omasa fixes (56307a6f) · Commits · SA4 / Audio / IVAS BASOP

lib_com/ivas_prot.h

+12 −0

Original line number	Diff line number	Diff line
		@@ -4763,6 +4763,18 @@ Word16 matrix_product_fx(
		Word32 Z_fx / o : resulting matrix after the matrix multiplication */
		);

		Word16 matrix_product_q30_fx(
		const Word32 X_fx, / i : left hand matrix */
		const Word16 rowsX, /* i : number of rows of the left hand matrix */
		const Word16 colsX, /* i : number of columns of the left hand matrix */
		const Word16 transpX, /* i : flag indicating the transposition of the left hand matrix prior to the multiplication */
		const Word32 Y_fx, / i : right hand matrix */
		const Word16 rowsY, /* i : number of rows of the right hand matrix */
		const Word16 colsY, /* i : number of columns of the right hand matrix */
		const Word16 transpY, /* i : flag indicating the transposition of the right hand matrix prior to the multiplication */
		Word32 Z_fx / o : resulting matrix after the matrix multiplication */
		);

		Word16 matrix_product_mant_exp(
		const Word32 X_fx, / i : left hand matrix */
		const Word16 X_e, / i : left hand matrix */

lib_com/ivas_tools.c

+121 −0

Original line number	Diff line number	Diff line
		@@ -1354,6 +1354,127 @@ Word16 matrix_product_fx(
		return EXIT_SUCCESS;
		}

		Word16 matrix_product_q30_fx(
		const Word32 X_fx, / i : left hand matrix */
		const Word16 rowsX, /* i : number of rows of the left hand matrix */
		const Word16 colsX, /* i : number of columns of the left hand matrix */
		const Word16 transpX, /* i : flag indicating the transposition of the left hand matrix prior to the multiplication */
		const Word32 Y_fx, / i : right hand matrix */
		const Word16 rowsY, /* i : number of rows of the right hand matrix */
		const Word16 colsY, /* i : number of columns of the right hand matrix */
		const Word16 transpY, /* i : flag indicating the transposition of the right hand matrix prior to the multiplication */
		Word32 Z_fx / o : resulting matrix after the matrix multiplication */
		)
		{
		Word16 i, j, k;
		Word32 *Zp_fx = Z_fx;
		Word64 W_tmp;

		/* Processing */
		test();
		test();
		test();
		IF( EQ_16( transpX, 1 ) && EQ_16( transpY, 0 ) ) /* We use X transpose */
		{
		IF( NE_16( rowsX, rowsY ) )
		{
		return EXIT_FAILURE;
		}
		FOR( j = 0; j < colsY; ++j )
		{
		FOR( i = 0; i < colsX; ++i )
		{
		//( *Zp_fx ) = 0;
		W_tmp = 0;
		move64();
		FOR( k = 0; k < rowsX; ++k )
		{
		//( Zp_fx ) = L_add( Zp_fx, Mpy_32_32( X_fx[k + i * rowsX], Y_fx[k + j * rowsY] ) );
		W_tmp = W_add( W_tmp, W_mult0_32_32( X_fx[k + i * rowsX], Y_fx[k + j * rowsY] ) );//Q56
		}
		W_tmp = W_shl( W_tmp, 6 );
		( *Zp_fx ) = L_sub(W_round64_L( W_tmp ), 64); //adjusting for precision
		Zp_fx++;
		}
		}
		}
		ELSE IF( EQ_16( transpX, 0 ) && EQ_16( transpY, 1 ) ) /* We use Y transpose */
		{
		IF( NE_16( colsX, colsY ) )
		{
		return EXIT_FAILURE;
		}
		FOR( j = 0; j < rowsY; ++j )
		{
		FOR( i = 0; i < rowsX; ++i )
		{
		//( *Zp_fx ) = 0;
		W_tmp = 0;
		move64();
		FOR( k = 0; k < colsX; ++k )
		{
		//( Zp_fx ) = L_add( Zp_fx, Mpy_32_32( X_fx[i + k * rowsX], Y_fx[j + k * rowsY] ) );
		W_tmp = W_add( W_tmp, W_mult0_32_32( X_fx[i + k * rowsX], Y_fx[j + k * rowsY] ) ); // Q56
		}
		W_tmp = W_shl( W_tmp, 6 );
		( *Zp_fx ) = L_sub( W_round64_L( W_tmp ), 64 ); // adjusting for precision
		Zp_fx++;
		}
		}
		}
		ELSE IF( EQ_16( transpX, 1 ) && EQ_16( transpY, 1 ) ) /* We use both transpose */
		{
		IF( NE_16( rowsX, colsY ) )
		{
		return EXIT_FAILURE;
		}
		FOR( j = 0; j < rowsY; ++j )
		{
		FOR( i = 0; i < colsX; ++i )
		{
		//( *Zp_fx ) = 0;
		W_tmp = 0;
		move64();
		FOR( k = 0; k < colsX; ++k )
		{
		//( Zp_fx ) = L_add( Zp_fx, Mpy_32_32( X_fx[k + i * rowsX], Y_fx[j + k * rowsY] ) );
		W_tmp = W_add( W_tmp, W_mult0_32_32( X_fx[k + i * rowsX], Y_fx[j + k * rowsY] ) ); // Q56
		}

