Commit d6f2a28e authored by Sandesh Venkatesh's avatar Sandesh Venkatesh
Browse files

Merge branch 'enc_float_code_cleanup_ltv_crash_fixes_2' into 'main' 

Float structure elements cleanup and LTV crash fixes for +10dB inputs

See merge request !822
parents b4076d35 fd1fccde

lib_enc/core_enc_init.c

+7 −4
Original line number Diff line number Diff line
@@ -853,8 +853,10 @@ static void init_sig_buffers( 
#ifndef IVAS_FLOAT_FIXED // Need to remove after cleanup of old_inp_12k8

            mvr2r( st->old_inp_12k8, st->buf_speech_enc_pe_flt + st->L_frame - L_INP_MEM, L_INP_MEM );

#else

            Word16 tmp;

            f2me_buf_16( st->old_inp_12k8, st->buf_speech_enc_pe + st->L_frame - L_INP_MEM, &tmp, L_INP_MEM );



            Word16 tmp = st->exp_old_inp_12k8;

            move16();

            Copy( st->old_inp_12k8_fx, st->buf_speech_enc_pe + st->L_frame - L_INP_MEM, L_INP_MEM );

            /* SCaling to common exponent*/

            Scale_sig( st->buf_speech_enc_pe + st->L_frame - L_INP_MEM, L_INP_MEM, sub( tmp, s_max( tmp, st->exp_buf_speech_enc_pe ) ) );

            Scale_sig( st->buf_speech_enc_pe, st->L_frame - L_INP_MEM, sub( st->exp_buf_speech_enc_pe, s_max( tmp, st->exp_buf_speech_enc_pe ) ) );
@@ -870,8 +872,9 @@ static void init_sig_buffers( 
#ifndef IVAS_FLOAT_FIXED // Need to remove after cleanup of old_inp_12k8

            mvr2r( st->old_inp_16k, st->buf_speech_enc_pe_flt + st->L_frame - L_INP_MEM, L_INP_MEM );

#else

            Word16 tmp;

            f2me_buf_16( st->old_inp_12k8, st->buf_speech_enc_pe + st->L_frame - L_INP_MEM, &tmp, L_INP_MEM );

            Word16 tmp = st->exp_old_inp_12k8;

            move16();

            Copy( st->old_inp_12k8_fx, st->buf_speech_enc_pe + st->L_frame - L_INP_MEM, L_INP_MEM );

            /* SCaling to common exponent*/

            Scale_sig( st->buf_speech_enc_pe + st->L_frame - L_INP_MEM, L_INP_MEM, sub( tmp, s_max( tmp, st->exp_buf_speech_enc_pe ) ) );

            Scale_sig( st->buf_speech_enc_pe, st->L_frame - L_INP_MEM, sub( st->exp_buf_speech_enc_pe, s_max( tmp, st->exp_buf_speech_enc_pe ) ) );

lib_enc/core_enc_switch.c

+8 −6
Original line number Diff line number Diff line
@@ -197,7 +197,6 @@ void core_coder_mode_switch_ivas_fx( 
    {

        st->igf = getIgfPresent_fx( st->element_mode, st->total_brate, st->bwidth, st->rf_mode );

        move16();

#ifdef IVAS_FLOAT_FIXED_CONVERSIONS

        IF( hTcxEnc != NULL )

        {

            hTcxEnc->L_frameTCX = extract_l( Mult_32_16( st->input_Fs, 0x0290 ) ); /* 0x0290  is 1/FRAMES_PER_SEC in Q15*/
@@ -205,11 +204,14 @@ void core_coder_mode_switch_ivas_fx( 
        }

        st->currEnergyHF_fx = 0;

        move32();



        /* Initialize Signal Buffers */

        f2me_buf_16( st->old_inp_12k8, st->old_inp_12k8_fx, &st->exp_old_inp_12k8, L_INP_MEM );

        f2me_buf_16( st->old_inp_16k, st->old_inp_16k_fx, &st->exp_old_inp_16k, L_INP_MEM );

#endif

        Word16 shift = getScaleFactor16( st->old_inp_16k_fx, L_INP_MEM );

        Scale_sig( st->old_inp_16k_fx, L_INP_MEM, shift );

        st->exp_old_inp_16k = sub( st->exp_old_inp_16k, shift );

        move16();

        shift = getScaleFactor16( st->old_inp_12k8_fx, L_INP_MEM );

        Scale_sig( st->old_inp_12k8_fx, L_INP_MEM, shift );

        st->exp_old_inp_12k8 = sub( st->exp_old_inp_12k8, shift );

        move16();

        init_coder_ace_plus_ivas_fx( st, last_total_brate, MCT_flag );

