Fix colormap segfault

The segfaults due to unaligned vector access should have happened before
the recent changes. Not sure why it worked before. Reverts colormap
vectorization. This is going to be restored once I rework the colormaps.

5 files changed, 29 insertions, 28 deletions

diff --git a/flam3.h b/flam3.h
index ac8278d..7129eee 100644
--- a/flam3.h
+++ b/flam3.h
@@ -41,7 +41,7 @@ char *flam3_version();
 //typedef double flam3_palette[256][3];
 typedef struct {
 	double index;
-	double4 color;
+	double color[4];
 } flam3_palette_entry;
 
 typedef flam3_palette_entry flam3_palette[256];
diff --git a/interpolation.c b/interpolation.c
index 279e1a2..7e1d482 100644
--- a/interpolation.c
+++ b/interpolation.c
@@ -159,8 +159,8 @@ void interpolate_cmap(flam3_palette cmap, double blend,
       double4 t, s;
 	  double t4, s4;
     
-      s = rgb2hsv(p0[i].color);
-      t = rgb2hsv(p1[i].color);
+      s = rgb2hsv(vector_d4 (p0[i].color));
+      t = rgb2hsv(vector_d4 (p1[i].color));
       
       s[3] = p0[i].color[3];
       t[3] = p1[i].color[3];
@@ -172,7 +172,10 @@ void interpolate_cmap(flam3_palette cmap, double blend,
          t[j] = ((1.0-blend) * s[j]) + (blend * t[j]);
       t4 = ((1.0-blend) * s4) + (blend * t4);
          
-      cmap[i].color = hsv2rgb(t);
+	  const double4 c = hsv2rgb(t);
+      cmap[i].color[0] = c[0];
+      cmap[i].color[1] = c[1];
+      cmap[i].color[2] = c[2];
       cmap[i].color[3] = t[3];
       cmap[i].index = t4;
    }
@@ -371,7 +374,7 @@ void flam3_interpolate_n(flam3_genome *result, int ncp,
          s[0] = s[1] = s[2] = s[3] = s4 = 0.0;
          
          for (k = 0; k < ncp; k++) {
-            t = rgb2hsv(cpi[k].palette[i].color);
+            t = rgb2hsv(vector_d4 (cpi[k].palette[i].color));
             for (j = 0; j < 3; j++)
                s[j] += c[k] * t[j];
             
@@ -385,7 +388,10 @@ void flam3_interpolate_n(flam3_genome *result, int ncp,
          if (alpha1 == 1)
             s[3] = 1.0;
        
-         result->palette[i].color = hsv2rgb(s);
+		 const double4 ret_color = hsv2rgb(s);
+         result->palette[i].color[0] = ret_color[0];
+         result->palette[i].color[1] = ret_color[1];
+         result->palette[i].color[2] = ret_color[2];
          result->palette[i].color[3] = s[3];
          result->palette[i].index = s4;
        
diff --git a/palettes.c b/palettes.c
index cf76003..32bef52 100644
--- a/palettes.c
+++ b/palettes.c
@@ -184,8 +184,9 @@ int flam3_get_palette(int n, flam3_palette c, double hue_rotation, randctx * con
 	         
 	         c[i].index = i;
 	       
-		     c[i].color = rgb;
-		         
+		     c[i].color[0] = rgb[0];
+		     c[i].color[1] = rgb[1];
+		     c[i].color[2] = rgb[2];
 		     c[i].color[3] = 1.0;
          }
 	   
diff --git a/rect.c b/rect.c
index 0d9a9db..d988ed1 100644
--- a/rect.c
+++ b/rect.c
@@ -31,9 +31,6 @@ typedef struct {
 	unsigned int sub_batch_size, fuse;
 	unsigned short *xform_distrib;
 
-	flam3_palette dmap;
-	unsigned int cmap_size;
-
 	/* camera stuff */
 	double ws0, wb0s0, hs1, hb1s1; /* shortcuts for indexing */
 	double bounds[4]; /* Corner coords of viewable area */
@@ -44,7 +41,7 @@ typedef struct {
 /*	Lookup color [0,1]
  */
 static double4 color_palette_lookup (const double color,
-		const color_palette_mode mode, const flam3_palette map,
+		const color_palette_mode mode, const flam3_palette_entry * const map,
 		const unsigned int map_count) {
 	assert (color >= 0.0 && color <= 1.0);
 
@@ -56,17 +53,18 @@ static double4 color_palette_lookup (const double color,
 			const unsigned int intix = bottomix;
 
 			if (intix == map_count-1) {
-				return map[intix].color;
+				return vector_d4 (map[intix].color);
 			} else {
-				return map[intix].color * (1.0-frac) +
-					map[intix+1].color * frac;
+				const double4 c1 = vector_d4 (map[intix].color);
+				const double4 c2 = vector_d4 (map[intix+1].color);
+				return c1 * (1.0-frac) + c2 * frac;
 			}
 			break;
 		}
 
 		case PALETTE_MODE_STEP: {
-			const unsigned int intix = nearbyint (color * map_count);
-			return map[intix].color;
+			const unsigned int intix = nearbyint (color * (map_count-1));
+			return vector_d4 (map[intix].color);
 			break;
 		}
 
@@ -137,7 +135,8 @@ static void iter_thread (flam3_genome * const input_genome,
 #endif
 
 					double4 interpcolor = color_palette_lookup (p[2],
-							genome.palette_mode, c->dmap, c->cmap_size);
+							genome.palette_mode, input_genome->palette,
+							256);
 
 					const double logvis = p[3];
 					if (logvis != 1.0) {
@@ -243,16 +242,6 @@ bool render_bucket (flam3_genome * const genome, bucket * const bucket,
 			};
 	assert (c.xform_distrib != NULL);
 
-	/* compute the colormap entries.                             */
-	/* the input colormap is 256 long with entries from 0 to 1.0 */
-	for (unsigned int j = 0; j < CMAP_SIZE; j++) {
-		c.dmap[j].index = genome->palette[(j * 256) / CMAP_SIZE].index / 256.0;
-		for (unsigned int k = 0; k < 4; k++) {
-			c.dmap[j].color[k] = genome->palette[(j * 256) / CMAP_SIZE].color[k];
-		}
-	}
-	c.cmap_size = 256;
-
 	/* compute camera */
 	compute_camera (genome, bucket, &c);
 
diff --git a/vector.h b/vector.h
index 2c4ff52..f914500 100644
--- a/vector.h
+++ b/vector.h
@@ -14,3 +14,8 @@ typedef double double4 __attribute__ ((vector_size (sizeof (double)*4)));
 typedef unsigned int uint2 __attribute__ ((vector_size (sizeof (unsigned int)*2)));
 #endif
 
+inline double4 vector_d4 (const double c[4]) {
+	return (double4) { c[0], c[1], c[2], c[3] };
+}
+
+