Add image subsampling function.

When drawing images, if they are much bigger than we need, quickly
subsample them. Makes images much more cachable, reduces time spent
in expensive smooth scaler.

4 files changed, 163 insertions, 7 deletions

diff --git a/fitz/fitz-internal.h b/fitz/fitz-internal.h
index 528fb50a..2f1d137e 100644
--- a/fitz/fitz-internal.h
+++ b/fitz/fitz-internal.h
@@ -836,6 +836,7 @@ fz_pixmap *fz_alpha_from_gray(fz_context *ctx, fz_pixmap *gray, int luminosity);
 unsigned int fz_pixmap_size(fz_context *ctx, fz_pixmap *pix);
 
 fz_pixmap *fz_scale_pixmap(fz_context *ctx, fz_pixmap *src, float x, float y, float w, float h, fz_bbox *clip);
+void fz_subsample_pixmap(fz_context *ctx, fz_pixmap *tile, int factor);
 
 fz_bbox fz_pixmap_bbox_no_ctx(fz_pixmap *src);
 
diff --git a/fitz/res_colorspace.c b/fitz/res_colorspace.c
index 7b36ac1e..edda776d 100644
--- a/fitz/res_colorspace.c
+++ b/fitz/res_colorspace.c
@@ -316,6 +316,31 @@ static void fast_cmyk_to_gray(fz_pixmap *dst, fz_pixmap *src)
 	}
 }
 
+#if 0
+static void
+fast_cmyk_to_rgb_ARM(unsigned char *dst, unsigned char *src, int n)
+{
+	asm volatile(
+	ENTER_ARM
+	"stmfd	r13!,{r4-r12,r14}				\n"
+	"@ r0 = dst						\n"
+	"@ r1 = src						\n"
+	"@ r2 = weights						\n"
+	"mov	r4, #0			@ r4 = CMYK = 0		\n"
+	"mvn	r5, #0xFF000000		@ r5 = RGB = FFFFFF	\n"
+	"1:							\n"
+	"ldr	r3, [r1], #4		@ r3 = cmyk		\n"
+	"cmp	r3, r4			@ if (cmyk == CMYK)	\n"
+	"beq	match			@   goto match		\n"
+	"cmp	r3, #0			@ if (cmyk = 0000)	\n"
+	"beq	black			@
+
+	"ldmfd	r13!,{r4-r7,r9,PC}	@ pop, return to thumb	\n"
+	ENTER_THUMB
+	);
+}
+#endif
+
 static void fast_cmyk_to_rgb(fz_context *ctx, fz_pixmap *dst, fz_pixmap *src)
 {
 	unsigned char *s = src->samples;
diff --git a/fitz/res_pixmap.c b/fitz/res_pixmap.c
index f7fbf239..a3ae7562 100644
--- a/fitz/res_pixmap.c
+++ b/fitz/res_pixmap.c
@@ -687,3 +687,123 @@ fz_drop_image(fz_context *ctx, fz_image *image)
 {
 	fz_drop_storable(ctx, &image->storable);
 }
+
+void
+fz_subsample_pixmap(fz_context *ctx, fz_pixmap *tile, int factor)
+{
+	int dst_w, dst_h, w, h, fwd, fwd2, fwd3, back, back2, x, y, n, xx, yy, nn, f;
+	unsigned char *s, *d;
+
+	if (!tile)
+		return;
+	s = d = tile->samples;
+	f = 1<<factor;
+	w = tile->w;
+	h = tile->h;
+	n = tile->n;
+	dst_w = (w + f-1)>>factor;
+	dst_h = (h + f-1)>>factor;
+	fwd = w*n;
+	back = f*fwd-n;
+	back2 = f*n-1;
+	fwd2 = (f-1)*n;
+	fwd3 = (f-1)*fwd;
+	factor *= 2;
+	for (y = h - f; y >= 0; y -= f)
+	{
+		for (x = w - f; x >= 0; x -= f)
+		{
+			for (nn = n; nn > 0; nn--)
+			{
+				int v = 0;
+				for (xx = f; xx > 0; xx--)
+				{
+					for (yy = f; yy > 0; yy--)
+					{
+						v += *s;
+						s += fwd;
+					}
+					s -= back;
+				}
+				*d++ = v >> factor;
+				s -= back2;
+			}
+			s += fwd2;
+		}
+		/* Do any strays */
+		x += f;
+		if (x > 0)
+		{
+			int div = x * f;
+			int fwd4 = (x-1) * n;
+			int back4 = x*n-1;
+			for (nn = n; nn > 0; nn--)
+			{
+				int v = 0;
+				for (;x > 0; x--)
+				{
+					for (yy = f; yy > 0; yy--)
+					{
+						v += *s;
+						s += fwd;
+					}
+					s -= back;
+				}
+				*d++ = v / div;
+				s -= back4;
+			}
+			s += fwd4;
+		}
+		s += fwd3;
+	}
+	/* Do any stray line */
+	y += f;
+	if (y > 0)
+	{
+		int div = y * f;
+		back = fwd * y - 1;
+		for (x = w - f; x >= 0; x -= f)
+		{
+			for (nn = n; nn > 0; nn--)
+			{
+				int v = 0;
+				for (xx = f; xx > 0; xx--)
+				{
+					for (yy = y; yy > 0; yy--)
+					{
+						v += *s;
+						s += fwd;
+					}
+					s -= back;
+				}
+				*d++ = v / div;
+				s -= back2;
+			}
+			s += fwd2;
+		}
+		/* Do any stray at the end of the stray line */
+		x += f;
+		if (x > 0)
+		{
+			int div = x * y;
+			for (nn = n; nn > 0; nn--)
+			{
+				int v = 0;
+				for (;x > 0; x--)
+				{
+					for (; y > 0; y--)
+					{
+						v += *s;
+						s += fwd;
+					}
+					s -= back;
+				}
+				*d++ = v / div;
+				s -= back2;
+			}
+		}
+	}
+	tile->w = dst_w;
+	tile->h = dst_h;
+	tile->samples = fz_resize_array(ctx, tile->samples, dst_w * n, dst_h);
+}
diff --git a/pdf/pdf_image.c b/pdf/pdf_image.c
index 27a13100..ead7db49 100644
--- a/pdf/pdf_image.c
+++ b/pdf/pdf_image.c
@@ -99,15 +99,15 @@ static fz_store_type pdf_image_store_type =
 };
 
