bring DCT_FX to FP (5ff54f09) · Commits · IVAS Codec Public Collaboration / IVAS Codec

lib_com/ivas_rom_com.c

+43 −1

Original line number	Diff line number	Diff line
		@@ -2778,6 +2778,48 @@ const int32_t sep_object_brate[][MAX_NUM_OBJECTS] =
		};

		/* column wise DCT matrices for 4 5, and 8 dim */
		#ifdef NONBE_1231_BASOP_819_THRESHOLD_MASA2TOTAL
		const Word32 dct4_fx[4 * 4] = { // Q31
		1073741824, 1402951040, 1073741824, 581109056,
		1073741824, 581109056, -1073741824, -1402951040,
		1073741824, -581109056, -1073741824, 1402951040,
		1073741824, -1402951040, 1073741824, -581109056
		};

		const Word32 dct5_fx[5 * 5] = { // Q31
		960354688, 1291711360, 1098867328, 798219648, 419618304,
		960354688, 798219648, -419618304, -1291711360, -1098867328,
		960354688, 0, -1358283392, 0, 1358283392,
		960354688, -798219648, -419618304, 1291711360, -1098867328,
		960354688, -1291711360, 1098867328, -798219648, 419618304
		};

		const Word32 dct8_fx[8 * 8] = { // Q31
		759350208, 1053125952, 991922688, 892708928, 759350208, 596570944, 410813632, 209379648,
		759350208, 892708928, 410813632, -209379648, -759350208, -1053125952, -991922688, -596570944,
		759350208, 596570944, -410813632, -1053125952, -759350208, 209379648, 991922688, 892708928,
		759350208, 209379648, -991922688, -596570944, 759350208, 892708928, -410813632, -1053125952,
		759350208, -209379648, -991922688, 596570944, 759350208, -892708928, -410813632, 1053125952,
		759350208, -596570944, -410813632, 1053125952, -759350208, -209379648, 991922688, -892708928,
		759350208, -892708928, 410813632, 209379648, -759350208, 1053125952, -991922688, 596570944,
		759350208, -1053125952, 991922688, -892708928, 759350208, -596570944, 410813632, -209379648
		};

		const Word32 dct12_fx[12 * 12] = { // Q31
		619978560, 869301376, 846752832, 810030848, 759350208, 695569984, 619978560, 533649696, 438301408, 335436960, 226989024, 114460880,
		619978560, 810030848, 619978560, 335436960, 0, -335436960, -619978560, -810030848, -876602816, -810030848, -619978560, -335436960,
		619978560, 695569984, 226989024, -335436960, -759350208, -869301376, -619978560, -114460880, 438301408, 810030848, 846752832, 533649696,
		619978560, 533649696, -226989024, -810030848, -759350208, -114460880, 619978560, 869301376, 438301408, -335436960, -846752832, -695569984,
		619978560, 335436960, -619978560, -810030848, 0, 810030848, 619978560, -335436960, -876602816, -335436960, 619978560, 810030848, 619978560,
		114460880, -846752832, -335436960, 759350208, 533649696, -619978560, -695569984, 438301408, 810030848, -226989024, -869301376, 619978560,
		-114460880, -846752832, 335436960, 759350208, -533649696, -619978560, 695569984, 438301408, -810030848, -226989024, 869301376, 619978560,
		-335436960, -619978560, 810030848, 0, -810030848, 619978560, 335436960, -876602816, 335436960, 619978560, -810030848, 619978560, -533649696,
		-226989024, 810030848, -759350208, 114460880, 619978560, -869301376, 438301408, 335436960, -846752832, 695569984, 619978560, -695569984,
		226989024, 335436960, -759350208, 869301376, -619978560, 114460880, 438301408, -810030848, 846752832, -533649696, 619978560, -810030848,
		619978560, -335436960, 0, 335436960, -619978560, 810030848, -876602816, 810030848, -619978560, 335436960, 619978560, -869301376, 846752832,
		-810030848, 759350208, -695569984, 619978560, -533649696, 438301408, -335436960, 226989024, -114460880
		};
		#else
		const float dct4[4*4] =
		{
		0.5000f, 0.6533f, 0.5000f, 0.2706f,
		@@ -2822,7 +2864,7 @@ const float dct12[12*12]=
		0.2887f, -0.3772f, 0.2887f, -0.1562f, -0.0000f, 0.1562f, -0.2887f, 0.3772f, -0.4082f, 0.3772f, -0.2887f, 0.1562f,
		0.2887f, -0.4048f, 0.3943f, -0.3772f, 0.3536f, -0.3239f, 0.2887f, -0.2485f, 0.2041f, -0.1562f, 0.1057f, -0.0533f
		};

