theora dev - Jun 2009 - Thusnelda - Floating point exception in encoder_toplevel.c:209

Hi all,

  I've been hitting the odd floating point exception errors using the 
new alpha encoders. They're encoding a live stream so I can't reproduce 
the crash on demand, and it might only happen once every few days so 
it's been fun trying to get a backtrace. The crash happens in a fairly 
scary block of code, I wouldn't have a clue where to begin trying to fix it!

Here's what gdb has to say about it:

--
#0  0x00007f8d0e3ce515 in oc_enc_select_qi (cpi=0x7f8d0c14f010,
     _qti=<value optimized out>, _trial=0) at enc/encoder_toplevel.c:209
209           rderiv=nframes[_qti]*KEY_RATIO[_qti]+drscale;
Missing separate debuginfos, use: debuginfo-install 
SDL-1.2.13-7.fc10.x86_64 alsa-lib-1.0.20-1.fc10.x86_64 
freetype-2.3.7-3.fc10.x86_64 glibc-2.9-3.x86_64 
libjpeg-6b-43.fc10.x86_64 libogg-1.1.3-9.fc9.x86_64 
libshout-2.2.2-3.fc9.x86_64 libvorbis-1.2.0-5.fc10.x86_64 
speex-1.2-0.10.rc1.fc10.x86_64
(gdb) print nframes
$1 = {0, 183}
(gdb) print KEY_RATIO
$2 = " \021"
(gdb) print _qti
$3 = <value optimized out>
(gdb) print drscale
No symbol "drscale" in current context.
(gdb) backtrace
#0  0x00007f8d0e3ce515 in oc_enc_select_qi (cpi=0x7f8d0c14f010,
     _qti=<value optimized out>, _trial=0) at enc/encoder_toplevel.c:209
#1  0x00007f8d0e3d036a in CompressFrame (cpi=0x7f8d0c14f010, recode=0)
     at enc/encoder_toplevel.c:278
#2  0x00007f8d0e3d08ce in theora_encode_YUVin (t=0x7fff1660eb00,
     yuv=0x7fff1660ecb8) at enc/encoder_toplevel.c:473
#3  0x0000000000403a29 in l3_theora_encode (s=0x7fff1660e8e0,
     im=<value optimized out>, eos=0) at live3.c:636
#4  0x0000000000405696 in main (argc=<value optimized out>,
     argv=<value optimized out>) at live3.c:1403
--

Any ideas?

-Phil

Philip Heron wrote:
> Any ideas?
My guess is that you're running out of bits in the buffer, and
rate_total is greater than zero, but less than buf_delay (e.g., less
than 1 bit per frame). This one is clearly my fault, try the following
patch, and let me know if it fixes the problem:

--- encoder_toplevel.c  (revision 16089)
+++ encoder_toplevel.c  (working copy)
@@ -177,7 +177,7 @@

if(cpi->dup_count>0)nframes[1]=(nframes[1]+cpi->dup_count)/(cpi->dup_count+1);
   /*If there aren't enough bits to achieve our desired fullness level,
use the
      minimum quality permitted.*/
-  if(rate_total<=0)log_qtarget=OC_QUANT_MAX_LOG;
+  if(rate_total<=buf_delay)log_qtarget=OC_QUANT_MAX_LOG;
   else{
     static const unsigned char KEY_RATIO[2]={32,17};
     ogg_int64_t   log_scale0;

I know it's difficult to collect, but a test case that can reliably
reproduce this would be enormously helpful.