Accelerate some of the blitting functions using SWAR.

1 files changed, 190 insertions, 20 deletions

diff --git a/draw/archport.c b/draw/archport.c
index 2b45b7c0..93e50fe9 100644
--- a/draw/archport.c
+++ b/draw/archport.c
@@ -1,25 +1,24 @@
 #include "fitz.h"
 
-/* This C implementation was a prototype of the algorithm used
- * in the ARM code in draw/archarm.c. It is conceivable that on some
- * architectures/compilers this may be preferable to the vanilla
- * version below. */
+typedef unsigned char byte;
+
+/* These C implementations use SWAR (SIMD-within-a-register) techniques. */
+
+#define MASK 0xFF00FF00;
 
 static void
-path_w4i1o4_32bit(unsigned char * restrict argb,
-	unsigned char * restrict src, unsigned char cov, int len,
-	unsigned char * restrict dst)
+path_w4i1o4_32bit(byte * restrict argb,
+	byte * restrict src, byte cov, int len,
+	byte * restrict dst)
 {
 	/* COLOR * coverage + DST * (256-coverage) = (COLOR - DST)*coverage + DST*256 */
 	unsigned int *dst32 = (unsigned int *)(void *)dst;
 	int alpha = argb[0];
-	unsigned int rb = argb[1] | argb[3] << 16;
-	unsigned int ag = 255 | argb[2] << 16;
-	const int MASK = 0xFF00FF00;
+	unsigned int rb = argb[1] | (argb[3] << 16);
+	unsigned int ag = 255 | (argb[2] << 16);
 
-	/* sanity test */
-	if (sizeof(int) != 4 || sizeof(unsigned int) != 4)
-		abort();
+	if (alpha == 0)
+		return;
 
 	if (alpha != 255)
 	{
@@ -28,7 +27,7 @@ path_w4i1o4_32bit(unsigned char * restrict argb,
 		{
 			unsigned int ca, drb, dag, crb, cag;
 			cov += *src; *src++ = 0;
-			ca = cov+(cov>>7); /* ca is in 0...256 range */
+			ca = cov + (cov>>7); /* ca is in 0...256 range */
 			ca = (ca*alpha)>>8; /* ca is is in 0...256 range */
 			dag = *dst32++;
 			if (ca != 0)
@@ -39,8 +38,8 @@ path_w4i1o4_32bit(unsigned char * restrict argb,
 				cag = ag - (dag>>8);
 				drb += crb * ca;
 				dag += cag * ca;
-				drb = drb & MASK;
-				dag = dag & MASK;
+				drb &= MASK;
+				dag &= MASK;
 				dag = drb | (dag>>8);
 				dst32[-1] = dag;
 			}
@@ -52,7 +51,7 @@ path_w4i1o4_32bit(unsigned char * restrict argb,
 		{
 			unsigned int ca, drb, dag, crb, cag;
 			cov += *src; *src++ = 0;
-			ca = cov+(cov>>7); /* ca is in 0...256 range */
+			ca = cov + (cov>>7); /* ca is in 0...256 range */
 			dag = *dst32++;
 			if (ca == 0)
 				continue;
@@ -68,8 +67,8 @@ path_w4i1o4_32bit(unsigned char * restrict argb,
 				cag = ag - (dag>>8);
 				drb += crb * ca;
 				dag += cag * ca;
-				drb = drb & MASK;
-				dag = dag & MASK;
+				drb &= MASK;
+				dag &= MASK;
 				dag = drb | (dag>>8);
 			}
 			dst32[-1] = dag;
@@ -77,11 +76,182 @@ path_w4i1o4_32bit(unsigned char * restrict argb,
 	}
 }
 