		W_tmp = W_shl( W_tmp, 6 );
		( *Zp_fx ) = L_sub( W_round64_L( W_tmp ), 64 ); // adjusting for precision
		Zp_fx++;
		}
		}
		}
		ELSE /* Regular case */
		{
		IF( NE_16( colsX, rowsY ) )
		{
		return EXIT_FAILURE;
		}

		FOR( j = 0; j < colsY; ++j )
		{
		FOR( i = 0; i < rowsX; ++i )
		{
		//( *Zp_fx ) = 0;
		W_tmp = 0;
		move64();
		FOR( k = 0; k < colsX; ++k )
		{
		//( Zp_fx ) = L_add( Zp_fx, Mpy_32_32( X_fx[i + k * rowsX], Y_fx[k + j * rowsY] ) );
		W_tmp = W_add( W_tmp, W_mult0_32_32( X_fx[i + k * rowsX], Y_fx[k + j * rowsY] ) ); // Q56
		}
		W_tmp = W_shl( W_tmp, 6 );
		( *Zp_fx ) = L_sub( W_round64_L( W_tmp ), 64 ); // adjusting for precision
		Zp_fx++;
		}
		}
		}

		return EXIT_SUCCESS;
		}
		/takes input matrices in mantissa and exponent forms/
		Word16 matrix_product_mant_exp(
		const Word32 X_fx, / i : left hand matrix */

lib_dec/ivas_qmetadata_dec.c

+9 −7

Original line number	Diff line number	Diff line
		@@ -8562,28 +8562,28 @@ void ivas_omasa_decode_masa_to_total_fx(
		SWITCH( len_stream )
		{
		case 4:
		matrix_product_fx( dct4_fx, nblocks, nblocks, 1, q_dct_data_fx, nblocks, 1, 0, dct_data_tmp_fx );
		matrix_product_q30_fx( dct4_fx, nblocks, nblocks, 1, q_dct_data_fx, nblocks, 1, 0, dct_data_tmp_fx );
		Copy32( dct_data_tmp_fx, q_dct_data_fx, nblocks );
		BREAK;
		case 5:
		matrix_product_fx( dct5_fx, nbands, nbands, 1, q_dct_data_fx, nbands, 1, 0, dct_data_tmp_fx );
		matrix_product_q30_fx( dct5_fx, nbands, nbands, 1, q_dct_data_fx, nbands, 1, 0, dct_data_tmp_fx );
		Copy32( dct_data_tmp_fx, q_dct_data_fx, nbands );
		BREAK;
		case 8:
		matrix_product_fx( dct8_fx, nbands, nbands, 1, q_dct_data_fx, nbands, 1, 0, dct_data_tmp_fx );
		matrix_product_q30_fx( dct8_fx, nbands, nbands, 1, q_dct_data_fx, nbands, 1, 0, dct_data_tmp_fx );
		Copy32( dct_data_tmp_fx, q_dct_data_fx, nbands );
		BREAK;
		case 12:
		matrix_product_fx( dct12_fx, nbands, nbands, 1, q_dct_data_fx, nbands, 1, 0, dct_data_tmp_fx );
		matrix_product_q30_fx( dct12_fx, nbands, nbands, 1, q_dct_data_fx, nbands, 1, 0, dct_data_tmp_fx );
		Copy32( dct_data_tmp_fx, q_dct_data_fx, nbands );
		BREAK;
		case 20:
		matrix_product_fx( dct5_fx, nbands, nbands, 1, q_dct_data_fx, nbands, nblocks, 0, dct_data_tmp_fx );
		matrix_product_fx( dct_data_tmp_fx, nbands, nblocks, 0, dct4_fx, nblocks, nblocks, 0, q_dct_data_fx ); /* reuse of variable*/
		matrix_product_q30_fx( dct_data_tmp_fx, nbands, nblocks, 0, dct4_fx, nblocks, nblocks, 0, q_dct_data_fx ); /* reuse of variable*/
		BREAK;
		case 32:
		matrix_product_fx( dct8_fx, nbands, nbands, 1, q_dct_data_fx, nbands, nblocks, 0, dct_data_tmp_fx );
		matrix_product_fx( dct_data_tmp_fx, nbands, nblocks, 0, dct4_fx, nblocks, nblocks, 0, q_dct_data_fx );
		matrix_product_q30_fx( dct_data_tmp_fx, nbands, nblocks, 0, dct4_fx, nblocks, nblocks, 0, q_dct_data_fx );
		BREAK;
		default:
		printf( "Incorrect number of coefficients for OMASA.\n" );
		@@ -8595,7 +8595,9 @@ void ivas_omasa_decode_masa_to_total_fx(
		{
		FOR( j = 0; j < nbands; j++ )
		{
		masa_to_total_energy_ratio_fx[i][j] = L_max( 0, L_shr( L_shl_sat( q_dct_data_fx[k], 31 - 25 ), 1 ) ); // Q30
		masa_to_total_energy_ratio_fx[i][j] = L_max( 0, q_dct_data_fx[k] ); // Q30
		move32();
		masa_to_total_energy_ratio_fx[i][j] = L_min( ONE_IN_Q30, masa_to_total_energy_ratio_fx[i][j] );
		move32();
		k++;
		}