Commit 6b8a3beb authored by Sandesh Venkatesh's avatar Sandesh Venkatesh
Browse files

Binaural renderer optimization - non bit exact changes 

Around 6 WMOPs reduction observed for multichannel test case.
MLD deviation < 1.
parent 9773847c

lib_com/options.h

+1 −0
Original line number Diff line number Diff line
@@ -81,6 +81,7 @@ 
/* Note: each compile switch (FIX_1101_...) is independent from the other ones */

#define OPT_MCT_ENC_V3_NBE

#define OPT_MCT_ENC_V2_BE

#define OPT_BIN_REND_V2_NBE

#define OPT_MCH_DEC_V1_NBE

#define OPT_MASA_DEC_V1_NBE

#define OPT_MASA_DEC_V2_NBE

lib_dec/ivas_binRenderer_internal_fx.c

+384 −338
Original line number Diff line number Diff line
@@ -72,12 +72,17 @@ static void ivas_binRenderer_filterModule_fx( 
    Q_filterStates = hBinRenderer->hBinRenConvModule->Q_filterStatesLeft;

    move16();



    // to be checked: feasibility with 32 bit buffers

#ifdef OPT_BIN_REND_V2_NBE

    shift_q = add( sub( Q_filterStates, Q_curr ), 1 );

    hBinRenderer->hBinRenConvModule->Q_filterStatesLeft = Q_curr;

#else  /* OPT_BIN_REND_V2_NBE */

    Word64 Cldfb_RealBuffer_64fx[BINAURAL_CHANNELS][MAX_PARAM_SPATIAL_SUBFRAMES];

    Word64 Cldfb_ImagBuffer_64fx[BINAURAL_CHANNELS][MAX_PARAM_SPATIAL_SUBFRAMES];

#endif /* OPT_BIN_REND_V2_NBE */



    FOR( bandIdx = 0; bandIdx < hBinRenderer->conv_band; bandIdx++ )

    {

#ifndef OPT_BIN_REND_V2_NBE

        set64_fx( &Cldfb_RealBuffer_64fx[0][0], 0, BINAURAL_CHANNELS * MAX_PARAM_SPATIAL_SUBFRAMES );

        set64_fx( &Cldfb_ImagBuffer_64fx[0][0], 0, BINAURAL_CHANNELS * MAX_PARAM_SPATIAL_SUBFRAMES );
@@ -91,6 +96,7 @@ static void ivas_binRenderer_filterModule_fx( 
            filterTapsRightRealPtr_fx = hBinRenderer->hBinRenConvModule->filterTapsRightReal_fx[bandIdx][chIdx]; // Q29

            filterTapsRightImagPtr_fx = hBinRenderer->hBinRenConvModule->filterTapsRightImag_fx[bandIdx][chIdx]; // Q29



#endif /* OPT_BIN_REND_V2_NBE */

            FOR( k = 0; k < numTimeSlots; k++ )

            {

                Word64 outRealLeft_fx = 0, outRealRight_fx = 0, outImagLeft_fx = 0, outImagRight_fx = 0;
@@ -98,6 +104,23 @@ static void ivas_binRenderer_filterModule_fx( 
                move64();

                move64();

                move64();

#ifdef OPT_BIN_REND_V2_NBE

                Word64 outRealLeft = 0, outRealRight = 0, outImagLeft = 0, outImagRight = 0;

                move64();

                move64();

                move64();

                move64();



                FOR( chIdx = 0; chIdx < hBinRenderer->nInChannels; chIdx++ )

                {

                    filterStatesLeftRealPtr_fx = (Word32 *) &( hBinRenderer->hBinRenConvModule->filterStatesLeftReal_fx[bandIdx][chIdx][0] );

                    filterStatesLeftImagPtr_fx = (Word32 *) &( hBinRenderer->hBinRenConvModule->filterStatesLeftImag_fx[bandIdx][chIdx][0] );



                    filterTapsLeftRealPtr_fx = hBinRenderer->hBinRenConvModule->filterTapsLeftReal_fx[bandIdx][chIdx];   // Q29

                    filterTapsLeftImagPtr_fx = hBinRenderer->hBinRenConvModule->filterTapsLeftImag_fx[bandIdx][chIdx];   // Q29

                    filterTapsRightRealPtr_fx = hBinRenderer->hBinRenConvModule->filterTapsRightReal_fx[bandIdx][chIdx]; // Q29

                    filterTapsRightImagPtr_fx = hBinRenderer->hBinRenConvModule->filterTapsRightImag_fx[bandIdx][chIdx]; // Q29