+static void
+duff_4o4_32bit(byte * restrict sp, int sw, byte * restrict dp, int dw, int w0, int h)
+{
+	unsigned int *sp32 = (unsigned int *)(void *)sp;
+	unsigned int *dp32 = (unsigned int *)(void *)dp;
+
+	/* duff_non(sp0, sw, 4, dp0, dw, w0, h); */
+
+	sw = (sw>>2)-w0;
+	dw = (dw>>2)-w0;
+	while (h--)
+	{
+		int w = w0;
+		while (w--)
+		{
+			unsigned int sag = *sp32++;
+			unsigned int dag = *dp32++;
+			unsigned int srb, drb;
+			int alpha = sag & 255;
+			if (alpha == 0)
+				continue;
+			alpha += alpha>>7;
+			sag |= 0xFF;
+			drb = dag & MASK;
+			dag = (dag<<8) & MASK;
+			srb = (sag>>8) & ~MASK;
+			sag = sag & ~MASK;
+			srb -= (drb>>8);
+			sag -= (dag>>8);
+			drb += srb * alpha;
+			dag += sag * alpha;
+			drb &= MASK;
+			dag &= MASK;
+			dag = drb | (dag>>8);
+			dp32[-1] = dag;
+		}
+		sp32 += sw;
+		dp32 += dw;
+	}
+}
+
+static void
+duff_4i1o4_32bit(byte * restrict sp, int sw,
+	byte * restrict mp, int mw,
+	byte * restrict dp, int dw, int w0, int h)
+{
+	unsigned int *sp32 = (unsigned int *)(void *)sp;
+	unsigned int *dp32 = (unsigned int *)(void *)dp;
+
+	/* duff_nimon(sp, sw, 4, mp, mw, 1, dp, dw, w0, h); */
+
+	sw = (sw>>2)-w0;
+	dw = (dw>>2)-w0;
+	mw -= w0;
+	while (h--)
+	{
+		int w = w0;
+		while (w--)
+		{
+			unsigned int sag = *sp32++;
+			unsigned int dag = *dp32++;
+			unsigned int srb, drb, alpha, ma;
+			alpha = sag & 255;
+			ma = *mp++;
+			if (alpha == 0)
+				continue;
+			ma += ma>>7;
+			if (ma == 0)
+				continue;
+			alpha += alpha>>7;
+			alpha = (alpha*ma)>>8;
+			sag |= 0xFF;
+			drb = dag & MASK;
+			dag = (dag<<8) & MASK;
+			srb = (sag>>8) & ~MASK;
+			sag = sag & ~MASK;
+			srb -= (drb>>8);
+			sag -= (dag>>8);
+			drb += srb * alpha;
+			dag += sag * alpha;
+			drb &= MASK;
+			dag &= MASK;
+			dag = drb | (dag>>8);
+			dp32[-1] = dag;
+		}
+		sp32 += sw;
+		mp += mw;
+		dp32 += dw;
+	}
+}
+
+static void
+text_w4i1o4_32bit(byte * restrict argb,
+	byte * restrict src, int srcw,
+	byte * restrict dst, int dstw, int w0, int h)
+{
+	unsigned int *dst32 = (unsigned int *)(void *)dst;
+	unsigned int alpha = argb[0];
+	unsigned int rb = argb[1] | (argb[3] << 16);
+	unsigned int ag = 255 | (argb[2] << 16);
+
+	if (alpha == 0)
+		return;
+
+	srcw -= w0;
+	dstw = (dstw>>2)-w0;
+
+	if (alpha != 255)
+	{
+		alpha += alpha>>7;
+		while (h--)
+		{
+			int w = w0;
+			while (w--)
+			{
+				unsigned int ca, drb, dag, crb, cag;
+				ca = *src++;
+				dag = *dst32++;
+				ca += ca>>7;
+				ca = (ca*alpha)>>8;
+				if (ca == 0)
+				 continue;
+				drb = dag & MASK;
+				dag = (dag<<8) & MASK;
+				crb = rb - (drb>>8);
+				cag = ag - (dag>>8);
+				drb += crb * ca;
+				dag += cag * ca;
+				drb &= MASK;
+				dag &= MASK;
+				dag = drb | (dag>>8);
+				dst32[-1] = dag;
+			}
+			src += srcw;
+			dst32 += dstw;
+		}
+	}
+	else
+	{
+		alpha += alpha>>7;
+		while (h--)
+		{
+			int w = w0;
+			while (w--)
+			{
+				unsigned int ca, drb, dag, crb, cag;
+				ca = *src++;
+				dag = *dst32++;
+				ca += ca>>7;
+				if (ca == 0)
+				 continue;
+				drb = dag & MASK;
+				dag = (dag<<8) & MASK;
+				crb = rb - (drb>>8);
+				cag = ag - (dag>>8);
+				drb += crb * ca;
+				dag += cag * ca;
+				drb &= MASK;
+				dag &= MASK;
+				dag = drb | (dag>>8);
+				dst32[-1] = dag;
+			}
+			src += srcw;
+			dst32 += dstw;
+		}
+	}
+}
+
 void fz_accelerate(void)
 {
-	if (sizeof(int) == 4)
+	if (sizeof(int) == 4 && sizeof(unsigned int) == 4)
 	{
+		fz_duff_4o4 = duff_4o4_32bit;
+		fz_duff_4i1o4 = duff_4i1o4_32bit;
 		fz_path_w4i1o4 = path_w4i1o4_32bit;
+		fz_text_w4i1o4 = text_w4i1o4_32bit;
 	}
 
 	if (sizeof(int) == 8)