zdoom/fluidsynth
0
0
Fork 0
mirror of https://github.com/ZDoom/fluidsynth.git synced 2025-01-31 21:50:37 +00:00
Code Issues Projects Releases Packages Wiki Activity

Clarify and cleanup the denormal handling in fluid_rev

Browse source
This commit is contained in:
Marcus Weseloh 2017-11-23 11:09:55 +01:00
parent b742b9a58e
commit 58f5b87074
1 changed files with 14 additions and 40 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

54
src/rvoice/fluid_rev.c
View file

@ -19,49 +19,23 @@
/* Denormalising:
*
* According to music-dsp thread 'Denormalise', Pentium processors
* have a hardware 'feature', that is of interest here, related to
* numeric underflow. We have a recursive filter. The output decays
* exponentially, if the input stops. So the numbers get smaller and
* smaller... At some point, they reach 'denormal' level. This will
* lead to drastic spikes in the CPU load. The effect was reproduced
* with the reverb - sometimes the average load over 10 s doubles!!.
* We have a recursive filter. The output decays exponentially, if the input
* stops. So the numbers get smaller and smaller... At some point, they reach
* 'denormal' level. On some platforms this will lead to drastic spikes in the
* CPU load. This is especially noticable on some older Pentium (especially
* Pentium 3) processors, but even more modern Intel Core processors still show
* reduced performance with denormals. While there are compile-time switches to
* treat denormals as zero for a lot of processors, those are not available or
* effective on all platforms.
*
* The 'undenormalise' macro fixes the problem: As soon as the number
* is close enough to denormal level, the macro forces the number to
* 0.0f. The original macro is:
*
* #define undenormalise(sample) if(((*(unsigned int*)&sample)&0x7f800000)==0) sample=0.0f
*
* This will zero out a number when it reaches the denormal level.
* Advantage: Maximum dynamic range Disadvantage: We'll have to check
* every sample, expensive. The alternative macro comes from a later
* mail from Jon Watte. It will zap a number before it reaches
* denormal level. Jon suggests to run it once per block instead of
* every sample.
* The fix used here: Use a small DC-offset in the filter calculations. Now
* the signals converge not against 0, but against the offset. The constant
* offset is invisible from the outside world (i.e. it does not appear at the
* output. There is a very small turn-on transient response, which should not
* cause problems.
*/
# if defined(WITH_FLOATX)
# define zap_almost_zero(sample) (((*(unsigned int*)&(sample))&0x7f800000) < 0x08000000)?0.0f:(sample)
# else
/* 1e-20 was chosen as an arbitrary (small) threshold. */
#define zap_almost_zero(sample) fabs(sample)<1e-10 ? 0 : sample;
#endif
/* Denormalising part II:
*
* Another method fixes the problem cheaper: Use a small DC-offset in
* the filter calculations. Now the signals converge not against 0,
* but against the offset. The constant offset is invisible from the
* outside world (i.e. it does not appear at the output. There is a
* very small turn-on transient response, which should not cause
* problems.
*/
//#define DC_OFFSET 0
#define DC_OFFSET 1e-8
//#define DC_OFFSET 0.001f
typedef struct _fluid_allpass fluid_allpass;
typedef struct _fluid_comb fluid_comb;