Remove attempts to optimize normalized additions. (dc90a03f) · Commits · SA4 / Audio / IVAS BASOP

lib_dec/ivas_svd_dec.c

+5 −105

Original line number	Diff line number	Diff line
		@@ -66,8 +66,6 @@
		#if 1
		#define OPT_DIV
		#define OPT_DIV_NORM /* 5 dB SNR precision improvement */
		#define OPT_SUM /* Very little WMOPS savings */
		#define OPT_SUM2 /* Very little WMOPS savings */

		#define OPT_GIVENS
		#define OPT_GIVENS_INV
		@@ -1406,67 +1404,6 @@ Word32 BASOP_Util_Inv32(Word32 x, Word16 *px_e)
		}
		#endif

		#ifdef OPT_SUM
		static
		Word32 BASOP_Util_Accu_Mant32Exp /* o : normalized result mantissa */
		( Word32 a_m, /* i : Mantissa of 1st operand a */
		Word16 p_a_e, / i/o : Exponent of 1st operand a */
		Word32 b_m, /* i : Mantissa of 2nd operand b */
		Word16 b_e /* i : Exponent of 2nd operand b */
		)
		{
		Word32 L_tmp;
		Word16 shift, a_e = *p_a_e;

		/* Compare exponents: the difference is limited to +/- 30
		The Word32 mantissa of the operand with lower exponent is shifted right by the exponent difference.
		Then, the unshifted mantissa of the operand with the higher exponent is added. The addition result
		is normalized and the result represents the mantissa to return. The returned exponent takes into
		account all shift operations.
		*/

		#if 0
		if ( !a_m )
		a_e = add( b_e, 0 );
		#endif
		if ( !b_m )
		b_e = add( a_e, 0 );

		shift = sub( a_e, b_e );
		#if 0
		shift = s_max( -31, shift );
		shift = s_min( 31, shift );
		#endif
		if ( shift < 0 )
		{
		/* exponent of b is greater than exponent of a, shr a_m */
		a_m = L_shl( a_m, shift );
		}
		if ( shift > 0 )
		{
		/* exponent of a is greater than exponent of b */
		b_m = L_shr( b_m, shift );
		}
		a_e = add( s_max( a_e, b_e ), 1 );
		L_tmp = L_add( L_shr( a_m, 1 ), L_shr( b_m, 1 ) );
		#if 1
		shift = norm_l( L_tmp );
		if ( shift )
		L_tmp = L_shl( L_tmp, shift );
		#if 0
		if ( L_tmp == 0 )
		a_e = add( 0, 0 );
		#endif
		if ( L_tmp != 0 )
		a_e = sub( a_e, shift );
		#endif
		*p_a_e = a_e;

		return ( L_tmp );
		}

		#endif

		#ifdef IVAS_FLOAT_FIXED
		static void biDiagonalReductionLeft_fx(
		Word32 singularVectors[][MAX_OUTPUT_CHANNELS], /* exp(singularVectors_e) */
		@@ -1523,21 +1460,12 @@ move32();

		IF( LT_16( currChannel, nChannelsL ) ) /* i <= m */
		{
		#ifdef OPT_SUM2
		*sig_x_e = -31;
		move16();
		#endif

		idx = currChannel;
		move16();

		FOR( jCh = idx; jCh < nChannelsL; jCh++ ) /* nChannelsL */
		{
		#ifdef OPT_SUM2
		( sig_x ) = BASOP_Util_Accu_Mant32Exp( sig_x, sig_x_e, L_abs( singularVectors[jCh][currChannel] ), singularVectors_e ); / exp(sig_x_e) */
		#else
		( sig_x ) = BASOP_Util_Add_Mant32Exp( sig_x, sig_x_e, L_abs( singularVectors[jCh][currChannel] ), singularVectors_e, sig_x_e ); /* exp(sig_x_e) */
		#endif
		}

