not crashing.... (5a74ba82) · Commits · SA4 / Audio / IVAS BASOP

lib_dec/ivas_svd_dec_fx.c

+59 −98

Original line number	Diff line number	Diff line
		@@ -76,9 +76,8 @@ static void biDiagonalReductionLeft_fx(
		const Word16 nChannelsL, /* Q0 */
		const Word16 nChannelsC, /* Q0 */
		const Word16 currChannel, /* Q0 */
		Word32 sig_x, / exp(sig_x_e) */
		Word16 *sig_x_e,
		Word32 g / Q31 */
		Word32 g, /* Q31 */
		Word16 g_e
		);

		static void biDiagonalReductionRight_fx(
		@@ -89,10 +88,9 @@ static void biDiagonalReductionRight_fx(
		const Word16 nChannelsL, /* Q0 */
		const Word16 nChannelsC, /* Q0 */
		const Word16 currChannel, /* Q0 */
		Word32 sig_x, / exp(sig_x_e) */
		Word16 *sig_x_e,
		Word32 g / Q31 */
		); // Q31
		Word32 g, / Q31 */
		Word16 *g_e
		);

		static void singularVectorsAccumulationLeft_fx(
		Word32 singularVectors_Left[][MAX_OUTPUT_CHANNELS], /* exp(singularVectors_e) as Input, Q31 as output */
		@@ -822,10 +820,9 @@ static void HouseholderReduction_fx(
		{
		Word16 nCh;
		// float g = 0.0f, sig_x = 0.0f;// to be removed
		Word32 g_fx = 0, sig_x_fx = 0;
		move32();
		Word32 g_fx = 0;
		Word16 g_e = 0;
		move32();
		Word16 sig_x_fx_e = 0;
		move16();

		Word16 iCh, jCh;
		@@ -842,8 +839,8 @@ static void HouseholderReduction_fx(
		/* Bidiagonal Reduction for every channel */
		FOR( nCh = 0; nCh < nChannelsC; nCh++ ) /* nChannelsC */
		{
		biDiagonalReductionLeft_fx( singularVectors_Left_fx, singularValues_fx, secDiag_fx, singularVectors_Left_fx_e, singularValues_fx_e, secDiag_fx_e, nChannelsL, nChannelsC, nCh, &sig_x_fx, &sig_x_fx_e, &g_fx );
		biDiagonalReductionRight_fx( singularVectors_Left_fx, secDiag_fx, singularVectors_Left_fx_e, secDiag_fx_e, nChannelsL, nChannelsC, nCh, &sig_x_fx, &sig_x_fx_e, &g_fx );
		biDiagonalReductionLeft_fx( singularVectors_Left_fx, singularValues_fx, secDiag_fx, singularVectors_Left_fx_e, singularValues_fx_e, secDiag_fx_e, nChannelsL, nChannelsC, nCh, g_fx, g_e );
		biDiagonalReductionRight_fx( singularVectors_Left_fx, secDiag_fx, singularVectors_Left_fx_e, secDiag_fx_e, nChannelsL, nChannelsC, nCh, &g_fx, &g_e );

		Word16 L_temp_e;
		Word32 L_temp = BASOP_Util_Add_Mant32Exp( L_abs( singularValues_fx[nCh] ), singularValues_fx_e[nCh], L_abs( secDiag_fx[nCh] ), secDiag_fx_e[nCh], &L_temp_e ); /* exp(L_temp_e) */
		@@ -881,9 +878,8 @@ static void biDiagonalReductionLeft_fx(
		const Word16 nChannelsL, /* Q0 */
		const Word16 nChannelsC, /* Q0 */
		const Word16 currChannel, /* Q0 */
		Word32 sig_x, / exp(sig_x_e) */
		Word16 *sig_x_e,
		Word32 g / Q31 */
		Word32 g, /* Q31 */
		Word16 g_e
		)
		{
		Word16 iCh, jCh, idx;
		@@ -892,68 +888,54 @@ static void biDiagonalReductionLeft_fx(
		Word32 L_temp;
		Word16 L_temp_e;

		secDiag[currChannel] = Mpy_32_32( sig_x, g ); /* exp(sig_x_e) */
		secDiag[currChannel] = g ; /* exp(sig_x_e) */
		move32();
		secDiag_e[currChannel] = *sig_x_e;
		secDiag_e[currChannel] = g_e;
		move16();

		/* Setting values to 0 */
		( *sig_x ) = 0;
		move32();
		( *g ) = 0;
		g = 0;
		move32();
		g_e = 0;
		move16();

		IF( LT_16( currChannel, nChannelsL ) ) /* i <= m */
		{
		idx = currChannel;
		move16();

		FOR( jCh = idx; jCh < nChannelsL; jCh++ ) /* nChannelsL */
		{
		( sig_x ) = BASOP_Util_Add_Mant32Exp( sig_x, sig_x_e, L_abs( singularVectors[jCh][currChannel] ), singularVectors2_e[jCh][currChannel], sig_x_e ); / exp(sig_x_e) */
		}

