Merge branch '1227-basop-encoder-improve-wmops-performance-of-ivas_band_cov' into 'main' (5bb960bf) · Commits · SA4 / Audio / IVAS BASOP

lib_enc/ivas_enc_cov_handler.c

+21 −14

Original line number	Diff line number	Diff line
		@@ -513,9 +513,11 @@ static void ivas_band_cov_fx(
		Word32 pV_re[L_FRAME48k];
		Word64 pV_re_64bit[L_FRAME48k];
		Word64 cov_real_64bit[IVAS_SPAR_MAX_CH][IVAS_SPAR_MAX_CH][IVAS_MAX_NUM_BANDS];
		Word16 q_shift;
		Word16 q_shift, q_shift_tmp;
		Word16 m, start_bin, active_bins;
		Word16 num_blocks;

		num_blocks = idiv1616( num_bins, stride ); /* Q0 */
		FOR( i = 0; i < num_chans; i++ )
		{
		FOR( j = i; j < num_chans; j++ )
		@@ -536,14 +538,16 @@ static void ivas_band_cov_fx(
		move16();
		FOR( k = 0; k < num_bins; k++ )
		{
		IF( pV_re_64bit[k] != 0 )
		q_shift_tmp = W_norm( pV_re_64bit[k] );
		if ( pV_re_64bit[k] != 0 )
		{
		q_shift = s_min( q_shift, W_norm( pV_re_64bit[k] ) );
		q_shift = s_min( q_shift, q_shift_tmp );
		}
		}
		q_shift_tmp = sub( q_shift, 32 );
		FOR( k = 0; k < num_bins; k++ )
		{
		pV_re[k] = W_extract_l( W_shl_nosat( pV_re_64bit[k], sub( q_shift, 32 ) ) ); //(q_In_FR[i1] + q_In_FR[j1]) + (q_shift - 32)
		pV_re[k] = W_extract_l( W_shl_nosat( pV_re_64bit[k], q_shift_tmp ) ); //(q_In_FR[i1] + q_In_FR[j1]) + (q_shift - 32)
		move32();
		/* perform rounding towards lower value for negative results */
		if ( pV_re[k] < 0 )
		@@ -556,31 +560,32 @@ static void ivas_band_cov_fx(
		{
		Word64 temp;
		const Word32 *p_bin_to_band = pFb_bin_to_band[k]; // Q22
		Word32 *cov_ptr = pV_re;
		Word16 num_blocks;
		Word32 *cov_ptr;
		Word16 blk;

		temp = 0;
		move64();
		num_blocks = idiv1616( num_bins, stride ); /* Q0 */
		move16();
		start_bin = pFb_start_bin_per_band[k]; /* Q0 */
		move16();
		active_bins = pFb_active_bins_per_band[k]; /* Q0 */
		move16();

		cov_ptr = &pV_re[start_bin];
		move16();
		FOR( blk = 0; blk < num_blocks; blk++ )
		{
		/* optional: add temporal weight here */
		FOR( m = start_bin; m < add( start_bin, active_bins ); m++ )
		FOR( m = 0; m < active_bins; m++ )
		{
		temp = W_add( temp, W_mult0_32_32( cov_ptr[m], p_bin_to_band[sub( m, start_bin )] ) ); // ((q_In_FR[i1] + q_In_FR[j1]) + (q_shift - 32), Q22) -> (q_In_FR[i1] + q_In_FR[j1] + (q_shift - 10)
		temp = W_add( temp, W_mult0_32_32( cov_ptr[m], p_bin_to_band[m] ) ); // ((q_In_FR[i1] + q_In_FR[j1]) + (q_shift - 32), Q22) -> (q_In_FR[i1] + q_In_FR[j1] + (q_shift - 10)
		}
		cov_ptr += stride;
		move16();
		}
		// What basop to add below????
		cov_real_64bit[i][j][k] = temp * (Word64) ( num_blocks ); // (q_In_FR[i1] + q_In_FR[j1] + (q_shift - 10) - guard_bits
		move64();
		move64(); // conservative estimation of a 64 bit multiplication
		}
		q_cov_real[i][j] = add( add( q_In_FR, q_In_FR ), sub( q_shift, Q10 ) );
		move16();
		@@ -595,17 +600,19 @@ static void ivas_band_cov_fx(
		move16();
		FOR( k = start_band; k < end_band; k++ )
		{
		IF( cov_real_64bit[i][j][k] != 0 )
		q_shift_tmp = W_norm( cov_real_64bit[i][j][k] );
		if ( cov_real_64bit[i][j][k] != 0 )
		{
		q_shift = s_min( q_shift, W_norm( cov_real_64bit[i][j][k] ) );
		q_shift = s_min( q_shift, q_shift_tmp );
		}
		}
		q_shift_tmp = sub( q_shift, 32 );
		FOR( k = start_band; k < end_band; k++ )
		{
		cov_real[i][j][k] = W_extract_l( W_shl_nosat( cov_real_64bit[i][j][k], sub( q_shift, 32 ) ) ); /* q_cov_real[i][j] + q_shift - 32 */
		cov_real[i][j][k] = W_extract_l( W_shl_nosat( cov_real_64bit[i][j][k], q_shift_tmp ) ); /* q_cov_real[i][j] + q_shift - 32 */
		move32();
		}
		q_cov_real[i][j] = add( q_cov_real[i][j], sub( q_shift, 32 ) );
		q_cov_real[i][j] = add( q_cov_real[i][j], q_shift_tmp );
		move16();
		}
		}