Threaded Shadowbuffers Creation

===============================

- One shadowbuffer per thread.
- Added more break tests in shadow buffer code.
- Removed R.clipcrop global, solution is not nice yet, but at
  least threadsafe.
- Fixed bug in strand render shadow buffer code.

9 files changed, 151 insertions, 54 deletions

diff --git a/source/blender/render/intern/include/render_types.h b/source/blender/render/intern/include/render_types.h
index 84230bd4f6d..4a731878ffb 100644
--- a/source/blender/render/intern/include/render_types.h
+++ b/source/blender/render/intern/include/render_types.h
@@ -446,6 +446,10 @@ typedef struct LampRen {
 	ObjectInstanceRen *obi_last[BLENDER_MAX_THREADS];
 	
 	struct MTex *mtex[MAX_MTEX];
+
+	/* threading */
+	int thread_assigned;
+	int thread_ready;
 } LampRen;
 
 /* **************** defines ********************* */
diff --git a/source/blender/render/intern/include/shadbuf.h b/source/blender/render/intern/include/shadbuf.h
index b4f196a89fa..20a266f8579 100644
--- a/source/blender/render/intern/include/shadbuf.h
+++ b/source/blender/render/intern/include/shadbuf.h
@@ -46,6 +46,8 @@ struct ObjectRen;
 void makeshadowbuf(struct Render *re, LampRen *lar);
 void freeshadowbuf(struct LampRen *lar);
 
+void threaded_makeshadowbufs(struct Render *re);
+
 /**
  * Determines the shadow factor for a face and lamp. There is some
  * communication with global variables here.
diff --git a/source/blender/render/intern/include/zbuf.h b/source/blender/render/intern/include/zbuf.h
index 5f1aa5679bf..4d30c1e0dfb 100644
--- a/source/blender/render/intern/include/zbuf.h
+++ b/source/blender/render/intern/include/zbuf.h
@@ -91,6 +91,8 @@ typedef struct ZSpan {
 	float shad_alpha;						/* copy from material, used by irregular shadbuf */
 	int mask, apsmcounter;					/* in use by apixbuf */
 
+	float clipcrop;							/* for shadow, was in R global before */
+
 	void *sss_handle;						/* used by sss */
 	void (*sss_func)(void *, int, int, int, int, int);
 	
@@ -108,7 +110,7 @@ void zspan_scanconvert(struct ZSpan *zpan, void *handle, float *v1, float *v2, f
 
 /* exported to edge render... */
 void zbufclip(struct ZSpan *zspan, int obi, int zvlnr, float *f1, float *f2, float *f3, int c1, int c2, int c3);
-void zbuf_alloc_span(struct ZSpan *zspan, int rectx, int recty);
+void zbuf_alloc_span(struct ZSpan *zspan, int rectx, int recty, float clipcrop);
 void zbufclipwire(struct ZSpan *zspan, int obi, int zvlnr, int ec, float *ho1, float *ho2, float *ho3, float *ho4, int c1, int c2, int c3, int c4);
 
 /* exported to shadeinput.c */
diff --git a/source/blender/render/intern/source/convertblender.c b/source/blender/render/intern/source/convertblender.c
index 4b683441833..c18409c1d00 100644
--- a/source/blender/render/intern/source/convertblender.c
+++ b/source/blender/render/intern/source/convertblender.c
@@ -4209,7 +4209,6 @@ void RE_Database_FromScene(Render *re, Scene *scene, int use_camera_view)
 	database_init_objects(re, lay, 0, 0, 0, 0);
 
 	if(!re->test_break()) {
-		LampRen *lar;
 		int tothalo;
 
 		set_material_lightgroups(re);
@@ -4237,13 +4236,7 @@ void RE_Database_FromScene(Render *re, Scene *scene, int use_camera_view)
 		re->stats_draw(&re->i);
 
 		/* SHADOW BUFFER */
-		for(lar=re->lampren.first; lar; lar= lar->next) {
-			if(re->test_break()) break;
-			if(lar->shb) {
-				/* if type is irregular, this only sets the perspective matrix and autoclips */
-				makeshadowbuf(re, lar);
-			}
-		}
+		threaded_makeshadowbufs(re);
 		
 		/* yafray: 'direct' radiosity, environment maps and raytree init not needed for yafray render */
 		/* although radio mode could be useful at some point, later */
@@ -4767,22 +4760,10 @@ void RE_Database_Baking(Render *re, Scene *scene, int type, Object *actob)
 
 	set_material_lightgroups(re);
 	
