Probably porting mistake in protoSignalComputation_shd_fx
Basic info
- Float reference:
- Decoder (float): dfed311fdc7c91f7f9c4204ba4f1c9b02441888a
- Fixed point:
- Decoder (fixed): 496d928d
Bug description
Looking at the rotation code in protoSignalComputation_shd_fx, there seems to be a possible porting mistake. See code pieces below:
In float we have
if ( p_Rmat != 0 )
{
assert( num_inputs == 4 && "This code block should never be run with num_inputs != 4!" );
for ( l = 0; l < num_freq_bands; l++ )
{
*( p_k[0] ) = RealBuffer[0][0][l];
reference_power[l + num_freq_bands] = *( p_k[0] ) * *( p_k[0] );
p_k[0]++;
*( p_k[0] ) = ImagBuffer[0][0][l];
reference_power[l + num_freq_bands] += *( p_k[0] ) * *( p_k[0] );
p_k[0]++;
reference_power[l] = 0.5f * reference_power[l + num_freq_bands];
for ( k = 1; k < 4; k++ )
{
*( p_k[k] ) = p_Rmat[3 * Rmat_k[k] + 1] * RealBuffer[1][0][l] + p_Rmat[3 * Rmat_k[k] + 2] * RealBuffer[2][0][l] + p_Rmat[3 * Rmat_k[k] + 0] * RealBuffer[3][0][l];
reference_power[l + ( k + 1 ) * num_freq_bands] = *( p_k[k] ) * *( p_k[k] );
p_k[k]++;
*( p_k[k] ) = p_Rmat[3 * Rmat_k[k] + 1] * ImagBuffer[1][0][l] + p_Rmat[3 * Rmat_k[k] + 2] * ImagBuffer[2][0][l] + p_Rmat[3 * Rmat_k[k] + 0] * ImagBuffer[3][0][l];
reference_power[l + ( k + 1 ) * num_freq_bands] += *( p_k[k] ) * *( p_k[k] );
p_k[k]++;
reference_power[l] += 0.5f * ( reference_power[l + ( k + 1 ) * num_freq_bands] );
}
for ( k = 1; k < 4; k++ )
{
RealBuffer[k][0][l] = p_proto_direct_buffer[k * 2 * num_freq_bands + 2 * l];
ImagBuffer[k][0][l] = p_proto_direct_buffer[k * 2 * num_freq_bands + 2 * l + 1];
}
}
}
In BASOP, the corresponding part is this
IF( p_Rmat_fx != 0 )
{
assert( EQ_16( num_inputs, 4 ) && "This code block should never be run with num_inputs != 4!" );
FOR( l = 0; l < num_freq_bands; l++ )
{
*p_k_fx[0] = L_shl( RealBuffer_fx[0][0][l], min_q_shift ); /*Q(q_cldfb+min_q_shift)*/
move32();
reference_power_fx[l + num_freq_bands] = Mpy_32_32( *p_k_fx[0], *p_k_fx[0] ); /*2*Q(q_cldfb+min_q_shift)-31*/
move32();
p_k_fx[0]++;
*p_k_fx[0] = L_shl( ImagBuffer_fx[0][0][l], min_q_shift ); /*Q(q_cldfb+min_q_shift)*/
move32();
reference_power_fx[l + num_freq_bands] = Madd_32_32( reference_power_fx[l + num_freq_bands], *p_k_fx[0], *p_k_fx[0] ); /*2*Q(q_cldfb+min_q_shift)-31*/
move32();
p_k_fx[0]++;
reference_power_fx[l] = L_shr( reference_power_fx[l + num_freq_bands], 1 ); /*2*Q(q_cldfb+min_q_shift)-31-1*/
move32();
re1 = L_shl( RealBuffer_fx[1][0][l], min_q_shift ); /*Q(q_cldfb+min_q_shift)*/
re2 = L_shl( RealBuffer_fx[2][0][l], min_q_shift ); /*Q(q_cldfb+min_q_shift)*/
re3 = L_shl( RealBuffer_fx[3][0][l], min_q_shift ); /*Q(q_cldfb+min_q_shift)*/
im1 = L_shl( ImagBuffer_fx[1][0][l], min_q_shift ); /*Q(q_cldfb+min_q_shift)*/
im2 = L_shl( ImagBuffer_fx[2][0][l], min_q_shift ); /*Q(q_cldfb+min_q_shift)*/
im3 = L_shl( ImagBuffer_fx[3][0][l], min_q_shift ); /*Q(q_cldfb+min_q_shift)*/
FOR( k = 1; k < 4; k++ )
{
idx = i_mult( 3, Rmat_k[k] );
idx1 = add( l, i_mult( add( k, 1 ), num_freq_bands ) );
*p_k_fx[k] = Madd_32_32( Madd_32_32( Mpy_32_32( p_Rmat_fx[idx + 1], re1 ), p_Rmat_fx[idx + 2], re2 ), p_Rmat_fx[idx], re3 ); /*Q(30 + q_cldfb+min_q_shift-31)=>Q(q_cldfb+min_q_shift-1)*/
move32();
*p_k_fx[k] = L_shl( *p_k_fx[k], Q1 ); // left shift is done to maintain constant Q factor for p_k_fx Q(q_cldfb+min_q_shift)
move32();
reference_power_fx[idx1] = Mpy_32_32( *p_k_fx[k], *p_k_fx[k] ); // Q(2*(q_cldfb + min_q_shift)-31)
move32();
p_k_fx[k]++;
*p_k_fx[k] = Madd_32_32( Madd_32_32( Mpy_32_32( p_Rmat_fx[idx + 1], im1 ), p_Rmat_fx[idx + 2], im2 ), p_Rmat_fx[idx], im3 ); /*Q(q_cldfb+min_q_shift-1)*/
move32();
*p_k_fx[k] = L_shl( *p_k_fx[k], Q1 ); // left shift is done to maintain constant Q factor Q(q_cldfb+min_q_shift)
move32();
BUG HERE? -> reference_power_fx[idx1] = Mpy_32_32( *p_k_fx[k], *p_k_fx[k] ); // Q(2*(q_cldfb + min_q_shift)-31)
move32();
p_k_fx[k]++;
reference_power_fx[l] = L_add( reference_power_fx[l], L_shr( reference_power_fx[idx1], 1 ) ); /*2*Q(q_cldfb+min_q_shift)-31-1*/
move32();
}
*proto_direct_buffer_f_q = add( q_cldfb, min_q_shift );
move16();
Word16 qidx = s_min( 1, s_max( 0, sub( l, CLDFB_NO_CHANNELS_HALF - 1 ) ) );
reference_power_q[qidx] = sub( add( *proto_direct_buffer_f_q, *proto_direct_buffer_f_q ), 31 );
move16();
Word16 shift = sub( *proto_direct_buffer_f_q, q_cldfb );
FOR( k = 1; k < 4; k++ )
{
RealBuffer_fx[k][0][l] = L_shr( p_proto_direct_buffer_fx[2 * ( k * num_freq_bands + l )], shift ); // proto_direct_buffer_f_q -> q_cldfb
move32();
ImagBuffer_fx[k][0][l] = L_shr( p_proto_direct_buffer_fx[2 * ( k * num_freq_bands + l ) + 1], shift ); // proto_direct_buffer_f_q -> q_cldfb
move32();
}
}
}I have marked the potential bug point above. Comparing to float, this probably should have accumulation instead of replacing the value. Effect should be probably wrong rendering when rotating SHD for SBA.