Reinsert Robin Watts' specialized pixel unpacking loops.

1 files changed, 131 insertions, 18 deletions

diff --git a/draw/imageunpack.c b/draw/imageunpack.c
index b899a8ad..577eac21 100644
--- a/draw/imageunpack.c
+++ b/draw/imageunpack.c
@@ -10,38 +10,151 @@
 #define get8(buf,x) (buf[x])
 #define get16(buf,x) (buf[x << 1])
 
+static unsigned char get1tab1[256][8];
+static unsigned char get1tab1p[256][16];
+static unsigned char get1tab255[256][8];
+static unsigned char get1tab255p[256][16];
+
+static void
+initget1tables(void)
+{
+	static int once = 0;
+	unsigned char bits[1];
+	int i, k, x;
+
+	/* TODO: mutex lock here */
+
+	if (once)
+		return;
+
+	for (i = 0; i < 256; i++)
+	{
+		bits[0] = i;
+		for (k = 0; k < 8; k++)
+		{
+			x = get1(bits, k);
+
+			get1tab1[i][k] = x;
+			get1tab1p[i][k * 2] = x;
+			get1tab1p[i][k * 2 + 1] = 1;
+
+			get1tab255[i][k] = x * 255;
+			get1tab255p[i][k * 2] = x * 255;
+			get1tab255p[i][k * 2 + 1] = 255;
+		}
+	}
+
+	once = 1;
+}
+
 void
-fz_unpacktile(fz_pixmap *dst, unsigned char * restrict src, int n, int depth, int stride)
+fz_unpacktile(fz_pixmap *dst, unsigned char * restrict src, int n, int depth, int stride, int scale)
 {
 	int pad, x, y, k;
+	int w = dst->w;
 
 	pad = 0;
 	if (dst->n > n)
-		pad = (1 << depth) - 1;
+		pad = ((1 << depth) - 1) * scale;
 
-	/* TODO: reinsert Robin Watts' specialized loops here */
+	initget1tables();
 
 	for (y = 0; y < dst->h; y++)
 	{
 		unsigned char *sp = src + y * stride;
 		unsigned char *dp = dst->samples + y * (dst->w * dst->n);
-		int b = 0;
-		for (x = 0; x < dst->w; x++)
+
+		/* Specialized loops */
+
+		if (n == 1 && depth == 1 && scale == 1 && !pad)
+		{
+			int w3 = w >> 3;
+			for (x = 0; x < w3; x++)
+			{
+				memcpy(dp, get1tab1[*sp++], 8);
+				dp += 8;
+			}
+			x = x << 3;
+			if (x < w)
+				memcpy(dp, get1tab1[*sp], w - x);
+		}
+
+		else if (n == 1 && depth == 1 && scale == 255 && !pad)
+		{
+			int w3 = w >> 3;
+			for (x = 0; x < w3; x++)
+			{
+				memcpy(dp, get1tab255[*sp++], 8);
+				dp += 8;
+			}
+			x = x << 3;
+			if (x < w)
+				memcpy(dp, get1tab255[*sp], w - x);
+		}
+
+		else if (n == 1 && depth == 1 && scale == 1 && pad == 1)
+		{
+			int w3 = w >> 3;
+			for (x = 0; x < w3; x++)
+			{
+				memcpy(dp, get1tab1p[*sp++], 16);
+				dp += 16;
+			}
+			x = x << 3;
+			if (x < w)
+				memcpy(dp, get1tab1p[*sp], (w - x) << 1);
+		}
+
+		else if (n == 1 && depth == 1 && scale == 255 && pad == 255)
+		{
+			int w3 = w >> 3;
+			for (x = 0; x < w3; x++)
+			{
+				memcpy(dp, get1tab255p[*sp++], 16);
+				dp += 16;
+			}
+			x = x << 3;
+			if (x < w)
+				memcpy(dp, get1tab255p[*sp], (w - x) << 1);
+		}
+
+		else if (depth == 8 && !pad)
+		{
+			int len = w * n;
+			while (len--)
+				*dp++ = *sp++;
+		}
+
+		else if (depth == 8 && pad == 255)
+		{
+			for (x = 0; x < w; x++)
+			{
+				for (k = 0; k < n; k++)
+					*dp++ = *sp++;
+				*dp++ = 255;
+			}
+		}
+
+		else
 		{
-			for (k = 0; k < n; k++)
+			int b = 0;
+			for (x = 0; x < w; x++)
 			{
-				switch (depth)
+				for (k = 0; k < n; k++)
 				{
-				case 1: *dp++ = get1(sp, b); break;
-				case 2: *dp++ = get2(sp, b); break;
-				case 4: *dp++ = get4(sp, b); break;
-				case 8: *dp++ = get8(sp, b); break;
-				case 16: *dp++ = get16(sp, b); break;
+					switch (depth)
+					{
+					case 1: *dp++ = get1(sp, b) * scale; break;
+					case 2: *dp++ = get2(sp, b) * scale; break;
+					case 4: *dp++ = get4(sp, b) * scale; break;
+					case 8: *dp++ = get8(sp, b); break;
+					case 16: *dp++ = get16(sp, b); break;
+					}
+					b++;
 				}
-				b++;
+				if (pad)
+					*dp++ = pad;
 			}
-			if (pad)
-				*dp++ = pad;
 		}
 	}
 }
@@ -51,7 +164,7 @@ fz_unpacktile(fz_pixmap *dst, unsigned char * restrict src, int n, int depth, in
  */
 
 void
-fz_decodetile(fz_pixmap *pix, float *decode, int scale)
+fz_decodetile(fz_pixmap *pix, float *decode)
 {
 	int min[FZ_MAXCOLORS + 2];
 	int max[FZ_MAXCOLORS + 2];
@@ -62,7 +175,7 @@ fz_decodetile(fz_pixmap *pix, float *decode, int scale)
 	int needed;
 	int k;
 
-	needed = scale != 1;
+	needed = 0;
 	for (k = 0; k < n; k++)
 	{
 		min[k] = decode[k * 2] * 255;
@@ -77,7 +190,7 @@ fz_decodetile(fz_pixmap *pix, float *decode, int scale)
 	while (len--)
 	{
 		for (k = 0; k < n; k++)
-			p[k] = min[k] + fz_mul255(sub[k], p[k] * scale);
+			p[k] = min[k] + fz_mul255(sub[k], p[k]);
 		p += n;
 	}
 }