From 395b742050af148137758f1d4492a3922569fb91 Mon Sep 17 00:00:00 2001
From: Nick Bowler <nbowler@draconx.ca>
Date: Wed, 22 Sep 2021 01:24:25 -0400
Subject: [PATCH] lbximg: Improve performance of PAM output.

The PAM output currently calls fwrite for every pixel, which amounts to
about half a million fwrite calls to output a full screen image.  For a
fullscreen animation like logo.lbx.001, the majority of the execution
time is spent in fwrite.

Let's reduce this a lot by formatting and outputting an entire row at
a time, which almost completely eliminates fwrite from the profile.
---
 src/pnm.c | 87 ++++++++++++++++++++++++++++---------------------------
 1 file changed, 45 insertions(+), 42 deletions(-)

diff --git a/src/pnm.c b/src/pnm.c
index d70d8b5..1596c5a 100644
--- a/src/pnm.c
+++ b/src/pnm.c
@@ -322,58 +322,61 @@ int img_output_pam(FILE *f, const char *filename,
                    struct lbx_colour *palette)
 {
 	bool masked = img_is_masked(mask, width, height);
-	unsigned x, y, depth;
+	unsigned x, y, depth, maxval = palette ? 63 : 255;
+	unsigned char *rowbuf;
+	int rc;
 
-	if (fprintf_ascii(f, "P7\nWIDTH %u\nHEIGHT %u\n", width, height) < 0)
-		goto err;
+	depth = (palette ? 3 : 1) + masked;
+	if (width > (size_t)-1 / depth || !(rowbuf = malloc(width * depth))) {
+		tool_err(-1, "failed to allocate memory");
+		return -1;
+	}
 
-	if (palette) {
-		depth = masked ? 4 : 3;
+	rc = fprintf_ascii(f, "P7\nWIDTH %u\nHEIGHT %u\nDEPTH %u\n",
+	                      width, height, depth);
+	if (rc < 0)
+		goto write_err;
 
-		if (fprintf_ascii(f, "DEPTH %u\nMAXVAL 63\n"
-		                     "TUPLTYPE RGB%s\nENDHDR\n",
-		                     depth, masked ? "_ALPHA" : "") < 0)
-			goto err;
+	rc = fprintf_ascii(f, "MAXVAL %u\nTUPLTYPE %s",
+	                      maxval, palette ? "RGB" : "GRAYSCALE");
+	if (rc < 0)
+		goto write_err;
 
-		for (x = y = 0; y < height; ++x < width || (x = 0, y++)) {
-			unsigned long offset = (unsigned long) y * width + x;
-			struct lbx_colour *c = &palette[pixels[offset]];
-			unsigned char buf[4];
-			bool vis;
-
-			vis = mask[offset/CHAR_BIT] & (1u << offset%CHAR_BIT);
-			buf[0] = c->red;
-			buf[1] = c->green;
-			buf[2] = c->blue;
-			buf[3] = vis ? 63 : 0;
-
-			if (fwrite(buf, 1, depth, f) < depth)
-				goto err;
-		}
-	} else {
-		depth = masked ? 2 : 1;
+	rc = fprintf_ascii(f, "_ALPHA\nENDHDR\n" + 6*!masked);
+	if (rc < 0)
+		goto write_err;
 
-		if (fprintf_ascii(f, "DEPTH %u\nMAXVAL 255\n"
-		                     "TUPLTYPE GRAYSCALE%s\nENDHDR\n",
-		                     depth, masked ? "_ALPHA" : "") < 0)
-			goto err;
-
-		for (x = y = 0; y < height; ++x < width || (x = 0, y++)) {
+	for (y = 0; y < height; y++) {
+		for (x = 0; x < width; x++) {
 			unsigned long offset = (unsigned long) y * width + x;
-			unsigned char buf[2];
-			bool vis;
-
-			vis = mask[offset/CHAR_BIT] & (1u << offset%CHAR_BIT);
-			buf[0] = pixels[offset];
-			buf[1] = vis ? 255 : 0;
-
-			if (fwrite(buf, 1, depth, f) < depth)
-				goto err;
+			size_t p = (size_t)depth * x; // pixel start
+
+			if (palette) {
+				struct lbx_colour *c = &palette[pixels[offset]];
+				rowbuf[p+0] = c->red;
+				rowbuf[p+1] = c->green;
+				rowbuf[p+2] = c->blue;
+			} else {
+				rowbuf[p+0] = pixels[offset];
+			}
+
+			if (masked) {
+				bool vis = 1 & ( mask[offset/CHAR_BIT]
+				                 >> (offset%CHAR_BIT) );
+
+				rowbuf[p+depth-1] = vis*maxval;
+			}
 		}
+
+		if (fwrite(rowbuf, depth, width, f) < width)
+			goto write_err;
 	}
 
+	free(rowbuf);
 	return 0;
-err:
+write_err:
 	tool_err(0, "error writing %s", filename);
+
+	free(rowbuf);
 	return -1;
 }
-- 
2.43.2