From c220edbe4acfffc6feb29eea8b977506edac6d99 Mon Sep 17 00:00:00 2001
From: John Tsiombikas <nuclear@member.fsf.org>
Date: Thu, 22 Apr 2021 10:26:31 +0300
Subject: [PATCH] wonky sweep

---
 Makefile       |    4 +-
 src/gamescr.c  |   54 +++++++++------------
 tools/tungen.c |  142 +++++++++++++++++++++++++++++++++++---------------------
 3 files changed, 115 insertions(+), 85 deletions(-)

diff --git a/Makefile b/Makefile
index 26ffc65..68b1f63 100644
--- a/Makefile
+++ b/Makefile
@@ -55,7 +55,7 @@ tools/lutgen: tools/lutgen.c
 	cc -o $@ $< -lm
 
 tools/tungen: tools/tungen.c
-	cc -o $@ $< -lm
+	cc -o $@ -O3 -fopenmp $< -lm -lgomp -lpthread
 
 tools/mmutil/mmutil:
 	$(MAKE) -C tools/mmutil
@@ -73,7 +73,7 @@ data/lut.s: tools/lutgen
 	tools/lutgen >$@
 
 data/tun.map: tools/tungen
-	tools/tungen -s 160x160 >$@
+	tools/tungen -s 240x160 -y -c 400 -n 32 >$@
 
 data/snd.bin: $(audata) tools/mmutil/mmutil
 	tools/mmutil/mmutil -o$@ -hdata/snd.h $(audata)
diff --git a/src/gamescr.c b/src/gamescr.c
index 8d5f344..aef4966 100644
--- a/src/gamescr.c
+++ b/src/gamescr.c
@@ -78,15 +78,6 @@ void gamescr(void)
 		vblperf_end();
 		wait_vblank();
 		present(backbuf);
-
-		tunrot = nframes;
-		bgmat[0] = (COS(tunrot) * 140) >> 8;
-		bgmat[1] = (-SIN(tunrot) * 140) >> 8;
-		bgmat[2] = (SIN(tunrot) * 140) >> 8;
-		bgmat[3] = (COS(tunrot) * 140) >> 8;
-		bgx = (120 << 8) - bgmat[0] * 120 + bgmat[1] * -80;
-		bgy = (80 << 8) - bgmat[2] * 120 + bgmat[3] * -80;
-
 		vblperf_begin();
 	}
 }
@@ -94,29 +85,36 @@ void gamescr(void)
 __attribute__((noinline, target("arm"), section(".iwram")))
 static void draw_tunnel(void)
 {
-	int i, j, tx, ty, angle, depth, zoffs, uoffs;
+	int i, j, tx, ty, angle, depth, zoffs, uoffs, flip;
+	int tunsweep;
 	uint16_t pptop, ppbot;
 	uint16_t *top, *bot;
 	uint32_t tun, *tunptr;
 
+	tunsweep = (SIN(nframes) >> 2) + 64;
+	if(tunsweep > 127) tunsweep = 127;
+
+	flip = tunsweep >= 64;
+
+	tunsweep = (tunsweep & 0x3f);
+	if(tunsweep >= 32) tunsweep = 63 - tunsweep;
+
 	zoffs = nframes;
 
 	uoffs = tunrot >> 1;
 
-	top = vram[backbuf] + 20;
-	bot = vram[backbuf] + 159 * 240 / 2 + 20;
-	tunptr = tunmap;
+	top = vram[backbuf];
+	bot = vram[backbuf] + 159 * 240 / 2;
+	tunptr = tunmap + tunsweep * 9600;
+	if(flip) tunptr += 240/2;
 	for(i=0; i<80; i++) {
-		for(j=0; j<160/2; j++) {
-#ifdef VBLBAR
-			if(j == 160/4) {
-				tunptr++;
-				top++;
-				bot++;
-				continue;
+		for(j=0; j<240/2; j++) {
+			if(flip) {
+				tun = *--tunptr;
+				tun = (tun >> 16) | (tun << 16);
+			} else {
+				tun = *tunptr++;
 			}
-#endif
-			tun = *tunptr++;
 
 			angle = tun & 0xff;
 			depth = (tun >> 8) & 0xff;
@@ -137,8 +135,9 @@ static void draw_tunnel(void)
 			*top++ = pptop;
 			*bot++ = ppbot;
 		}
-		top += 40;
-		bot -= 240 - 40;
+		bot -= 240;
+
+		if(flip) tunptr += 240;
 	}
 }
 
