Adds Neon assembly for correlation/convolution

Optimizing celt_pitch_xcorr()/xcorr_kernel() which also speeds up
FIRs, IIRs and auto-correlations

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

5 files changed, 19 insertions, 12 deletions

diff --git a/silk/float/SigProc_FLP.h b/silk/float/SigProc_FLP.h
index e7131f14..f0cb3733 100644
--- a/silk/float/SigProc_FLP.h
+++ b/silk/float/SigProc_FLP.h
@@ -94,7 +94,8 @@ opus_int silk_pitch_analysis_core_FLP(      /* O    Voicing estimate: 0 voiced,
     const silk_float    search_thres2,      /* I    Final threshold for lag candidates 0 - 1                    */
     const opus_int      Fs_kHz,             /* I    sample frequency (kHz)                                      */
     const opus_int      complexity,         /* I    Complexity setting, 0-2, where 2 is highest                 */
-    const opus_int      nb_subfr            /* I    Number of 5 ms subframes                                    */
+    const opus_int      nb_subfr,           /* I    Number of 5 ms subframes                                    */
+    int                 arch                /* I    Run-time architecture                                       */
 );
 
 void silk_insertion_sort_decreasing_FLP(
diff --git a/silk/float/encode_frame_FLP.c b/silk/float/encode_frame_FLP.c
index df75db91..d54e2686 100644
--- a/silk/float/encode_frame_FLP.c
+++ b/silk/float/encode_frame_FLP.c
@@ -129,7 +129,7 @@ opus_int silk_encode_frame_FLP(
         /*****************************************/
         /* Find pitch lags, initial LPC analysis */
         /*****************************************/
-        silk_find_pitch_lags_FLP( psEnc, &sEncCtrl, res_pitch, x_frame );
+        silk_find_pitch_lags_FLP( psEnc, &sEncCtrl, res_pitch, x_frame, psEnc->sCmn.arch );
 
         /************************/
         /* Noise shape analysis */
diff --git a/silk/float/find_pitch_lags_FLP.c b/silk/float/find_pitch_lags_FLP.c
index 455db82f..f3b22d25 100644
--- a/silk/float/find_pitch_lags_FLP.c
+++ b/silk/float/find_pitch_lags_FLP.c
@@ -37,7 +37,8 @@ void silk_find_pitch_lags_FLP(
     silk_encoder_state_FLP          *psEnc,                             /* I/O  Encoder state FLP                           */
     silk_encoder_control_FLP        *psEncCtrl,                         /* I/O  Encoder control FLP                         */
     silk_float                      res[],                              /* O    Residual                                    */
-    const silk_float                x[]                                 /* I    Speech signal                               */
+    const silk_float                x[],                                /* I    Speech signal                               */
+    int                             arch                                /* I    Run-time architecture                       */
 )
 {
     opus_int   buf_len;
@@ -116,7 +117,7 @@ void silk_find_pitch_lags_FLP(
         /*****************************************/
         if( silk_pitch_analysis_core_FLP( res, psEncCtrl->pitchL, &psEnc->sCmn.indices.lagIndex,
             &psEnc->sCmn.indices.contourIndex, &psEnc->LTPCorr, psEnc->sCmn.prevLag, psEnc->sCmn.pitchEstimationThreshold_Q16 / 65536.0f,
-            thrhld, psEnc->sCmn.fs_kHz, psEnc->sCmn.pitchEstimationComplexity, psEnc->sCmn.nb_subfr ) == 0 )
+            thrhld, psEnc->sCmn.fs_kHz, psEnc->sCmn.pitchEstimationComplexity, psEnc->sCmn.nb_subfr, arch ) == 0 )
         {
             psEnc->sCmn.indices.signalType = TYPE_VOICED;
         } else {
diff --git a/silk/float/main_FLP.h b/silk/float/main_FLP.h
index dbd34aef..71d58c20 100644
--- a/silk/float/main_FLP.h
+++ b/silk/float/main_FLP.h
@@ -71,7 +71,8 @@ opus_int silk_encode_frame_FLP(
 
 /* Initializes the Silk encoder state */
 opus_int silk_init_encoder(
-    silk_encoder_state_FLP          *psEnc                              /* I/O  Encoder state FLP                           */
+    silk_encoder_state_FLP          *psEnc,                             /* I/O  Encoder state FLP                           */
+    int                              arch                               /* I    Run-tim architecture                        */
 );
 