		#endif
		/----------------------------------------------------------------------------------
		* ISM ROM tables
		----------------------------------------------------------------------------------/

lib_com/ivas_rom_com.h

+7 −1

Original line number	Diff line number	Diff line
		@@ -325,11 +325,17 @@ extern const float McMASA_LFEGain_vectors[64];
		----------------------------------------------------------------------------------/

		extern const int32_t sep_object_brate[][MAX_NUM_OBJECTS];
		#ifdef NONBE_1231_BASOP_819_THRESHOLD_MASA2TOTAL
		extern const Word32 dct4_fx[];
		extern const Word32 dct5_fx[];
		extern const Word32 dct8_fx[];
		extern const Word32 dct12_fx[];
		#else
		extern const float dct4[];
		extern const float dct5[];
		extern const float dct8[];
		extern const float dct12[];

		#endif
		/----------------------------------------------------------------------------------
		* ISM ROM tables
		----------------------------------------------------------------------------------/

lib_com/ivas_tools.c

+245 −0

Original line number	Diff line number	Diff line
		@@ -1274,3 +1274,248 @@ float rand_triangular_signed(
		return 0.5f - 0.5f * sqrtf( 1.0f - rand_val );
		}
		}
		#ifdef NONBE_1231_BASOP_819_THRESHOLD_MASA2TOTAL
		Word16 matrix_product_fx(
		const Word32 X_fx, / i : left hand matrix Qx*/
		const Word16 rowsX, /* i : number of rows of the left hand matrix Q0*/
		const Word16 colsX, /* i : number of columns of the left hand matrix Q0*/
		const Word16 transpX, /* i : flag indicating the transposition of the left hand matrix prior to the multiplication Q0*/
		const Word32 Y_fx, / i : right hand matrix Qy*/
		const Word16 rowsY, /* i : number of rows of the right hand matrix Q0*/
		const Word16 colsY, /* i : number of columns of the right hand matrix Q0*/
		const Word16 transpY, /* i : flag indicating the transposition of the right hand matrix prior to the multiplication Q0*/
		Word32 Z_fx / o : resulting matrix after the matrix multiplication Qx + Qy - 31*/
		)
		{
		Word16 i, j, k;
		Word16 x_idx, y_idx;
		Word32 *Zp_fx = Z_fx;

		/* Processing */

		if( transpX == 1 && transpY == 0 ) /* We use X transpose */
		{
		if( rowsX != rowsY )
		{
		return EXIT_FAILURE;
		}
		for( j = 0; j < colsY; ++j )
		{
		for( i = 0; i < colsX; ++i )
		{
		( *Zp_fx ) = 0;

		for( k = 0; k < rowsX; ++k )
		{
		x_idx = k + irowsX; /Q0*/
		y_idx = k + jrowsY; /Q0*/
		( Zp_fx ) = Zp_fx + ( X_fx[x_idx] * Y_fx[y_idx] ); /Qx + Qy - 31/
		}
		Zp_fx++;
		}
		}
		}
		else if( transpX == 0 && transpY == 1 ) /* We use Y transpose */
		{
		if( colsX != colsY )
		{
		return EXIT_FAILURE;
		}
		for( j = 0; j < rowsY; ++j )
		{
		for( i = 0; i < rowsX; ++i )
		{
		( *Zp_fx ) = 0;
		for( k = 0; k < colsX; ++k )
		{
		x_idx = i + krowsX; /Q0*/
		y_idx = j + krowsY; /Q0*/
		( Zp_fx ) = (Zp_fx) + ( X_fx[x_idx] * Y_fx[y_idx] ); /Qx + Qy - 31/
		}
		Zp_fx++;
		}
		}
		}
		else if( transpX == 1 && transpY == 1 ) /* We use both transpose */
		{
		if( rowsX != colsY )
		{
		return EXIT_FAILURE;
		}
		for( j = 0; j < rowsY; ++j )
		{
		for( i = 0; i < colsX; ++i )
		{
		( *Zp_fx ) = 0;
		move32();
		for( k = 0; k < colsX; ++k )
		{
		x_idx = add( k, imult1616( i, rowsX ) ); /Q0/
		y_idx = add( j, imult1616( k, rowsY ) ); /Q0/
		( Zp_fx ) = L_add( Zp_fx, Mpy_32_32( X_fx[x_idx], Y_fx[y_idx] ) ); /Qx + Qy - 31/
		move32();
		}

