Fixed several annoyances with halo render + unified render (bug 1989+2382)

- Maximum faces/halos per pixel was 500, which wasn't correctly applied in
  all cases, causing errors in AA
- Moved maximum up to 1000 now
- made halos become clipped away from filling in buffers when behind a
  solid face. That saves a lot of rendertime!

Unified remains weak with halos...

5 files changed, 67 insertions, 93 deletions

diff --git a/source/blender/render/intern/include/vanillaRenderPipe_types.h b/source/blender/render/intern/include/vanillaRenderPipe_types.h
index cf4f2ffb289..df24ed01f5e 100644
--- a/source/blender/render/intern/include/vanillaRenderPipe_types.h
+++ b/source/blender/render/intern/include/vanillaRenderPipe_types.h
@@ -37,7 +37,7 @@
 
 /* Render defines */
 #define  RE_MAX_OSA_COUNT 16 /* The max. number of possible oversamples     */
-#define  RE_MAX_FACES_PER_PIXEL 500 /* max. nr of faces rendered behind one */
+#define  RE_MAX_FACES_PER_PIXEL 1000 /* max. nr of faces rendered behind one */
                              /* pixel                                       */
 
 enum RE_SkyAlphaBlendingType {
diff --git a/source/blender/render/intern/include/zbufferdatastruct.h b/source/blender/render/intern/include/zbufferdatastruct.h
index 790cce2cb77..2cef53dc9f6 100644
--- a/source/blender/render/intern/include/zbufferdatastruct.h
+++ b/source/blender/render/intern/include/zbufferdatastruct.h
@@ -71,7 +71,6 @@ RE_APixstrExt  *addpseA(void);
  * Add an object to a zbuffer entry.
  */
 void insertObject(int teller,
-/*  				  int opaque, */
 				  int obindex,
 				  int obtype, 
 				  int dist, 
diff --git a/source/blender/render/intern/include/zbufferdatastruct_types.h b/source/blender/render/intern/include/zbufferdatastruct_types.h
index 0607f19bbc9..5048adfa8bc 100644
--- a/source/blender/render/intern/include/zbufferdatastruct_types.h
+++ b/source/blender/render/intern/include/zbufferdatastruct_types.h
@@ -55,10 +55,12 @@ typedef struct RE_APixstrExt {
 /* but this should definitely be done in a better way. An enum may   */
 /* be some help, but masking is still a nice feature...              */
 /* object types to buffer in the z buffer */
-#define RE_NONE 0
-#define RE_POLY 1
-#define RE_HALO 2
-#define RE_SKY  4
+/* RE_SOLID is flag for RE_POLY, as speedup */
+#define RE_NONE		0
+#define RE_POLY		1
+#define RE_HALO		2
+#define RE_SKY		4
+#define RE_SOLID	8
 
 /* unique indices for each field */
 #define RE_ZMIN           0
diff --git a/source/blender/render/intern/source/vanillaRenderPipe.c b/source/blender/render/intern/source/vanillaRenderPipe.c
index 500da84dedb..91a5173c353 100644
--- a/source/blender/render/intern/source/vanillaRenderPipe.c
+++ b/source/blender/render/intern/source/vanillaRenderPipe.c
@@ -197,6 +197,7 @@ static void freeRenderBuffers(void) {
 static void zBufferFillFace(unsigned int zvlnr, float *v1, float *v2, float *v3)  
 {
 	/* Coordinates of the vertices are specified in ZCS */
+	VlakRen *vlr;
 	int apteller, apoffsetteller;
 	double z0; /* used as temp var*/
 	double xx1;
@@ -205,13 +206,17 @@ static void zBufferFillFace(unsigned int zvlnr, float *v1, float *v2, float *v3)
 	register int zverg,zvlak,x;
 	int my0,my2,sn1,sn2,rectx,zd;
 	int y,omsl,xs0,xs1,xs2,xs3, dx0,dx1,dx2, mask;
-
+	int obtype;
 	/* These used to be doubles.  We may want to change them back if the     */
 	/* loss of accuracy proves to be a problem? There does not seem to be    */
 	/* any performance issues here, so I'll just keep the doubles.           */
 	/*  	float vec0[3], vec1[3], vec2[3]; */
 	double vec0[3], vec1[3], vec2[3];
-
+	
+	vlr= RE_findOrAddVlak( (zvlnr-1) & 0x7FFFFF);
+	if(vlr->mat->mode & MA_ZTRA) obtype= RE_POLY;
+	else obtype= RE_POLY|RE_SOLID;
+	
 	/* MIN MAX */
 	/* sort vertices for min mid max y value */
 	if(v1[1]<v2[1]) {
@@ -351,7 +356,7 @@ static void zBufferFillFace(unsigned int zvlnr, float *v1, float *v2, float *v3)
 		zverg-= Azvoordeel;
 		
