Added assembly versions of HQ2x and HQ3x scalers.

svn-id: r13844

Makefile.common

@@ -158,6 +158,13 @@ else
 	$(CXX) -Wp,-MMD,"$(*D)/$(DEPDIR)/$(*F).d",-MQ,"$@",-MP $(CXXFLAGS) $(CPPFLAGS) -c $(<) -o $*.o
 endif
 
+ifdef HAVE_NASM
+.SUFFIXES: .asm
+
+.asm.o:
+	$(NASM) -O1 $(NASMFLAGS) -g -o $*.o $(<)
+endif
+
 # Include the dependency tracking files. We add /dev/null at the end
 # of the list to avoid a warning/error if no .d file exist
 -include $(wildcard $(addsuffix /*.d,$(DEPDIRS))) /dev/null

NEWS

@@ -7,6 +7,7 @@ For a more comprehensive changelog for the latest experimental CVS code, see:
    - Added an 'On Screen Display' to the SDL backend
    - Rewrote the backend API partially
    - Comments in config files are preserved now
+   - Added MMX i386 assembler HQ2x and HQ3x scalers
 
  SCUMM:
    - Added graphics decoders for 3DO Humongous Entertainment games

README

@@ -999,6 +999,11 @@ Visual C++ are supported. If you wish to use MP3-compressed CD tracks or
 USE_MAD. Tools for compressing .SOU files to .SO3 files can be
 found in the 'tools' CVS module, or in the 'scummvm-tools' package.
 
+Some parts of ScummVM, particularly scalers, have highly optimized versions 
+written in assembler. If you wish to use this option, you will need to install 
+nasm assembler (see http://nasm.sf.net). Note, that currently we have only x86
+MMX optimized versions, and they will not compile on other processors.
+
 On Win9x/NT/XP you can define USE_WINDBG and attach WinDbg to browse debug 
 messages (see http://www.sysinternals.com/ntw2k/freeware/debugview.shtml).
 

common/module.mk

@@ -16,6 +16,12 @@ MODULE_OBJS := \
 	common/scaler/hq2x.o \
 	common/scaler/hq3x.o
 
+ifdef HAVE_NASM
+MODULE_OBJS += \
+	common/scaler/hq2x_i386.o \
+	common/scaler/hq3x_i386.o
+endif
+
 MODULE_DIRS += \
 	common \
 	common/scaler

common/scaler.cpp

@@ -27,7 +27,22 @@
 int gBitFormat = 565;
 
 // RGB-to-YUV lookup table
-int RGBtoYUV[65536];
+extern "C" {
+
+#ifdef USE_NASM
+// NOTE: if your compiler uses different mangled names, add another
+//       condition here
+
+#ifndef _MSC_VER
+#define RGBtoYUV _RGBtoYUV
+#define LUT16to32 _LUT16to32
+#endif
+
+#endif
+
+uint RGBtoYUV[65536];
+uint LUT16to32[65536];
+}
 
 static const uint16 dotmatrix_565[16] = {
 	0x01E0, 0x0007, 0x3800, 0x0000,
@@ -63,6 +78,10 @@ void InitLUT(uint32 BitFormat) {
 	int Y, u, v;
 	int gInc, gShift;
 	
+	for (int i = 0; i < 65536; i++) {
+		LUT16to32[i] = ((i & 0xF800) << 8) + ((i & 0x07E0) << 5) + ((i & 0x001F) << 3);
+	}
+
 	if (BitFormat == 565) {
 		gInc = 256 >> 6;
 		gShift = 6 - 3;
@@ -76,8 +95,8 @@ void InitLUT(uint32 BitFormat) {
 			for (b = 0; b < 256; b += 8) {
 				Y = (r + g + b) >> 2;
 				u = 128 + ((r - b) >> 2);
-				v = 128 + ((-r + 2 * g -b) >> 3);
-				RGBtoYUV[ (r << (5+gShift)) + (g << gShift) + (b >> 3) ] = (Y << 16) + (u << 8) + v;
+				v = 128 + ((-r + 2 * g - b) >> 3);
+				RGBtoYUV[ (r << (5 + gShift)) + (g << gShift) + (b >> 3) ] = (Y << 16) + (u << 8) + v;
 			}
 		}
 	}

common/scaler/hq2x.cpp

@@ -22,6 +22,25 @@
 
 #include "common/scaler/intern.h"
 