        IF( st->hLPDmem != NULL )

        {

lib_enc/fd_cng_enc.c

+20 −11
Original line number Diff line number Diff line
@@ -701,9 +701,18 @@ void FdCng_encodeSID_ivas_fx( 
    move32();

    FOR( i = 0; i < N; i++ )

    {

        v_fx[i] = Mpy_32_32( 1342177280, BASOP_Util_Log10( L_add_sat( E_fx[i], L_shr( 214748 /* 1e-4f in Q31*/, hFdCngEnc->msNoiseEst_fx_exp ) ), hFdCngEnc->msNoiseEst_fx_exp ) ); // Q21 = 27+25-31

        IF( E_fx[i] == 0 )

        {

            /* 10 * log(1e-4) = 10 * (-4) = -40 */

            v_fx[i] = -41943040; // -40.0 in Q20

            move32();

        }

        ELSE

        {

            v_fx[i] = Mpy_32_32( 671088640 /*10 in Q26*/, BASOP_Util_Log10( E_fx[i], hFdCngEnc->msNoiseEst_fx_exp ) ); // Q20 = 26+25-31

            move32();

        e_fx = L_add( e_fx, v_fx[i] ); // Q21

        }

        e_fx = L_add( e_fx, v_fx[i] ); // Q20

    }



    /* Normalize MSVQ input */
@@ -711,19 +720,19 @@ void FdCng_encodeSID_ivas_fx( 
    move32();

    FOR( i = N_GAIN_MIN; i < N_GAIN_MAX; i++ )

    {

        gain_fx = L_add( gain_fx, v_fx[i] ); // Q21

        gain_fx = L_add( gain_fx, v_fx[i] ); // Q20

    }



    /*gain /= (float) ( N_GAIN_MAX - N_GAIN_MIN );*/

    gain_fx = Mpy_32_32( gain_fx, 165191050 /* 1/13 in Q31*/ ); // Q21

    gain_fx = Mpy_32_32( gain_fx, 165191050 /* 1/13 in Q31*/ ); // Q20



    FOR( i = 0; i < N; i++ )

    {

        v_fx[i] = L_sub( v_fx[i], gain_fx ); // Q21

        v_fx[i] = L_sub( v_fx[i], gain_fx ); // Q20

        move32();

    }



    v_e = 10; // Q21

    v_e = 11; // Q20

    move16();



    /* MSVQ encoder */
@@ -740,12 +749,12 @@ void FdCng_encodeSID_ivas_fx( 
            /* truncated DCT21 analysis */

            create_IDCT_N_Matrix_fx( invTrfMatrix_fx, N, FDCNG_VQ_DCT_MAXTRUNC, sizeof( tmpRAM_fx ) / ( sizeof( Word32 ) ) ); // Q31



            dctT2_N_apply_matrix_fx( v_fx /*Q21*/, dct_target_fx, FDCNG_VQ_DCT_MAXTRUNC, N, invTrfMatrix_fx, FDCNG_VQ_DCT_MAXTRUNC, DCT_T2_21_XX ); // Q21

            dctT2_N_apply_matrix_fx( v_fx /*Q20*/, dct_target_fx, FDCNG_VQ_DCT_MAXTRUNC, N, invTrfMatrix_fx, FDCNG_VQ_DCT_MAXTRUNC, DCT_T2_21_XX ); // Q20



            /* truncated IDCT21 extension to 24 bands  */

            extend_dctN_input_fx( v_fx, dct_target_fx, N, tot_sig_ext_fx, FDCNG_VQ_MAX_LEN, invTrfMatrix_fx, FDCNG_VQ_DCT_MAXTRUNC, IDCT_T2_XX_21 ); // Q21

            extend_dctN_input_fx( v_fx, dct_target_fx, N, tot_sig_ext_fx, FDCNG_VQ_MAX_LEN, invTrfMatrix_fx, FDCNG_VQ_DCT_MAXTRUNC, IDCT_T2_XX_21 ); // Q20



            Copy32( tot_sig_ext_fx, v_fx, FDCNG_VQ_MAX_LEN ); /*  write  extended result as input to  VQ stage #1 */ // Q21

            Copy32( tot_sig_ext_fx, v_fx, FDCNG_VQ_MAX_LEN ); /*  write  extended result as input to  VQ stage #1 */ // Q20

        }



        create_IDCT_N_Matrix_fx( invTrfMatrix_fx, FDCNG_VQ_MAX_LEN, FDCNG_VQ_DCT_MAXTRUNC, sizeof( tmpRAM_fx ) / ( sizeof( Word32 ) ) ); // Q31
@@ -773,7 +782,7 @@ void FdCng_encodeSID_ivas_fx( 
        gain_fx = L_add( gain_fx, v_fx[i] ); // Q = 31 - v_e

