Speed tweaks to muraster.

Divide up the work amongst the threads to solve the odd situation
of larger buffers producing slower runtimes.

Avoid creating (and hence clearing) larger bitmaps than required
at the end of the run.

1 files changed, 18 insertions, 1 deletions

diff --git a/source/tools/muraster.c b/source/tools/muraster.c
index ed52b01a..5f56b63e 100644
--- a/source/tools/muraster.c
+++ b/source/tools/muraster.c
@@ -767,6 +767,9 @@ static int dodrawpage(fz_context *ctx, int pagenum, fz_cookie *cookie, render_de
 				w->tbounds = tbounds;
 				memset(&w->cookie, 0, sizeof(fz_cookie));
 				w->list = render->list;
+				if (remaining_height < band_height)
+					ibounds.y1 = ibounds.y0 + remaining_height;
+				remaining_height -= band_height;
 				w->pix = fz_new_pixmap_with_bbox(ctx, colorspace, &ibounds, 0);
 				fz_set_pixmap_resolution(ctx, w->pix, x_resolution, y_resolution);
 				DEBUG_THREADS(("Worker %d, Pre-triggering band %d\n", band, band));
@@ -1112,7 +1115,7 @@ static void
 initialise_banding(fz_context *ctx, render_details *render, int color)
 {
 	size_t min_band_mem;
-	int bpp, w, reps;
+	int bpp, h, w, reps;
 
 	render->colorspace = output_cs;
 	render->format = output_format;
@@ -1156,6 +1159,20 @@ initialise_banding(fz_context *ctx, render_details *render, int color)
 	if (reps < 1)
 		reps = 1;
 
+	/* Adjust reps to even out the work between threads */
+	if (render->num_workers > 0)
+	{
+		int runs, num_bands;
+		h = render->ibounds.y1 - render->ibounds.y0;
+		num_bands = (h + min_band_height - 1) / min_band_height;
+		/* num_bands = number of min_band_height bands */
+		runs = (num_bands + reps-1) / reps;
+		/* runs = number of worker runs of reps min_band_height bands */
+		runs = ((runs + render->num_workers - 1) / render->num_workers) * render->num_workers;
+		/* runs = number of worker runs rounded up to make use of all our threads */
+		reps = (num_bands + runs - 1) / runs;
+	}
+
 	render->band_height_multiple = reps;
 	render->bands_rendered = 0;
 }