 		while(x>=0) {
-			insertObject(apteller, /*  RE_treat_face_as_opaque, */ zvlnr, RE_POLY, zverg, mask);
+			insertObject(apteller, zvlnr, obtype, zverg, mask);
 			zverg+= zd;
 			apteller++;
 			x--;
@@ -392,7 +397,7 @@ static void zBufferFillFace(unsigned int zvlnr, float *v1, float *v2, float *v3)
 		zverg-= Azvoordeel;
       
 		while(x>=0) {
-			insertObject(apteller, /*  RE_treat_face_as_opaque, */ zvlnr, RE_POLY, zverg, mask);
+			insertObject(apteller, zvlnr, obtype, zverg, mask);
 			zverg+= zd;
 			apteller++;
 			x--;
@@ -455,7 +460,7 @@ static void zBufferFillEdge(unsigned int zvlnr, float *vec1, float *vec2)
 			}
 			
 			if(x>=0 && y>=Aminy && y<=Amaxy) {
-				insertObject(apteller, /*  RE_treat_face_as_opaque, */ zvlnr, RE_POLY, vergz, mask);
+				insertObject(apteller, zvlnr, RE_POLY, vergz, mask);
 			}
 			
 			v1[1]+= dy;
@@ -507,7 +512,7 @@ static void zBufferFillEdge(unsigned int zvlnr, float *vec1, float *vec2)
 			}
 			
 			if(x>=0 && y>=Aminy && (x < zBufferWidth)) {
-				insertObject(apteller, /*  RE_treat_face_as_opaque, */ zvlnr, RE_POLY, vergz, mask);
+				insertObject(apteller, zvlnr, RE_POLY, vergz, mask);
 			}
 			
 			v1[0]+= dx;
@@ -522,7 +527,7 @@ static void zBufferFillEdge(unsigned int zvlnr, float *vec1, float *vec2)
  * Count and sort the list behind ap into buf. Sorts on min. distance.
  * Low index <=> high z
  */
-static int countAndSortPixelFaces(int zrow[RE_MAX_FACES_PER_PIXEL][RE_PIXELFIELDSIZE], 
+static int countAndSortPixelFaces(int zrow[][RE_PIXELFIELDSIZE], 
                            RE_APixstrExt *ap)
 {
     int totvlak;          /* face counter                          */
@@ -582,7 +587,7 @@ static int            VR_cbuf[RE_MAX_FACES_PER_PIXEL][2];
 /** 
  * Analyze the z-buffer, and pre-sample the colours.
  */
-static int composeStack(int zrow[RE_MAX_FACES_PER_PIXEL][RE_PIXELFIELDSIZE], RE_COLBUFTYPE *collector, 
+static int composeStack(int zrow[][RE_PIXELFIELDSIZE], RE_COLBUFTYPE *collector, 
 				 struct RE_faceField* stack, int ptr,
 				 int totvlak, float x, float y, int osaNr) 
 {
@@ -974,10 +979,18 @@ static void integratePerSubStack(float *sampcol, struct RE_faceField* stack,
 /* an RE_APixstrExt* array                                                   */
 /* - redo the numbering to something more logical                            */
 