		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;

		for( k = 0; k < colsX; ++k )
		{
		x_idx = add( i, imult1616( k, rowsX ) ); /Q0/
		y_idx = add( k, imult1616( j, rowsY ) ); /Q0/
		( Zp_fx ) = L_add_sat( Zp_fx, Mpy_32_32( X_fx[x_idx], Y_fx[y_idx] ) ); /Qx + Qy - 31/
		// TODO: overflow of Z_fx to be checked
		move32();
		}
		Zp_fx++;
		}
		}
		}

		return EXIT_SUCCESS;
		}

		Word16 matrix_product_q30_fx(
		const Word32 X_fx, / i : left hand matrix Q31*/
		const Word16 rowsX, /* i : number of rows of the left hand matrix Q0*/
		const Word16 colsX, /* i : number of columns of the left hand matrix Q0*/
		const Word16 transpX, /* i : flag indicating the transposition of the left hand matrix prior to the multiplication Q0*/
		const Word32 Y_fx, / i : right hand matrix Q25*/
		const Word16 rowsY, /* i : number of rows of the right hand matrix Q0*/
		const Word16 colsY, /* i : number of columns of the right hand matrix Q0*/
		const Word16 transpY, /* i : flag indicating the transposition of the right hand matrix prior to the multiplication Q0*/
		Word32 Z_fx / o : resulting matrix after the matrix multiplication Q30*/
		)
		{
		Word16 i, j, k;
		Word16 x_idx, y_idx;
		Word32 *Zp_fx = Z_fx;
		long long int W_tmp;

		/* Processing */
		test();
		test();
		test();
		if( transpX == 1 && transpY == 0 ) /* We use X transpose */
		{
		if( rowsX != rowsY )
		{
		return EXIT_FAILURE;
		}
		for( j = 0; j < colsY; ++j )
		{
		for( i = 0; i < colsX; ++i )
		{
		//( *Zp_fx ) = 0;
		W_tmp = 0;
		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] ) );
		x_idx = k + i * rowsX; /* Q0 */
		y_idx = k + j * rowsY; /* Q0 */
		W_tmp += X_fx[x_idx] * Y_fx[y_idx]; /* Q56 */
		}
		W_tmp = W_tmp << 6; /* W_shl( W_tmp, 6 ); / /Q62*/
		( Zp_fx ) = ( W_tmp + 0x80000000 ) >> 32; / W_round64_L( W_tmp ); / /Q30*/
		Zp_fx++;
		}
		}
		}
		else if( transpX == 0 && transpY ==1 ) /* We use Y transpose */
		{
		if( colsX != colsY )
		{
		return EXIT_FAILURE;
		}
		for( j = 0; j < rowsY; ++j )
		{
		for( i = 0; i < rowsX; ++i )
		{
		//( *Zp_fx ) = 0;
		W_tmp = 0;

		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] ) );
		x_idx = i + k * rowsX; /Q0/
		y_idx = j + k * rowsY; /Q0/
		W_tmp += X_fx[x_idx]Y_fx[y_idx]; / Q56 */
		}
		W_tmp = W_tmp << 6; /Q62/
		( Zp_fx ) = ( W_tmp + 0x80000000 ) >> 32; /Q30*/
		Zp_fx++;
		}
		}
		}
		else if( transpX == 1 && transpY ==1 ) /* We use both transpose */
		{
		if( rowsX != colsY )
		{
		return EXIT_FAILURE;
		}
		for( j = 0; j < rowsY; ++j )
		{
		for( i = 0; i < colsX; ++i )
		{
		//( *Zp_fx ) = 0;
		W_tmp = 0;
		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] ) );
		x_idx = k + i * rowsX; /Q0/
		y_idx = j + k * rowsY; /Q0/
		W_tmp += X_fx[x_idx] * Y_fx[y_idx]; // Q56
		}
		W_tmp = W_tmp << 6; /Q62/
		( Zp_fx ) = ( W_tmp + 0x80000000 ) >> 32; /Q30*/

