X-Git-Url: http://git.mutantstargoat.com/user/nuclear/?p=fbgfx;a=blobdiff_plain;f=src%2Ftunnel.c;h=85ca7b42fa27f2a6576e07db94657c66e6d6bbac;hp=a348431ab3c0d6d3e9550257dac91c7ab3f04269;hb=6c34c443c62d0b40c91aee5f28985ee99c944caa;hpb=4b522caf5387f8075b7bbb2a2b0475c012157456

diff --git a/src/tunnel.c b/src/tunnel.c
index a348431..85ca7b4 100644
--- a/src/tunnel.c
+++ b/src/tunnel.c
@@ -2,67 +2,213 @@
 #include <stdlib.h>
 #include <math.h>
 #include <imago2.h>
+#include "tpool.h"
 #include "tunnel.h"
 
-static int xsz, ysz;
+#define TEX_FNAME	"data/grid.png"
+#define TEX_USCALE	4
+#define TEX_VSCALE	2
+
+#define USCALE	2
+#define VSCALE	1
+
+extern unsigned long time_msec;
+
+static void draw_tunnel_range(unsigned short *pixels, int starty, int num_lines);
+static int count_bits(unsigned int x);
+static int count_zeros(unsigned int x);
+
+static int xsz, ysz, vxsz, vysz;
 static unsigned int *tunnel_map;
+static unsigned char *tunnel_fog;
+
+static int tex_xsz, tex_ysz;
+static unsigned int *tex_pixels;
+static int tex_xshift, tex_yshift;
+static unsigned int tex_xmask, tex_ymask;
+
+static struct thread_pool *tpool;
 
 
 int init_tunnel(int x, int y)
 {
-	int i, j;
+	int i, j, n;
 	unsigned int *tmap;
+	unsigned char *fog;
+	float aspect = (float)x / (float)y;
 
 	xsz = x;
 	ysz = y;
+	vxsz = xsz / USCALE;
+	vysz = ysz / VSCALE;
 
-	printf("precalculating tunnel map...\n");
-
-	if(!(tunnel_map = malloc(xsz * ysz * sizeof *tunnel_map))) {
+	if(!(tunnel_map = malloc(vxsz * vysz * sizeof *tunnel_map))) {
 		fprintf(stderr, "failed to allocate tunnel map\n");
 		return -1;
 	}
+	if(!(tunnel_fog = malloc(vxsz * vysz))) {
+		fprintf(stderr, "failed to allocate tunnel fog map\n");
+		return -1;
+	}
+
 	tmap = tunnel_map;
+	fog = tunnel_fog;
 
-	for(i=0; i<ysz; i++) {
-		float y = 2.0 * (float)i / (float)ysz - 0.5;
-		for(j=0; j<xsz; j++) {
-			float x = 2.0 * (float)j / (float)xsz - 0.5;
+	for(i=0; i<vysz; i++) {
+		float y = 2.0 * (float)i / (float)vysz - 1.0;
+		for(j=0; j<vxsz; j++) {
+			float x = aspect * (2.0 * (float)j / (float)vxsz - 1.0);
 			float tu = atan2(y, x) / M_PI * 0.5 + 0.5;
-			float tv = sqrt(x*x + y*y);
+			float d = sqrt(x * x + y * y);
+			float tv = d == 0.0 ? 0.0 : 1.0 / d;
+
+			int tx = (int)(tu * 65535.0 * TEX_USCALE) & 0xffff;
+			int ty = (int)(tv * 65535.0 * TEX_VSCALE) & 0xffff;
 
-			int tx = (int)(tu * 65535.0) & 0xffff;
-			int ty = (int)(tv * 65535.0) & 0xffff;
+			int f = (int)(d * 95.0);
 
 			*tmap++ = (tx << 16) | ty;
+			*fog++ = f > 255 ? 255 : f;
 		}
 	}
 
+	if(!(tex_pixels = img_load_pixels(TEX_FNAME, &tex_xsz, &tex_ysz, IMG_FMT_RGBA32))) {
+		fprintf(stderr, "failed to load image " TEX_FNAME "\n");
+		return -1;
+	}
+	if((count_bits(tex_xsz) | count_bits(tex_ysz)) != 1) {
+		fprintf(stderr, "non-pow2 image (%dx%d)\n", tex_xsz, tex_ysz);
+		return -1;
+	}
+
+	n = count_zeros(tex_xsz);
+	for(i=0; i<n; i++) {
+		tex_xmask |= 1 << i;
+	}
+	tex_xshift = n;
+
+	n = count_zeros(tex_ysz);
+	for(i=0; i<n; i++) {
+		tex_ymask |= 1 << i;
+	}
+	tex_yshift = n;
+
+	if(!(tpool = tpool_create(0))) {
+		fprintf(stderr, "failed to create thread pool\n");
+		return -1;
+	}
+
 	return 0;
 }
 