@@ -149,13 +148,6 @@ static void vblank(void)
 	int16_t mat[4];
 	static short gate_speed;
 
-	REG_BG2PA = bgmat[0];
-	REG_BG2PB = bgmat[1];
-	REG_BG2PC = bgmat[2];
-	REG_BG2PD = bgmat[3];
-	REG_BG2X = bgx;
-	REG_BG2Y = bgy;
-
 	bnstate = ~REG_KEYINPUT;
 	if(bnstate & BN_DPAD) {
 		if(gate_speed < 5) {
diff --git a/tools/tungen.c b/tools/tungen.c
index ec21217..49b3f16 100644
--- a/tools/tungen.c
+++ b/tools/tungen.c
@@ -9,10 +9,14 @@ struct vec2 {
 
 int main(int argc, char **argv)
 {
-	int i, j, imgrad, xsz = 240, ysz = 160, texsz = 128;
+	int i, j, frm, imgrad, out_nlines, xsz = 240, ysz = 160, texsz = 128;
+	int half_y = 0;
+	int center = 0;
 	struct vec2 *tunbuf, *tun;
-	float aspect, prev_r;
+	float prev_r;
 	struct vec2 *buf, *ptr;
+	char *endp;
+	int num_frames = 1;
 
 	for(i=1; i<argc; i++) {
 		if(argv[i][0] == '-') {
@@ -32,6 +36,33 @@ int main(int argc, char **argv)
 				}
 				break;
 
+			case 'y':
+				half_y = 1;
+				break;
+
+			case 'c':
+				if(!argv[++i]) {
+					fprintf(stderr, "-c must be followed by a center pixel\n");
+					return 1;
+				}
+				center = strtol(argv[i], &endp, 10);
+				if(endp == argv[i]) {
+					fprintf(stderr, "-c invalid center position: %s\n", argv[i]);
+					return 1;
+				}
+				break;
+
+			case 'n':
+				if(!argv[++i]) {
+					fprintf(stderr, "-n must be followed by the number of frames\n");
+					return 1;
+				}
+				if(!(num_frames = atoi(argv[i]))) {
+					fprintf(stderr, "-n invalid number of frames: %s\n", argv[i]);
+					return 1;
+				}
+				break;
+
 			default:
 				goto invalopt;
 			}
@@ -41,71 +72,78 @@ invalopt:	fprintf(stderr, "invalid argument: %s\n", argv[i]);
 		}
 	}
 
+	out_nlines = half_y ? ysz / 2 : ysz;
+
 	if(!(buf = malloc(xsz * ysz * sizeof *buf))) {
 		perror("failed to allocate buffer");
 		return 1;
 	}
 	imgrad = sqrt(xsz * xsz + ysz * ysz);
 
