ref: 6adea4b93ce15b5c6eab1be29da6e01806ee8853
parent: fc7ce7d588afcd6c537e5063171156c9cbf16445
author: Koen Vos <[email protected]>
date: Fri Feb 19 06:11:20 EST 2016
pitch analysis (in FIX) now scales the input down at the start, instead of at every stage
--- a/silk/fixed/pitch_analysis_core_FIX.c
+++ b/silk/fixed/pitch_analysis_core_FIX.c
@@ -80,7 +80,7 @@
/* FIXED POINT CORE PITCH ANALYSIS FUNCTION */
/*************************************************************/
opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 voiced, 1 unvoiced */
- const opus_int16 *frame, /* I Signal of length PE_FRAME_LENGTH_MS*Fs_kHz */
+ const opus_int16 *frame_unscaled, /* I Signal of length PE_FRAME_LENGTH_MS*Fs_kHz */
opus_int *pitch_out, /* O 4 pitch lag values */
opus_int16 *lagIndex, /* O Lag Index */
opus_int8 *contourIndex, /* O Pitch contour Index */
@@ -94,10 +94,11 @@
int arch /* I Run-time architecture */
)
{- VARDECL( opus_int16, frame_8kHz );
+ VARDECL( opus_int16, frame_8kHz_buf );
VARDECL( opus_int16, frame_4kHz );
+ VARDECL( opus_int16, frame_scaled );
opus_int32 filt_state[ 6 ];
- const opus_int16 *input_frame_ptr;
+ const opus_int16 *frame, *frame_8kHz;
opus_int i, k, d, j;
VARDECL( opus_int16, C );
VARDECL( opus_int32, xcorr32 );
@@ -119,6 +120,7 @@
opus_int32 delta_lag_log2_sqr_Q7, lag_log2_Q7, prevLag_log2_Q7, prev_lag_bias_Q13;
const opus_int8 *Lag_CB_ptr;
SAVE_STACK;
+
/* Check for valid sampling frequency */
silk_assert( Fs_kHz == 8 || Fs_kHz == 12 || Fs_kHz == 16 );
@@ -137,17 +139,33 @@
min_lag = PE_MIN_LAG_MS * Fs_kHz;
max_lag = PE_MAX_LAG_MS * Fs_kHz - 1;
+ /* Downscale input if necessary */
+ silk_sum_sqr_shift( &energy, &shift, frame_unscaled, frame_length );
+ shift += 2 - silk_CLZ32( energy ); /* at least one bit headroom */
+ ALLOC( frame_scaled, frame_length, opus_int16 );
+ if( shift > 0 ) {+ shift = silk_RSHIFT( shift + 1, 1 );
+ for( i = 0; i < frame_length; i++ ) {+ frame_scaled[ i ] = silk_RSHIFT( frame_unscaled[ i ], shift );
+ }
+ frame = frame_scaled;
+ } else {+ frame = frame_unscaled;
+ }
+
+ ALLOC( frame_8kHz_buf, ( Fs_kHz == 8 ) ? 1 : frame_length_8kHz, opus_int16 );
/* Resample from input sampled at Fs_kHz to 8 kHz */
- ALLOC( frame_8kHz, frame_length_8kHz, opus_int16 );
if( Fs_kHz == 16 ) {silk_memset( filt_state, 0, 2 * sizeof( opus_int32 ) );
- silk_resampler_down2( filt_state, frame_8kHz, frame, frame_length );
+ silk_resampler_down2( filt_state, frame_8kHz_buf, frame, frame_length );
+ frame_8kHz = frame_8kHz_buf;
} else if( Fs_kHz == 12 ) {silk_memset( filt_state, 0, 6 * sizeof( opus_int32 ) );
- silk_resampler_down2_3( filt_state, frame_8kHz, frame, frame_length );
+ silk_resampler_down2_3( filt_state, frame_8kHz_buf, frame, frame_length );
+ frame_8kHz = frame_8kHz_buf;
} else {silk_assert( Fs_kHz == 8 );
- silk_memcpy( frame_8kHz, frame, frame_length_8kHz * sizeof(opus_int16) );
+ frame_8kHz = frame;
}
/* Decimate again to 4 kHz */
