the 64bit part of the householder matrix compiles. (890836c1) · Commits · SA4 / Audio / IVAS BASOP

lib_dec/ivas_svd_dec_fx.c

+55 −31

Original line number	Diff line number	Diff line
		@@ -88,7 +88,7 @@ static void biDiagonalReductionRight_fx(
		Word16 *g_e
		);
		static void biDiagonalReductionLeft_64(
		Word64 singularVectors_Left_64,
		Word64 singularVectors_Left_64[MAX_OUTPUT_CHANNELS][MAX_OUTPUT_CHANNELS],
		Word16 singularVectors_e,
		Word32 singularValues[MAX_OUTPUT_CHANNELS], /* exp(singularValues_e) */
		Word16 singularValues_e[MAX_OUTPUT_CHANNELS],
		@@ -98,7 +98,7 @@ static void biDiagonalReductionLeft_64(
		);

		static void biDiagonalReductionRight_64(
		Word64 singularVectors_Left_64,
		Word64 singularVectors_Left_64[MAX_OUTPUT_CHANNELS][MAX_OUTPUT_CHANNELS],
		Word16 singularVectors_e,
		Word32 secDiag[MAX_OUTPUT_CHANNELS], /* exp(singularValues_e) */
		Word16 secDiag_e[MAX_OUTPUT_CHANNELS],
		@@ -893,26 +893,22 @@ static void HouseholderReduction_fx(
		for (nCh=0;nCh<nChannelsC;nCh++)
		{
		biDiagonalReductionLeft_64(
		singularVectors_Left_64,currChannel,
		singularValues,singularValues_e,
		singularVectors_Left_64,nCh,
		singularValues_fx,singularValues_fx_e,
		nChannelsL,
		nChannelsC,
		currChannel
		nCh
		);
		biDiagonalReductionRight_64(
		singularVectors_Left_64,currChannel,
		secDiag,secDiag_exp,
		singularVectors_Left_64,nCh,
		secDiag_fx,secDiag_fx_e,
		nChannelsL,
		nChannelsC,
		currChannel,
		nCh,
		&g_fx,
		&g_e
		);
		}




		#endif
		FOR( jCh = 0; jCh < nChannelsL; jCh++ )
		{
		@@ -978,6 +974,7 @@ static void biDiagonalReductionLeft_64(
		Word64 r_64;
		Word32 tmp;
		Word16 tmpe;
		Word64 norm_64;
		norm_x=0;
		move32();
		IF( LT_16( currChannel, nChannelsL ) ) /* i <= m */
		@@ -987,12 +984,12 @@ static void biDiagonalReductionLeft_64(
		tmpe=add(singularVectors_e,1);
		FOR( jCh = currChannel; jCh < nChannelsL; jCh++ ) /* nChannelsL */
		{
		tmp=E_extract_l(W_shr(singularVectors_Left_64[jCh][currChannel],tmpe));
		tmp=W_extract_l(W_shr(singularVectors_Left_64[jCh][currChannel],tmpe));
		norm_64=W_add(norm_64,W_mult0_32_32(tmp,tmp));
		}
		tmpe=W_norm(norm_64);
		norm_x=W_extract_h(W_shl(norm_64, tmpe ));
		norm_x_e = add(sub(shl(singularVectors_e, 1), tmp3), 3 );
		norm_x_e = add(sub(shl(singularVectors_e, 1), tmpe), 3 );
		}
		IF ( norm_x )
		{
		@@ -1003,8 +1000,8 @@ static void biDiagonalReductionLeft_64(
		Word64 tmp64;
		L_temp_e = norm_x_e;
		move16();
		L_temp = Sqrt( norm_x, &L_temp_e);
		IF ( GE_64( singularVectors_Left_64[currChannel][currChannel] ) )
		L_temp = Sqrt32( norm_x, &L_temp_e);
		IF ( GE_64( singularVectors_Left_64[currChannel][currChannel], 0 ) )
		{
		L_temp = L_negate( L_temp );
		}
		@@ -1021,13 +1018,13 @@ static void biDiagonalReductionLeft_64(

		tmpe2=W_norm(r_64);
		r=W_extract_h(W_shl(r_64,tmpe2));
		r_e = sub(sub(add(add(1,singularVectors_e),g_e),tmpe,tmpe2));
		r_e = sub(sub(add(add(1,singularVectors_e),g_e),tmpe),tmpe2);

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

		tmpe=add(32,sub(singularVectors_e,g_e)); // TODO: maybe the other way around??
		tmp64=W_shr(W_deposit_h(g),tmpe);
		tmp64=W_shr(W_deposit32_h(g),tmpe);
		singularVectors_Left_64[currChannel][currChannel]=W_sub(singularVectors_Left_64[currChannel][currChannel],tmp64); // exponent +1


		@@ -1062,7 +1059,7 @@ static void biDiagonalReductionLeft_64(
		}

		static void biDiagonalReductionRight_64(
		Word64 singularVectors_Left_64,
		Word64 singularVectors_Left_64[MAX_OUTPUT_CHANNELS][MAX_OUTPUT_CHANNELS],
		Word16 singularVectors_e,
		Word32 secDiag[MAX_OUTPUT_CHANNELS], /* exp(singularValues_e) */
		Word16 secDiag_e[MAX_OUTPUT_CHANNELS],
		@@ -1075,7 +1072,7 @@ static void biDiagonalReductionRight_64(
		{
		secDiag[currChannel] = ( *g );
		move32();
		secDiag_exp[currChannel] = ( *g_e );
		secDiag_e[currChannel] = ( *g_e );
		move16();

