--- a/dmblitfunc.h	Tue Oct 02 18:14:44 2012 +0300
+++ b/dmblitfunc.h	Tue Oct 02 18:51:16 2012 +0300
@@ -309,5 +309,237 @@
 #include "dmscaledblit.h"
 
 
+
 // =======================================================================
 // =======================================================================
+
+
+
+// =======================================================================
+// DMD_NONE
+// =======================================================================
+
+#define DM_BLITFUNC_NAME dmUnscaledBlitSurface8to8
+#define DM_BLITFUNC_SRC_BYTES 1
+#define DM_BLITFUNC_DST_BYTES 1
+#define DM_BLITFUNC_SRC_TYPE Uint8
+#define DM_BLITFUNC_DST_TYPE Uint8
+
+#define DM_BLITFUNC_INNER *dp++ = sp[xv];
+
+#include "dmunscaledblit.h"
+
+// -----------------------------------------------------------------------
+
+#define DM_BLITFUNC_NAME dmUnscaledBlitSurface8to32
+#define DM_BLITFUNC_SRC_BYTES 1
+#define DM_BLITFUNC_DST_BYTES 4
+#define DM_BLITFUNC_SRC_TYPE Uint8
+#define DM_BLITFUNC_DST_TYPE Uint32
+
+#define DM_BLITFUNC_VARS const Uint32 *pal;
+#define DM_BLITFUNC_INIT \
+    if (src->format->palette == NULL || src->format->palette->ncolors < 256) return -2;	\
+    pal = (Uint32 *) src->format->palette->colors;
+
+#define DM_BLITFUNC_INNER *dp++ = pal[sp[xv]];
+
+#include "dmunscaledblit.h"
+
+// -----------------------------------------------------------------------
+
+#define DM_BLITFUNC_NAME dmUnscaledBlitSurface32to32
+#define DM_BLITFUNC_SRC_BYTES 4
+#define DM_BLITFUNC_DST_BYTES 4
+#define DM_BLITFUNC_SRC_TYPE Uint32
+#define DM_BLITFUNC_DST_TYPE Uint32
+#define DM_BLITFUNC_INIT
+#define DM_BLITFUNC_INNER *dp++ = sp[xv];
+
+#include "dmunscaledblit.h"
+
+
+// =======================================================================
+// DMD_TRANSPARENT
+// =======================================================================
+
+#define DM_BLITFUNC_NAME dmUnscaledBlitSurface8to8Transparent
+#define DM_BLITFUNC_SRC_BYTES 1
+#define DM_BLITFUNC_DST_BYTES 1
+#define DM_BLITFUNC_SRC_TYPE Uint8
+#define DM_BLITFUNC_DST_TYPE Uint8
+
+#define DM_BLITFUNC_INNER \
+    *dp = sp[xv] ? sp[xv] : *dp; dp++;
+
+#include "dmunscaledblit.h"
+
+// -----------------------------------------------------------------------
+
+#define DM_BLITFUNC_NAME dmUnscaledBlitSurface8to32Transparent
+#define DM_BLITFUNC_SRC_BYTES 1
+#define DM_BLITFUNC_DST_BYTES 4
+#define DM_BLITFUNC_SRC_TYPE Uint8
+#define DM_BLITFUNC_DST_TYPE DMRGBA32
+#define DM_BLITFUNC_VARS const DMRGBA32 *pal;
+#define DM_BLITFUNC_INIT \
+    if (src->format->palette == NULL || src->format->palette->ncolors < 256) return -2;	\
+    pal = (DMRGBA32 *) src->format->palette->colors;
+
+#define DM_BLITFUNC_INNER \
+    const DMRGBA32 q = pal[sp[xv]]; \
+    dp->r += ((q.r - dp->r) * q.a) >> 8; \
+    dp->g += ((q.g - dp->g) * q.a) >> 8; \
+    dp->b += ((q.b - dp->b) * q.a) >> 8; \
+    dp->a  = q.a; \
+    dp++;
+
+#include "dmunscaledblit.h"
+
+
+// -----------------------------------------------------------------------
+
+#define DM_BLITFUNC_NAME dmUnscaledBlitSurface32to32Transparent
+#define DM_BLITFUNC_SRC_BYTES 4
+#define DM_BLITFUNC_DST_BYTES 4
+
+#ifdef DM_USE_SIMD
+#define DM_BLITFUNC_VARS \
+    const Uint32 qpdmask =       0xff000000; \
+    const Uint64 qpdrm   = 0xff00ff00ff00ffULL;
+    
+#define DM_BLITFUNC_SRC_TYPE Uint32
+#define DM_BLITFUNC_DST_TYPE Uint32
+#define DM_BLITFUNC_INNER				\
+    asm(					\
+        "movd        %2,     %%mm1\n"		\
+        \
+        "movd        %3,     %%mm2\n"		\
+        "movq        %%mm1,  %%mm5\n"		\
+        "pand        %%mm2,  %%mm5\n"		\
+        "psrlw       $8,     %%mm5\n"		\
+        "punpcklwd   %%mm5,  %%mm5\n"		\
+        "punpckhwd   %%mm5,  %%mm5\n"		\
+        \
+        "pxor        %%mm2,  %%mm2\n"		\
+        "movd        %1,     %%mm3\n"		\