		IF( ( sig_x ) ) /(fabsf(sig_x) > EPSILON fabsf(sig_x)) { /
		{
		Word16 invVal_e;
		Word32 invVal;
		invVal = BASOP_Util_Divide3232_Scale_newton( MAXVAL_WORD32, maxWithSign_fx( *sig_x ), &invVal_e );
		norm_x = 0;
		move32();
		norm_x_e = 0;
		move16();
		FOR( jCh = idx; jCh < nChannelsL; jCh++ ) /* nChannelsL */
		{
		Word16 temp_e = norm_l( singularVectors[jCh][currChannel] );
		singularVectors[jCh][currChannel] = Mpy_32_32( L_shl( singularVectors[jCh][currChannel], temp_e ), invVal ); /* exp(sing_exp + (singularVectors_e - sig_x_e) */
		move32();
		singularVectors2_e[jCh][currChannel] = sub( add( invVal_e, sub( singularVectors2_e[jCh][currChannel], *sig_x_e ) ), temp_e );
		move16();
		norm_x = BASOP_Util_Add_Mant32Exp( norm_x, norm_x_e, Mpy_32_32( singularVectors[jCh][currChannel], singularVectors[jCh][currChannel] ), shl( singularVectors2_e[jCh][currChannel], 1 ), &norm_x_e ); /* exp(norm_x_e) */
		}
		IF( GT_16( norm_x_e, 0 ) )

		IF( norm_x ) /(fabsf(sig_x) > EPSILON * fabsf(sig_x)) { /
		{
		norm_x = MAX_32;
		move32();
		norm_x_e = 0;
		move16();
		}
		Word16 invVal_e;
		Word32 invVal;
		L_temp_e = norm_x_e;
		move16();
		L_temp = Sqrt32( norm_x, &L_temp_e );
		L_temp = L_shl_r( L_temp, L_temp_e ); // Q31
		// L_temp = L_shl_r( L_temp, L_temp_e ); // Q31
		//( *g ) = L_negate( GE_32( singularVectors[currChannel][idx], 0 ) ? L_temp : L_negate( L_temp ) );
		if ( singularVectors[currChannel][idx] >= 0 )
		{
		L_temp = L_negate( L_temp );
		}
		( *g ) = L_temp;
		g = L_temp;
		g_e = L_temp_e;
		move32();
		move16();

		r = BASOP_Util_Add_Mant32Exp( Mpy_32_32( ( g ), singularVectors[currChannel][idx] ), singularVectors2_e[currChannel][idx], -norm_x, norm_x_e, &r_e ); / exp(r_e) */
		singularVectors[currChannel][idx] = BASOP_Util_Add_Mant32Exp( singularVectors[currChannel][idx], singularVectors2_e[currChannel][idx], -( g ), 0, &singularVectors2_e[currChannel][idx] ); / sing_exp */
		r = BASOP_Util_Add_Mant32Exp( Mpy_32_32( ( g ), singularVectors[currChannel][idx] ), add( singularVectors2_e[currChannel][idx], g_e), -norm_x, norm_x_e, &r_e ); /* exp(r_e) */
		singularVectors[currChannel][idx] = BASOP_Util_Add_Mant32Exp( singularVectors[currChannel][idx], singularVectors2_e[currChannel][idx], -( g ), g_e, &singularVectors2_e[currChannel][idx] ); /* sing_exp */
		move32();

		invVal_e = r_e;
		move16();
		invVal = BASOP_Util_Divide3232_Scale_newton( MAXVAL_WORD32, maxWithSign_fx( r ), &invVal_e );

		FOR( iCh = currChannel + 1; iCh < nChannelsC; iCh++ ) /* nChannelsC */
		@@ -978,19 +960,12 @@ static void biDiagonalReductionLeft_fx(
		}


		FOR( jCh = idx; jCh < nChannelsL; jCh++ ) /* nChannelsL */
		{
		singularVectors[jCh][currChannel] = Mpy_32_32( singularVectors[jCh][currChannel], ( sig_x ) ); / sing_exp + sig_x_e */
		move32();
		singularVectors2_e[jCh][currChannel] = add( singularVectors2_e[jCh][currChannel], *sig_x_e );
		move16();
		}
		}

		// rescaling block
		singularValues[currChannel] = Mpy_32_32( ( sig_x ), ( g ) ); /* sig_x_e */
		singularValues[currChannel] = g;
		move32();
		singularValues_e[currChannel] = *sig_x_e;
		singularValues_e[currChannel] = g_e;
		move16();
		}

		@@ -1011,9 +986,8 @@ static void biDiagonalReductionRight_fx(
		const Word16 nChannelsL, /* Q0 */
		const Word16 nChannelsC, /* Q0 */
		const Word16 currChannel, /* Q0 */
		Word32 sig_x, / exp(sig_x_e) */
		Word16 *sig_x_e,
		Word32 g / Q31 */
		Word32 g, / Q31 */
		Word16 *g_e
		)
		{
		Word16 iCh, jCh, idx;
		@@ -1022,51 +996,39 @@ static void biDiagonalReductionRight_fx(
		Word32 L_temp;
		Word16 L_temp_e;

