no-thread option, coalesced tunnel lut, more accurate fps counter

diff --git a/Makefile b/Makefile
index b1ba5c1..7ead672 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -3,7 +3,13 @@ obj = $(src:.c=.o)
 bin = fbgfx
 
 CFLAGS = -pedantic -Wall -g -O3
-LDFLAGS = -limago -lm -lpthread
+LDFLAGS = -limago -lm $(ldflags_pthread)
+
+ifeq ($(NOTHREADS), 1)
+       CFLAGS += -DNO_THREADS
+else
+       ldflags_pthread = -lpthread
+endif
 
 $(bin): $(obj)
        $(CC) -o $@ $(obj) $(LDFLAGS)

diff --git a/src/main.c b/src/main.c
index 4d63726..7c41e2e 100644 (file)
--- a/src/main.c
+++ b/src/main.c
@@ -61,7 +61,9 @@ int main(void)
 
                scr_update();
                scr_draw();
-               ++num_frames;
+               if(time_msec >= 4000) {
+                       ++num_frames;
+               }
        }
 
        time_msec = get_time_msec() - start_msec;
@@ -70,7 +72,7 @@ end:
        fbev_shutdown();
        fbgfx_restore_video_mode();
        if(num_frames && time_msec) {
-               printf("\ravg framerate: %.1f\n", (float)num_frames / ((float)time_msec / 1000.0));
+               printf("\ravg framerate: %.1f\n", (float)num_frames / ((float)(time_msec - 4000) / 1000.0));
        }
        return 0;
 }

diff --git a/src/tpool.c b/src/tpool.c
index 9e73f44..4d06570 100644 (file)
--- a/src/tpool.c
+++ b/src/tpool.c
@@ -1,3 +1,4 @@
+#ifndef NO_THREADS
 /* worker thread pool based on POSIX threads
  * author: John Tsiombikas <nuclear@member.fsf.org>
  * This code is public domain.
@@ -311,3 +312,5 @@ int tpool_num_processors(void)
        return info.dwNumberOfProcessors;
 #endif
 }
+
+#endif /* !def NO_THREADS */

diff --git a/src/tunnel.c b/src/tunnel.c
index 221fadc..e6a53d5 100644 (file)
--- a/src/tunnel.c
+++ b/src/tunnel.c
@@ -3,7 +3,9 @@
 #include <math.h>
 #include <assert.h>
 #include <imago2.h>
+#ifndef NO_THREADS
 #include "tpool.h"
+#endif
 #include "demo.h"
 #include "screen.h"
 
@@ -44,17 +46,23 @@ static struct screen scr = {
        draw
 };
 
+struct tunmap {
+       int u, v;
+       unsigned char fog;
+};
+
 static int xsz, ysz, vxsz, vysz;
 static int pan_width, pan_height;
-static unsigned int *tunnel_map;
-static unsigned char *tunnel_fog;
+static struct tunmap *tunnel_map;
 
 static int tex_xsz, tex_ysz;
 static unsigned int *tex_pixels;
 static int tex_xshift, tex_yshift;
 static unsigned int tex_xmask, tex_ymask;
 
+#ifndef NO_THREADS
 static struct thread_pool *tpool;
+#endif
 
 static long trans_start, trans_dur;
 static int trans_dir;
@@ -69,8 +77,7 @@ struct screen *tunnel_screen(void)
 static int init(void)
 {
        int i, j, n;
-       unsigned int *tmap;
-       unsigned char *fog;
+       struct tunmap *tmap;
        float aspect = (float)fb_width / (float)fb_height;
 
        switch(fb_depth) {
@@ -97,13 +104,8 @@ static int init(void)
                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<vysz; i++) {
                float y = 2.0 * (float)i / (float)vysz - 1.0;
@@ -113,13 +115,12 @@ static int init(void)
                        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 f = (int)(d * 128.0);
+                       int fog = (int)(d * 128.0f);
 
-                       *tmap++ = (tx << 16) | ty;
-                       *fog++ = f > 255 ? 255 : f;
+                       tmap->u = (int)(tu * 65535.0 * TEX_USCALE) & 0xffff;
+                       tmap->v = (int)(tv * 65535.0 * TEX_VSCALE) & 0xffff;
+                       tmap->fog = fog > 255 ? 255 : fog;
+                       tmap++;
                }
        }
 
@@ -144,10 +145,12 @@ static int init(void)
        }
        tex_yshift = n;
 
