PPC970 patch + cpu-specific tuning support by (Romain Dolbeau <dolbeau at irisa dot fr>)

git-svn-id: file:///var/local/repositories/ffmpeg/trunk@1997 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b

configure

@@ -26,6 +26,7 @@ ranlib="ranlib"
 make="make"
 strip="strip"
 cpu=`uname -m`
+tune="generic"
 mmx="default"
 altivec="default"
 mmi="default"
@@ -272,6 +273,8 @@ for opt do
   ;;
   --cpu=*) cpu=`echo $opt | cut -d '=' -f 2`
   ;;
+  --tune=*) tune=`echo $opt | cut -d '=' -f 2`
+  ;;
   --disable-mmx) mmx="no"
   ;;
   --disable-altivec) altivec="no"
@@ -351,6 +354,58 @@ if test $altivec = "default"; then
     fi
 fi
 
+# Add processor-specific flags
+TUNECPU="generic"
+if test $tune != "generic"; then
+    case $tune in
+	601|ppc601|PowerPC601)
+	    CFLAGS="$CFLAGS -mcpu=601"
+	    if test $altivec = "yes"; then
+		echo "WARNING: tuning for PPC601 but altivec enabled !";
+	    fi
+	    TUNECPU=ppc601
+	;;
+	603*|ppc603*|PowerPC603*)
+	    CFLAGS="$CFLAGS -mcpu=603"
+	    if test $altivec = "yes"; then
+		echo "WARNING: tuning for PPC603 but altivec enabled !";
+	    fi
+	    TUNECPU=ppc603
+	;;
+	604*|ppc604*|PowerPC604*)
+	    CFLAGS="$CFLAGS -mcpu=604"
+	    if test $altivec = "yes"; then
+		echo "WARNING: tuning for PPC604 but altivec enabled !";
+	    fi
+	    TUNECPU=ppc604
+	;;
+	G3|75*|ppc75*|PowerPC75*)
+	    CFLAGS="$CFLAGS -mcpu=750"
+	    if test $altivec = "yes"; then
+		echo "WARNING: tuning for PPC75x but altivec enabled !";
+	    fi
+	    TUNECPU=ppc750
+	;;
+	G4|74*|ppc74*|PowerPC74*)
+	    CFLAGS="$CFLAGS -mcpu=7400"
+	    if test $altivec = "no"; then
+		echo "WARNING: tuning for PPC74xx but altivec disabled !";
+	    fi
+	    TUNECPU=ppc7400
+	;;
+	G5|970|ppc970|PowerPC970|power4*|Power4*)
+	    CFLAGS="$CFLAGS -mcpu=970 -mtune=970 -mpowerpc64 -force_cpusubtype_ALL "
+	    if test $altivec = "no"; then
+		echo "WARNING: tuning for PPC970 but altivec disabled !";
+	    fi
+	    TUNECPU=ppc970
+	;;
+	*)
+	echo "WARNING: unknown CPU "$tune", ignored"
+	;;
+    esac
+fi
+
 # See if we have <altivec.h>
 cat > $TMPC << EOF
 #include <altivec.h>
@@ -686,6 +741,7 @@ echo "  --extra-cflags=ECFLAGS   add ECFLAGS to CFLAGS [$CFLAGS]"
 echo "  --extra-ldflags=ELDFLAGS add ELDFLAGS to LDFLAGS [$LDFLAGS]"
 echo "  --extra-libs=ELIBS       add ELIBS [$ELIBS]"
 echo "  --cpu=CPU                force cpu to CPU  [$cpu]"
+echo "  --tune=PROCESSOR         tune code for a particular CPU (may fails or misperforms on other CPUs)"
 echo "  --disable-mmx            disable mmx usage"
 echo "  --disable-altivec        disable AltiVec usage"
 echo "  --disable-audio-oss      disable OSS audio support [default=no]"
@@ -712,7 +768,7 @@ echo "Install prefix   $prefix"
 echo "Source path      $source_path"
 echo "C compiler       $cc"
 echo "make             $make"
-echo "CPU              $cpu"
+echo "CPU              $cpu ($tune)"
 echo "Big Endian       $bigendian"
 if test $cpu = "x86"; then
 echo "MMX enabled      $mmx"
@@ -792,6 +848,7 @@ elif test "$cpu" = "mips" ; then
   echo "TARGET_ARCH_MIPS=yes" >> config.mak
   echo "#define ARCH_MIPS 1" >> $TMPH
 fi
+echo "#define TUNECPU $TUNECPU" >> $TMPH
 if test "$bigendian" = "yes" ; then
   echo "WORDS_BIGENDIAN=yes" >> config.mak
   echo "#define WORDS_BIGENDIAN 1" >> $TMPH

