From 5d560b63e6fb85b8051a6b1eb452f2e34a37654e Mon Sep 17 00:00:00 2001 From: Sandesh Venkatesh Date: Tue, 13 Aug 2024 10:30:30 +0530 Subject: [PATCH] MLD improvement changes [x] Fixes in PCA tools [x] Bug fix in ProtoComputation2 --- lib_com/ivas_pca_tools.c | 142 +++++++++++++++++++++++++------------ lib_rend/ivas_dirac_rend.c | 27 ++++--- 2 files changed, 110 insertions(+), 59 deletions(-) diff --git a/lib_com/ivas_pca_tools.c b/lib_com/ivas_pca_tools.c index 9ee54da4f..5966d9dc5 100644 --- a/lib_com/ivas_pca_tools.c +++ b/lib_com/ivas_pca_tools.c @@ -560,6 +560,22 @@ void dquat2mat( return; } +static Word16 check_bound( Word32 tmp ) +{ + IF( GT_32( tmp, MAX16B ) ) + { + return MAX16B; + } + ELSE IF( LT_32( tmp, MIN16B ) ) + { + return MIN16B; + } + ELSE + { + return extract_l( tmp ); + } +} + #ifdef IVAS_FLOAT_FIXED void dquat2mat_fx( const Word16 *ql, @@ -605,37 +621,37 @@ void dquat2mat_fx( dx = mult( d, x ); dy = mult( d, y ); dz = mult( d, z ); - m[0] = sub( sub( aw, bx ), add( cy, dz ) ); + m[0] = check_bound( L_sub( L_sub( aw, bx ), L_add( cy, dz ) ) ); move16(); - m[1] = sub( sub( cz, dy ), add( ax, bw ) ); + m[1] = check_bound( L_sub( L_sub( cz, dy ), L_add( ax, bw ) ) ); move16(); - m[2] = add( sub( sub( negate( ay ), bz ), cw ), dx ); + m[2] = check_bound( L_add( L_sub( L_sub( negate( ay ), bz ), cw ), dx ) ); move16(); - m[3] = sub( sub( by, az ), add( cx, dw ) ); + m[3] = check_bound( L_sub( L_sub( by, az ), L_add( cx, dw ) ) ); move16(); - m[4] = add( sub( add( bw, ax ), dy ), cz ); + m[4] = check_bound( L_add( L_sub( L_add( bw, ax ), dy ), cz ) ); move16(); - m[5] = add( add( add( negate( bx ), aw ), dz ), cy ); + m[5] = check_bound( L_add( L_add( L_add( negate( bx ), aw ), dz ), cy ) ); move16(); - m[6] = sub( sub( add( negate( by ), az ), dw ), cx ); + m[6] = check_bound( L_sub( L_sub( L_add( negate( by ), az ), dw ), cx ) ); move16(); - m[7] = add( sub( sub( negate( bz ), ay ), dx ), cw ); + m[7] = check_bound( L_add( L_sub( L_sub( negate( bz ), ay ), dx ), cw ) ); move16(); - m[8] = sub( add( add( cw, dx ), ay ), bz ); + m[8] = check_bound( L_sub( L_add( L_add( cw, dx ), ay ), bz ) ); move16(); - m[9] = sub( sub( add( negate( cx ), dw ), az ), by ); + m[9] = check_bound( L_sub( L_sub( L_add( negate( cx ), dw ), az ), by ) ); move16(); - m[10] = add( add( add( negate( cy ), dz ), aw ), bx ); + m[10] = check_bound( L_add( L_add( L_add( negate( cy ), dz ), aw ), bx ) ); move16(); - m[11] = sub( add( sub( negate( cz ), dy ), ax ), bw ); + m[11] = check_bound( L_sub( L_add( L_sub( negate( cz ), dy ), ax ), bw ) ); move16(); - m[12] = add( add( sub( dw, cx ), by ), az ); + m[12] = check_bound( L_add( L_add( L_sub( dw, cx ), by ), az ) ); move16(); - m[13] = add( sub( sub( negate( dx ), cw ), bz ), ay ); + m[13] = check_bound( L_add( L_sub( L_sub( negate( dx ), cw ), bz ), ay ) ); move16(); - m[14] = sub( add( sub( negate( dy ), cz ), bw ), ax ); + m[14] = check_bound( L_sub( L_add( L_sub( negate( dy ), cz ), bw ), ax ) ); move16(); - m[15] = add( add( add( negate( dz ), cy ), bx ), aw ); + m[15] = check_bound( L_add( L_add( L_add( negate( dz ), cy ), bx ), aw ) ); move16(); return; @@ -819,6 +835,24 @@ static void norm_quat( } #ifdef IVAS_FLOAT_FIXED + +static Word32 dotp16_fixed_guarded( const Word16 x[], /* i : vector x[] */ + const Word16 y[], /* i : vector y[] */ + const Word16 n /* i : vector length */ ) +{ + Word16 i; + Word32 suma; + Word16 guarded_bits = find_guarded_bits_fx( n ); + suma = L_shr( L_mult( x[0], y[0] ), guarded_bits ); + + FOR( i = 1; i < n; i++ ) + { + suma = L_add( suma, L_shr( L_mult( x[i], y[i] ), guarded_bits ) ); + } + + return suma; +} + static void norm_quat_fx( Word16 *q ) { @@ -826,13 +860,13 @@ static void norm_quat_fx( Word16 i, exp1; exp1 = 0; move16(); - norm_q = dotp_fx( q, q, IVAS_PCA_INTERP, &exp1 ); - exp1 = sub( 31, add( exp1, 2 ) ); - norm_q = ISqrt32( norm_q, &exp1 ); /*q(15 - exp)*/ + norm_q = dotp16_fixed_guarded( q, q, IVAS_PCA_INTERP ); + exp1 = find_guarded_bits_fx( IVAS_PCA_INTERP ); + norm_q = ISqrt32( norm_q, &exp1 ); /*Q(31 - exp)*/ FOR( i = 0; i < IVAS_PCA_INTERP; i++ ) { - q[i] = mult( q[i], extract_l( L_shr( norm_q, 15 ) ) ); // todo : recheck + q[i] = round_fx( L_shl( Mpy_32_16_1( norm_q, q[i] ), exp1 ) ); /* Q(15) */ move16(); } @@ -922,7 +956,7 @@ void pca_interp_preproc_fx( { Word16 alpha; Word16 j; - Word16 tmp, tmp2, tmp3; + Word16 tmp, tmp2, tmp3, tmp_e; FOR( j = 0; j < len; j++ ) { tmp = sub( len, 1 ); @@ -933,11 +967,12 @@ void pca_interp_preproc_fx( } ELSE { - alpha = idiv1616( j, tmp ); // the increment can be updated by simple delta //q15 + alpha = BASOP_Util_Divide1616_Scale( j, tmp, &tmp_e ); // the increment can be updated by simple delta + alpha = shl_sat( alpha, tmp_e ); /* Q15 */ } - tmp2 = mult( EVS_PI_FX, alpha ); // q13 - tmp3 = getCosWord16( tmp2 ); // q14 - alpha = shr( sub( ONE_IN_Q14, tmp3 ), 1 ); // q15 + tmp2 = mult( EVS_PI_FX, alpha ); /* Q13 */ + tmp3 = getCosWord16( tmp2 ); /* Q14 */ + alpha = sub_sat( ONE_IN_Q14, tmp3 ); /* Q15 */ alpha = sub( MAX_16, alpha ); quat_nlerp_preproc_fx( prev_ql, ql, alpha, &ql_interp[j * IVAS_PCA_INTERP] ); quat_nlerp_preproc_fx( prev_qr, qr, alpha, &qr_interp[j * IVAS_PCA_INTERP] ); @@ -986,16 +1021,30 @@ static void sp2cart_fx( Word16 *q ) { Word16 s1, s2, s1s2; - s1 = getSinWord16( ph1 ); // q15 - s2 = getSinWord16( ph2 ); // q15 - s1s2 = mult( s1, s2 ); // q15 - q[3] = mult( getSinWord16( ph3 ), s1s2 ); // q15 + Word16 sin_ph3, cos_ph3; + + sin_ph3 = cos_ph3 = ph3; move16(); - q[2] = mult( getCosWord16( ph3 ), s1s2 ); // q15 move16(); - q[1] = mult( getCosWord16( ph2 ), s1 ); // q15 + + IF( GT_16( ph3, 12868 /* PI in Q12 */ ) ) + { + sin_ph3 = sub( 12868, ph3 ); /* sin(x) = sin(PI - x) */ + cos_ph3 = sub( 25736, ph3 ); /* cos(x) = cos(2*PI - x) */ + } + sin_ph3 = shl( sin_ph3, 1 ); /* Q12 -> Q13 */ + cos_ph3 = shl( cos_ph3, 1 ); /* Q12 -> Q13 */ + + s1 = getSinWord16( ph1 ); /* Q15 */ + s2 = getSinWord16( ph2 ); /* Q15 */ + s1s2 = mult( s1, s2 ); /* Q15 */ + q[3] = mult( getSinWord16( sin_ph3 ), s1s2 ); /* Q15 */ move16(); - q[0] = shl_sat( getCosWord16( ph1 ), 1 ); // q15 + q[2] = shl_sat( mult( getCosWord16( cos_ph3 ), s1s2 ), 1 ); /* Q15 */ + move16(); + q[1] = shl_sat( mult( getCosWord16( ph2 ), s1 ), 1 ); /* Q15 */ + move16(); + q[0] = shl_sat( getCosWord16( ph1 ), 1 ); /* Q15 */ move16(); return; @@ -1021,9 +1070,9 @@ static Word16 calc_n2_fx( { Word16 n2; Word32 temp; - temp = L_mult( 23040, getSinWord16( ph1 ) ); // q8 - n2 = round_fx( temp ); - n2 = shr( n2, 7 ); + temp = L_mult( 23040, getSinWord16( ph1 ) ); /* Q8 + Q15 + Q1 = Q24 */ + n2 = round_fx( temp ); /* Q24 -> Q8 */ + n2 = shr( n2, 8 ); IF( s_and( n2, 1 ) == 0 ) { n2 = add( n2, 1 ); @@ -1061,10 +1110,10 @@ static Word16 calc_n3_fx( const Word16 ph2 ) { Word16 n3; - Word16 temp1 = mult( getSinWord16( ph2 ), getSinWord16( ph1 ) ); // q7 + q15 - q15 - n3 = round_fx( L_mult( temp1, getSinWord16( ph2 ) ) ); // q7 + q15 - q15 + Word16 temp1 = mult( getSinWord16( ph2 ), getSinWord16( ph1 ) ); /* Q15 */ + n3 = round_fx( L_mult( temp1, 23040 /* 180.0f in Q7 */ ) ); /* Q15 + Q7 + Q1 - Q16 -> Q7*/ - n3 = shr( n3, 8 ); + n3 = shr( n3, 7 ); IF( n3 == 0 ) { @@ -1489,8 +1538,10 @@ void pca_dec_s3_fx( num_fx = 12868; move16(); - d_fx = idiv1616( num_fx, n1 ); // Q12 - ph1_q_fx = i_mult( index1, d_fx ); // Q12 + d_fx = idiv1616( num_fx, sub( n1, 1 ) ); // Q12 + ph1_q_fx = i_mult( index1, d_fx ); // Q12 + + ph1_q_fx = shl( ph1_q_fx, 1 ); /* Q12 -> Q13 */ n2 = calc_n2_fx( ph1_q_fx ); @@ -1521,6 +1572,7 @@ void pca_dec_s3_fx( move16(); d_fx = idiv1616( num_fx, sub( n2, 1 ) ); // Q12 ph2_q_fx = i_mult( index2, d_fx ); // Q12 + ph2_q_fx = shl( ph2_q_fx, 1 ); /* Q12 -> Q13 */ } j = L_sub( j, ivas_pca_offset_index2[add( index2, get_pca_offset_n2_fx( index1 ) )] ); @@ -1536,11 +1588,13 @@ void pca_dec_s3_fx( } ELSE { - num_fx = 6434; + Word16 exp; + num_fx = 25736; /* PI2 in Q12 */ move16(); - d_fx = idiv1616( num_fx, n3 ); // Q10 - ph3_q_fx = round_fx( L_mult( index3, d_fx ) ); // Q10 - // ph3_q_fx = shl(ph3_q_fx, 2); + d_fx = BASOP_Util_Divide1616_Scale( num_fx, n3, &exp ); /* Q12 */ + exp = add( 3 - 15, exp ); + d_fx = shl( d_fx, sub( exp, 3 ) ); /* Q12 */ + ph3_q_fx = i_mult( index3, d_fx ); /* Q12 */ } sp2cart_fx( ph1_q_fx, ph2_q_fx, ph3_q_fx, q_fx ); diff --git a/lib_rend/ivas_dirac_rend.c b/lib_rend/ivas_dirac_rend.c index 59a8ef6b6..c6ef71a46 100644 --- a/lib_rend/ivas_dirac_rend.c +++ b/lib_rend/ivas_dirac_rend.c @@ -2857,7 +2857,6 @@ void protoSignalComputation2_fx( Word32 tempSpaced_fx, tempDmx_fx; Word16 q_shift, min_q_shift, exp, q_temp, temp_q_shift, q_temp2; Word32 temp; - /* Calculate maximum possible shift for the buffers RealBuffer_fx and ImagBuffer_fx */ min_q_shift = Q31; move16(); @@ -3187,7 +3186,6 @@ void protoSignalComputation2_fx( { Real_aux_fx = L_shr( Real_aux_fx, 1 ); Imag_aux_fx = L_shr( Imag_aux_fx, 1 ); - temp = Madd_32_32( Mpy_32_32( Real_aux_fx, Real_aux_fx ), Imag_aux_fx, Imag_aux_fx ); IF( LT_16( q_temp, *q_proto_power_smooth ) ) { @@ -3434,8 +3432,8 @@ void protoSignalComputation2_fx( } q_lr_bb_power = sub( q_lr_bb_power, 1 ); /* = (lr_bb_power_fx * 2) */ - temp = BASOP_Util_Divide3232_Scale( lr_bb_power_fx, L_add( stereo_type_detect->total_bb_power_fx, EPSILON_FX ), &exp ); - exp = sub( 31, add( sub( 15, exp ), sub( q_lr_bb_power, stereo_type_detect->q_total_bb_power ) ) ); + temp = BASOP_Util_Divide3232_Scale_cadence( lr_bb_power_fx, L_add( stereo_type_detect->total_bb_power_fx, EPSILON_FX ), &exp ); + exp = sub( 31, add( sub( 31, exp ), sub( q_lr_bb_power, stereo_type_detect->q_total_bb_power ) ) ); temp = BASOP_Util_Log2( temp ); IF( NE_32( temp, MIN_32 ) ) { @@ -3500,8 +3498,8 @@ void protoSignalComputation2_fx( } q_lr_hi_power = sub( q_lr_hi_power, 1 ); /* = (q_lr_hi_power * 2) */ - temp = BASOP_Util_Divide3232_Scale( lr_hi_power_fx, L_add( stereo_type_detect->total_hi_power_fx, EPSILON_FX ), &exp ); - exp = sub( 31, add( sub( 15, exp ), sub( q_lr_hi_power, stereo_type_detect->q_total_hi_power ) ) ); + temp = BASOP_Util_Divide3232_Scale_cadence( lr_hi_power_fx, L_add( stereo_type_detect->total_hi_power_fx, EPSILON_FX ), &exp ); + exp = sub( 31, add( sub( 31, exp ), sub( q_lr_hi_power, stereo_type_detect->q_total_hi_power ) ) ); temp = BASOP_Util_Log2( temp ); IF( NE_32( temp, MIN_32 ) ) { @@ -3596,7 +3594,6 @@ void protoSignalComputation2_fx( move32(); } } - *q_proto_frame_f = add( q_cldfb, min_q_shift ); move16(); *q_proto_direct_buffer_f = add( q_cldfb, min_q_shift ); @@ -4698,7 +4695,7 @@ void ivas_masa_stereotype_detection_fx( { /* subtract_temp = ( -subtract_target_ratio_db - 3.0f ) / 3.0f; */ temp = L_sub( L_shr( -subtract_target_ratio_db_fx, 1 ), L_shr( THREE_Q21, 1 ) ); - subtract_temp_fx = Mpy_32_32( temp, 715827883 /* 1 / 3.0f in Q31 */ ); + subtract_temp_fx = Mpy_32_32( temp, 708669604 /* 1 / 3.0f in Q31 */ ); subtract_temp_fx = L_shl( subtract_temp_fx, 1 ); /* Q21 */ /* min_sum_temp = max( -min_sum_total_ratio_db / 6.0f, 0.0f ); */ @@ -4710,7 +4707,7 @@ void ivas_masa_stereotype_detection_fx( change_to_spaced_fx = L_add( L_add( subtract_temp_fx, min_sum_temp_fx ), lr_total_bb_temp_fx ); /* Q21 */ - IF( GE_32( change_to_spaced_fx, ONE_IN_Q15 ) ) + IF( GE_32( change_to_spaced_fx, ONE_IN_Q21 ) ) { change_to_spaced_selection = 1; move16(); @@ -4723,7 +4720,7 @@ void ivas_masa_stereotype_detection_fx( IF( subtract_target_ratio_db_fx > 0 ) { /* subtract_temp = subtract_target_ratio_db / 3.0f; */ - subtract_temp_fx = Mpy_32_32( subtract_target_ratio_db_fx, 715827883 /* 1 / 3.0f in Q31 */ ); /* Q21 */ + subtract_temp_fx = Mpy_32_32( subtract_target_ratio_db_fx, 708669604 /* 1 / 3.0f in Q31 */ ); /* Q21 */ /* min_sum_temp = ( min_sum_total_ratio_db + 1.0f ) / 6.0f; */ min_sum_temp_fx = Mpy_32_32_r( L_add( min_sum_total_ratio_db_fx, ONE_IN_Q21 ), 357913941 /* 1 / 6.0f in Q31 */ ); /* Q21 */ @@ -4733,7 +4730,7 @@ void ivas_masa_stereotype_detection_fx( change_to_downmix_fx = L_add( L_add( subtract_temp_fx, min_sum_temp_fx ), lr_total_bb_temp_fx ); /* Q21 */ - IF( GE_32( change_to_downmix_fx, ONE_IN_Q15 ) ) + IF( GE_32( change_to_downmix_fx, ONE_IN_Q21 ) ) { change_to_downmix_selection = 1; move16(); @@ -4744,17 +4741,17 @@ void ivas_masa_stereotype_detection_fx( IF( LT_32( lr_total_hi_ratio_db_fx, -25165824 ) ) // 25165824 = 12.0 in Q21 { /* subtract_temp = ( subtract_target_ratio_db + 4.0f ) / 3.0f; */ - subtract_temp_fx = Mpy_32_32( L_add( subtract_target_ratio_db_fx, 8388608 /* 4.0 in Q21 */ ), 715827883 /* 1 / 3.0f in Q31 */ ); + subtract_temp_fx = Mpy_32_32( L_add( subtract_target_ratio_db_fx, 8388608 /* 4.0 in Q21 */ ), 708669604 /* 1 / 3.0f in Q31 */ ); /* min_sum_temp = min_sum_total_ratio_db / 6.0f; */ min_sum_temp_fx = Mpy_32_32_r( min_sum_total_ratio_db_fx, 357913941 /* 1 / 6.0f in Q31 */ ); /* lr_total_hi_temp = ( -lr_total_hi_ratio_db - 12.0f ) / 3.0f; */ - lr_total_hi_temp_fx = Mpy_32_32( L_sub( -lr_total_hi_ratio_db_fx, 25165824 /* 12.0 in Q21 */ ), 715827883 /* 1 / 3.0f in Q31 */ ); + lr_total_hi_temp_fx = Mpy_32_32( L_sub( -lr_total_hi_ratio_db_fx, 25165824 /* 12.0 in Q21 */ ), 708669604 /* 1 / 3.0f in Q31 */ ); - change_to_downmix2_fx = L_add( L_add( subtract_temp_fx, min_sum_temp_fx ), lr_total_hi_temp_fx ); // Q15 + change_to_downmix2_fx = L_add( L_add( subtract_temp_fx, min_sum_temp_fx ), lr_total_hi_temp_fx ); // Q21 - IF( GE_32( change_to_downmix2_fx, ONE_IN_Q15 ) ) + IF( GE_32( change_to_downmix2_fx, ONE_IN_Q21 ) ) { change_to_downmix_selection = 1; move16(); -- GitLab