resampler: Work at double precision on x86-64 machines.

We get audio artifacts if we don't work at the higher precision, but
this is painful on CPUs that have to use a software fallback for this,
so for now (that is, until we have a better solution), get better output
on amd64 chips, where the cost is less painful.

(cherry picked from commit 1e5e8e2fda3796e76e6f7b1c39683925a3e9fed9)
This commit is contained in:
Ryan C. Gordon 2023-02-21 12:56:02 -05:00
parent 13969f8567
commit 9403c9e95a
No known key found for this signature in database
GPG Key ID: FA148B892AB48044

View File

@ -198,10 +198,18 @@ static int SDL_ResampleAudio(const int chans, const int inrate, const int outrat
const float *inbuf, const int inbuflen, const float *inbuf, const int inbuflen,
float *outbuf, const int outbuflen) float *outbuf, const int outbuflen)
{ {
/* !!! FIXME: this produces artifacts if we don't work at double precision, but this turns out to
be a big performance hit. Until we can resolve this better, we force this to double
for amd64 CPUs, which should be able to take the hit for now, vs small embedded
things that might end up in a software fallback here. */
/* Note that this used to be double, but it looks like we can get by with float in most cases at /* Note that this used to be double, but it looks like we can get by with float in most cases at
almost twice the speed on Intel processors, and orders of magnitude more almost twice the speed on Intel processors, and orders of magnitude more
on CPUs that need a software fallback for double calculations. */ on CPUs that need a software fallback for double calculations. */
#if defined(_M_X64) || defined(__x86_64__)
typedef double ResampleFloatType;
#else
typedef float ResampleFloatType; typedef float ResampleFloatType;
#endif
const ResampleFloatType finrate = (ResampleFloatType)inrate; const ResampleFloatType finrate = (ResampleFloatType)inrate;
const ResampleFloatType ratio = ((float)outrate) / ((float)inrate); const ResampleFloatType ratio = ((float)outrate) / ((float)inrate);