		Zp_fx++;
		}
		}
		}
		else /* Regular case */
		{
		if( colsX != rowsY )
		{
		return EXIT_FAILURE;
		}

		for( j = 0; j < colsY; ++j )
		{
		for( i = 0; i < rowsX; ++i )
		{
		//( *Zp_fx ) = 0;
		W_tmp = 0;

		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] ) );
		x_idx = i + k * rowsX; /Q0/
		y_idx = k + j * rowsY; /Q0/
		W_tmp += X_fx[x_idx] * Y_fx[y_idx]; // Q56
		}
		W_tmp = W_tmp << 6; /Q62/
		( Zp_fx ) = ( W_tmp + 0x80000000 ) >> 32; /Q30*/

		Zp_fx++;
		}
		}
		}

		return EXIT_SUCCESS;
		}
		#endif
		No newline at end of file

lib_dec/ivas_qmetadata_dec.c

+109 −8

Original line number	Diff line number	Diff line
		@@ -97,7 +97,13 @@ static int16_t read_surround_coherence_hr( uint16_t bitstream, int16_t p_bit_p

		static int16_t read_coherence_data_hr_512( uint16_t bitstream, int16_t p_bit_pos, IVAS_QMETADATA *hQMetaData, const int16_t idx_dir, const int16_t nbits_coh );

		static void read_stream_dct_coeffs_omasa( int16_t q_idx, float q_dct_data, const int16_t len_stream, uint16_t bit_stream, int16_t index, const int16_t first_line );
		static void read_stream_dct_coeffs_omasa( int16_t *q_idx,
		#ifdef NONBE_1231_BASOP_819_THRESHOLD_MASA2TOTAL
		Word32 * q_dct_data_fx,
		#else
		float *q_dct_data,
		#endif
		const int16_t len_stream, uint16_t bit_stream, int16_t index, const int16_t first_line );


		/-----------------------------------------------------------------------
		@@ -4248,7 +4254,11 @@ static void decode_combined_index(

		static void read_stream_dct_coeffs_omasa(
		int16_t *q_idx,
		#ifdef NONBE_1231_BASOP_819_THRESHOLD_MASA2TOTAL
		Word32 * q_dct_data_fx,
		#else
		float *q_dct_data,
		#endif
		const int16_t len_stream,
		uint16_t *bit_stream,
		int16_t *index,
		@@ -4325,6 +4335,20 @@ static void read_stream_dct_coeffs_omasa(
		}

		/* deindex */
		#ifdef NONBE_1231_BASOP_819_THRESHOLD_MASA2TOTAL
		q_dct_data_fx[0] = L_shl( Mpy_32_16_1( step, shl( q_idx[0], 7 ) ), 2 ); // q = 25
		FOR( i = 1; i < len_stream; i++ )
		{
		IF( s_and( q_idx[i], 1 ) == 0 )
		{
		q_dct_data_fx[i] = L_shl( Mpy_32_16_1( step, negate( shl( q_idx[i], 6 ) ) ), 2 ); /Q25/
		}
		ELSE
		{
		q_dct_data_fx[i] = L_shl( Mpy_32_16_1( step, shl( q_idx[i] + 1, 6 ) ), 2 ); /Q25/
		}
		}
		#else
		q_dct_data[0] = q_idx[0] * step;
		for ( i = 1; i < len_stream; i++ )
		{
		@@ -4337,7 +4361,7 @@ static void read_stream_dct_coeffs_omasa(
		q_dct_data[i] = ( ( q_idx[i] + 1 ) >> 1 ) * step;
		}
		}