 static fz_pixmap *
-decomp_image_from_stream(fz_context *ctx, fz_stream *stm, pdf_image *image, int in_line, int indexed, int l2factor, int cache)
+decomp_image_from_stream(fz_context *ctx, fz_stream *stm, pdf_image *image, int in_line, int indexed, int l2factor, int native_l2factor, int cache)
 {
 	fz_pixmap *tile = NULL;
 	fz_pixmap *existing_tile;
 	int stride, len, i;
 	unsigned char *samples = NULL;
-	int f = 1<<l2factor;
-	int w = (image->base.w + f-1) >> l2factor;
-	int h = (image->base.h + f-1) >> l2factor;
+	int f = 1<<native_l2factor;
+	int w = (image->base.w + f-1) >> native_l2factor;
+	int h = (image->base.h + f-1) >> native_l2factor;
 	pdf_image_key *key;
 
 	fz_var(tile);
@@ -195,6 +195,14 @@ decomp_image_from_stream(fz_context *ctx, fz_stream *stm, pdf_image *image, int
 		fz_rethrow(ctx);
 	}
 
+	/* Now apply any extra subsampling required */
+	if (l2factor - native_l2factor > 0)
+	{
+		if (l2factor - native_l2factor > 8)
+			l2factor = native_l2factor + 8;
+		fz_subsample_pixmap(ctx, tile, l2factor - native_l2factor);
+	}
+
 	if (!cache)
 		return tile;
 
@@ -249,6 +257,7 @@ pdf_image_get_pixmap(fz_context *ctx, fz_image *image_, int w, int h)
 	fz_stream *stm;
 	int l2factor;
 	pdf_image_key key;
+	int native_l2factor;
 
 	/* Check for 'simple' images which are just pixmaps */
 	if (image->buffer == NULL)
@@ -285,9 +294,10 @@ pdf_image_get_pixmap(fz_context *ctx, fz_image *image_, int w, int h)
 	while (key.l2factor >= 0);
 
 	/* We need to make a new one. */
-	stm = fz_open_image_decomp_stream(ctx, image->buffer, &l2factor);
+	native_l2factor = l2factor;
+	stm = fz_open_image_decomp_stream(ctx, image->buffer, &native_l2factor);
 
-	return decomp_image_from_stream(ctx, stm, image, 0, 0, l2factor, 1);
+	return decomp_image_from_stream(ctx, stm, image, 0, 0, l2factor, native_l2factor, 1);
 }
 
 static pdf_image *
@@ -450,7 +460,7 @@ pdf_load_image_imp(pdf_document *xref, pdf_obj *rdb, pdf_obj *dict, fz_stream *c
 			stm = pdf_open_stream(xref, pdf_to_num(dict), pdf_to_gen(dict));
 		}
 
-		image->tile = decomp_image_from_stream(ctx, stm, image, cstm != NULL, indexed, 0, 0);
+		image->tile = decomp_image_from_stream(ctx, stm, image, cstm != NULL, indexed, 0, 0, 0);
 	}
 	fz_catch(ctx)
 	{