-	if(type!=RE_BAKE_LIGHT) {
-		if(re->r.mode & R_SHADOW) {
-			LampRen *lar;
-			
-			/* SHADOW BUFFER */
-			for(lar=re->lampren.first; lar; lar= lar->next) {
-				
-				if(re->test_break()) break;
-				if(lar->shb) {
-					/* if type is irregular, this only sets the perspective matrix and autoclips */
-					/* but, that's not supported for bake... */
-					makeshadowbuf(re, lar);
-				}
-			}
-		}
-	}
+	/* SHADOW BUFFER */
+	if(type!=RE_BAKE_LIGHT)
+		if(re->r.mode & R_SHADOW)
+			threaded_makeshadowbufs(re);
 
 	/* raytree */
 	if(!re->test_break())
diff --git a/source/blender/render/intern/source/pipeline.c b/source/blender/render/intern/source/pipeline.c
index 1c3867a702c..da2542aabaa 100644
--- a/source/blender/render/intern/source/pipeline.c
+++ b/source/blender/render/intern/source/pipeline.c
@@ -1251,7 +1251,6 @@ static void render_tile_processor(Render *re, int firsttile)
 	}
 	
 	freeparts(re);
-	R.strandbuckets= NULL;
 }
 
 /* calculus for how much 1 pixel rendered should rotate the 3d geometry */
@@ -1485,7 +1484,6 @@ static void threaded_tile_processor(Render *re)
 	BLI_end_threads(&threads);
 	freeparts(re);
 	re->viewplane= viewplane; /* restore viewplane, modified by pano render */
-	R.strandbuckets= NULL;
 }
 
 /* currently only called by preview renders and envmap */
diff --git a/source/blender/render/intern/source/rendercore.c b/source/blender/render/intern/source/rendercore.c
index 631a2c184ed..0a9078b32f0 100644
--- a/source/blender/render/intern/source/rendercore.c
+++ b/source/blender/render/intern/source/rendercore.c
@@ -2090,7 +2090,7 @@ static void shade_tface(BakeShade *bs)
 		bs->ibuf= BKE_image_get_ibuf(ima, NULL);
 		/* note, these calls only free/fill contents of zspan struct, not zspan itself */
 		zbuf_free_span(bs->zspan);
-		zbuf_alloc_span(bs->zspan, bs->ibuf->x, bs->ibuf->y);
+		zbuf_alloc_span(bs->zspan, bs->ibuf->x, bs->ibuf->y, R.clipcrop);
 	}				
 	
 	bs->rectx= bs->ibuf->x;
diff --git a/source/blender/render/intern/source/shadbuf.c b/source/blender/render/intern/source/shadbuf.c
index b424a0688fd..8e37606cdf9 100644
--- a/source/blender/render/intern/source/shadbuf.c
+++ b/source/blender/render/intern/source/shadbuf.c
@@ -42,6 +42,8 @@
 #include "BLI_memarena.h"
 #include "BLI_rand.h"
 
+#include "PIL_time.h"
+
 #include "renderpipeline.h"
 #include "render_types.h"
 #include "renderdatabase.h"
@@ -379,9 +381,6 @@ void makeshadowbuf(Render *re, LampRen *lar)
 	float wsize, *jitbuf, twozero[2]= {0.0f, 0.0f}, angle, temp;
 	int *rectz, samples;
 	
-	/* XXXX EVIL! this global is used in clippyra(), zbuf.c */
-	R.clipcrop= 1.0f;
-	
 	if(lar->bufflag & (LA_SHADBUF_AUTO_START|LA_SHADBUF_AUTO_END))
 		shadowbuf_autoclip(re, lar);
 	
@@ -402,8 +401,10 @@ void makeshadowbuf(Render *re, LampRen *lar)
 
 	if(ELEM(lar->buftype, LA_SHADBUF_REGULAR, LA_SHADBUF_HALFWAY)) {
 		/* jitter, weights - not threadsafe! */
+		BLI_lock_thread(LOCK_CUSTOM1);
 		shb->jit= give_jitter_tab(shb->samp);
 		make_jitter_weight_tab(shb, lar->filtertype);
+		BLI_unlock_thread(LOCK_CUSTOM1);
 		
 		shb->totbuf= lar->buffers;
 		if(shb->totbuf==4) jitbuf= give_jitter_tab(2);
@@ -417,6 +418,9 @@ void makeshadowbuf(Render *re, LampRen *lar)
 			zbuffer_shadow(re, shb->persmat, lar, rectz, shb->size, jitbuf[2*samples], jitbuf[2*samples+1]);
 			/* create Z tiles (for compression): this system is 24 bits!!! */
 			compress_shadowbuf(shb, rectz, lar->mode & LA_SQUARE);
+
+			if(re->test_break())
+				break;
 		}
 		