		( *g ) =0;
		@@ -1088,6 +1085,11 @@ static void biDiagonalReductionRight_64(
		Word64 norm_64;
		Word64 abs_x;
		Word16 idx;
		Word32 tmp;
		Word16 tmpe;
		Word16 iCh,jCh;
		Word32 norm_x;
		Word16 norm_x_e;


		idx=add(currChannel,1);
		@@ -1099,13 +1101,13 @@ static void biDiagonalReductionRight_64(
		tmpe=add(singularVectors_e,1);
		for (jCh=idx;jCh<nChannelsC; jCh++)
		{
		tmp=E_extract_l(W_shr(singularVectors_Left_64[jCh][currChannel],tmpe));
		tmp=W_extract_l(W_shr(singularVectors_Left_64[jCh][currChannel],tmpe));
		norm_64=W_add(norm_64,W_mult0_32_32(tmp,tmp));
		abs_x=W_add(abs_x, W_abs(singularVectors_Left_64[jCh][currChannel]));
		}
		tmpe=W_norm(norm_64);
		norm_x=W_extract_h(W_shl(norm_64, tmpe ));
		norm_x_e = add(sub(shl(singularVectors_e, 1), tmp3), 3 );
		norm_x_e = add(sub(shl(singularVectors_e, 1), tmpe), 3 );
		IF (norm_x)
		{
		Word16 invVal_e;
		@@ -1113,33 +1115,42 @@ static void biDiagonalReductionRight_64(
		Word64 tmpmul;
		Word16 tmpe2;
		Word64 tmp64;
		Word32 f;
		Word16 f_e;
		Word32 L_temp;
		Word16 L_temp_e;
		Word64 r_64;
		Word32 r;
		Word16 r_e;


		L_temp_e = norm_x_e;
		move16();
		L_temp = Sqrt( norm_x, &L_temp_e);
		IF ( GE_64( singularVectors_Left_64[currChannel][idx] ) )
		L_temp = Sqrt32( norm_x, &L_temp_e);
		IF ( GE_64( singularVectors_Left_64[currChannel][idx],0 ) )
		{
		L_temp = L_negate( L_temp );
		}
		g=L_temp;
		*g=L_temp;
		move32();
		g_e = L_temp_e;
		*g_e = L_temp_e;
		move16();

		tmp=W_extract_l(W_shr(singularVectors_Left_64[currChannel][idx],singularVectors_e) );
		tmpe=sub(sub(g_e, singularVectors_e),1);
		tmpmul=W_mult0_32_32(g,tmp);
		tmpe=sub(sub(*g_e, singularVectors_e),1);
		tmpmul=W_mult0_32_32(*g,tmp);
		tmpmul=W_shl(tmpmul,tmpe);
		r_64=W_sub(tmpmul, norm_64 );

		tmpe2=W_norm(r_64);
		r=W_extract_h(W_shl(r_64,tmpe2));
		r_e = sub(sub(add(add(1,singularVectors_e),g_e),tmpe,tmpe2));
		r_e = sub(sub(add(add(1,singularVectors_e),*g_e),tmpe),tmpe2);

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

		tmpe=add(32,sub(singularVectors_e,g_e)); // TODO: maybe the other way around??
		tmp64=W_shr(W_deposit_h(g),tmpe);
		tmp64=W_shr(W_deposit32_h(g),tmpe);
		singularVectors_Left_64[currChannel][idx]=W_sub(singularVectors_Left_64[currChannel][idx],tmp64); // exponent +1


		@@ -1164,12 +1175,25 @@ static void biDiagonalReductionRight_64(
		f=Mpy_32_32(norm_x,invVal);
		f_e=add(invVal_e, sub(norm_x_e, r_e ));

		for (jCh=currChannel;jCh<nChannelsL; jCh++)
		for (jCh=idx;jCh<nChannelsL; jCh++)
		{
		tmp2=W_extract_l(W_shr(singularVectors_Left_64[jCh][currChannel],tmpe));
		singularVectors_Left_64[jCh][iCh]=W_add(singularVectors_Left_64[jCh][iCh],W_mult0_32_32(f,tmp2)); // exponent +1
		}
		}
		invVal_e = 0;
		move16();
		tmpe = W_norm(abs_x);
		invVal = BASOP_Util_Divide3232_Scale_newton( W_extract_h(W_shl(abs_x,tmpe)), r, &invVal_e);
		invVal_e = add(invVal_e, sub(tmpe, r_e));
		tmpe = add(1,singularVectors_e);
		for ( jCh = idx; jCh < nChannelsC ; jCh++)
		{
		secDiag[jCh] = Mpy_32_32( W_extract_l(W_shr(singularVectors_Left_64[currChannel][jCh],tmpe)), invVal);
		move32();
		secDiag_e[jCh]=add(invVal_e, tmpe);
		move16();
		}
		}
		}
		return;