rewritten all cov_smooth loops with better precision - with optional debug output (a6ea0a05) · Commits · SA4 / Audio / IVAS BASOP

lib_com/ivas_cov_smooth_fx.c

100644 → 100755

+106 −33

Original line number	Diff line number	Diff line
		@@ -42,6 +42,13 @@
		-----------------------------------------------------------------------------------------/

		#define BAND_SMOOTH_REST_START_IDX ( 2 )
		#define ONEeN20_Q97 0x5E728433 // 1e-20 in Q97 1.584.563.251

		//#define DEBUG_ivas_compute_smooth_cov_fx
		#ifdef DEBUG_ivas_compute_smooth_cov_fx
		#include <math.h>
		#endif


		/-----------------------------------------------------------------------------------------
		* Function ivas_calculate_update_factor_fx()
		@@ -87,8 +94,14 @@ static void ivas_calculate_smoothning_factor_fx(
		move16();

		tmp = BASOP_Util_Divide3232_Scale( update_factor, L_shl( L_deposit_l( min_pool_size ), Q22 ), &exp_diff ); // Q(31 - exp_diff)
		*Smoothing_factor = L_shl_sat( L_deposit_l( tmp ), add( Q16, exp_diff ) ); // Q31
		*Smoothing_factor = L_shl_sat( L_deposit_h( tmp ), exp_diff ); // Q31
		move32();
		#ifdef DEBUG_ivas_compute_smooth_cov_fx
		printf( "Smoothing_factor %g update_factor %g min_pool_size %d j=%d\n",
		(double) Smoothing_factor pow (2.0, -31),
		(double) update_factor + pow (2.0, -22.0), min_pool_size, j );
		#endif


		IF( NE_32( smooth_mode, COV_SMOOTH_MC ) )
		{
		@@ -264,19 +277,17 @@ void ivas_spar_covar_smooth_enc_close_fx(
		return;
		}


		/-----------------------------------------------------------------------------------------
		* Function ivas_compute_smooth_cov_fx()
		*
		* Compute smooth covariance real/imag.
		-----------------------------------------------------------------------------------------/

		static void ivas_compute_smooth_cov_fx(
		ivas_cov_smooth_state_t *hCovState,
		ivas_filterbank_t *pFb,
		Word32 *pCov_buf[IVAS_SPAR_MAX_CH][IVAS_SPAR_MAX_CH], // i/o: Q(q_cov[i][j])
		Word32 *pPrior_cov_buf[IVAS_SPAR_MAX_CH][IVAS_SPAR_MAX_CH], // i: hCovState->q_cov_real_per_band[i][j][k]
		const Word32 fac,
		const Word32 fac, // i: ONEeN20_Q97 (1e-20 in Q97)
		const Word16 start_band,
		const Word16 end_band,
		const Word16 num_ch,
		@@ -285,10 +296,15 @@ static void ivas_compute_smooth_cov_fx(
		{
		Word16 i, j, k;
		Word16 prev_idx = hCovState->prior_bank_idx;
		Word32 factor = 0, L_tmp, L_tmp1;
		Word32 factor = 0, L_tmp0, L_tmp1;

		const Word16 q_fac = 97;
		const Word16 fac_e = sub(Q31, q_fac);

		Word16 cov_buf_e;

		Word16 sm_b;
		Word16 non_sm_b_idx;
		Word16 q_tmp[IVAS_MAX_NUM_BANDS];

		sm_b = BAND_SMOOTH_REST_START_IDX;
		move16();
		@@ -296,7 +312,9 @@ static void ivas_compute_smooth_cov_fx(
		move32();

		assert( end_band <= pFb->filterbank_num_bands );