 		MEM_freeN(rectz);
@@ -425,6 +429,103 @@ void makeshadowbuf(Render *re, LampRen *lar)
 	}
 }
 
+static void *do_shadow_thread(void *re_v)
+{
+	Render *re= (Render*)re_v;
+	LampRen *lar;
+
+	do {
+		BLI_lock_thread(LOCK_CUSTOM1);
+		for(lar=re->lampren.first; lar; lar=lar->next) {
+			if(lar->shb && !lar->thread_assigned) {
+				lar->thread_assigned= 1;
+				break;
+			}
+		}
+		BLI_unlock_thread(LOCK_CUSTOM1);
+
+		/* if type is irregular, this only sets the perspective matrix and autoclips */
+		if(lar) {
+			makeshadowbuf(re, lar);
+			BLI_lock_thread(LOCK_CUSTOM1);
+			lar->thread_ready= 1;
+			BLI_unlock_thread(LOCK_CUSTOM1);
+		}
+	} while(lar && !re->test_break());
+
+	return NULL;
+}
+
+static volatile int g_break= 0;
+static int thread_break(void)
+{
+	return g_break;
+}
+
+void threaded_makeshadowbufs(Render *re)
+{
+	ListBase threads;
+	LampRen *lar;
+	int a, totthread= 0;
+	int (*test_break)(void);
+
+	/* count number of threads to use */
+	if(G.rendering) {
+		for(lar=re->lampren.first; lar; lar= lar->next)
+			if(lar->shb)
+				totthread++;
+		
+		totthread= MIN2(totthread, re->r.threads);
+	}
+	else
+		totthread= 1; /* preview render */
+
+	if(totthread <= 1) {
+		for(lar=re->lampren.first; lar; lar= lar->next) {
+			if(re->test_break()) break;
+			if(lar->shb) {
+				/* if type is irregular, this only sets the perspective matrix and autoclips */
+				makeshadowbuf(re, lar);
+			}
+		}
+	}
+	else {
+		/* swap test break function */
+		test_break= re->test_break;
+		re->test_break= thread_break;
+
+		for(lar=re->lampren.first; lar; lar= lar->next) {
+			lar->thread_assigned= 0;
+			lar->thread_ready= 0;
+		}
+
+		BLI_init_threads(&threads, do_shadow_thread, totthread);
+		
+		for(a=0; a<totthread; a++)
+			BLI_insert_thread(&threads, re);
+
+		/* keep rendering as long as there are shadow buffers not ready */
+		do {
+			if((g_break=test_break()))
+				break;
+
+			PIL_sleep_ms(50);
+
+			BLI_lock_thread(LOCK_CUSTOM1);
+			for(lar=re->lampren.first; lar; lar= lar->next)
+				if(lar->shb && !lar->thread_ready)
+					break;
+			BLI_unlock_thread(LOCK_CUSTOM1);
+		} while(lar);
+	
+		BLI_end_threads(&threads);
+
+		/* unset threadsafety */
+		re->test_break= test_break;
+		g_break= 0;
+	}
+}
+
 void freeshadowbuf(LampRen *lar)
 {
 	if(lar->shb) {
@@ -1413,7 +1514,7 @@ static void isb_bsp_fillfaces(Render *re, LampRen *lar, ISBBranch *root)
 	if(lar->mode & LA_LAYER) lay= lar->lay;
 	
 	/* (ab)use zspan, since we use zbuffer clipping code */
-	zbuf_alloc_span(&zspan, size, size);
+	zbuf_alloc_span(&zspan, size, size, re->clipcrop);
 	
 	zspan.zmulx=  ((float)size)/2.0f;
 	zspan.zmuly=  ((float)size)/2.0f;
diff --git a/source/blender/render/intern/source/strand.c b/source/blender/render/intern/source/strand.c
index bbc2b62536f..cadde9530ed 100644
--- a/source/blender/render/intern/source/strand.c
+++ b/source/blender/render/intern/source/strand.c
@@ -965,7 +965,7 @@ unsigned short *zbuffer_strands_shade(Render *re, RenderPart *pa, RenderLayer *r
 	spart.ssamp2.tot= 1;
 	spart.ssamp.tot= 1;
 
-	zbuf_alloc_span(&zspan, pa->rectx, pa->recty);
+	zbuf_alloc_span(&zspan, pa->rectx, pa->recty, re->clipcrop);
 
