From f054ce9dad41febee3dd4b4842a60e7d35445093 Mon Sep 17 00:00:00 2001
From: pabr <pabr@pabr.org>
Date: Sat, 25 Mar 2017 10:34:11 +0100
Subject: [PATCH] fir_sampler dot product: Even simpler form that triggers NEON
 vectorization with gcc-4.9.2

---
 src/leansdr/sdr.h | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)

diff --git a/src/leansdr/sdr.h b/src/leansdr/sdr.h
index 47361d1..8e0bcba 100644
--- a/src/leansdr/sdr.h
+++ b/src/leansdr/sdr.h
@@ -485,14 +485,15 @@ namespace leansdr {
       if ( subsampling == 1 ) {
 	// Special case for heavily oversampled signals,
 	// where filtering is expensive.
-	while ( pc+16 < pcend ) {
-	  // gcc-4.9.2 can vectorize this form with NEON on ARM
-	  for ( int i=0; i<16; ++i)
-	    acc += (*pc++)*(*pin++);
-	}
+	// gcc-4.9.2 can vectorize this form with NEON on ARM.
+	while ( pc < pcend )
+	  acc += (*pc++)*(*pin++);
+      } else {
+	// Not vectorized because the coefficients are not
+	// guaranteed to be contiguous in memory.
+	for ( ; pc<pcend; pc+=subsampling,++pin ) 
+	  acc += (*pc)*(*pin);
       }
-      for ( ; pc<pcend; pc+=subsampling,++pin ) 
-	acc += (*pc)*(*pin);
       // Derotate
       return trig.expi(-phase) * acc;
     }