@@ -156,24 +174,17 @@
silk_resampler_down2( filt_state, frame_4kHz, frame_8kHz, frame_length_8kHz );
/* Low-pass filter */
- for( i = frame_length_4kHz - 1; i > 0; i-- ) {- frame_4kHz[ i ] = silk_ADD_SAT16( frame_4kHz[ i ], frame_4kHz[ i - 1 ] );
+ silk_assert( (frame_length_4kHz & 3) == 0 );
+ for( i = frame_length_4kHz - 1; i > 4; i -= 4 ) {+ frame_4kHz[ i - 0 ] = silk_ADD_SAT16( frame_4kHz[ i - 0 ], frame_4kHz[ i - 1 ] );
+ frame_4kHz[ i - 1 ] = silk_ADD_SAT16( frame_4kHz[ i - 1 ], frame_4kHz[ i - 2 ] );
+ frame_4kHz[ i - 2 ] = silk_ADD_SAT16( frame_4kHz[ i - 2 ], frame_4kHz[ i - 3 ] );
+ frame_4kHz[ i - 3 ] = silk_ADD_SAT16( frame_4kHz[ i - 3 ], frame_4kHz[ i - 4 ] );
}
+ frame_4kHz[ i - 0 ] = silk_ADD_SAT16( frame_4kHz[ i - 0 ], frame_4kHz[ i - 1 ] );
+ frame_4kHz[ i - 1 ] = silk_ADD_SAT16( frame_4kHz[ i - 1 ], frame_4kHz[ i - 2 ] );
+ frame_4kHz[ i - 2 ] = silk_ADD_SAT16( frame_4kHz[ i - 2 ], frame_4kHz[ i - 3 ] );
- /*******************************************************************************
- ** Scale 4 kHz signal down to prevent correlations measures from overflowing
- ** find scaling as max scaling for each 8kHz(?) subframe
- *******************************************************************************/
-
- /* Inner product is calculated with different lengths, so scale for the worst case */
- silk_sum_sqr_shift( &energy, &shift, frame_4kHz, frame_length_4kHz );
- if( shift > 0 ) {- shift = silk_RSHIFT( shift, 1 );
- for( i = 0; i < frame_length_4kHz; i++ ) {- frame_4kHz[ i ] = silk_RSHIFT( frame_4kHz[ i ], shift );
- }
- }
-
/******************************************************************************
* FIRST STAGE, operating in 4 khz
******************************************************************************/
@@ -311,18 +322,6 @@
** SECOND STAGE, operating at 8 kHz, on lag sections with high correlation
*************************************************************************************/
- /******************************************************************************
- ** Scale signal down to avoid correlations measures from overflowing
- *******************************************************************************/
- /* find scaling as max scaling for each subframe */
- silk_sum_sqr_shift( &energy, &shift, frame_8kHz, frame_length_8kHz );
- if( shift > 0 ) {- shift = silk_RSHIFT( shift, 1 );
- for( i = 0; i < frame_length_8kHz; i++ ) {- frame_8kHz[ i ] = silk_RSHIFT( frame_8kHz[ i ], shift );
- }
- }
-
/*********************************************************************************
* Find energy of each subframe projected onto its history, for a range of delays
*********************************************************************************/
@@ -462,24 +461,6 @@
silk_assert( *LTPCorr_Q15 >= 0 );
if( Fs_kHz > 8 ) {- VARDECL( opus_int16, scratch_mem );
- /***************************************************************************/
