[FFmpeg-devel] [PATCH 7/9] avcodec/mdct_template: use lrint instead of floor hack

[Ganesh Ajjanagadde]

On Sun, Dec 6, 2015 at 8:20 AM, Ganesh Ajjanagadde
<gajjanagadde at gmail.com> wrote:
> On Tue, Dec 1, 2015 at 7:27 PM, Ganesh Ajjanagadde
> <gajjanagadde at gmail.com> wrote:
>> Signed-off-by: Ganesh Ajjanagadde <gajjanagadde at gmail.com>
>> ---
>>  libavcodec/mdct_template.c | 5 +++--
>>  1 file changed, 3 insertions(+), 2 deletions(-)
>>
>> diff --git a/libavcodec/mdct_template.c b/libavcodec/mdct_template.c
>> index e7e5f62..ecdeb54 100644
>> --- a/libavcodec/mdct_template.c
>> +++ b/libavcodec/mdct_template.c
>> @@ -23,6 +23,7 @@
>>  #include <string.h>
>>  #include "libavutil/common.h"
>>  #include "libavutil/mathematics.h"
>> +#include "libavutil/libm.h"
>>  #include "fft.h"
>>  #include "fft-internal.h"
>>
>> @@ -82,8 +83,8 @@ av_cold int ff_mdct_init(FFTContext *s, int nbits, int inverse, double scale)
>>      for(i=0;i<n4;i++) {
>>          alpha = 2 * M_PI * (i + theta) / n;
>>  #if FFT_FIXED_32
>> -        s->tcos[i*tstep] = (FFTSample)floor(-cos(alpha) * 2147483648.0 + 0.5);
>> -        s->tsin[i*tstep] = (FFTSample)floor(-sin(alpha) * 2147483648.0 + 0.5);
>> +        s->tcos[i*tstep] = lrint(-cos(alpha) * 2147483648.0);
>> +        s->tsin[i*tstep] = lrint(-sin(alpha) * 2147483648.0);
>>  #else
>>          s->tcos[i*tstep] = FIX15(-cos(alpha) * scale);
>>          s->tsin[i*tstep] = FIX15(-sin(alpha) * scale);
>> --
>> 2.6.2
>>
>
> Ping for this one, it is relatively important among the lrint patches
> as this is frequently used and table may be large, making speed
> benefit somewhat useful.
> I have noted the header alphabetical include order mistake above.

This is going in soon, as it actually fixes a bug since cos, sin can
be positive in the relevant ranges here. Last call.
Note: I think there might be undefined behavior here (before and after
the patch) since cos(alpha) can be -1 (or very close it) here. That is
a separate issue that I will examine some time.