Speex dev - Nov 2005 - Noisy sound quality with Blackfin in WB-mode

Hello Jean-Marc,
> Got any luck with the Blackfin stuff? One thing I forgot mentioning that
> would probably help a lot in narrowing down the bug is to simply disable
> assembly functions one by one and see which one breaks Speex.
sorry that I didn't reply for some days. Unfortunately I had some other
serious
problems with my software which are not related to Speex. I hope to solve my
other problems and to get back to the Speex code during the next week. I'll
try
disabling the assembly functions then.

So far,

Bernhard


----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.

Hello Jean-Marc,

I disabled the Blackfin optimization functions one by one and found  
the error in libspeex/lpc_bfin.h in the autocorrelation computation.  
Compared to the fixed-point-variant of that function (from lpc.c),  
the blackfin-code produces correct values except for the zero offset  
value. Additionally, the loop after the ASM-code assigned values  
(ac32) wrong to the functions output structure (ac). I added a loop  
to compute ac0-value correctly and another loop, that moves ac32- 
values to ac. I checked my code with debug printf's comparing  
Blackfin-ASM-computed values against fixed-point-variant values and  
found the values to be generally the equal. Some  Blackfin-ASM-values  
differ slightly from fixed-point-values what may be the result of  
rounding errors. I can hear no difference any more between fixed- 
point-only- and fixed-point+blackfin-encoded files. I tested for some  
quality and complexity values in wideband and ultra-wideband mode.

Find attached the diff against current svn with my changes.

Best regards,

Bernhard




Index: lpc_bfin.h
==================================================================--- lpc_bfin.h
+++ lpc_bfin.h
@@ -109,23 +110,19 @@
     : : "m" (xs), "m" (x), "m" (ac32top),
"m" (N_lag), "m" (lag_1),
"m" (nshift)
     : "A0", "A1", "P0", "P1",
"P2", "P3", "P4", "R0", "R1",
"R2",
"R3", "R4", "I0", "I1", "L0",
"L1", "B0", "B1", "memory"
     );
+
+
+   for ( i=1; i<lag; i++ )
+   {
+        ac[i]=SHR32(ac32[i+(10-(lag-1))], ac_shift);
+   }
+
+
     d=0;
     for (j=0;j<n;j++)
     {
        d = ADD32(d,SHR32(MULT16_16(x[j],x[j]), shift));
     }
-   ac32[0] = d;
-
-   for (i=0;i<lag;i++)
-   {
-      d=0;
-      for (j=i;j<lag_1;j++)
-      {
-         d = ADD32(d,SHR32(MULT16_16(x[j],x[j-i]), shift));
-      }
-      if (i)
-         ac32[i] += d;
-      ac[i] = SHR32(ac32[i], ac_shift);
-   }
+   ac[0] = SHR32(d, ac_shift);
}

Hi Bernhard,

It seems like I messed up somewhere in the lpc code. However, it seems
like your patch is also removing some code, which is probably why you
see small differences between your asm version and the fixed point.

I don't have my setup working at the moment, but could you test the
following patch:

Index: lpc_bfin.h
==================================================================--- lpc_bfin.h
(r?vision 10333)
+++ lpc_bfin.h  (copie de travail)
@@ -47,7 +47,7 @@
    spx_word32_t ac0=1;
    spx_word32_t ac32[11], *ac32top;
    int shift, ac_shift;
-   ac32top = ac32+10;
+   ac32top = ac32+lag-1;
    int lag_1, N_lag;
    int nshift;
    lag_1 = lag-1;

Let me know if it works.

	Jean-Marc

Le mercredi 09 novembre 2005 ? 14:02 +0100, Bernhard Gerlach a ?crit
:
> Hello Jean-Marc,
> 
> I disabled the Blackfin optimization functions one by one and found  
> the error in libspeex/lpc_bfin.h in the autocorrelation computation.  
> Compared to the fixed-point-variant of that function (from lpc.c),  
> the blackfin-code produces correct values except for the zero offset  
> value. Additionally, the loop after the ASM-code assigned values  
> (ac32) wrong to the functions output structure (ac). I added a loop  
> to compute ac0-value correctly and another loop, that moves ac32- 
> values to ac. I checked my code with debug printf's comparing  
> Blackfin-ASM-computed values against fixed-point-variant values and  
> found the values to be generally the equal. Some  Blackfin-ASM-values  
> differ slightly from fixed-point-values what may be the result of  
> rounding errors. I can hear no difference any more between fixed- 
> point-only- and fixed-point+blackfin-encoded files. I tested for some  
> quality and complexity values in wideband and ultra-wideband mode.
> 
> Find attached the diff against current svn with my changes.
> 
> Best regards,
> 
> Bernhard
> 
> 
> 
> 
> Index: lpc_bfin.h
> ==================================================================> ---
lpc_bfin.h
> +++ lpc_bfin.h
> @@ -109,23 +110,19 @@
>      : : "m" (xs), "m" (x), "m" (ac32top),
"m" (N_lag), "m" (lag_1),
> "m" (nshift)
>      : "A0", "A1", "P0", "P1",
"P2", "P3", "P4", "R0", "R1",
"R2",
> "R3", "R4", "I0", "I1",
"L0", "L1", "B0", "B1",
"memory"
>      );
> +
> +
> +   for ( i=1; i<lag; i++ )
> +   {
> +        ac[i]=SHR32(ac32[i+(10-(lag-1))], ac_shift);
> +   }
> +
> +
>      d=0;
>      for (j=0;j<n;j++)
>      {
>         d = ADD32(d,SHR32(MULT16_16(x[j],x[j]), shift));
>      }
> -   ac32[0] = d;
> -
> -   for (i=0;i<lag;i++)
> -   {
> -      d=0;
> -      for (j=i;j<lag_1;j++)
> -      {
> -         d = ADD32(d,SHR32(MULT16_16(x[j],x[j-i]), shift));
> -      }
> -      if (i)
> -         ac32[i] += d;
> -      ac[i] = SHR32(ac32[i], ac_shift);
> -   }
> +   ac[0] = SHR32(d, ac_shift);
> }
> 
> _______________________________________________
> Speex-dev mailing list
> Speex-dev@xiph.org
> http://lists.xiph.org/mailman/listinfo/speex-dev
>