dla2csv.xex: optimizations WIP.

3 files changed, 135 insertions, 44 deletions

diff --git a/dla.atr b/dla.atr
index 247fe9d..1af905c 100644
--- a/dla.atr
+++ b/dla.atr
diff --git a/dla2csv.c b/dla2csv.c
index a7a4e6e..d8341a8 100644
--- a/dla2csv.c
+++ b/dla2csv.c
@@ -20,6 +20,7 @@
 
 #define HEIGHT 170
 #define WIDTH 176
+#define BYTEWIDTH 22
 
 /* cc65 doesn't fold constants, it won't let us do this:
 	#define INBUF_SIZE (WIDTH * HEIGHT)
@@ -35,7 +36,7 @@ void print_id(void) {
 #ifdef __ATARI__
 
 /* Uncomment this to see elapsed time in convert() */
-	// #define PROFILE
+	#define PROFILE
 
 /* cc65 doesn't "localize" \b to the Atari backspace character, so: */
 	#define BS CH_DEL
@@ -214,6 +215,12 @@ char *eoltypes[][2] = {
 	{ NULL, NULL }
 };
 
+char *eol; /* gets assigned one of the eoltypes */
+
+/* using a table of masks instead of calculating them saves maybe 2s
+	on the Atari. */
+unsigned char masks[] = { 0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01 };
+
 /* Read a string from stdin (E: on the Atari).
 	Exit if we get EOF on stdin (e.g. ^D on Linux, ^3 on Atari).
 	See exit_eof_stdin() comments above.
@@ -320,11 +327,16 @@ char *prompt_eol(void) {
 	return eoltypes[i][1];
 }
 
-void backspace3(void) {
-	putchar(BS);
+#ifdef __ATARI__
+void backspace2(void) {
+	OS.colcrs -= 2;
+}
+#else
+void backspace2(void) {
 	putchar(BS);
 	putchar(BS);
 }
+#endif
 
 /* check_dla() returns 1 if all is well, 0 if there's a problem. */
 int check_dla(int bytes) {
@@ -351,7 +363,6 @@ int check_dla(int bytes) {
 		}
 	}
 
-
 	return ok;
 }
 
@@ -375,39 +386,114 @@ int read_file(void) {
 	return bytes;
 }
 
+#if 0
 /* convert() returns the number of particles on success, 0 on failure. */
 int convert(char *eol) {
+	div_t d;
+	unsigned int particles;
 #ifdef __ATARI__
 	register
 #endif
-	unsigned char x, y;
+	unsigned int i;
+	unsigned char j, x, pct = 0;
+
+	outf = prompt_filename("Output CSV", "wb");
+#ifdef __ATARI__
+	printf("Press Ctrl-C to abort conversion.\n");
+#endif
+	printf("\nConverting...   ");
+	fprintf(outf, "x,y%s", eol);
+
+	for(i = 0; i < INBUF_SIZE; ++i) {
+		if(i % 38 == 0) {
+			backspace3();
+			printf("%02d%%", pct++);
+			fflush(stdout);
+		}
+		if(inbuf[i]) {
+			d = div(i, 22);
+			x = d.rem * 8;
+			for(j = 0; j < 8; ++x, ++j) {
+				if(inbuf[i] & masks[j]) {
+					if(fprintf(outf, "%d,%d%s", x, d.quot, eol) < 0) {
+						putchar('\n');
+						PERROR(stringbuf);
+						fclose(outf);
+						remove(stringbuf);
+						return 0;
+					}
+					particles++;
+				}
+			}
+		}
+	}
+
+	fclose(outf);
+
+
+	return particles;
+}
+#endif
+
+#ifdef __ATARI__
+/* using utoa() rather than fprintf() saves ~12 sec on the Atari. */
+void writepoint(unsigned char x, unsigned char y) {
+	static char ubuf[10];
+	static char fbuf[16];
+
+	utoa(x, fbuf, 10);
+	strcat(fbuf, ",");
+	strcat(fbuf, utoa(y, ubuf, 10));
+	strcat(fbuf, eol);
+	fputs(fbuf, outf);
+	/*
+	utoa(x, buf, 10);
+	fputs(buf, outf);
+	fputc(',', outf);
+	utoa(y, buf, 10);
+	fputs(buf, outf);
+	fputs(eol, outf);
+	*/
+}
+#else
+/* since utoa() only exists on cc65, can't use it on non-Atari. */
+void writepoint(unsigned char x, unsigned char y) {
+	fprintf(outf, "%d,%d%s", x, y, eol);
+}
+#endif
+
+/* currently, on the 800 with atariserver, convert() takes:
+	5s for a DLA with no points (a file of zeroes).
+	16s for a DLA with 1000 points.
+	9s for the same DLA, with the fputs() in writepoint() commented out.
+*/
+int convert(void) {
+	unsigned char x, y, j, pixx;
 #ifdef __ATARI__
 	register
 #endif
 	char *inp = inbuf;
-	unsigned char xmask = 0x80;
 	int particles = 0;
 
 	outf = prompt_filename("Output CSV", "wb");
 
