[opus] Make check failure on clone from 31 January

[Marcello Caramma (mcaramma)]

After a few experiments, I found that both alternatives are very similar, and 2~5% slower compared to the following:


diff --git a/celt/mdct.c b/celt/mdct.c
index 1634e8e..e490c3b 100644
--- a/celt/mdct.c
+++ b/celt/mdct.c
@@ -277,7 +277,7 @@ void clt_mdct_backward(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_scala
       it in-place. */
    {
       kiss_fft_scalar * OPUS_RESTRICT yp0 = out+(overlap>>1);
-      kiss_fft_scalar * OPUS_RESTRICT yp1 = out+(overlap>>1)+N2-2;
+      kiss_fft_scalar * yp1 = yp0+N2-2;
       const kiss_twiddle_scalar *t = &trig[0];
       /* Loop to (N4+1)>>1 to handle odd N4. When N4 is odd, the
          middle pair will be computed twice. */


This should always be safe because by making yp1 directly derived from yp0 the compiler should be able to determine that the addresses could overlap.
I am not sure why this is faster than removing restrict from both, though. Mysteries of gcc I guess...

I also tried replacing yp1[k] with yp0[k+N2-2-4*i] with k={1,2}, but that seemed to make no measurable difference.
Also, replacing the loop with a first loop (to N4) for the rotation and a second loop (to N4>>1) for the shuffle made no measurable impact with my compiler (i.e. same as the result after the last suggested patch). It might still be worth considering, as I personally found that easier to read (it also avoids calculating the central value twice for odd N4).

Marcello


________________________________________
From: Jean-Marc Valin [jmvalin at jmvalin.ca]
Sent: 22 February 2014 15:56
To: Marcello Caramma (mcaramma); opus at xiph.org
Subject: Re: [opus] Make check failure on clone from 31 January

Good catch! I think gcc is just unable to make use of restrict for
anything useful. Apparently your compiler does, so could you benchmark
both patches below and see if it makes a difference in terms of speed.
If it's worth it, I can keep the restrict and add the odd case handling,
but if not I'll just remove the restrict.

Thanks,

        Jean-Marc

On 21/02/14 05:54 PM, Marcello Caramma (mcaramma) wrote:
> I tracked down the bug to an incorrect use of restrict.
>
> I would not consider this a compiler bug: we are lying to the optimizer by
> telling it that a pointer is restrict when in fact it isn't.
>
> This can be fixed like so:
>
>
> diff --git a/celt/mdct.c b/celt/mdct.c
> index 1634e8e..fa5098c 100644
> --- a/celt/mdct.c
> +++ b/celt/mdct.c
> @@ -276,8 +276,8 @@ void clt_mdct_backward(const mdct_lookup *l,
> kiss_fft_scalar *in, kiss_fft_scala
>      /* Post-rotate and de-shuffle from both ends of the buffer at once to
> make
>         it in-place. */
>      {
> -      kiss_fft_scalar * OPUS_RESTRICT yp0 = out+(overlap>>1);
> -      kiss_fft_scalar * OPUS_RESTRICT yp1 = out+(overlap>>1)+N2-2;
> +      kiss_fft_scalar * yp0 = out+(overlap>>1);
> +      kiss_fft_scalar * yp1 = out+(overlap>>1)+N2-2;
>         const kiss_twiddle_scalar *t = &trig[0];
>         /* Loop to (N4+1)>>1 to handle odd N4. When N4 is odd, the
>            middle pair will be computed twice. */
>
>
> Or like so:
>
>
> diff --git a/celt/mdct.c b/celt/mdct.c
> index 1634e8e..cdd053f 100644
> --- a/celt/mdct.c
> +++ b/celt/mdct.c
> @@ -279,9 +279,8 @@ void clt_mdct_backward(const mdct_lookup *l,
> kiss_fft_scalar *in, kiss_fft_scala
>         kiss_fft_scalar * OPUS_RESTRICT yp0 = out+(overlap>>1);
>         kiss_fft_scalar * OPUS_RESTRICT yp1 = out+(overlap>>1)+N2-2;
>         const kiss_twiddle_scalar *t = &trig[0];
> -      /* Loop to (N4+1)>>1 to handle odd N4. When N4 is odd, the
> -         middle pair will be computed twice. */
> -      for(i=0;i<(N4+1)>>1;i++)
> +      /* Loop to N4>>1 to make sure pointers never overlap */
> +      for(i=0;i<N4>>1;i++)
>         {
>            kiss_fft_scalar re, im, yr, yi;
>            kiss_twiddle_scalar t0, t1;
> @@ -309,6 +308,20 @@ void clt_mdct_backward(const mdct_lookup *l,
> kiss_fft_scalar *in, kiss_fft_scala
>            yp0 += 2;
>            yp1 -= 2;
>         }
> +      // Handle the odd case
> +      if(N4 & 1)
> +      {
> +         kiss_fft_scalar re, im;
> +         kiss_twiddle_scalar t0, t1;
> +         /* We swap real and imag because we're using an FFT instead of
> an IFFT. */
> +         re = yp0[1];
> +         im = yp0[0];
> +         t0 = t[i];
> +         t1 = t[N4+i];
> +         /* We'd scale up by 2 here, but instead it's done when mixing
> the windows */
> +         yp0[0] = S_MUL(re,t0) + S_MUL(im,t1);
> +         yp0[1] = S_MUL(re,t1) - S_MUL(im,t0);
> +      }
>      }
>
>      /* Mirror on both sides for TDAC */
>
>
> Regards,
>
> Marcello
>
>
>
> On 05/02/2014 18:46, "Gregory Maxwell" <gmaxwell at gmail.com> wrote:
>
>> On Wed, Feb 5, 2014 at 8:05 AM, Marcello Caramma (mcaramma)
>> <mcaramma at cisco.com> wrote:
>>> Hi,
>>>
>>> Apologies if this is a known issue, but running make on revision
>>> e3187444692195957eb66989622c7b1ad8448b06 fails one of the tests when
>>> using fixed point configuration (floating point is ok) on my linux x86.
>>> Note that libopus1.1, as extracted from the tar ball, is OK.
>>
>> I can't reproduce with Fedora 19, gcc version 4.8.2 20131212 (Red Hat
>> 4.8.2-7) (GCC) compiled with
>> CFLAGS='-m32 -O2 -g'  ./configure --enable-fixed-point ; make clean ;
>> make celt/tests/test_unit_mdct
>>
>> or with Clang 3.5 on the same system.
>>
>> None of the fixed point builds on or ci system are throwing errors on
>> this either.
>>
>> Can you try compiling without optimizations? Smells like a compiler bug.
>
> _______________________________________________
> opus mailing list
> opus at xiph.org
> http://lists.xiph.org/mailman/listinfo/opus
>