[xiph-commits] r16212 - in branches/theora-thusnelda/lib/enc: x86 x86_vc

[tterribe at svn.xiph.org]

Author: tterribe
Date: 2009-07-07 04:55:25 -0700 (Tue, 07 Jul 2009)
New Revision: 16212

Modified:
   branches/theora-thusnelda/lib/enc/x86/mmxfdct.c
   branches/theora-thusnelda/lib/enc/x86/sse2fdct.c
   branches/theora-thusnelda/lib/enc/x86_vc/mmxfdct.c
Log:
Fix a potential overflow in the fDCT.


Modified: branches/theora-thusnelda/lib/enc/x86/mmxfdct.c
===================================================================
--- branches/theora-thusnelda/lib/enc/x86/mmxfdct.c	2009-07-07 01:43:19 UTC (rev 16211)
+++ branches/theora-thusnelda/lib/enc/x86/mmxfdct.c	2009-07-07 11:55:25 UTC (rev 16212)
@@ -177,12 +177,14 @@
  "packssdw %%mm5,%%mm4\n\t" \
  "movq "_r5"(%[y]),%%mm1\n\t" \
  "paddw %%mm2,%%mm4\n\t" \
- /*mm2=t6'', mm0=_y[0]=u=r+s>>1*/ \
+ /*mm2=t6'', mm0=_y[0]=u=r+s>>1 \
+   The naive implementation could cause overflow, so we use u=s+(r-s>>1).*/ \
  "mov $0x7FFF54DC,%[a]\n\t" \
- "paddw %%mm4,%%mm0\n\t" \
+ "psubw %%mm4,%%mm0\n\t" \
  "movq "_r3"(%[y]),%%mm2\n\t" \
+ "psraw $1,%%mm0\n\t" \
  "movd %[a],%%mm7\n\t" \
- "psraw $1,%%mm0\n\t" \
+ "paddw %%mm4,%%mm0\n\t" \
  /*mm7={54491-0x7FFF,0x7FFF}x2 \
    mm4=_y[4]=v=r-u*/ \
  "psubw %%mm0,%%mm4\n\t" \

Modified: branches/theora-thusnelda/lib/enc/x86/sse2fdct.c
===================================================================
--- branches/theora-thusnelda/lib/enc/x86/sse2fdct.c	2009-07-07 01:43:19 UTC (rev 16211)
+++ branches/theora-thusnelda/lib/enc/x86/sse2fdct.c	2009-07-07 11:55:25 UTC (rev 16212)
@@ -152,10 +152,12 @@
  "psubw %%xmm14,%%xmm1\n\t" \
  "mov $0x7FFF6C84,%[a]\n\t" \
  "paddw %%xmm1,%%xmm4\n\t" \
- /*xmm0=_y[0]=u=r+s>>1*/ \
- "paddw %%xmm4,%%xmm0\n\t" \
+ /*xmm0=_y[0]=u=r+s>>1 \
+   The naive implementation could cause overflow, so we use u=s+(r-s>>1).*/ \
+ "psubw %%xmm4,%%xmm0\n\t" \
  "movd %[a],%%xmm13\n\t" \
  "psraw $1,%%xmm0\n\t" \
+ "paddw %%xmm4,%%xmm0\n\t" \
  /*xmm4=_y[4]=v=r-u*/ \
  "pshufd $00,%%xmm13,%%xmm13\n\t" \
  "psubw %%xmm0,%%xmm4\n\t" \

Modified: branches/theora-thusnelda/lib/enc/x86_vc/mmxfdct.c
===================================================================
--- branches/theora-thusnelda/lib/enc/x86_vc/mmxfdct.c	2009-07-07 01:43:19 UTC (rev 16211)
+++ branches/theora-thusnelda/lib/enc/x86_vc/mmxfdct.c	2009-07-07 11:55:25 UTC (rev 16212)
@@ -178,12 +178,14 @@
 __asm  packssdw mm4,mm5 \
 __asm  movq mm1,[Y+_r5] \
 __asm  paddw mm4,mm2 \
- /*mm2=t6'', mm0=_y[0]=u=r+s>>1*/ \
+ /*mm2=t6'', mm0=_y[0]=u=r+s>>1 \
+   The naive implementation could cause overflow, so we use u=s+(r-s>>1).*/ \
 __asm  mov A,0x7FFF54DC \
-__asm  paddw mm0,mm4 \
+__asm  psubw mm0,mm4 \
 __asm  movq mm2,[Y+_r3] \
+__asm  psraw mm0,1 \
 __asm  movd mm7,A \
-__asm  psraw mm0,1 \
+__asm  paddw mm0,mm4 \
  /*mm7={54491-0x7FFF,0x7FFF}x2 \
    mm4=_y[4]=v=r-u*/ \
 __asm  psubw mm4,mm0 \