+#ifdef USE_NASM
+// Assembly version of HQ2x
+
+extern "C" {
+
+#ifndef _MSC_VER
+#define hq2x_16 _hq2x_16
+#endif
+
+void hq2x_16(const byte *, byte *, uint32, uint32, uint32, uint32);
+
+}
+
+void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, int width, int height) {
+	hq2x_16(srcPtr, dstPtr, width, height, srcPitch, dstPitch);
+}
+
+#else
+
 #ifdef HAS_ALTIVEC
 #include <sys/sysctl.h> 
 
@@ -120,7 +139,6 @@ void HQ2x_555(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPit
 	#undef bitFormat
 #endif
 
-
 void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, int width, int height) {
 #ifdef HAS_ALTIVEC
 	if (isAltiVecAvailable()) {
@@ -131,8 +149,11 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 		return;
 	}
 #endif
+
 	if (gBitFormat == 565)
 		HQ2x_565(srcPtr, srcPitch, dstPtr, dstPitch, width, height);
 	else
 		HQ2x_555(srcPtr, srcPitch, dstPtr, dstPitch, width, height);
 }
+
+#endif //Assembly version

common/scaler/hq3x.cpp

@@ -22,6 +22,25 @@
 
 #include "common/scaler/intern.h"
 
+#ifdef USE_NASM
+// Assembly version of HQ3x
+
+extern "C" {
+
+#ifndef _MSC_VER
+#define hq3x_16 _hq3x_16
+#endif
+
+void hq3x_16(const byte *, byte *, uint32, uint32, uint32, uint32);
+
+}
+
+void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, int width, int height) {
+	hq3x_16(srcPtr, dstPtr, width, height, srcPitch, dstPitch);
+}
+
+#else
+
 #ifdef HAS_ALTIVEC
 #include <sys/sysctl.h> 
 
@@ -122,7 +141,6 @@ void HQ3x_555(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPit
 	#undef bitFormat
 #endif
 
-
 void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, int width, int height) {
 #ifdef HAS_ALTIVEC
 	if (isAltiVecAvailable()) {
@@ -133,8 +151,11 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 		return;
 	}
 #endif
+
 	if (gBitFormat == 565)
 		HQ3x_565(srcPtr, srcPitch, dstPtr, dstPitch, width, height);
 	else
 		HQ3x_555(srcPtr, srcPitch, dstPtr, dstPitch, width, height);
 }
+
+#endif

common/scaler/intern.h

@@ -153,7 +153,7 @@ static inline bool diffYUV(int yuv1, int yuv2) {
  * 16bit RGB to YUV conversion table. This table is setup by InitLUT().
  * Used by the hq scaler family.
  */
-extern int   RGBtoYUV[65536];
+extern "C" uint   RGBtoYUV[65536];
 
 /** Auxiliary macro to simplify creating those template function wrappers. */
 #define MAKE_WRAPPER(FUNC) \

configure

@@ -42,12 +42,16 @@ _build_kyra=no
 _build_saga=no
 _need_memalign=no
 _build_plugins=no
+_nasm=auto
 # more defaults
 _backend=sdl
 _ranlib=ranlib
 _install=install
 _sdlconfig=sdl-config
 _sdlpath="$PATH"
+_nasmpath="$PATH"
+NASMFLAGS=""
+NASM=""
 _prefix=/usr/local
 
 _srcdir=`dirname $0`
@@ -177,6 +181,45 @@ rm -f tmp_find_type_with_size$EXEEXT tmp_find_type_with_size.cpp
 echo $datatype
 }
 
+CheckNASM()
+{
+    echocheck "nasm"
+    if test "$_nasm" = no ; then
+	echo "disabled"
+	return;
+    fi
+
+    IFS="${IFS= 	}"; ac_save_ifs="$IFS"; IFS=":"
+	
+    for path_dir in $_nasmpath; do
+	if test -x "$path_dir/nasm" ; then
+	    NASM="$path_dir/nasm"
+	    echo $NASM
+	    break
+	fi
+    done
+	
+    IFS="$ac_save_ifs"
+
+    if test x$NASM = x -o x$NASM = x'"$NASM"'; then
+	echo "not found"
+	_nasm=no
+    else
+	case $_host_os in
+	    mingw* | cygwin*)
+		NASMFLAGS="-f win32"
+	    ;;
+	    openbsd*)
+                NASMFLAGS="-f aoutb"
+            ;;
+            *)
+                NASMFLAGS="-f elf"
+            ;;
+	esac
+	_nasm=yes
+    fi
+}
+
 #
 # Greet user
 #
@@ -245,7 +288,10 @@ Optional Libraries:
   --with-mpeg2-prefix=DIR  Prefix where libmpeg2 is installed (optional)
   --disable-mpeg2          disable mpeg2 codec for cutscenes [autodetect]
 