 	/* needed for transform from hoco to zbuffer co */
 	zspan.zmulx= ((float)re->winx)/2.0;
diff --git a/source/blender/render/intern/source/zbuf.c b/source/blender/render/intern/source/zbuf.c
index 9e9a407472c..57d2bbb3489 100644
--- a/source/blender/render/intern/source/zbuf.c
+++ b/source/blender/render/intern/source/zbuf.c
@@ -84,7 +84,7 @@ extern struct Render R;
 /* ****************** Spans ******************************* */
 
 /* each zbuffer has coordinates transformed to local rect coordinates, so we can simply clip */
-void zbuf_alloc_span(ZSpan *zspan, int rectx, int recty)
+void zbuf_alloc_span(ZSpan *zspan, int rectx, int recty, float clipcrop)
 {
 	memset(zspan, 0, sizeof(ZSpan));
 	
@@ -93,6 +93,8 @@ void zbuf_alloc_span(ZSpan *zspan, int rectx, int recty)
 	
 	zspan->span1= MEM_mallocN(recty*sizeof(float), "zspan");
 	zspan->span2= MEM_mallocN(recty*sizeof(float), "zspan");
+
+	zspan->clipcrop= clipcrop;
 }
 
 void zbuf_free_span(ZSpan *zspan)
@@ -1550,7 +1552,7 @@ void zspan_scanconvert(ZSpan *zspan, void *handle, float *v1, float *v2, float *
  * @param a index for coordinate (x, y, or z)
  */
 
-static void clippyra(float *labda, float *v1, float *v2, int *b2, int *b3, int a)
+static void clippyra(float *labda, float *v1, float *v2, int *b2, int *b3, int a, float clipcrop)
 {
 	float da,dw,u1=0.0,u2=1.0;
 	float v13;
@@ -1566,9 +1568,8 @@ static void clippyra(float *labda, float *v1, float *v2, int *b2, int *b3, int a
 		v13= v1[3];
 	}
 	else {
-		/* XXXXX EVIL! this is a R global, whilst this function can be called for shadow, before R was set */
-		dw= R.clipcrop*(v2[3]-v1[3]);
-		v13= R.clipcrop*v1[3];
+		dw= clipcrop*(v2[3]-v1[3]);
+		v13= clipcrop*v1[3];
 	}	
 	/* according the original article by Liang&Barsky, for clipping of
 	 * homogenous coordinates with viewplane, the value of "0" is used instead of "-w" .
@@ -1826,9 +1827,9 @@ void zbufclip(ZSpan *zspan, int obi, int zvlnr, float *f1, float *f2, float *f3,
 							else if (b==1) arg= 0;
 							else arg= 1;
 							
-							clippyra(labda[0], vlzp[v][0],vlzp[v][1], &b2,&b3, arg);
-							clippyra(labda[1], vlzp[v][1],vlzp[v][2], &b2,&b3, arg);
-							clippyra(labda[2], vlzp[v][2],vlzp[v][0], &b2,&b3, arg);
+							clippyra(labda[0], vlzp[v][0],vlzp[v][1], &b2,&b3, arg, zspan->clipcrop);
+							clippyra(labda[1], vlzp[v][1],vlzp[v][2], &b2,&b3, arg, zspan->clipcrop);
+							clippyra(labda[2], vlzp[v][2],vlzp[v][0], &b2,&b3, arg, zspan->clipcrop);
 
 							if(b2==0 && b3==1) {
 								/* completely 'in', but we copy because of last for() loop in this section */;
@@ -1944,7 +1945,7 @@ void zbuffer_solid(RenderPart *pa, unsigned int lay, short layflag, void(*fillfu
 		zspan= &zspans[zsample];
 
 		zbuffer_part_bounds(&R, pa, bounds);
-		zbuf_alloc_span(zspan, pa->rectx, pa->recty);
+		zbuf_alloc_span(zspan, pa->rectx, pa->recty, R.clipcrop);
 		
 		/* needed for transform from hoco to zbuffer co */
 		zspan->zmulx= ((float)R.winx)/2.0;
@@ -2142,10 +2143,8 @@ void RE_zbufferall_radio(struct RadView *vw, RNode **rg_elem, int rg_totelem, Re
 	/* needed for projectvert */
 	MTC_Mat4MulMat4(winmat, vw->viewmat, vw->winmat);
 