		#ifdef DEBUG_ivas_compute_smooth_cov_fx
		printf( "\n" );
		#endif
		test();
		IF( EQ_16( prev_idx, -1 ) \|\| EQ_16( transient_det[1], 1 ) )
		{
		@@ -309,11 +327,28 @@ static void ivas_compute_smooth_cov_fx(
		}
		FOR( i = 0; i < num_ch; i++ )
		{
		#ifdef DEBUG_ivas_compute_smooth_cov_fx
		Word16 cov_buf_e = sub( Q31, q_cov[i][i] );
		#endif
		FOR( k = start_band; k < end_band; k++ )
		{
		L_tmp = Mpy_32_32( hCovState->pSmoothing_factor_fx[k], fac ); // (Q31, Q31) -> Q31
		pCov_buf[i][i][k] = L_add( pCov_buf[i][i][k], L_shl( L_tmp, sub( hCovState->q_cov_real_per_band[i][i][k], Q31 ) ) ); // hCovState->q_cov_real_per_band[i][j][k]
		/* ref: pCov_buf[i][i][k] += ( hCovState->pSmoothing_factor[k] * fac ); */
		#ifdef DEBUG_ivas_compute_smooth_cov_fx
		Word32 L_buf = pCov_buf[i][i][k]; /* debug !! */
		#endif
		pCov_buf[i][i][k] = BASOP_Util_Add_Mant32Exp( pCov_buf[i][i][k], cov_buf_e, Mpy_32_32( hCovState->pSmoothing_factor_fx[k], fac ), fac_e, &cov_buf_e );
		move32();
		hCovState->q_cov_real_per_band[i][i][k] = sub( Q31, cov_buf_e );
		move16();
		#ifdef DEBUG_ivas_compute_smooth_cov_fx
		printf( "A: pCov_buf[%d][%d][%d] %g (0x%08X 0x%04X) fac %g sm-factor %g buf %g\n", i, i, k,
		(double) pCov_buf[i][i][k] * pow( 2.0, (double) -hCovState->q_cov_real_per_band[i][i][k] ),
		pCov_buf[i][i][k],
		hCovState->q_cov_real_per_band[i][i][k] & 0xFFFF,
		(double) fac * pow( 2.0, (double) -q_fac ),
		(double) hCovState->pSmoothing_factor_fx[k] *pow (2.0, -31.0),
		(double) L_buf * pow(2.0, (double) -q_cov[i][i]));
		#endif
		}
		}
		}
		@@ -326,38 +361,63 @@ static void ivas_compute_smooth_cov_fx(
		{
		IF( EQ_16( i, j ) )
		{
		factor = fac; // Q31
		factor = fac; // Q97
		move32();
		}
		ELSE
		{
		factor = 0;
		factor = 0; // Q97
		move32();
		}

		set16_fx( q_tmp, q_cov[i][j], sub( end_band, start_band ) );
		FOR( k = start_band; k < non_sm_b_idx; k++ )
		{
		L_tmp = Mpy_32_32( hCovState->pSmoothing_factor_fx[k], pCov_buf[i][j][k] ); // (Q31, q_cov[i][j]) -> q_cov[i][j]
		L_tmp1 = Mpy_32_32( L_sub( ONE_IN_Q31, hCovState->pSmoothing_factor_fx[k] ), pPrior_cov_buf[i][j][k] ); // (Q31, hCovState->q_cov_real_per_band[i][j][k]) -> hCovState->q_cov_real_per_band[i][j][k]
		pCov_buf[i][j][k] = BASOP_Util_Add_Mant32Exp( L_tmp, sub( Q31, q_cov[i][j] ), L_tmp1, sub( Q31, hCovState->q_prior_cov_real_per_band[i][j][k] ), &q_tmp[k] ); // Q(31 - q_tmp[k])
		/* ref: pCov_buf[i][j][k] = pPrior_cov_buf[i][j][k] + ( hCovState->pSmoothing_factor[k] * ( pCov_buf[i][j][k] - pPrior_cov_buf[i][j][k] + factor ) ); */
		/* mod: pCov_buf[i][j][k] = pCov_buf[i][j][k] * hCovState->pSmoothing_factor[k] +
		pPrior_cov_buf[i][j][k] * (1.0 - hCovState->pSmoothing_factor[k]) +
		factor * hCovState->pSmoothing_factor[k] ); */
		#ifdef DEBUG_ivas_compute_smooth_cov_fx
		Word32 L_buf = pCov_buf[i][j][k];
		#endif
		L_tmp0 = Mpy_32_32( hCovState->pSmoothing_factor_fx[k], pCov_buf[i][j][k] ); // q_tmp0: q_cov[i][j] // (Q31, q_cov[i][j]) -> q_cov[i][j]
		L_tmp1 = Mpy_32_32( L_sub( ONE_IN_Q31, hCovState->pSmoothing_factor_fx[k] ), pPrior_cov_buf[i][j][k] ); // q_tmp1: hCovState->q_prior_cov_real_per_band[i][j][k]
		L_tmp0 = BASOP_Util_Add_Mant32Exp( L_tmp0, sub( Q31, q_cov[i][j] ), L_tmp1, sub( Q31, hCovState->q_prior_cov_real_per_band[i][j][k] ), &cov_buf_e );
		pCov_buf[i][j][k] = BASOP_Util_Add_Mant32Exp( L_tmp0, cov_buf_e, Mpy_32_32( hCovState->pSmoothing_factor_fx[k], factor ), fac_e, &cov_buf_e );
		move32();
		q_tmp[k] = sub( Q31, q_tmp[k] ); // Q of pConv_buf[i][j][k] is now q_tmp[k]
		hCovState->q_cov_real_per_band[i][j][k] = sub( Q31, cov_buf_e ); // Q of pCov_buf[i][j][k]
		move16();
		L_tmp = L_shl( Mpy_32_32( hCovState->pSmoothing_factor_fx[k], factor ), sub( q_tmp[k], Q31 ) ); // ((Q31, Q31) -> Q31) -> q_tmp[k]
		pCov_buf[i][j][k] = L_add( pCov_buf[i][j][k], L_tmp ); // q_tmp[k]
		move32();