+        "punpcklbw   %%mm2,  %%mm1\n"		\
+        "punpcklbw   %%mm2,  %%mm3\n"		\
+        \
+        "psubw       %%mm3,  %%mm1\n"		\
+        "pmullw      %%mm5,  %%mm1\n"		\
+        "psraw       $8,     %%mm1\n"		\
+        "paddw       %%mm3,  %%mm1\n"		\
+        "pand        %4,     %%mm1\n"		\
+        "packuswb    %%mm2,  %%mm1\n"		\
+        "movd        %%mm1,  %0\n"		\
+        : "=m" (*dp)				\
+        : "m" (*dp), "m" (sp[xv]), "m" (qpdmask), "m" (qpdrm)	\
+        : "memory", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5" ); dp++;
+
+#define DM_BLITFUNC_FINISH asm("emms\n");
+
+#else
+
+#define DM_BLITFUNC_SRC_TYPE DMRGBA32
+#define DM_BLITFUNC_DST_TYPE DMRGBA32
+#define DM_BLITFUNC_INNER \
+    const DMRGBA32 q = sp[xv]; \
+    dp->r += ((q.r - dp->r) * q.a) >> 8; \
+    dp->g += ((q.g - dp->g) * q.a) >> 8; \
+    dp->b += ((q.b - dp->b) * q.a) >> 8; \
+    dp->a  = q.a; \
+    dp++;
+#endif
+
+#include "dmunscaledblit.h"
+
+
+// =======================================================================
+// DMD_SATURATE
+// =======================================================================
+
+#define DM_BLITFUNC_NAME dmUnscaledBlitSurface8to8Saturate
+#define DM_BLITFUNC_SRC_BYTES 1
+#define DM_BLITFUNC_DST_BYTES 1
+#define DM_BLITFUNC_SRC_TYPE Uint8
+#define DM_BLITFUNC_DST_TYPE Uint8
+#define DM_BLITFUNC_INNER \
+    const int q = sp[xv] + *dp; \
+    *(dp++) = q < 256 ? q : 255;
+
+#include "dmunscaledblit.h"
+
+// -----------------------------------------------------------------------
+
+#define DM_BLITFUNC_NAME dmUnscaledBlitSurface8to32Saturate
+#define DM_BLITFUNC_SRC_BYTES 1
+#define DM_BLITFUNC_DST_BYTES 4
+#define DM_BLITFUNC_SRC_TYPE Uint8
+#define DM_BLITFUNC_DST_TYPE DMRGBA32
+#define DM_BLITFUNC_VARS const DMRGBA32 *pal;
+#define DM_BLITFUNC_INIT \
+    if (src->format->palette == NULL || src->format->palette->ncolors < 256) return -2;	\
+    pal = (DMRGBA32 *) src->format->palette->colors;
+
+#ifdef DM_USE_SIMD
+#define DM_BLITFUNC_INNER				\
+    asm("movd        %2,     %%mm1\n"		\
+        "movd        %1,     %%mm2\n"		\
+        "paddusb     %%mm2,  %%mm1\n"		\
+        "movd        %%mm1,  %0\n"		\
+        : "=m" (*dp)				\
+        : "m" (*dp), "m" (pal[sp[xv]])	\
+        : "memory", "%mm1", "%mm2" ); dp++;
+
+#define DM_BLITFUNC_FINISH asm("emms\n");
+#else
+#define DM_BLITFUNC_INNER \
+    const DMRGBA32 q = pal[sp[xv]]; \
+    const int qr = dp->r + q.r, qg = dp->g + q.g, qb = dp->b + q.b; \
+    dp->r = qr < 256 ? qr : 255; \
+    dp->g = qg < 256 ? qg : 255; \
+    dp->b = qb < 256 ? qb : 255; \
+    dp->a = q.a; \
+    dp++;
+#endif
+
+#include "dmunscaledblit.h"
+
+// -----------------------------------------------------------------------
+
+#define DM_BLITFUNC_NAME dmUnscaledBlitSurface32to32Saturate
+#define DM_BLITFUNC_SRC_BYTES 4
+#define DM_BLITFUNC_DST_BYTES 4
+#define DM_BLITFUNC_INIT
+
+#ifdef DM_USE_SIMD
+#define DM_BLITFUNC_SRC_TYPE Uint32
+#define DM_BLITFUNC_DST_TYPE Uint32
+#define DM_BLITFUNC_INNER				\
+    asm("movd        %2,     %%mm1\n"		\
+        "movd        %1,     %%mm2\n"		\
+        "paddusb     %%mm2,  %%mm1\n"		\
+        "movd        %%mm1,  %0\n"		\
+        : "=m" (*dp)				\
+        : "m" (*dp), "m" (sp[xv])	\
+        : "memory", "%mm1", "%mm2" ); dp++;
+
+#define DM_BLITFUNC_FINISH asm("emms\n");
+#else
+#define DM_BLITFUNC_SRC_TYPE DMRGBA32
+#define DM_BLITFUNC_DST_TYPE DMRGBA32
+#define DM_BLITFUNC_INNER \
+    const DMRGBA32 q = sp[xv]; \
+    const int qr = dp->r + q.r, qg = dp->g + q.g, qb = dp->b + q.b; \
+    dp->r = qr < 256 ? qr : 255; \
+    dp->g = qg < 256 ? qg : 255; \
+    dp->b = qb < 256 ? qb : 255; \
+    dp->a = q.a; \
+    dp++;
+#endif
+
+#include "dmunscaledblit.h"
+
+
+// =======================================================================