 void destroy_tunnel(void)
 {
+	tpool_destroy(tpool);
 	free(tunnel_map);
+	free(tunnel_fog);
+}
+
+#define NUM_WORK_ITEMS	32
+
+static struct work {
+	unsigned short *pixels;
+	int starty, num_lines;
+} work[NUM_WORK_ITEMS];
+
+static void work_func(void *cls)
+{
+	struct work *w = (struct work*)cls;
+	draw_tunnel_range(w->pixels, w->starty, w->num_lines);
 }
 
 void draw_tunnel(unsigned short *pixels)
 {
-	int i, j, r, g, b;
-	unsigned int *tmap = tunnel_map;
+	int i, num_lines = vysz / NUM_WORK_ITEMS;
+	for(i=0; i<NUM_WORK_ITEMS; i++) {
+		work[i].pixels = pixels;
+		work[i].starty = i * num_lines;
+		work[i].num_lines = num_lines;
+
+		tpool_enqueue(tpool, work + i, work_func, 0);
+	}
+	tpool_wait(tpool);
+}
+
+#define PACK_RGB16(r, g, b) \
+	(((((r) >> 3) & 0x1f) << 11) | ((((g) >> 2) & 0x3f) << 5) | ((b) & 0x1f))
+
+static void draw_tunnel_range(unsigned short *pixels, int starty, int num_lines)
+{
+	int i, j, k, r, g, b;
+	unsigned int *tmap = tunnel_map + starty * vxsz;
+	unsigned char *fog = tunnel_fog + starty * vxsz;
+
+	long toffs = time_msec / 4;
+	pixels += starty * xsz * VSCALE;
 
-	for(i=0; i<ysz; i++) {
-		for(j=0; j<xsz; j++) {
-			unsigned int tx = (*tmap >> 16) & 0xffff;
-			unsigned int ty = *tmap & 0xffff;
+	for(i=0; i<num_lines; i++) {
+		for(j=0; j<vxsz; j++) {
+			unsigned short *ptr;
+			unsigned int col;
+			unsigned int tx = (((*tmap >> 16) & 0xffff) << tex_xshift) >> 16;
+			unsigned int ty = ((*tmap & 0xffff) << tex_yshift) >> 16;
 			++tmap;
 
-			r = tx >> 8;
-			g = ty >> 8;
+			tx += toffs;
+			ty += toffs << 1;
+
+			tx &= tex_xmask;
+			ty &= tex_ymask;
 
-			*pixels++ = ((((r >> 3) & 0x1f) << 11) |
-					(((g >> 2) & 0x3f) << 5));/* |
-					((b >> 3) & 0x1f));*/
+			col = tex_pixels[(ty << tex_xshift) + tx];
+			r = col & 0xff;
+			g = (col >> 8) & 0xff;
+			b = (col >> 16) & 0xff;
+
+			r = (r * *fog) >> 8;
+			g = (g * *fog) >> 8;
+			b = (b * *fog) >> 8;
+			++fog;
+
+			col = ((((r >> 3) & 0x1f) << 11) | (((g >> 2) & 0x3f) << 5) | ((b >> 3) & 0x1f));
+
+			ptr = pixels;
+			for(k=0; k<VSCALE; k++) {
+				switch(USCALE) {
+				case 4:
+					ptr[3] = col;
+				case 3:
+					ptr[2] = col;
+				case 2:
+					ptr[1] = col;
+				case 1:
+					*ptr = col;
+				}
+				ptr += xsz;
+			}
+			pixels += USCALE;
 		}
+		pixels += xsz * (VSCALE - 1);
+	}
+}
+
+static int count_bits(unsigned int x)
+{
+	int i, nbits = 0;
+	for(i=0; i<32; i++) {
+		if(x & 1) ++nbits;
+		x >>= 1;
+	}
+	return nbits;
+}
+
+static int count_zeros(unsigned int x)
+{
+	int i, num = 0;
+	for(i=0; i<32; i++) {
+		if(x & 1) break;
+		++num;
+		x >>= 1;
 	}
+	return num;
 }