-  --with-sdl-prefix=DIR    Prefix where the sdl-config script is installed
+  --with-sdl-prefix=DIR    Prefix where the sdl-config script is installed (optional)
+
+  --with-nasm-prefix=DIR   Prefix where nasm executable is installed (optional)
+  --disable-nasm           disable assembly language optimizations [autodetect]
 
 Some influential environment variables:
   LDFLAGS	linker flags, e.g. -L<lib dir> if you have libraries in a
@@ -282,6 +328,8 @@ for ac_option in $@; do
       --disable-mad)		_mad=no		;;
       --enable-zlib)		_zlib=yes	;;
       --disable-zlib)		_zlib=no	;;
+      --enable-nasm)		_nasm=yes	;;
+      --disable-nasm)		_nasm=no	;;
       --disable-mpeg2)          _mpeg2=no       ;;
       --enable-plugins)         _build_plugins=yes ;;
       --with-mpeg2-prefix=*)
@@ -332,6 +380,10 @@ for ac_option in $@; do
 	arg=`echo $ac_option | cut -d '=' -f 2`
 	_sdlpath="$arg:$arg/bin"
 	;;
+      --with-nasm-prefix=*)
+	arg=`echo $ac_option | cut -d '=' -f 2`
+	_nasmpath="$arg:$arg/bin"
+	;;
       --host=*)
 	_host=`echo $ac_option | cut -d '=' -f 2`
 	;;
@@ -509,7 +561,6 @@ else
 	_mak_saga='# DISABLE_SAGA = 1'
 fi
 
-
 if test -n "$_host"; then
     # Cross-compiling mode - add your target here if needed
     case "$_host" in
@@ -847,6 +898,19 @@ fi
 echo "$_mpeg2"
 rm -f $TMPC $TMPO$EXEEXT
 
+#
+# Check for nasm
+#
+CheckNASM
+
+if test "$_nasm" = yes ; then
+    _def_nasm='#define USE_NASM'
+    _make_def_HAVE_NASM='HAVE_NASM = 1'
+else
+    _def_nasm='#undef USE_NASM'
+    _make_def_HAVE_NASM='# HAVE_NASM = 1'
+fi
+
 #
 # figure out installation directories
 #
@@ -892,10 +956,17 @@ fi
 if test "$_build_kyra" = yes ; then
   echo "    Legend of Kyrandia"
 fi
+
 echo
 
 echo_n "Backend... "
-echo "$_backend"
+echo_n "$_backend"
+
+if test "$_nasm" = yes ; then
+    echo ", assembly routines"
+else
+    echo
+fi
 
 #
 # Backend related stuff
@@ -954,6 +1025,9 @@ $_def_alsa
 $_def_zlib
 $_def_mpeg2
 
+/* Should we use i386 assembly routines */
+$_def_nasm
+
 #endif /* CONFIG_H */
 EOF
 
@@ -970,6 +1044,8 @@ BACKEND := $_backend
 MODULES += $MODULES
 MODULE_DIRS += $MODULE_DIRS
 EXEEXT := $EXEEXT
+NASM := $NASM
+NASMFLAGS := $NASMFLAGS
 
 PREFIX := $_prefix
 BINDIR := $_bindir
@@ -977,6 +1053,7 @@ MANDIR := $_mandir
 
 $_mak_plugins
 $_make_def_HAVE_GCC3
+$_make_def_HAVE_NASM
 $_mak_scumm
 $_mak_simon
 $_mak_sky

doc/09.tex

@@ -13,6 +13,11 @@ Visual C++ are supported. If you wish to use MP3-compressed CD tracks or
 USE\_MAD. Tools for compressing .SOU files to .SO3 files can be
 found in the 'tools' CVS module, or in the 'scummvm-tools' package.
 
+Some parts of ScummVM, particularly scalers, have highly optimized versions 
+written in assembler. If you wish to use this option, you will need to install 
+nasm assembler (see \url{http://nasm.sf.net}). Note, that currently we have only x86
+MMX optimized versions, and they will not compile on other processors.
+
 On Win9x/NT/XP you can define USE\_WINDBG and attach WinDbg to browse debug 
 messages (see \url{http://www.sysinternals.com/ntw2k/freeware/debugview.shtml}).