+
+/* threadsafe global arrays, too large for stack */
+typedef struct zbufline {
+	int  zrow[RE_MAX_FACES_PER_PIXEL][RE_PIXELFIELDSIZE];
+	struct RE_faceField osastack[RE_MAX_FACES_PER_PIXEL + 1]; 
+} zbufline;
+
+static zbufline zb1, zb2;
+
 static void renderZBufLine(int y, RE_COLBUFTYPE *colbuf1, RE_COLBUFTYPE *colbuf2, RE_COLBUFTYPE *colbuf3) 
 {
-    int  zrow[RE_MAX_FACES_PER_PIXEL][RE_PIXELFIELDSIZE];
-    RE_APixstrExt *ap;       /* iterator for the face-lists                  */
+     RE_APixstrExt *ap;       /* iterator for the face-lists                  */
 	RE_COLBUFTYPE  collector[4];
 	RE_COLBUFTYPE  sampcol[RE_MAX_OSA_COUNT * 4];
     RE_COLBUFTYPE *j = NULL; /* generic pixel pointer                        */
@@ -985,9 +998,12 @@ static void renderZBufLine(int y, RE_COLBUFTYPE *colbuf1, RE_COLBUFTYPE *colbuf2
     int x;                   /* pixel counter                                */
     int i;                   /* yet another counter                          */
     int stackDepth;          /* faces-behind-this-pixel counter              */
-	struct RE_faceField RE_OSAstack[RE_MAX_FACES_PER_PIXEL + 1]; 
-	int RE_OSAstack_ptr;    /* Points to the lowest empty field. The indexed */
-                            /* field is NOT readable.                        */
+	int osastack_ptr;	/* Points to the lowest empty field. The indexed */
+	zbufline *zbl;
+
+	/* thread safe row buffers */
+	if(y & 1) zbl= &zb1;
+	else zbl= &zb2;
 	
     /* Prepare iterators */
     ap        = APixbufExt + (zBufferWidth * (y - Aminy));
@@ -1004,15 +1020,13 @@ static void renderZBufLine(int y, RE_COLBUFTYPE *colbuf1, RE_COLBUFTYPE *colbuf2
             };
 
             /* a. count and sort number of faces */
-            stackDepth = countAndSortPixelFaces(zrow, ap);
+            stackDepth = countAndSortPixelFaces( zbl->zrow, ap);
 			
             /* b,c. oversample all subpixels, then integrate                 */
-			RE_OSAstack_ptr = 0;
-			RE_OSAstack_ptr = composeStack(zrow, collector,
-										   RE_OSAstack, RE_OSAstack_ptr,
-										   stackDepth, x, y, osaNr);
-  			integratePerSubStack(sampcol, RE_OSAstack, RE_OSAstack_ptr, 
-  								 x, y, osaNr); 
+			osastack_ptr = 0;
+			osastack_ptr = composeStack(zbl->zrow, collector, zbl->osastack, osastack_ptr, 
+										stackDepth, x, y, osaNr);
+  			integratePerSubStack(sampcol, zbl->osastack, osastack_ptr,  x, y, osaNr); 
 			
 			/* d. Gamma corrected blending and Gaussian                      */
 			sampleFloatColV2FloatColVFilter(sampcol, colbuf1, colbuf2, colbuf3, osaNr);
@@ -1436,6 +1450,7 @@ static void calcZBufLine(int y)
 	
     /* (FORALL y: Aminy =< y =< Amaxy: y is buffered)                        */
     if( (y < Aminy) || (y > Amaxy)) {
+
         /* prepare buffer */
         part  = (y/RE_ZBUFLEN);     /* These two lines are mystifying me...  */
         Aminy = part * RE_ZBUFLEN;  /* Possibly for rounding things?         */
@@ -1458,7 +1473,7 @@ static void calcZBufLine(int y)
             if(RE_local_test_break()) keepLooping = 0;
             Zsample++;
         }
-    };
+    }
     
 } 
 
diff --git a/source/blender/render/intern/source/zbufferdatastruct.c b/source/blender/render/intern/source/zbufferdatastruct.c
index 319393290bd..4ab8a401c20 100644
--- a/source/blender/render/intern/source/zbufferdatastruct.c
+++ b/source/blender/render/intern/source/zbufferdatastruct.c
@@ -61,6 +61,7 @@
 
 #include "MEM_guardedalloc.h"
 #include "zbufferdatastruct.h"
+#include "vanillaRenderPipe_types.h"
 #include "render.h"
 