-#define UDIV	2048
-#define VDIV	32768
-	prev_r = 0.0f;
-	for(i=0; i<VDIV; i++) {
-		float v = (float)(VDIV - i) / (float)VDIV;
-		float r = 1.5 / v + 16;
-		float z = v * 200.0f;
-
-		/* don't bother drawing rings < 1 pixel apart */
-		if(fabs(r - prev_r) < 0.05) continue;
-
-		for(j=0; j<UDIV; j++) {
-			float u = (float)j / (float)(UDIV - 1);
-			float theta = 2.0f * u * M_PI;
-
-			int x = (int)(cos(theta) * r - z) + xsz / 2;
-			int y = (int)(sin(theta) * r) + ysz / 2;
-
-			if(x >= 0 && x < xsz && y >= 0 && y < ysz) {
-				ptr = buf + y * xsz + x;
-				ptr->x = u;
-				ptr->y = v * 8;
-			}
-		}
-		prev_r = r;
-	}
-
 	FILE *fp = fopen("tun_preview.ppm", "wb");
 	if(fp) {
-		fprintf(fp, "P6\n%d %d\n255\n", xsz, ysz);
+		fprintf(fp, "P6\n%d %d\n255\n", xsz, out_nlines * num_frames);
 	}
 
-	aspect = (float)xsz / (float)ysz;
-
-	ptr = buf;
-	for(i=0; i<ysz; i++) {
-		for(j=0; j<xsz; j++) {
-			float u = ptr->x;
-			float v = ptr->y;
-			int r = (int)(u * 8.0 * 255.0f) & 0xff;
-			int g = (int)(v * 8.0 * 255.0f) & 0xff;
-			int b = (~(int)(v * 0.5 * 255.0f) & 0xff) + 105;
-			if(b > 255) b = 255;
-			if(b < 0) b = 0;
-
-			/*if(v > 2.0) r = g = b = 0;*/
 
-			ptr++;
+	for(frm=0; frm<num_frames; frm++) {
+		int coffs = num_frames > 1 ? frm * center / (num_frames - 1) : center;
 
-			uint16_t out = ((uint16_t)(u * 255.0f) & 0xff) |
-				(((uint16_t)(v * 255.0f) & 0xff) << 8);
-			fwrite(&out, sizeof out, 1, stdout);
+#define UDIV	2048
+#define VDIV	32768
+		prev_r = 0.0f;
+#pragma omp parallel for private(i, j, prev_r, ptr) schedule(dynamic)
+		for(i=0; i<VDIV; i++) {
+			float v = (float)(VDIV - i) / (float)VDIV;
+			float r = 4.0 / v + 16;
+			float z = v * coffs;
+
+			/* don't bother drawing rings < 1 pixel apart */
+			if(fabs(r - prev_r) < 0.05) continue;
+
+			for(j=0; j<UDIV; j++) {
+				float u = (float)j / (float)(UDIV - 1);
+				float theta = 2.0f * u * M_PI;
+
+				int x = (int)(cos(theta) * r - z) + xsz / 2;
+				int y = (int)(sin(theta) * r) + ysz / 2;
+
+				if(x >= 0 && x < xsz && y >= 0 && y < ysz) {
+					ptr = buf + y * xsz + x;
+					ptr->x = u;
+					ptr->y = v * 8;
+				}
+			}
+			prev_r = r;
+		}
 
-			if(fp) {
-				fputc(r, fp);
-				fputc(g, fp);
-				fputc(b, fp);
+		ptr = buf;
+		for(i=0; i<out_nlines; i++) {
+			for(j=0; j<xsz; j++) {
+				float u = ptr->x;
+				float v = ptr->y;
+				int r = (int)(u * 8.0 * 255.0f) & 0xff;
+				int g = (int)(v * 8.0 * 255.0f) & 0xff;
+				int b = (~(int)(v * 0.5 * 255.0f) & 0xff) + 105;
+				if(b > 255) b = 255;
+				if(b < 0) b = 0;
+
+				/*if(v > 2.0) r = g = b = 0;*/
+
+				ptr++;
+
+				uint16_t out = ((uint16_t)(u * 255.0f) & 0xff) |
+					(((uint16_t)(v * 255.0f) & 0xff) << 8);
+				fwrite(&out, sizeof out, 1, stdout);
+
+				if(fp) {
+					fputc(r, fp);
+					fputc(g, fp);
+					fputc(b, fp);
+				}
 			}
 		}
+
 	}
 	fflush(stdout);
 
-- 
1.7.10.4