libavcodec/ppc/dsputil_ppc.c

@@ -57,7 +57,8 @@ static unsigned char* perfname[] = {
   "put_no_rnd_pixels8_xy2_altivec",
   "put_pixels16_xy2_altivec",
   "put_no_rnd_pixels16_xy2_altivec",
-  "clear_blocks_dcbz32_ppc"
+  "clear_blocks_dcbz32_ppc",
+  "clear_blocks_dcbz128_ppc"
 };
 #ifdef POWERPC_PERF_USE_PMC
 unsigned long long perfdata_miss[powerpc_perf_total][powerpc_data_total];
@@ -110,6 +111,18 @@ void powerpc_display_perf_report(void)
   It simply clear to zero a single cache line,
   so you need to know the cache line size to use it !
   It's absurd, but it's fast...
+
+  update 24/06/2003 : Apple released yesterday the G5,
+  with a PPC970. cache line size : 128 bytes. Oups.
+  The semantic of dcbz was changed, it always clear
+  32 bytes. so the function below will work, but will
+  be slow. So I fixed check_dcbz_effect to use dcbzl,
+  which is defined to clear a cache line (as dcbz before).
+  So we still can distinguish, and use dcbz (32 bytes)
+  or dcbzl (one cache line) as required.
+
+  see <http://developer.apple.com/technotes/tn/tn2087.html>
+  and <http://developer.apple.com/technotes/tn/tn2086.html>
 */
 void clear_blocks_dcbz32_ppc(DCTELEM *blocks)
 {
@@ -126,7 +139,7 @@ POWERPC_TBL_START_COUNT(powerpc_clear_blocks_dcbz32, 1);
       i += 16;
     }
     for ( ; i < sizeof(DCTELEM)*6*64 ; i += 32) {
-      asm volatile("dcbz %0,%1" : : "r" (blocks), "r" (i) : "memory");
+      asm volatile("dcbz %0,%1" : : "r" (i), "r" (blocks) : "memory");
     }
     if (misal) {
       ((unsigned long*)blocks)[188] = 0L;
@@ -141,8 +154,45 @@ POWERPC_TBL_START_COUNT(powerpc_clear_blocks_dcbz32, 1);
 POWERPC_TBL_STOP_COUNT(powerpc_clear_blocks_dcbz32, 1);
 }
 
+/* same as above, when dcbzl clear a whole 128B cache line
+   i.e. the PPC970 aka G5 */
+#ifndef NO_DCBZL
+void clear_blocks_dcbz128_ppc(DCTELEM *blocks)
+{
+POWERPC_TBL_DECLARE(powerpc_clear_blocks_dcbz128, 1);
+    register int misal = ((unsigned long)blocks & 0x0000007f);
+    register int i = 0;
+POWERPC_TBL_START_COUNT(powerpc_clear_blocks_dcbz128, 1);
+#if 1
+ if (misal) {
+   // we could probably also optimize this case,
+   // but there's not much point as the machines
+   // aren't available yet (2003-06-26)
+      memset(blocks, 0, sizeof(DCTELEM)*6*64);
+    }
+    else
+      for ( ; i < sizeof(DCTELEM)*6*64 ; i += 128) {
+	asm volatile("dcbzl %0,%1" : : "r" (i), "r" (blocks) : "memory");
+      }
+#else
+    memset(blocks, 0, sizeof(DCTELEM)*6*64);
+#endif
+POWERPC_TBL_STOP_COUNT(powerpc_clear_blocks_dcbz128, 1);
+}
+#else
+void clear_blocks_dcbz128_ppc(DCTELEM *blocks)
+{
+  memset(blocks, 0, sizeof(DCTELEM)*6*64);
+}
+#endif
+
+#ifndef NO_DCBZL
 /* check dcbz report how many bytes are set to 0 by dcbz */
-long check_dcbz_effect(void)
+/* update 24/06/2003 : replace dcbz by dcbzl to get
+   the intended effect (Apple "fixed" dcbz)
+   unfortunately this cannot be used unless the assembler
+   knows about dcbzl ... */
+long check_dcbzl_effect(void)
 {
   register char *fakedata = (char*)av_malloc(1024);
   register char *fakedata_middle;
@@ -159,7 +209,7 @@ long check_dcbz_effect(void)
 
   memset(fakedata, 0xFF, 1024);
 
-  asm volatile("dcbz %0, %1" : : "r" (fakedata_middle), "r" (zero));
+  asm volatile("dcbzl %0, %1" : : "r" (fakedata_middle), "r" (zero));
 
   for (i = 0; i < 1024 ; i ++)
   {
@@ -171,15 +221,24 @@ long check_dcbz_effect(void)
   
   return count;
 }
+#else
+long check_dcbzl_effect(void)
+{
+  return 0;
+}
+#endif
 
 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx)
 {
-    // Common optimisations whether Altivec or not
+    // Common optimizations whether Altivec is available or not
 
-  switch (check_dcbz_effect()) {
+  switch (check_dcbzl_effect()) {
   case 32:
     c->clear_blocks = clear_blocks_dcbz32_ppc;
     break;
+  case 128:
+    c->clear_blocks = clear_blocks_dcbz128_ppc;
+    break;
   default:
     break;
   }

libavcodec/ppc/dsputil_ppc.h

@@ -37,6 +37,7 @@ enum powerpc_perf_index {
   altivec_put_pixels16_xy2_num,
   altivec_put_no_rnd_pixels16_xy2_num,
   powerpc_clear_blocks_dcbz32,
+  powerpc_clear_blocks_dcbz128,
   powerpc_perf_total
 };
 enum powerpc_data_index {

libavcodec/ppc/gcc_fixes.h

@@ -13,9 +13,15 @@
 
 #ifdef CONFIG_DARWIN
 #define AVV(x...) (x)
+/* The Apple assembler shipped w/ gcc-3.3 knows about DCBZL, previous assemblers don't
+   We assume here that the Darwin GCC is from Apple.... */
+#if (__GNUC__ * 100 + __GNUC_MINOR__ < 303)
+#define NO_DCBZL
+#endif
 #else
 #define AVV(x...) {x}
-
+/* I don't think any non-Apple assembler knows about DCBZL */
+#define NO_DCBZL
 #if (__GNUC__ * 100 + __GNUC_MINOR__ < 303)  
 
 /* This code was provided to me by Bartosch Pixa