    }



    e_fx = L_shl( e_fx, sub( 10, v_e ) );                           // Q = 31 - v_e

    e_fx = L_shl( e_fx, sub( 11, v_e ) );                           // Q = 31 - v_e

    gain_fx = Mpy_32_16_1( L_sub( e_fx, gain_fx ), div_s( 1, N ) ); // Q = 31 - v_e

    gain_fx = L_shl( gain_fx, sub( v_e, 8 ) );                      // Q23

lib_enc/init_enc.c

+21 −13
Original line number Diff line number Diff line
@@ -446,9 +446,8 @@ ivas_error init_encoder( 
        st->Bin_E_old = st->hSignalBuf->Bin_E_old;

        st->mem_decim = st->hSignalBuf->mem_decim_flt;

        st->mem_decim16k = st->hSignalBuf->mem_decim16k_flt;

#endif

        st->old_inp_12k8 = st->hSignalBuf->old_inp_12k8_flt;

        st->old_inp_16k = st->hSignalBuf->old_inp_16k_flt;

#endif

#ifdef IVAS_FLOAT_FIXED

        st->mem_decim_fx = st->hSignalBuf->mem_decim_fx;

        st->mem_decim16k_fx = st->hSignalBuf->mem_decim16k_fx;
@@ -456,6 +455,7 @@ ivas_error init_encoder( 
        st->old_inp_16k_fx = st->hSignalBuf->old_inp_16k_fx;

#endif

#ifndef IVAS_FLOAT_FIXED

        st->old_inp_16k = st->hSignalBuf->old_inp_16k_flt;

        st->buf_synth_flt = st->hSignalBuf->buf_synth_flt;

        st->buf_speech_enc_flt = st->hSignalBuf->buf_speech_enc_flt;

        st->buf_wspeech_enc_flt = st->hSignalBuf->buf_wspeech_enc_flt;
@@ -479,12 +479,13 @@ ivas_error init_encoder( 
        set_f( st->Bin_E_old, 0, L_FFT / 2 );

        set_f( st->mem_decim, 0, 2 * L_FILT_MAX );

        set_f( st->mem_decim16k, 0, 2 * L_FILT_MAX );

#endif

        set_f( st->old_inp_12k8, 0, L_INP_MEM );

        set_f( st->old_inp_16k, 0, L_INP_MEM );

        set_f( st->old_inp_12k8, 0, L_INP_MEM );

#endif

#ifdef IVAS_FLOAT_FIXED

        set16_fx( st->old_inp_12k8_fx, 0, L_INP_MEM );

        set16_fx( st->old_inp_16k_fx, 0, L_INP_MEM );

        st->exp_old_inp_16k = 0;

        st->input_buff_fx = st->hSignalBuf->input_buff;

        st->input_buff32_fx = st->hSignalBuf->input_buff32;

        set32_fx( st->hSignalBuf->input_buff32, 0, L_FRAME48k + L_FRAME48k + NS2SA( 48000, DELAY_FIR_RESAMPL_NS ) );
@@ -533,9 +534,9 @@ ivas_error init_encoder( 
        st->mem_decim_fx = NULL;

        st->mem_decim16k_fx = NULL;

#endif

#ifndef IVAS_FLOAT_FIXED

        st->old_inp_12k8 = NULL;

        st->old_inp_16k = NULL;

#ifndef IVAS_FLOAT_FIXED

        st->buf_synth_flt = NULL;

        st->buf_speech_enc_flt = NULL;

        st->buf_speech_enc = NULL;
@@ -1629,6 +1630,10 @@ ivas_error init_encoder_ivas_fx( 
        set16_fx( st->mem_decim16k_fx, 0, shl( L_FILT_MAX, 1 ) );

        set16_fx( st->old_inp_12k8_fx, 0, L_INP_MEM );

        set16_fx( st->old_inp_16k_fx, 0, L_INP_MEM );

        st->exp_old_inp_16k = 0;

        st->exp_old_inp_12k8 = 0;

        move16();

        move16();



        st->input_buff_fx = st->hSignalBuf->input_buff;

        st->input_buff32_fx = st->hSignalBuf->input_buff32;
@@ -1655,16 +1660,14 @@ ivas_error init_encoder_ivas_fx( 
            st->input_fx = st->input_buff_fx + frame_length;

        }

#if 1 // TODO: To be removed later

        st->old_inp_12k8 = st->hSignalBuf->old_inp_12k8_flt;

        st->old_inp_16k = st->hSignalBuf->old_inp_16k_flt;