		#ifdef DEBUG_ivas_compute_smooth_cov_fx
		printf( "B1: pCov_buf[%d][%d][%d] %g fac %g sm-factor %g prior %g buf %g (q=%d)\n", i, j, k,
		(double) pCov_buf[i][j][k] * pow( 2.0, (double) -hCovState->q_cov_real_per_band[i][j][k] ),
		(double) factor * pow( 2.0, (double) -q_fac ),
		(double) hCovState->pSmoothing_factor_fx[k] * pow( 2.0, -Q31 ),
		(double) pPrior_cov_buf[i][j][k] * pow( 2.0, (double) -hCovState->q_prior_cov_real_per_band[i][j][k] ),
		(double) L_buf * pow( 2.0, (double) -q_cov[i][j] ), q_cov[i][j] );
		#endif

		}
		Copy( q_tmp, hCovState->q_cov_real_per_band[i][j], sub( end_band, start_band ) ); // Q of pCov_buf[i][j][k] is hCovState->q_cov_real_per_band[i][j][j]
		}
		}
		FOR( i = 0; i < num_ch; i++ )
		{
		FOR( k = non_sm_b_idx; k < end_band; k++ )
		{
		L_tmp = Mpy_32_32( hCovState->pSmoothing_factor_fx[k], fac ); // (Q31, Q31) -> Q31
		pCov_buf[i][i][k] = L_add( pCov_buf[i][i][k], L_shl( L_tmp, sub( hCovState->q_cov_real_per_band[i][i][k], Q31 ) ) ); // hCovState->q_cov_real_per_band[i][j][j]
		/* ref: pCov_buf[i][i][k] += ( hCovState->pSmoothing_factor[k] * fac ); */
		#ifdef DEBUG_ivas_compute_smooth_cov_fx
		Word32 L_buf = pCov_buf[i][i][k];
		#endif
		pCov_buf[i][i][k] = BASOP_Util_Add_Mant32Exp( pCov_buf[i][i][k], sub( Q31, q_cov[i][i] ), Mpy_32_32( hCovState->pSmoothing_factor_fx[k], fac ), fac_e , &cov_buf_e);
		move32();
		hCovState->q_cov_real_per_band[i][i][k] = sub( Q31, cov_buf_e ); // Q of pCov_buf[i][i][k]
		move16();
		#ifdef DEBUG_ivas_compute_smooth_cov_fx
		printf( "B2: pCov_buf[%d][%d][%d] %g fac %g sm-factor %g buf %g (q=%d)\n", i, i, k,
		(double) pCov_buf[i][i][k] * pow( 2.0, (double) -hCovState->q_cov_real_per_band[i][i][k] ),
		(double) fac * pow( 2.0, (double) -q_fac ),
		(double) hCovState->pSmoothing_factor_fx[k] * pow( 2.0, -Q31 ),
		(double) L_buf * pow( 2.0, (double) -q_cov[i][i] ), q_cov[i][i]);
		#endif
		}
		}
		}
		@@ -369,28 +429,41 @@ static void ivas_compute_smooth_cov_fx(
		{
		IF( EQ_16( i, j ) )
		{
		factor = fac; // Q31
		factor = fac; // Q97
		move32();
		}
		ELSE
		{
		factor = 0;
		factor = 0; // Q97
		move32();
		}

