replaced 64-bit correlations with scaled 32-bit ones, in Burg

2 files changed, 13 insertions, 5 deletions

diff --git a/silk/fixed/burg_modified_FIX.c b/silk/fixed/burg_modified_FIX.c
index 7d072a46..769205bd 100644
--- a/silk/fixed/burg_modified_FIX.c
+++ b/silk/fixed/burg_modified_FIX.c
@@ -84,8 +84,7 @@ void silk_burg_modified_c(
         for( s = 0; s < nb_subfr; s++ ) {
             x_ptr = x + s * subfr_length;
             for( n = 1; n < D + 1; n++ ) {
-                c[ n ] += (opus_int32)silk_RSHIFT64(
-                    silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n, arch ), rshifts );
+                c[ n ] += silk_inner_prod_aligned_scale( x_ptr, x_ptr + n, rshifts, subfr_length - n );
             }
         }
     } else {
diff --git a/silk/inner_prod_aligned.c b/silk/inner_prod_aligned.c
index 257ae9e0..e192e946 100644
--- a/silk/inner_prod_aligned.c
+++ b/silk/inner_prod_aligned.c
@@ -39,9 +39,18 @@ opus_int32 silk_inner_prod_aligned_scale(
 )
 {
     opus_int   i;
-    opus_int32 sum = 0;
-    for( i = 0; i < len; i++ ) {
-        sum = silk_ADD_RSHIFT32( sum, silk_SMULBB( inVec1[ i ], inVec2[ i ] ), scale );
+    opus_int32 sum, bias;
+
+	/* The bias serves two purposes: */
+	/* - round to nearest (instead of rounding down) */
+	/* - enable two-fold unrolling without risk of wrap around when multiplying values of -32768 */
+	bias = silk_LSHIFT32( 1, scale - 1 ) - silk_LSHIFT32( 1, 15 );
+	sum = silk_LSHIFT32( ( len + 1 ) >> 1, 15 - scale );
+    for( i = 0; i < len - 1; i += 2 ) {
+        sum = silk_ADD_RSHIFT32( sum, silk_SMLABB( silk_SMLABB( bias, inVec1[ i ], inVec2[ i ] ), inVec1[ i + 1 ], inVec2[ i + 1 ] ), scale );
     }
+	if( i < len ) {
+        sum = silk_ADD_RSHIFT32( sum, silk_SMLABB( bias, inVec1[ i ], inVec2[ i ] ), scale );
+	}
     return sum;
 }