		/* Setting values to 0 */
		( *sig_x ) = 0;
		move32();
		( *g ) = 0;
		move32();
		( *g_e ) = 0;
		move16();

		IF( LT_16( currChannel, nChannelsL ) && NE_16( currChannel, sub( nChannelsC, 1 ) ) ) /* i <=m && i !=n */
		{
		idx = add( currChannel, 1 ); /* Q0 */

		FOR( jCh = idx; jCh < nChannelsC; jCh++ ) /* nChannelsC */
		{
		( sig_x ) = BASOP_Util_Add_Mant32Exp( sig_x, sig_x_e, L_abs( singularVectors[currChannel][jCh] ), singularVectors2_e[currChannel][jCh], sig_x_e ); / exp(sig_x_e) */
		}

		IF( ( sig_x ) ) /(fabsf(sig_x) > EPSILON fabsf(sig_x)) { /
		{
		norm_x = 0;
		move32();
		norm_x_e = 0;
		move16();

		Word16 invVal_e, temp_e;
		Word32 invVal;
		invVal = BASOP_Util_Divide3232_Scale_newton( MAXVAL_WORD32, maxWithSign_fx( *sig_x ), &invVal_e );
		FOR( jCh = idx; jCh < nChannelsC; jCh++ ) /nChannelsC /
		{
		temp_e = norm_l( singularVectors[currChannel][jCh] );
		singularVectors[currChannel][jCh] = Mpy_32_32( L_shl( singularVectors[currChannel][jCh], temp_e ), invVal ); /* exp(sing_exp + (singularVectors_e - sig_x_e) */
		move32();
		singularVectors2_e[currChannel][jCh] = add( sub( invVal_e, temp_e ), sub( singularVectors2_e[currChannel][jCh], *sig_x_e ) );
		move16();
		norm_x = BASOP_Util_Add_Mant32Exp( norm_x, norm_x_e, Mpy_32_32( singularVectors[currChannel][jCh], singularVectors[currChannel][jCh] ), shl( singularVectors2_e[currChannel][jCh], 1 ), &norm_x_e ); /* exp(norm_x_e) */
		}
		IF( GT_16( norm_x_e, 0 ) )

		IF( norm_x ) /(fabsf(sig_x) > EPSILON * fabsf(sig_x)) { /
		{
		norm_x = MAX_32;
		move32();
		norm_x_e = 0;
		move16();
		}

		Word16 invVal_e, temp_e;
		Word32 invVal;
		// IF( GT_16( norm_x_e, 0 ) )
		// {
		// norm_x = MAX_32;
		// move32();
		// norm_x_e = 0;
		// move16();
		// }
		L_temp_e = norm_x_e;
		move16();
		L_temp = Sqrt32( norm_x, &L_temp_e );
		L_temp = L_shl_r( L_temp, L_temp_e ); // Q31
		// L_temp = L_shl_r( L_temp, L_temp_e ); // Q31
		IF( singularVectors[currChannel][idx] >= 0 )
		{
		( g ) = L_negate( L_temp ); / exp(L_temp_e) */
		@@ -1077,11 +1039,17 @@ static void biDiagonalReductionRight_fx(
		( g ) = L_negate( L_negate( L_temp ) ); / exp(L_temp_e) */
		move32();
		}
		*g_e = L_temp_e;
		move16();


		r = BASOP_Util_Add_Mant32Exp( Mpy_32_32( ( g ), singularVectors[currChannel][idx] ), singularVectors2_e[currChannel][idx], -norm_x, norm_x_e, &r_e ); / exp(r_e) */
		singularVectors[currChannel][idx] = BASOP_Util_Add_Mant32Exp( singularVectors[currChannel][idx], singularVectors2_e[currChannel][idx], -( g ), 0, &singularVectors2_e[currChannel][idx] ); / exp(sing_exp) */
		singularVectors[currChannel][idx] = BASOP_Util_Add_Mant32Exp( singularVectors[currChannel][idx], singularVectors2_e[currChannel][idx], -( g ), ( g_e ), &singularVectors2_e[currChannel][idx] ); /* exp(sing_exp) */
		move32();


		invVal_e = r_e;
		move16();
		invVal = BASOP_Util_Divide3232_Scale_newton( MAXVAL_WORD32, maxWithSign_fx( r ), &invVal_e );

		FOR( jCh = idx; jCh < nChannelsC; jCh++ ) /* nChannelsC */
		@@ -1111,13 +1079,6 @@ static void biDiagonalReductionRight_fx(
		}
		}

		FOR( jCh = idx; jCh < nChannelsC; jCh++ ) /* nChannelsC */
		{
		singularVectors[currChannel][jCh] = Mpy_32_32( singularVectors[currChannel][jCh], ( sig_x ) ); / exp(sing_exp + sig_x_e) */
		move32();
		singularVectors2_e[currChannel][jCh] = add( singularVectors2_e[currChannel][jCh], *sig_x_e );
		move16();
		}
		}
		}