		#endif
		return;
		}

		@@ -4356,8 +4380,13 @@ void ivas_omasa_decode_masa_to_total(
		{
		int16_t i, j, k;
		int16_t q_idx[MAX_PARAM_SPATIAL_SUBFRAMES * MASA_FREQUENCY_BANDS];
		#ifdef NONBE_1231_BASOP_819_THRESHOLD_MASA2TOTAL
		Word32 q_dct_data_fx[MAX_PARAM_SPATIAL_SUBFRAMES * MASA_FREQUENCY_BANDS],
		dct_data_tmp_fx[MAX_PARAM_SPATIAL_SUBFRAMES * MASA_FREQUENCY_BANDS];
		#else
		float q_dct_data[MAX_PARAM_SPATIAL_SUBFRAMES * MASA_FREQUENCY_BANDS],
		dct_data_tmp[MAX_PARAM_SPATIAL_SUBFRAMES * MASA_FREQUENCY_BANDS];
		#endif
		int16_t n_streams, len_stream;
		#ifdef NONBE_1231_BASOP_819_THRESHOLD_MASA2TOTAL
		int32_t int_tmp;
		@@ -4374,10 +4403,86 @@ void ivas_omasa_decode_masa_to_total(
		set_s( q_idx, 0, nbands * nblocks );
		for ( i = 0; i < n_streams; i++ )
		{
		read_stream_dct_coeffs_omasa( &q_idx[i * len_stream], &q_dct_data[i * len_stream], len_stream, bit_stream, index, i == 0 );
		read_stream_dct_coeffs_omasa( &q_idx[i * len_stream],
		#ifdef NONBE_1231_BASOP_819_THRESHOLD_MASA2TOTAL
		&q_dct_data_fx[i * len_stream],
		#else
		&q_dct_data[i * len_stream],
		#endif
		len_stream, bit_stream, index, i == 0 );
		}

		/* inverse DCT2 transform */
		#ifdef NONBE_1231_BASOP_819_THRESHOLD_MASA2TOTAL
		SWITCH( len_stream )
		{
		case 4:
		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 ); /Q30/
		BREAK;
		case 5:
		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 ); /Q30/
		BREAK;
		case 8:
		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 ); /Q30/
		BREAK;
		case 12:
		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 ); /Q30/
		BREAK;
		case 20:
		matrix_product_fx( dct5_fx, nbands, nbands, 1, q_dct_data_fx, nbands, nblocks, 0, dct_data_tmp_fx );
		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_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" );
		BREAK;
		}
		k = 0;
		move16();
		FOR( i = 0; i < nblocks; i++ )
		{
		FOR( j = 0; j < nbands; j++ )
		{
		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] ); /Q30/
		move32();
		k = add( k, 1 );
		}
		}

		IF( EQ_16( nblocks, 1 ) )
		{
		FOR( i = 1; i < MAX_PARAM_SPATIAL_SUBFRAMES; i++ )
		{
		FOR( j = 0; j < nbands; j++ )
		{
		masa_to_total_energy_ratio_fx[i][j] = masa_to_total_energy_ratio_fx[0][j]; /Q30/
		move32();
		}
		}
		}

		IF( EQ_16( nbands, 1 ) )
		{
		FOR( j = 1; j < 5; j++ )
		{
		FOR( i = 0; i < nblocks; i++ )
		{
		masa_to_total_energy_ratio_fx[i][j] = masa_to_total_energy_ratio_fx[i][0]; /Q30/
		move32();
		}
		}
		}

		#else
		switch ( len_stream )
		{
		case 4:
		@@ -4446,13 +4551,9 @@ void ivas_omasa_decode_masa_to_total(
		for ( i = 0; i < nblocks; i++ )
		{
		masa_to_total_energy_ratio[i][j] = masa_to_total_energy_ratio[i][0];
		#ifdef NONBE_1231_BASOP_819_THRESHOLD_MASA2TOTAL
		int_tmp = (int32_t) ( MASA_SUR_COH_PRECISION * masa_to_total_energy_ratio[i][j] );
		masa_to_total_energy_ratio[i][j] = (float) ( int_tmp * MASA_SUR_COH_THRESHOLD );
		#endif
		}
		}
		}

		#endif
		return;
		}