		IF( ( sig_x ) ) /(fabsf(sig_x) > EPSILON fabsf(sig_x)) { /
		@@ -1553,11 +1481,7 @@ IF( LT_16( currChannel, nChannelsL ) ) /* i <= m */
		#endif
		norm_x = 0;
		move32();
		#ifdef OPT_SUM2
		norm_x_e = -31;
		#else
		norm_x_e = 0;
		#endif
		move16();
		FOR( jCh = idx; jCh < nChannelsL; jCh++ ) /* nChannelsL */
		{
		@@ -1578,11 +1502,7 @@ IF( LT_16( currChannel, nChannelsL ) ) /* i <= m */
		move32();
		sing_exp[jCh] = add( sing_exp[jCh], sub( singularVectors_e, sig_x_e ) );
		move16();
		#ifdef OPT_SUM2
		norm_x = BASOP_Util_Accu_Mant32Exp( norm_x, &norm_x_e, Mpy_32_32( singularVectors[jCh][currChannel], singularVectors[jCh][currChannel] ), shl( sing_exp[jCh], 1 ) ); /* exp(norm_x_e) */
		#else
		norm_x = BASOP_Util_Add_Mant32Exp( norm_x, norm_x_e, Mpy_32_32( singularVectors[jCh][currChannel], singularVectors[jCh][currChannel] ), shl( sing_exp[jCh], 1 ), &norm_x_e ); /* exp(norm_x_e) */
		#endif
		}
		IF( GT_16( norm_x_e, 0 ) )
		{
		@@ -1628,19 +1548,11 @@ IF( LT_16( currChannel, nChannelsL ) ) /* i <= m */
		{
		norm_x = 0;
		move32();
		#ifdef OPT_SUM2
		norm_x_e = -31;
		#else
		norm_x_e = 0;
		#endif
		move16();
		FOR( jCh = idx; jCh < nChannelsL; jCh++ ) /* nChannelsL */
		{
		#ifdef OPT_SUM2
		norm_x = BASOP_Util_Accu_Mant32Exp( norm_x, &norm_x_e, Mpy_32_32( singularVectors[jCh][currChannel], singularVectors[jCh][iCh] ), add( sing_exp[jCh], singularVectors_e )); / exp(norm_x_e) */
		#else
		norm_x = BASOP_Util_Add_Mant32Exp( norm_x, norm_x_e, Mpy_32_32( singularVectors[jCh][currChannel], singularVectors[jCh][iCh] ), add( sing_exp[jCh], singularVectors_e ), &norm_x_e ); / exp(norm_x_e) */
		#endif
		}

		#ifndef OPT_DIV
		@@ -1821,18 +1733,9 @@ static void biDiagonalReductionRight_fx(
		{
		idx = add( currChannel, 1 ); /* Q0 */

		#ifdef OPT_SUM
		*sig_x_e = -31;
		move16();
		#endif

		FOR( jCh = idx; jCh < nChannelsC; jCh++ ) /* nChannelsC */
		{
		#ifdef OPT_SUM
		( sig_x ) = BASOP_Util_Accu_Mant32Exp( sig_x, sig_x_e, L_abs( singularVectors[currChannel][jCh] ), singularVectors_e ); / exp(sig_x_e) */
		#else
		( sig_x ) = BASOP_Util_Add_Mant32Exp( sig_x, sig_x_e, L_abs( singularVectors[currChannel][jCh] ), singularVectors_e, sig_x_e ); /* exp(sig_x_e) */
		#endif
		}

		IF( ( sig_x ) ) /(fabsf(sig_x) > EPSILON fabsf(sig_x)) { /
		@@ -1932,19 +1835,11 @@ static void biDiagonalReductionRight_fx(
		{
		norm_x = 0;
		move32();
		#ifdef OPT_SUM2
		norm_x_e = -31;
		#else
		norm_x_e = 0;
		#endif
		move16();
		FOR( jCh = idx; jCh < nChannelsC; jCh++ ) /* nChannelsC */
		{
		#ifdef OPT_SUM2
		norm_x = BASOP_Util_Accu_Mant32Exp( norm_x, &norm_x_e, Mpy_32_32( singularVectors[iCh][jCh], singularVectors[currChannel][jCh] ), add( singularVectors_e, sing_exp[jCh] ) ); / exp(sig_x_e) */
		#else
		norm_x = BASOP_Util_Add_Mant32Exp( norm_x, norm_x_e, Mpy_32_32( singularVectors[iCh][jCh], singularVectors[currChannel][jCh] ), add( singularVectors_e, sing_exp[jCh] ), &norm_x_e ); / exp(norm_x_e) */
		#endif
		}

		FOR( jCh = idx; jCh < nChannelsC; jCh++ ) /* nChannelsC */
		@@ -2280,8 +2175,13 @@ static void singularVectorsAccumulationRight_fx(

		FOR( iCh = nCh + 1; iCh < nChannelsC; iCh++ ) /* nChannelsC*/
		{
		#ifdef OPT_DIVno
		ratio_float = BASOP_Util_Divide3232_Scale( singularVectors_Left[nCh][iCh], maxWithSign_fx( singularVectors_Left[nCh][nCh + 1] ), &temp_exp1 ); /* exp(temp_exp1) */
		singularVectors_Right[iCh][nCh] = BASOP_Util_Divide3232_Scale( ratio_float, maxWithSign_fx( t_ii ), &sing_right_exp[iCh][nCh] ); /* exp(sing_right_exp + (temp_exp1 - secDiag_e) */
		#else
		ratio_float = BASOP_Util_Divide3232_Scale_cadence( singularVectors_Left[nCh][iCh], maxWithSign_fx( singularVectors_Left[nCh][nCh + 1] ), &temp_exp1 ); /* exp(temp_exp1) */
		singularVectors_Right[iCh][nCh] = BASOP_Util_Divide3232_Scale_cadence( ratio_float, maxWithSign_fx( t_ii ), &sing_right_exp[iCh][nCh] ); /* exp(sing_right_exp + (temp_exp1 - secDiag_e) */
		#endif
		move32();
		sing_right_exp[iCh][nCh] = add( sing_right_exp[iCh][nCh], sub( temp_exp1, secDiag_e ) );
		move16();