e*/
		move32();
		x_idx = i + k * rowsX;
		y_idx = k + j * rowsY;
		temp = W_mac_32_32( temp, X_fx[x_idx], Y_fx[y_idx] ); // X_fx_e + Y_fx_e
		}
		/* Maximize accumulated value to 32-bit */
		temp_e = W_norm( temp );
		temp = W_shl( temp, temp_e );
		if ( 0 == temp )
		{
		temp_e = prod_e;
		}
		*Zp_fx_e = sub( prod_e, temp_e );
		move16();
		( *Zp_fx ) = W_extract_h( temp );
		move32();
		max_exp = s_max( max_exp, *Zp_fx_e ); // Find the max exp
		Zp_fx++;
		Zp_fx_e++;
		}
		@@ -1108,18 +1153,11 @@ Word16 matrix_product_mant_exp_fx(
		#endif
		}
		Zp_fx = Z_fx; /Q31 - Zp_fx_e/

		Zp_fx_e = out_e;
		Word16 max_exp = -31;
		move16();
		FOR( j = 0; j < row; ++j )
		{
		FOR( i = 0; i < col; ++i )
		{
		max_exp = s_max( max_exp, *Zp_fx_e );
		Zp_fx_e++;
		}
		}
		Zp_fx_e = out_e;


		*Z_fx_e = max_exp;
		move16();
		FOR( j = 0; j < row; ++j )

lib_com/options.h

+3 −1

Original line number	Diff line number	Diff line
		@@ -151,6 +151,8 @@
		#define FIX_ISSUE_1237_KEEP_EVS_BE /* VA: Fix to keep EVS bitexactness to 26.444 */
		#define FIX_ISSUE_1214 /* Ittiam: Fix for issue 1214: Energy leakage in IGF tiles for MDCT-stereo @64kbps SWB*/
		#define FIX_881_HILBERT_FILTER /* VA: improve the precision of the Hilbert filter to remove 2kHz unwanted tone */
		#endif
		#define FIX_ISSUE_1245 /* Ittiam: Fix for issue 1245: Basop Encoder: Audible noise for silent Stereo input DTX on @24.4 kbps, @32 kbps*/
		#define FIX_MINOR_SVD_WMOPS_MR1010X /* FhG: Minor WMOPS tuning, bit-exact to previous version, saves about 8.2 WMOPS for MR1010 */
		#define FIX_TUNE_IVAS_TOOLS_MATRIX_PRODUCT_WMOPS_MR1010X /* FhG: Major WMOPS tuning in HQ matrix multiplication, bit-exact to previous version, saves about 37 WMOPS for MR1010 */
		#define SVD_WMOPS_OPT /* Ittiam : SVD related optimizations */
		#endif