Fix clipping bug in float to fixed point conversion, causing

audible glitches when samples smaller than the quantification
step are converted.

diff --git a/usr.bin/aucat/dsp.c b/usr.bin/aucat/dsp.c
index 927e00f..4575b26 100644 (file)
--- a/usr.bin/aucat/dsp.c
+++ b/usr.bin/aucat/dsp.c
@@ -1,4 +1,4 @@
-/*     $OpenBSD: dsp.c,v 1.1 2015/01/21 08:43:55 ratchov Exp $ */
+/*     $OpenBSD: dsp.c,v 1.2 2015/05/04 12:51:13 ratchov Exp $ */
 /*
  * Copyright (c) 2008-2012 Alexandre Ratchov <alex@caoua.org>
  *
@@ -604,9 +604,19 @@ f32_to_adata(unsigned int x)
        s = (x >> 31);
        e = (x >> 23) & 0xff;
        m = (x << 8) | 0x80000000;
-       if (e < 127 - 24)
+
+       /*
+        * f32 exponent is (e - 127) and the point is after the 31-th
+        * bit, thus the shift is:
+        *
+        * 31 - (BITS - 1) - (e - 127)
+        *
+        * to ensure output is in the 0..(2^BITS)-1 range, the minimum
+        * shift is 31 - (BITS - 1), and maximum shift is 31
+        */
+       if (e < 127 - (ADATA_BITS - 1))
                y = 0;
-       else if (e > 127 - 1)
+       else if (e > 127)
                y = ADATA_UNIT - 1;
        else
                y = m >> (127 + (32 - ADATA_BITS) - e);