 #ifdef __ATARI__
 	printf("Press Ctrl-C to abort conversion.\n");
+	OS.crsinh = 1;
+#endif
+	printf("\nConverting...   %%");
+#ifdef __ATARI__
+	putchar(0xa0);
+	backspace2();
 #endif
-	printf("\nConverting...   ");
 
 	/* CSV file header row (column names) */
 	fprintf(outf, "x,y%s", eol);
 
-#ifdef __ATARI__
-	#ifdef PROFILE
-	OS.rtclok[0] = OS.rtclok[1] = OS.rtclok[2] = 0;
-	#endif
-#endif
-
 	/* write output file one line at a time */
 	/* loop over all the pixels, by row, then column. this
 		loop is rather slow. */
-	for(y = 0; y < HEIGHT; y++) {
+	for(y = 0; y < HEIGHT; ++y) {
 #ifdef __ATARI__
 		/* check for ^C */
 		if(OS.ch == (KEY_C | KEY_CTRL)) {
@@ -418,46 +504,37 @@ int convert(char *eol) {
 			return 0;
 		}
 #endif
-		backspace3();
-		printf("%02d%%", y * 100 / HEIGHT); /* percentage */
+		backspace2();
+		printf("%02d", y * 100 / HEIGHT); /* percentage */
 		fflush(stdout);
-		for(x = 0; x < WIDTH; x++) {
-			/* slight optimization, saves ~12 sec (30% speedup) on Atari */
-			if(!*inp) {
-				inp++;
-				x += 7;
-				continue;
-			}
-			if(*inp & xmask) {
-				if(fprintf(outf, "%d,%d%s", x, y, eol) < 0) {
-					putchar('\n');
-					PERROR(stringbuf);
-					fclose(outf);
-					remove(stringbuf);
-					return 0;
+		for(x = 0; x < BYTEWIDTH; ++x) {
+			if(*inp) {
+				pixx = x * 8;
+				for(j = 0; j < 8; ++j) {
+					if(*inp & masks[j]) {
+						writepoint(pixx + j, y);
+						if(ferror(outf)) {
+							PERROR(stringbuf);
+							fclose(outf);
+							remove(stringbuf);
+							return 0;
+						}
+						particles++;
+					}
 				}
-				particles++;
-			}
-			xmask >>= 1;
-			if(!xmask) {
-				xmask = 0x80;
-				inp++;
 			}
+			inp++;
 		}
 	}
 	fclose(outf);
 
 #ifdef __ATARI__
-	#ifdef PROFILE
-	printf("\nElapsed time: %ds\n", ((OS.rtclok[1] << 8) | OS.rtclok[2]) / 60);
-	#endif
+	OS.crsinh = 0;
 #endif
-
 	return particles;
 }
 
 int main(int argc, char **argv) {
-	char *eol;
 	int done, result = 0;
 
 	init_console();
@@ -483,10 +560,18 @@ int main(int argc, char **argv) {
 			without re-reading the input file. */
 		done = 0;
 		while(!done) {
-			result = convert(eol);
-			if(result) {
+
+#ifdef __ATARI__
+	#ifdef PROFILE
+			OS.rtclok[0] = OS.rtclok[1] = OS.rtclok[2] = 0;
+	#endif
+#endif
+
+			result = convert();
+
+			if(result >= 0) {
 				done = 1;
-				backspace3();
+				backspace2();
 				printf("100%%\n%d particles.\n", result);
 			} else {
 				if(!prompt_yn("Conversion failed, try again", 1))
@@ -494,6 +579,12 @@ int main(int argc, char **argv) {
 			}
 		}
 
+#ifdef __ATARI__
+	#ifdef PROFILE
+			printf("\nElapsed time: %ds\n", ((OS.rtclok[1] << 8) | OS.rtclok[2]) / 60);
+	#endif
+#endif
+
 		if(!prompt_yn("\nConvert another file", 1))
 			break;
 	}
diff --git a/dla2csv.xex b/dla2csv.xex
index c9bd19f..41b117c 100644
--- a/dla2csv.xex
+++ b/dla2csv.xex