 #ifdef HAVE_CONFIG_H
@@ -174,7 +175,6 @@ RE_APixstrExt *addpseA(void)
 /* ------------------------------------------------------------------------- */
 
 void insertObject(int apteller,
-/*  				  int opaque, */
 				  int obindex,
 				  int obtype, 
 				  int dist, 
@@ -182,21 +182,17 @@ void insertObject(int apteller,
 {
 	/* Guard the insertion if needed? */
   	RE_APixstrExt* apn = &APixbufExt[apteller]; 
-	int all_subpixels= 0;
-	
-	//if(obtype==RE_POLY) {
-	//	VlakRen *vlr= RE_findOrAddVlak( (obindex-1) & 0x7FFFFF);
-	//	if(vlr->flag & R_FULL_OSA) all_subpixels= 1;
-	//}
+	int all_subpixels= 0;	// not used now... (ton)
 	
 	while(apn) {
+		
 		if(apn->t[0] == RE_NONE) {
 			apn->p[0] = obindex; apn->t[0] = obtype;
 			apn->zmin[0] = dist; apn->zmax[0] = dist;
 			apn->mask[0] = mask;   
 			break; 
 		}
-		if(all_subpixels==0) {
+		else if(all_subpixels==0) {
 			if((apn->p[0] == obindex) && (apn->t[0] & obtype)) {
 				if(dist < apn->zmin[0]) apn->zmin[0] = dist;
 				else if(dist > apn->zmax[0]) apn->zmax[0] = dist;
@@ -204,13 +200,14 @@ void insertObject(int apteller,
 				break; 
 			} 
 		}
+		
 		if(apn->t[1] == RE_NONE) {
 			apn->p[1] = obindex; apn->t[1] = obtype;
 			apn->zmin[1] = dist; apn->zmax[1] = dist;
 			apn->mask[1] = mask;   
 			break; 
 		}
-		if(all_subpixels==0) {
+		else if(all_subpixels==0) {
 			if((apn->p[1] == obindex) && (apn->t[1] & obtype)) {
 				if(dist < apn->zmin[1]) apn->zmin[1] = dist;
 				else if(dist > apn->zmax[1]) apn->zmax[1] = dist;
@@ -218,13 +215,14 @@ void insertObject(int apteller,
 				break; 
 			} 
 		}
+		
 		if(apn->t[2] == RE_NONE) {
 			apn->p[2] = obindex; apn->t[2] = obtype;
 			apn->zmin[2] = dist; apn->zmax[2] = dist;
 			apn->mask[2] = mask;   
 			break; 
 		}
-		if(all_subpixels==0) {
+		else if(all_subpixels==0) {
 			if((apn->p[2] == obindex) && (apn->t[2] & obtype)) {
 				if(dist < apn->zmin[2]) apn->zmin[2] = dist;
 				else if(dist > apn->zmax[2]) apn->zmax[2] = dist;
@@ -232,13 +230,14 @@ void insertObject(int apteller,
 				break; 
 			} 
 		}
+		
 		if(apn->t[3] == RE_NONE) {
 			apn->p[3] = obindex; apn->t[3] = obtype;
 			apn->zmin[3] = dist; apn->zmax[3] = dist;
 			apn->mask[3] = mask;   
 			break; 
 		}
-		if(all_subpixels==0) {
+		else if(all_subpixels==0) {
 			if((apn->p[3] == obindex) && (apn->t[3] & obtype)) {
 				if(dist < apn->zmin[3]) apn->zmin[3] = dist;
 				else if(dist > apn->zmax[3]) apn->zmax[3] = dist;
@@ -246,66 +245,16 @@ void insertObject(int apteller,
 				break; 
 			} 
 		}
+		
 		if(apn->next==0) apn->next= addpseA();
 		apn= apn->next;
 	}				
-} /* end of insertObject(RE_APixstrExt*, int, int, int, int) */
+}
 