- /* Scale input signal down to avoid correlations measures from overflowing */
- /***************************************************************************/
- /* find scaling as max scaling for each subframe */
- silk_sum_sqr_shift( &energy, &shift, frame, frame_length );
- ALLOC( scratch_mem, shift > 0 ? frame_length : ALLOC_NONE, opus_int16 );
- if( shift > 0 ) {- /* Move signal to scratch mem because the input signal should be unchanged */
- shift = silk_RSHIFT( shift, 1 );
- for( i = 0; i < frame_length; i++ ) {- scratch_mem[ i ] = silk_RSHIFT( frame[ i ], shift );
- }
- input_frame_ptr = scratch_mem;
- } else {- input_frame_ptr = frame;
- }
-
/* Search in original signal */
CBimax_old = CBimax;
@@ -519,14 +500,14 @@
/* Calculate the correlations and energies needed in stage 3 */
ALLOC( energies_st3, nb_subfr * nb_cbk_search, silk_pe_stage3_vals );
ALLOC( cross_corr_st3, nb_subfr * nb_cbk_search, silk_pe_stage3_vals );
- silk_P_Ana_calc_corr_st3( cross_corr_st3, input_frame_ptr, start_lag, sf_length, nb_subfr, complexity, arch );
- silk_P_Ana_calc_energy_st3( energies_st3, input_frame_ptr, start_lag, sf_length, nb_subfr, complexity, arch );
+ silk_P_Ana_calc_corr_st3( cross_corr_st3, frame, start_lag, sf_length, nb_subfr, complexity, arch );
+ silk_P_Ana_calc_energy_st3( energies_st3, frame, start_lag, sf_length, nb_subfr, complexity, arch );
lag_counter = 0;
silk_assert( lag == silk_SAT16( lag ) );
contour_bias_Q15 = silk_DIV32_16( SILK_FIX_CONST( PE_FLATCONTOUR_BIAS, 15 ), lag );
- target_ptr = &input_frame_ptr[ PE_LTP_MEM_LENGTH_MS * Fs_kHz ];
+ target_ptr = &frame[ PE_LTP_MEM_LENGTH_MS * Fs_kHz ];
energy_target = silk_ADD32( silk_inner_prod_aligned( target_ptr, target_ptr, nb_subfr * sf_length, arch ), 1 );
for( d = start_lag; d <= end_lag; d++ ) { for( j = 0; j < nb_cbk_search; j++ ) {--- a/silk/float/pitch_analysis_core_FLP.c
+++ b/silk/float/pitch_analysis_core_FLP.c
@@ -158,9 +158,16 @@
silk_short2float_array( frame_4kHz, frame_4_FIX, frame_length_4kHz );
/* Low-pass filter */
- for( i = frame_length_4kHz - 1; i > 0; i-- ) {- frame_4kHz[ i ] += frame_4kHz[ i - 1 ];
+ silk_assert( (frame_length_4kHz & 3) == 0 );
+ for( i = frame_length_4kHz - 1; i > 4; i -= 4 ) {+ frame_4kHz[ i - 0 ] = silk_ADD_SAT16( frame_4kHz[ i - 0 ], frame_4kHz[ i - 1 ] );
+ frame_4kHz[ i - 1 ] = silk_ADD_SAT16( frame_4kHz[ i - 1 ], frame_4kHz[ i - 2 ] );
+ frame_4kHz[ i - 2 ] = silk_ADD_SAT16( frame_4kHz[ i - 2 ], frame_4kHz[ i - 3 ] );
+ frame_4kHz[ i - 3 ] = silk_ADD_SAT16( frame_4kHz[ i - 3 ], frame_4kHz[ i - 4 ] );
}
+ frame_4kHz[ i - 0 ] = silk_ADD_SAT16( frame_4kHz[ i - 0 ], frame_4kHz[ i - 1 ] );
+ frame_4kHz[ i - 1 ] = silk_ADD_SAT16( frame_4kHz[ i - 1 ], frame_4kHz[ i - 2 ] );
+ frame_4kHz[ i - 2 ] = silk_ADD_SAT16( frame_4kHz[ i - 2 ], frame_4kHz[ i - 3 ] );
/******************************************************************************
* FIRST STAGE, operating in 4 khz