-	/* for clipping... bad stuff actually */
-	R.clipcrop= 1.0f;
-	
-	zbuf_alloc_span(&zspan, vw->rectx, vw->recty);
+	/* 1.0f for clipping in clippyra()... bad stuff actually */
+	zbuf_alloc_span(&zspan, vw->rectx, vw->recty, 1.0f);
 	zspan.zmulx=  ((float)vw->rectx)/2.0;
 	zspan.zmuly=  ((float)vw->recty)/2.0;
 	zspan.zofsx= -0.5f;
@@ -2249,7 +2248,8 @@ void zbuffer_shadow(Render *re, float winmat[][4], LampRen *lar, int *rectz, int
 
 	if(lar->mode & LA_LAYER) lay= lar->lay;
 
-	zbuf_alloc_span(&zspan, size, size);
+	/* 1.0f for clipping in clippyra()... bad stuff actually */
+	zbuf_alloc_span(&zspan, size, size, 1.0f);
 	zspan.zmulx=  ((float)size)/2.0;
 	zspan.zmuly=  ((float)size)/2.0;
 	zspan.zofsx= jitx;
@@ -2315,6 +2315,9 @@ void zbuffer_shadow(Render *re, float winmat[][4], LampRen *lar, int *rectz, int
 						zbufclip(&zspan, 0, 0, ho1, ho2, ho3, c1, c2, c3);
 				}
 			}
+
+			if((a & 255)==255 && re->test_break()) 
+				break;
 		}
 
 		/* strands */
@@ -2345,16 +2348,22 @@ void zbuffer_shadow(Render *re, float winmat[][4], LampRen *lar, int *rectz, int
 					sseg.v[2]= svert+1;
 					sseg.v[3]= (b < strand->totvert-2)? svert+2: svert+1;
 
-					c1= zbuf_shadow_project(cache, strand->vert-sseg.v[0], obwinmat, sseg.v[0]->co, ho1);
-					c2= zbuf_shadow_project(cache, strand->vert-sseg.v[1], obwinmat, sseg.v[1]->co, ho2);
-					c3= zbuf_shadow_project(cache, strand->vert-sseg.v[2], obwinmat, sseg.v[2]->co, ho3);
-					c4= zbuf_shadow_project(cache, strand->vert-sseg.v[3], obwinmat, sseg.v[3]->co, ho4);
+					c1= zbuf_shadow_project(cache, sseg.v[0]-strand->vert, obwinmat, sseg.v[0]->co, ho1);
+					c2= zbuf_shadow_project(cache, sseg.v[1]-strand->vert, obwinmat, sseg.v[1]->co, ho2);
+					c3= zbuf_shadow_project(cache, sseg.v[2]-strand->vert, obwinmat, sseg.v[2]->co, ho3);
+					c4= zbuf_shadow_project(cache, sseg.v[3]-strand->vert, obwinmat, sseg.v[3]->co, ho4);
 
 					if(!(c1 & c2 & c3 & c4))
 						render_strand_segment(NULL, winmat, NULL, &zspan, &sseg);
 				}
 			}
+
+			if((a & 255)==255 && re->test_break()) 
+				break;
 		}
+
+		if(re->test_break()) 
+			break;
 	}
 	
 	/* merge buffers */
@@ -2455,7 +2464,7 @@ void zbuffer_sss(RenderPart *pa, unsigned int lay, void *handle, void (*func)(vo
 	short nofill=0, env=0, wire=0;
 	
 	zbuffer_part_bounds(&R, pa, bounds);
-	zbuf_alloc_span(&zspan, pa->rectx, pa->recty);
+	zbuf_alloc_span(&zspan, pa->rectx, pa->recty, R.clipcrop);
 
 	zspan.sss_handle= handle;
 	zspan.sss_func= func;
@@ -2766,7 +2775,7 @@ void RE_zbuf_accumulate_vecblur(NodeBlurData *nbd, int xsize, int ysize, float *
 	static int firsttime= 1;
 	char *rectmove, *dm;
 	
-	zbuf_alloc_span(&zspan, xsize, ysize);
+	zbuf_alloc_span(&zspan, xsize, ysize, 1.0f);
 	zspan.zmulx=  ((float)xsize)/2.0;
 	zspan.zmuly=  ((float)ysize)/2.0;
 	zspan.zofsx= 0.0f;
@@ -3096,7 +3105,7 @@ static int zbuffer_abuf(RenderPart *pa, APixstr *APixbuf, ListBase *apsmbase, un
 	for(zsample=0; zsample<samples; zsample++) {
 		zspan= &zspans[zsample];
 
-		zbuf_alloc_span(zspan, pa->rectx, pa->recty);
+		zbuf_alloc_span(zspan, pa->rectx, pa->recty, R.clipcrop);
 		
 		/* needed for transform from hoco to zbuffer co */
 		zspan->zmulx=  ((float)R.winx)/2.0;