+#ifndef NO_THREADS
        if(!(tpool = tpool_create(0))) {
                fprintf(stderr, "failed to create thread pool\n");
                return -1;
        }
+#endif
 
        /* initialize the constant part of all work items */
        for(i=0; i<NUM_WORK_ITEMS; i++) {
@@ -161,9 +164,10 @@ static int init(void)
 
 static void destroy(void)
 {
+#ifndef NO_THREADS
        tpool_destroy(tpool);
+#endif
        free(tunnel_map);
-       free(tunnel_fog);
 }
 
 static void start(long trans_time)
@@ -221,17 +225,22 @@ static void draw(void)
                work[i].xoffs = xoffs;
                work[i].yoffs = yoffs;
 
+#ifdef NO_THREADS
+               work_func(work + i);
+       }
+#else
                tpool_enqueue(tpool, work + i, work_func, 0);
        }
        tpool_wait(tpool);
+#endif
 }
 
-static void tunnel_color(int *rp, int *gp, int *bp, long toffs, unsigned int tpacked, int fog)
+static void tunnel_color(int *rp, int *gp, int *bp, long toffs, struct tunmap *tmap)
 {
        int r, g, b;
        unsigned int col;
-       unsigned int tx = (((tpacked >> 16) & 0xffff) << tex_xshift) >> 16;
-       unsigned int ty = ((tpacked & 0xffff) << tex_yshift) >> 16;
+       unsigned int tx = (tmap->u << tex_xshift) >> 16;
+       unsigned int ty = (tmap->v << tex_yshift) >> 16;
        tx += toffs;
        ty += toffs << 1;
 
@@ -243,9 +252,9 @@ static void tunnel_color(int *rp, int *gp, int *bp, long toffs, unsigned int tpa
        g = (col >> 8) & 0xff;
        b = (col >> 16) & 0xff;
 
-       *rp = (r * fog) >> 8;
-       *gp = (g * fog) >> 8;
-       *bp = (b * fog) >> 8;
+       *rp = (r * tmap->fog) >> 8;
+       *gp = (g * tmap->fog) >> 8;
+       *bp = (b * tmap->fog) >> 8;
 }
 
 #define PACK_RGB16(r, g, b) \
@@ -256,8 +265,7 @@ static void tunnel_color(int *rp, int *gp, int *bp, long toffs, unsigned int tpa
 static void draw_tunnel_range16(void *pix, int xoffs, int yoffs, int starty, int num_lines, long tm)
 {
        int i, j;
-       unsigned int *tmap = tunnel_map + (starty + yoffs) * vxsz + xoffs;
-       unsigned char *fog = tunnel_fog + (starty + yoffs) * vxsz + xoffs;
+       struct tunmap *tmap = tunnel_map + (starty + yoffs) * vxsz + xoffs;
 
        long toffs = tm / 4;
        unsigned int *pixels = (unsigned int*)pix + starty * (fb_width >> 1);
@@ -267,20 +275,18 @@ static void draw_tunnel_range16(void *pix, int xoffs, int yoffs, int starty, int
                        unsigned int col;
                        int r, g, b;
 
-                       tunnel_color(&r, &g, &b, toffs, tmap[j], fog[j]);
+                       tunnel_color(&r, &g, &b, toffs, tmap + j);
                        col = PACK_RGB16(r, g, b);
                        *pixels++ = (col << 16) | col;
                }
                tmap += vxsz;
-               fog += vxsz;
        }
 }
 
 static void draw_tunnel_range32(void *pix, int xoffs, int yoffs, int starty, int num_lines, long tm)
 {
        int i, j;
-       unsigned int *tmap = tunnel_map + (starty + yoffs) * vxsz + xoffs;
-       unsigned char *fog = tunnel_fog + (starty + yoffs) * vxsz + xoffs;
+       struct tunmap *tmap = tunnel_map + (starty + yoffs) * vxsz + xoffs;
 
        long toffs = tm / 4;
        unsigned int *pixels = (unsigned int*)pix + starty * fb_width;
@@ -290,14 +296,13 @@ static void draw_tunnel_range32(void *pix, int xoffs, int yoffs, int starty, int
                        unsigned int col;
                        int r, g, b;
 
-                       tunnel_color(&r, &g, &b, toffs, tmap[j], fog[j]);
+                       tunnel_color(&r, &g, &b, toffs, tmap + j);
                        col = PACK_RGB32(r, g, b);
 
                        *pixels++ = col;
                        *pixels++ = col;
                }
                tmap += vxsz;
-               fog += vxsz;
        }
 }