		FOR( k = start_band; k < end_band; k++ )
		{
		L_tmp = Mpy_32_32( hCovState->pSmoothing_factor_fx[k], pCov_buf[i][j][k] ); // (Q31, q_cov[i][j]) -> q_cov[i][j]
		L_tmp1 = Mpy_32_32( L_sub( ONE_IN_Q31, hCovState->pSmoothing_factor_fx[k] ), pPrior_cov_buf[i][j][k] ); // (Q31, hCovState->q_cov_real_per_band[i][j][k]) -> hCovState->q_cov_real_per_band[i][j][k]
		pCov_buf[i][j][k] = BASOP_Util_Add_Mant32Exp( L_tmp, sub( Q31, q_cov[i][j] ), L_tmp1, sub( Q31, hCovState->q_prior_cov_real_per_band[i][j][k] ), &q_tmp[k] ); // Q(31 - q_tmp[k])
		/* ref: pCov_buf[i][j][k] = pPrior_cov_buf[i][j][k] + ( hCovState->pSmoothing_factor[k] * ( pCov_buf[i][j][k] - pPrior_cov_buf[i][j][k] + factor ) ); */
		/* mod: pCov_buf[i][j][k] = pCov_buf[i][j][k] * hCovState->pSmoothing_factor[k] +
		pPrior_cov_buf[i][j][k] * (1.0 - hCovState->pSmoothing_factor[k]) +
		factor * hCovState->pSmoothing_factor[k] ); */
		#ifdef DEBUG_ivas_compute_smooth_cov_fx
		Word32 L_buf = pCov_buf[i][j][k]; /* debug !! */
		#endif
		L_tmp0 = Mpy_32_32( hCovState->pSmoothing_factor_fx[k], pCov_buf[i][j][k] ); // q_tmp0: q_cov[i][j] // (Q31, q_cov[i][j]) -> q_cov[i][j]
		L_tmp1 = Mpy_32_32( L_sub( ONE_IN_Q31, hCovState->pSmoothing_factor_fx[k] ), pPrior_cov_buf[i][j][k] ); // q_tmp1: hCovState->q_prior_cov_real_per_band[i][j][k]
		L_tmp0 = BASOP_Util_Add_Mant32Exp( L_tmp0, sub( Q31, q_cov[i][j] ), L_tmp1, sub( Q31, hCovState->q_prior_cov_real_per_band[i][j][k] ), &cov_buf_e );
		pCov_buf[i][j][k] = BASOP_Util_Add_Mant32Exp( L_tmp0, cov_buf_e, Mpy_32_32( hCovState->pSmoothing_factor_fx[k], factor ), fac_e, &cov_buf_e );
		move32();
		q_tmp[k] = sub( Q31, q_tmp[k] ); // Q of pConv_buf[i][j][k] is now q_tmp[k]
		hCovState->q_cov_real_per_band[i][j][k] = sub( Q31, cov_buf_e ); // Q of pCov_buf[i][j][k]
		move16();
		L_tmp = L_shl( Mpy_32_32( hCovState->pSmoothing_factor_fx[k], factor ), sub( q_tmp[k], Q31 ) ); // ((Q31, Q31) -> Q31) -> q_tmp[k]
		pCov_buf[i][j][k] = L_add( pCov_buf[i][j][k], L_tmp ); // q_tmp[k]
		move32();
		#ifdef DEBUG_ivas_compute_smooth_cov_fx
		printf( "C: pCov_buf[%d][%d][%d] %g factor %g sm-factor %g prior %g buf %g (q=%d)\n", i, j, k,
		(double) pCov_buf[i][j][k] * pow(2.0, (double) -hCovState->q_cov_real_per_band[i][j][k]),
		(double) factor * pow(2.0, (double) -q_fac),
		(double) hCovState->pSmoothing_factor_fx[k] * pow(2.0, -Q31),
		(double) pPrior_cov_buf[i][j][k] * pow(2.0, (double) -hCovState->q_prior_cov_real_per_band[i][j][k] ),
		(double) L_buf * pow( 2.0, (double) -q_cov[i][j] ), q_cov[i][j] );
		#endif

		}
		Copy( q_tmp, hCovState->q_cov_real_per_band[i][j], sub( end_band, start_band ) ); // Q of pCov_buf[i][j][k] is hCovState->q_cov_real_per_band[i][j][j]
		}
		}
		}
		@@ -418,7 +491,7 @@ void ivas_cov_smooth_process_fx(
		Word16 i, j, k;
		Word16 num_bands = sub( end_band, start_band );

		ivas_compute_smooth_cov_fx( hCovState, pFb, cov_real, hCovState->pPrior_cov_real_fx, 0, start_band, end_band, num_ch, transient_det, q_cov );
		ivas_compute_smooth_cov_fx( hCovState, pFb, cov_real, hCovState->pPrior_cov_real_fx, ONEeN20_Q97, start_band, end_band, num_ch, transient_det, q_cov );

		FOR( i = 0; i < num_ch; i++ )
		{