 /* ------------------------------------------------------------------------- */
 
-void insertFlatObject(RE_APixstrExt *apn, int obindex, int obtype, int dist, int mask)
-{
-	while(apn) {
-		if(apn->t[0] == RE_NONE) {
-			apn->p[0] = obindex; apn->zmin[0] = dist; 
-			apn->zmax[0] = dist; apn->mask[0] = mask; 
-			apn->t[0] = obtype;  
-			break; 
-		}
-#ifndef RE_INDIVIDUAL_SUBPIXELS
-		if( (apn->t[0] & obtype) && (apn->p[0] == obindex)) {
-			apn->mask[0]|= mask; break; 
-		}
-#endif
-		if(apn->t[1] == RE_NONE) {
-			apn->p[1] = obindex; apn->zmin[1] = dist; 
-			apn->zmax[1] = dist; apn->mask[1] = mask; 
-			apn->t[1] = obtype;  
-			break;
-		}
-#ifndef RE_INDIVIDUAL_SUBPIXELS
-		if( (apn->t[1] & obtype) && (apn->p[1] == obindex)) {
-			apn->mask[1]|= mask; break; 
-		}
-#endif
-		if(apn->t[2] == RE_NONE) {
-			apn->p[2] = obindex; apn->zmin[2] = dist; 
-			apn->zmax[2] = dist; apn->mask[2] = mask; 
-			apn->t[2] = obtype;  
-			break;
-		}
-#ifndef RE_INDIVIDUAL_SUBPIXELS
-		if( (apn->t[2] & obtype) && (apn->p[2] == obindex)) {
-			apn->mask[2]|= mask; break; 
-		}
-#endif
-		if(apn->t[3] == RE_NONE) {
-			apn->p[3] = obindex; apn->zmin[3] = dist; 
-			apn->zmax[3] = dist; apn->mask[3] = mask; 
-			apn->t[3] = obtype;  
-			break;
-		}
-#ifndef RE_INDIVIDUAL_SUBPIXELS
-		if( (apn->t[3] & obtype) && (apn->p[3] == obindex)) {
-			apn->mask[3]|= mask; break; 
-		}
-#endif
-		if(apn->next==0) apn->next= addpseA();
-		apn= apn->next;
-	};                    
-} /* end of void insertFlatObject(RE_APixstrExt, int, int, int, int)*/
-
 /* ------------------------------------------------------------------------- */
+
 /* This function might be helped by an end-of-list marker                    */
 void insertFlatObjectNoOsa(RE_APixstrExt *ap, 
 						   int obindex, 
@@ -313,6 +262,8 @@ void insertFlatObjectNoOsa(RE_APixstrExt *ap,
 						   int dist, 
 						   int mask)
 {
+	int counter;
+	
 	while(ap) {
 		if(ap->t[0] == RE_NONE) {
 			ap->p[0] = obindex; ap->zmin[0] = dist; 
@@ -320,6 +271,7 @@ void insertFlatObjectNoOsa(RE_APixstrExt *ap,
 			ap->t[0] = obtype;  
 			break; 
 		}
+		else if(ap->t[0] & RE_SOLID) if( dist > ap->zmin[0] ) break;
 
 		if(ap->t[1] == RE_NONE) {
 			ap->p[1] = obindex; ap->zmin[1] = dist; 
@@ -327,6 +279,7 @@ void insertFlatObjectNoOsa(RE_APixstrExt *ap,
 			ap->t[1] = obtype;  
 			break;
 		}
+		else if(ap->t[1] & RE_SOLID) if( dist > ap->zmin[1] ) break;
 
 		if(ap->t[2] == RE_NONE) {
 			ap->p[2] = obindex; ap->zmin[2] = dist; 
@@ -334,6 +287,7 @@ void insertFlatObjectNoOsa(RE_APixstrExt *ap,
 			ap->t[2] = obtype;  
 			break;
 		}
+		else if(ap->t[2] & RE_SOLID) if( dist > ap->zmin[2] ) break;
 
 		if(ap->t[3] == RE_NONE) {
 			ap->p[3] = obindex; ap->zmin[3] = dist; 
@@ -341,11 +295,15 @@ void insertFlatObjectNoOsa(RE_APixstrExt *ap,
 			ap->t[3] = obtype;  
 			break;
 		}
-
+		else if(ap->t[3] & RE_SOLID) if( dist > ap->zmin[3] ) break;
+		
+		counter+= 4;
+		if(counter > RE_MAX_FACES_PER_PIXEL) break;
+		
 		if(ap->next==0) ap->next= addpseA();
 		ap= ap->next;
 	};                    
-} /* end of void insertFlatObjectNoOsa(RE_APixstrExt, int, int, int, int)*/
+}
 
 /* ------------------------------------------------------------------------- */