 /* Control the Silk encoder */
@@ -129,7 +130,8 @@ void silk_find_pitch_lags_FLP(
     silk_encoder_state_FLP          *psEnc,                             /* I/O  Encoder state FLP                           */
     silk_encoder_control_FLP        *psEncCtrl,                         /* I/O  Encoder control FLP                         */
     silk_float                      res[],                              /* O    Residual                                    */
-    const silk_float                x[]                                 /* I    Speech signal                               */
+    const silk_float                x[],                                /* I    Speech signal                               */
+    int                             arch                                /* I    Run-time architecture                       */
 );
 
 /* Find LPC and LTP coefficients */
diff --git a/silk/float/pitch_analysis_core_FLP.c b/silk/float/pitch_analysis_core_FLP.c
index b6bafe81..e58f041b 100644
--- a/silk/float/pitch_analysis_core_FLP.c
+++ b/silk/float/pitch_analysis_core_FLP.c
@@ -48,7 +48,8 @@ static void silk_P_Ana_calc_corr_st3(
     opus_int            start_lag,          /* I start lag                                                      */
     opus_int            sf_length,          /* I sub frame length                                               */
     opus_int            nb_subfr,           /* I number of subframes                                            */
-    opus_int            complexity          /* I Complexity setting                                             */
+    opus_int            complexity,         /* I Complexity setting                                             */
+    int                 arch                /* I Run-time architecture                                          */
 );
 
 static void silk_P_Ana_calc_energy_st3(
@@ -74,7 +75,8 @@ opus_int silk_pitch_analysis_core_FLP(      /* O    Voicing estimate: 0 voiced,
     const silk_float    search_thres2,      /* I    Final threshold for lag candidates 0 - 1                    */
     const opus_int      Fs_kHz,             /* I    sample frequency (kHz)                                      */
     const opus_int      complexity,         /* I    Complexity setting, 0-2, where 2 is highest                 */
-    const opus_int      nb_subfr            /* I    Number of 5 ms subframes                                    */
+    const opus_int      nb_subfr,           /* I    Number of 5 ms subframes                                    */
+    int                 arch                /* I    Run-time architecture                                       */
 )
 {
     opus_int   i, k, d, j;
@@ -176,7 +178,7 @@ opus_int silk_pitch_analysis_core_FLP(      /* O    Voicing estimate: 0 voiced,
         silk_assert( basis_ptr >= frame_4kHz );
         silk_assert( basis_ptr + sf_length_8kHz <= frame_4kHz + frame_length_4kHz );
 
-        celt_pitch_xcorr( target_ptr, target_ptr-max_lag_4kHz, xcorr, sf_length_8kHz, max_lag_4kHz - min_lag_4kHz + 1 );
+        celt_pitch_xcorr( target_ptr, target_ptr-max_lag_4kHz, xcorr, sf_length_8kHz, max_lag_4kHz - min_lag_4kHz + 1, arch );
 
         /* Calculate first vector products before loop */
         cross_corr = xcorr[ max_lag_4kHz - min_lag_4kHz ];
@@ -409,7 +411,7 @@ opus_int silk_pitch_analysis_core_FLP(      /* O    Voicing estimate: 0 voiced,
         CCmax = -1000.0f;
 
         /* Calculate the correlations and energies needed in stage 3 */
-        silk_P_Ana_calc_corr_st3( cross_corr_st3, frame, start_lag, sf_length, nb_subfr, complexity );
+        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 );
 
         lag_counter = 0;
@@ -493,7 +495,8 @@ static void silk_P_Ana_calc_corr_st3(
     opus_int            start_lag,          /* I start lag                                                      */
     opus_int            sf_length,          /* I sub frame length                                               */
     opus_int            nb_subfr,           /* I number of subframes                                            */
-    opus_int            complexity          /* I Complexity setting                                             */
+    opus_int            complexity,         /* I Complexity setting                                             */
+    int                 arch                /* I Run-time architecture                                          */
 )
 {
     const silk_float *target_ptr;
@@ -527,7 +530,7 @@ static void silk_P_Ana_calc_corr_st3(
         lag_low  = matrix_ptr( Lag_range_ptr, k, 0, 2 );
         lag_high = matrix_ptr( Lag_range_ptr, k, 1, 2 );
         silk_assert(lag_high-lag_low+1 <= SCRATCH_SIZE);
-        celt_pitch_xcorr( target_ptr, target_ptr - start_lag - lag_high, xcorr, sf_length, lag_high - lag_low + 1 );
+        celt_pitch_xcorr( target_ptr, target_ptr - start_lag - lag_high, xcorr, sf_length, lag_high - lag_low + 1, arch );
         for( j = lag_low; j <= lag_high; j++ ) {
             silk_assert( lag_counter < SCRATCH_SIZE );
             scratch_mem[ lag_counter ] = xcorr[ lag_high - j ];