#endif                                                                                                                   /* OPT_BIN_REND_V2_NBE */



                    FOR( tapIdx = hBinRenderer->hBinRenConvModule->numTapsArray[bandIdx] - 1; tapIdx > 0; tapIdx-- )

                    {
@@ -121,6 +144,7 @@ static void ivas_binRenderer_filterModule_fx( 
                        outImagRight_fx = W_mac_32_32( outImagRight_fx, filterStatesLeftImagPtr_fx[tapIdx], filterTapsRightRealPtr_fx[tapIdx] ); // Q30 + Q_filterStates

                    }



#ifndef OPT_BIN_REND_V2_NBE

                    shift_q = add( sub( Q_filterStates, Q_curr ), 1 );



                    IF( shift_q != 0 )
@@ -132,13 +156,34 @@ static void ivas_binRenderer_filterModule_fx( 
                        hBinRenderer->hBinRenConvModule->Q_filterStatesLeft = Q_curr;

                        move16();

                    }



#endif /* OPT_BIN_REND_V2_NBE */

                    filterStatesLeftRealPtr_fx[0] = CLDFB_real[chIdx][k][bandIdx];

                    move32();

                    filterStatesLeftImagPtr_fx[0] = CLDFB_imag[chIdx][k][bandIdx];

                    move32();





#ifdef OPT_BIN_REND_V2_NBE

                    /* Left Real and Imag */

                    Word32 neg_temp2 = L_negate( filterStatesLeftImagPtr_fx[0] ); // Q_curr -1



                    outRealLeft = W_mac_32_32( W_mac_32_32( outRealLeft, filterStatesLeftRealPtr_fx[0], filterTapsLeftRealPtr_fx[0] ), neg_temp2, filterTapsLeftImagPtr_fx[0] );

                    outImagLeft = W_mac_32_32( W_mac_32_32( outImagLeft, filterStatesLeftImagPtr_fx[0], filterTapsLeftRealPtr_fx[0] ), filterStatesLeftRealPtr_fx[0], filterTapsLeftImagPtr_fx[0] );



                    /* Right Real and Imag */

                    outRealRight = W_mac_32_32( W_mac_32_32( outRealRight, neg_temp2, filterTapsRightImagPtr_fx[0] ), filterStatesLeftRealPtr_fx[0], filterTapsRightRealPtr_fx[0] );

                    outImagRight = W_mac_32_32( W_mac_32_32( outImagRight, filterStatesLeftImagPtr_fx[0], filterTapsRightRealPtr_fx[0] ), filterStatesLeftRealPtr_fx[0], filterTapsRightImagPtr_fx[0] );

                }

                out_Conv_CLDFB_real[0][k][bandIdx] = W_shl_sat_l( W_add( W_shr( outRealLeft_fx, shift_q ), W_shr( outRealLeft, 1 ) ), shift_q6 );

                out_Conv_CLDFB_real[1][k][bandIdx] = W_shl_sat_l( W_add( W_shr( outRealRight_fx, shift_q ), W_shr( outRealRight, 1 ) ), shift_q6 );

                out_Conv_CLDFB_imag[0][k][bandIdx] = W_shl_sat_l( W_add( W_shr( outImagLeft_fx, shift_q ), W_shr( outImagLeft, 1 ) ), shift_q6 );

                out_Conv_CLDFB_imag[1][k][bandIdx] = W_shl_sat_l( W_add( W_shr( outImagRight_fx, shift_q ), W_shr( outImagRight, 1 ) ), shift_q6 );

                move32();

                move32();

                move32();

                move32();

            }

#else  /* OPT_BIN_REND_V2_NBE */

            /* Left Real and Imag */



            Word32 temp1 = L_shr( filterStatesLeftRealPtr_fx[0], 1 ); // Q_curr -1
@@ -178,6 +223,7 @@ static void ivas_binRenderer_filterModule_fx( 
            move32();

        }

    }

#endif /* OPT_BIN_REND_V2_NBE */

        }