From 5a191f0896681f375401824fdc806c8e69fcafe9 Mon Sep 17 00:00:00 2001
From: Nach <Nach@a31d4220-a93d-0410-bf67-fe4944624d44>
Date: Tue, 20 Nov 2007 16:03:49 +0000
Subject: [PATCH] HQXx now all working nicely in SDL. Can use assembly
 optimized in SDL port if compiled with USEASM=yes.

---
 Makefile        | 10 ++++++++--
 src/hq3x32.cpp  | 29 +++++++++++++++++++++++++----
 src/hqxx.h      |  4 ++--
 src/sdl/SDL.cpp |  8 ++++----
 4 files changed, 39 insertions(+), 12 deletions(-)

diff --git a/Makefile b/Makefile
index 41955ed1..57ed5c23 100644
--- a/Makefile
+++ b/Makefile
@@ -28,7 +28,6 @@ ifeq ($(PLATFORM),win-cross)
   OUT=vba.exe
 endif
 
-
 MAINDIR=src
 SDLDIR=src/sdl
 DMGDIR=src/gb
@@ -59,9 +58,16 @@ ${DMGDIR}/gbSound${OE}
 
 SDLOBJ=${SDLDIR}/debugger${OE} ${SDLDIR}/SDL${OE} ${SDLDIR}/dummy${OE}
 
-OBJECTS=${MAINOBJ} ${DMGOBJ} ${SDLOBJ} ${CALTERNOBJ}
+OBJECTS=${MAINOBJ} ${DMGOBJ} ${SDLOBJ}
 LIB=${RESAMPLEDIR}/filterkit${OE} ${RESAMPLEDIR}/resample${OE} ${RESAMPLEDIR}/resamplesubs${OE}
 
+ifeq ($(USEASM),yes)
+OBJECTS+=${ASMOBJ}
+else
+OBJECTS+=${CALTERNOBJ}
+endif
+
+
 .SUFFIXES: .c .cpp .asm
 
 %${OE}: %.c
diff --git a/src/hq3x32.cpp b/src/hq3x32.cpp
index 0e50a9b3..f8d5cd14 100644
--- a/src/hq3x32.cpp
+++ b/src/hq3x32.cpp
@@ -55,13 +55,12 @@ void InitLUTs(void)
 int hq3xinited=0;
 extern int realsystemRedShift, realsystemBlueShift;
 
+//16 bit input, see below for 32 bit input
 void hq3x32(unsigned char * pIn,  unsigned int srcPitch,
 			unsigned char *,
 			unsigned char * pOut, unsigned int dstPitch,
 			int Xres, int Yres)
 {
-	// NOTICE!  This driver wants 16 bit, not 32 bit input!
-
 	if (!hq3xinited)
 	{
 		InitLUTs();
@@ -114,13 +113,12 @@ void hq4x16(unsigned char * pIn,  unsigned int srcPitch,
 	hq4x_16( pIn, pOut, Xres, Yres, dstPitch, srcPitch - (Xres *2));
 }
 
+//16 bit input, see below for 32 bit input
 void hq4x32(unsigned char * pIn,  unsigned int srcPitch,
 			unsigned char *,
 			unsigned char * pOut, unsigned int dstPitch,
 			int Xres, int Yres)
 {
-	// NOTICE!  This driver wants 16 bit, not 32 bit input!
-
 	if (!hq3xinited)
 	{
 		InitLUTs();
@@ -145,3 +143,26 @@ void hq4x32(unsigned char * pIn,  unsigned int srcPitch,
 		}
 	}
 }
+
+static inline void convert32bpp_16bpp(unsigned char *pIn, unsigned int width)
+{
+  for (unsigned int i = 0; i < width; i+=4)
+  {
+    unsigned int p4 = ((unsigned int)pIn[i+2] << 16) | (unsigned int) (pIn[i+1] << 8) | pIn[i+0];
+    unsigned short p2 = ((p4 >> 8)&0xF800) | ((p4 >> 5)&0x07E0) | ((p4 >> 3)&0x001F);
+    pIn[i/2] = (p2 >> 0);
+    pIn[i/2+1] = (p2 >> 8);
+  }
+}
+
+void hq3x32_32(unsigned char *pIn,  unsigned int srcPitch, unsigned char *, unsigned char *pOut, unsigned int dstPitch, int Xres, int Yres)
+{
+  convert32bpp_16bpp(pIn, srcPitch*Yres);
+  hq3x32(pIn, srcPitch/2, 0, pOut, dstPitch, Xres, Yres);
+}
+
+void hq4x32_32(unsigned char *pIn,  unsigned int srcPitch, unsigned char *, unsigned char *pOut, unsigned int dstPitch, int Xres, int Yres)
+{
+  convert32bpp_16bpp(pIn, srcPitch*Yres);
+  hq4x32(pIn, srcPitch/2, 0, pOut, dstPitch, Xres, Yres);
+}
diff --git a/src/hqxx.h b/src/hqxx.h
index 2555c27d..43ccc789 100644
--- a/src/hqxx.h
+++ b/src/hqxx.h
@@ -271,7 +271,7 @@ void hq3x16(unsigned char * pIn,  unsigned int srcPitch,
 			unsigned char * pOut, unsigned int dstPitch,
 			int Xres, int Yres)
 #elif defined(_32BIT)
-void hq3x32(unsigned char * pIn,  unsigned int srcPitch,
+void hq3x32_32(unsigned char * pIn,  unsigned int srcPitch,
 			unsigned char *,
 			unsigned char * pOut, unsigned int dstPitch,
 			int Xres, int Yres)
@@ -283,7 +283,7 @@ void hq4x16(unsigned char * pIn,  unsigned int srcPitch,
 			unsigned char * pOut, unsigned int dstPitch,
 			int Xres, int Yres)
 #elif defined(_32BIT)
-void hq4x32(unsigned char * pIn,  unsigned int srcPitch,
+void hq4x32_32(unsigned char * pIn,  unsigned int srcPitch,
 			unsigned char *,
 			unsigned char * pOut, unsigned int dstPitch,
 			int Xres, int Yres)
diff --git a/src/sdl/SDL.cpp b/src/sdl/SDL.cpp
index 180abb77..c073a11a 100644
--- a/src/sdl/SDL.cpp
+++ b/src/sdl/SDL.cpp
@@ -90,9 +90,9 @@ extern void hq2x32(u8*,u32,u8*,u8*,u32,int,int);
 extern void lq2x(u8*,u32,u8*,u8*,u32,int,int);
 extern void lq2x32(u8*,u32,u8*,u8*,u32,int,int);
 extern void hq3x16(u8*,u32,u8*,u8*,u32,int,int);
-extern void hq3x32(u8*,u32,u8*,u8*,u32,int,int);
+extern void hq3x32_32(u8*,u32,u8*,u8*,u32,int,int);
 extern void hq4x16(u8*,u32,u8*,u8*,u32,int,int);
-extern void hq4x32(u8*,u32,u8*,u8*,u32,int,int);
+extern void hq4x32_32(u8*,u32,u8*,u8*,u32,int,int);
 
 extern void SmartIB(u8*,u32,int,int);
 extern void SmartIB32(u8*,u32,int,int);
@@ -2538,10 +2538,10 @@ int main(int argc, char **argv)
       filterFunction = lq2x32;
       break;
     case 14:
-      filterFunction = hq3x32;
+      filterFunction = hq3x32_32;
       break;
     case 15:
-      filterFunction = hq4x32;
+      filterFunction = hq4x32_32;
       break;
     default:
       filterFunction = NULL;