#ifndef IVAS_FLOAT_FIXED

        st->old_inp_12k8 = st->hSignalBuf->old_inp_12k8_flt;

        st->buf_synth_flt = st->hSignalBuf->buf_synth_flt;

        st->buf_speech_enc_flt = st->hSignalBuf->buf_speech_enc_flt;

        st->buf_wspeech_enc_flt = st->hSignalBuf->buf_wspeech_enc_flt;

#endif

        set_f( st->old_inp_12k8, 0, L_INP_MEM );

        set_f( st->old_inp_16k, 0, L_INP_MEM );

#ifndef IVAS_FLOAT_FIXED

        set_f( st->old_inp_12k8, 0, L_INP_MEM );

        set_f( st->hSignalBuf->buf_speech_enc_flt, 0, L_PAST_MAX_32k + L_FRAME32k + L_NEXT_MAX_32k );

        set_f( st->hSignalBuf->buf_wspeech_enc_flt, 0, L_FRAME16k + L_SUBFR + L_FRAME16k + L_NEXT_MAX_16k + 320 );

#endif
@@ -1684,9 +1687,9 @@ ivas_error init_encoder_ivas_fx( 
        st->buf_wspeech_enc = NULL;

        st->input_buff_fx = NULL;

#if 1

#ifndef IVAS_FLOAT_FIXED

        st->old_inp_12k8 = NULL;

        st->old_inp_16k = NULL;

#ifndef IVAS_FLOAT_FIXED

        st->buf_synth_flt = NULL;

        st->buf_speech_enc_flt = NULL;

        st->buf_wspeech_enc_flt = NULL;
@@ -2328,9 +2331,14 @@ ivas_error init_encoder_ivas_fx( 
#endif



    /* Initialize Signal Buffers */

    f2me_buf_16( st->old_inp_12k8, st->old_inp_12k8_fx, &st->exp_old_inp_12k8, L_INP_MEM );

    f2me_buf_16( st->old_inp_16k, st->old_inp_16k_fx, &st->exp_old_inp_16k, L_INP_MEM );



    Word16 shift = getScaleFactor16( st->old_inp_16k_fx, L_INP_MEM );

    Scale_sig( st->old_inp_16k_fx, L_INP_MEM, shift );

    st->exp_old_inp_16k = sub( st->exp_old_inp_16k, shift );

    move16();

    shift = getScaleFactor16( st->old_inp_12k8_fx, L_INP_MEM );

    Scale_sig( st->old_inp_12k8_fx, L_INP_MEM, shift );

    st->exp_old_inp_12k8 = sub( st->exp_old_inp_12k8, shift );

    move16();

    /* Initialize ACELP */

#endif

    init_coder_ace_plus_ivas_fx( st, st->last_total_brate, 0 );

lib_enc/ivas_core_pre_proc.c

+8 −7
Original line number Diff line number Diff line
@@ -542,8 +542,14 @@ ivas_error pre_proc_ivas( 
    Word16 Q_old_inp_16k = -1;

    Word16 Q_r[2] = { 0 };

    inp_16k_fx = old_inp_16k_fx + L_INP_MEM - L_look;

    f2me_buf_16( st->old_inp_12k8, st->old_inp_12k8_fx, &st->exp_old_inp_12k8, 240 );

    floatToFixed_arr16( st->old_inp_16k, st->old_inp_16k_fx, Q_old_inp_16k, L_INP_MEM );



    Word16 shift = getScaleFactor16( st->old_inp_12k8_fx, 240 );

    Scale_sig( st->old_inp_12k8_fx, 240, shift );

    st->exp_old_inp_12k8 = sub( st->exp_old_inp_12k8, shift );

    move16();



    Scale_sig( st->old_inp_16k_fx, L_INP_MEM, sub( Q_old_inp_16k, sub( Q15, Q_old_inp_16k ) ) );

    st->exp_old_inp_16k = sub( Q15, Q_old_inp_16k );

    if ( st->hBWE_FD != NULL )

    {

        Scale_sig( st->hBWE_FD->L_old_wtda_swb_fx, L_FRAME48k, sub( -1, st->Q_old_wtda ) );
@@ -621,11 +627,6 @@ ivas_error pre_proc_ivas( 
    st->exp_buf_speech_enc_pe = sub( Q15, *Q_new - 1 );

    move16();



#ifdef IVAS_FLOAT_FIXED_CONVERSIONS

    // No need to do fixed to float for signal_in_fx



    fixedToFloat_arr( st->old_inp_16k_fx, st->old_inp_16k, *Q_new - 1, L_INP_MEM );

#endif

    /* set the pointer of the current frame for the ACELP core */

    IF( EQ_16( st->L_frame, L_FRAME ) )

    {