implement a (partially) concurrent pipeline.

4 files changed, 234 insertions, 134 deletions

diff --git a/camera.c b/camera.c
index eb6f181..67bca08 100644
--- a/camera.c
+++ b/camera.c
@@ -79,15 +79,27 @@ reloadcamera(Camera *c)
 }
 
 void
-shootcamera(Camera *c, Shader *s)
+shootcamera(Camera *c, Shadertab *s)
 {
+	Renderjob *job;
 	uvlong t0, t1;
 
+	job = emalloc(sizeof *job);
+	memset(job, 0, sizeof *job);
+	job->fb = c->vp->fbctl->fb[c->vp->fbctl->idx^1];	/* address the back buffer */
+	job->scene = c->s;
+	job->shaders = s;
+	job->donec = chancreate(sizeof(void*), 0);
+
 	c->vp->fbctl->reset(c->vp->fbctl);
 	t0 = nanosec();
-	shade(c->vp->fbctl->fb[c->vp->fbctl->idx^1], c->s, s);	/* address the back buffer */
+	sendp(c->rctl->c, job);
+	recvp(job->donec);
 	t1 = nanosec();
 	c->vp->fbctl->swap(c->vp->fbctl);
 
+	chanfree(job->donec);
+	free(job);
+
 	updatestats(c, t1-t0);
 }
diff --git a/graphics.h b/graphics.h
index af8cdd0..fdd1463 100644
--- a/graphics.h
+++ b/graphics.h
@@ -27,7 +27,9 @@ typedef struct Scene Scene;
 typedef struct VSparams VSparams;
 typedef struct FSparams FSparams;
 typedef struct SUparams SUparams;
-typedef struct Shader Shader;
+typedef struct Shadertab Shadertab;
+typedef struct Renderer Renderer;
+typedef struct Renderjob Renderjob;
 typedef struct Framebuf Framebuf;
 typedef struct Framebufctl Framebufctl;
 typedef struct Viewport Viewport;
@@ -150,8 +152,8 @@ struct SUparams
 	int id;
 	Memimage *frag;
 	Channel *donec;
+	Renderjob *job;
 
-	/* TODO replace with a Scene */
 	Entity *entity;
 
 	uvlong uni_time;
@@ -160,13 +162,32 @@ struct SUparams
 	Color (*fshader)(FSparams*);
 };
 
-struct Shader
+struct Shadertab
 {
 	char *name;
 	Point3 (*vshader)(VSparams*);	/* vertex shader */
 	Color (*fshader)(FSparams*);	/* fragment shader */
 };
 
+struct Renderer
+{
+	Channel *c;
+};
+
+struct Renderjob
+{
+	Framebuf *fb;
+	Scene *scene;
+	Shadertab *shaders;
+	Channel *donec;
+
+	ulong nrem;		/* remaining entities to process */
+	ulong lastid;
+	uvlong time0;
+
+	Renderjob *next;
+};
+
 struct Framebuf
 {
 	Memimage *cb;	/* color buffer */
@@ -201,6 +222,7 @@ struct Camera
 	RFrame3;		/* VCS */
 	Viewport *vp;
 	Scene *s;
+	Renderer *rctl;
 	double fov;		/* vertical FOV */
 	struct {
 		double n, f;	/* near and far clipping planes */
@@ -219,13 +241,14 @@ void configcamera(Camera*, Viewport*, double, double, double, Projection);
 void placecamera(Camera*, Point3, Point3, Point3);
 void aimcamera(Camera*, Point3);
 void reloadcamera(Camera*);
-void shootcamera(Camera*, Shader*);
+void shootcamera(Camera*, Shadertab*);
 
 /* viewport */
 Viewport *mkviewport(Rectangle);
 void rmviewport(Viewport*);
 
 /* render */
+Renderer *initgraphics(void);
 Point3 model2world(Entity*, Point3);
 Point3 world2vcs(Camera*, Point3);
 Point3 vcs2clip(Camera*, Point3);
diff --git a/internal.h b/internal.h
index 5fcd6e8..7f0d648 100644
--- a/internal.h
+++ b/internal.h
@@ -1,3 +1,10 @@
+typedef struct Jobqueue Jobqueue;
+
+struct Jobqueue
+{
+	Renderjob *hd, *tl;
+};
+
 /* alloc */
 void *emalloc(ulong);
 void *erealloc(void*, ulong);
@@ -9,9 +16,6 @@ void rmfb(Framebuf*);
 Framebufctl *mkfbctl(Rectangle);
 void rmfbctl(Framebufctl*);
 
-/* render */
-void shade(Framebuf*, Scene*, Shader*);
-
 /* vertex */
 Vertex dupvertex(Vertex*);
 void lerpvertex(Vertex*, Vertex*, Vertex*, double);
diff --git a/render.c b/render.c
index d88f6c3..be19582 100644
--- a/render.c
+++ b/render.c
@@ -354,8 +354,9 @@ rasterize(SUparams *params, Triangle t)
 }
 
 static void
-shaderunit(void *arg)
+entityproc(void *arg)
 {
+	Channel *paramsc;
 	SUparams *params;
 	VSparams vsp;
 	OBJVertex *verts, *tverts, *nverts;	/* geometric, texture and normals vertices */
@@ -365,144 +366,204 @@ shaderunit(void *arg)
 	Triangle *t;				/* triangles to raster */
 	int i, nt;
 
-	params = arg;
-	vsp.su = params;
-
-	threadsetname("shader unit #%d", params->id);
+	threadsetname("entityproc");
 
+	paramsc = arg;
 	t = emalloc(sizeof(*t)*16);
-	verts = params->entity->mdl->obj->vertdata[OBJVGeometric].verts;
-	tverts = params->entity->mdl->obj->vertdata[OBJVTexture].verts;
-	nverts = params->entity->mdl->obj->vertdata[OBJVNormal].verts;
-	eb = params->entity->mdl->elems;
-	ee = eb + params->entity->mdl->nelems;
-
-	for(ep = eb; ep != ee; ep++){
-		nt = 1;	/* start with one. after clipping it might change */
-
-		idxtab = &(*ep)->indextab[OBJVGeometric];
-		t[0][0].p = Pt3(verts[idxtab->indices[0]].x,
-				verts[idxtab->indices[0]].y,
-				verts[idxtab->indices[0]].z,
-				verts[idxtab->indices[0]].w);
-		t[0][1].p = Pt3(verts[idxtab->indices[1]].x,
-				verts[idxtab->indices[1]].y,
-				verts[idxtab->indices[1]].z,
-				verts[idxtab->indices[1]].w);
-		t[0][2].p = Pt3(verts[idxtab->indices[2]].x,
-				verts[idxtab->indices[2]].y,
-				verts[idxtab->indices[2]].z,
-				verts[idxtab->indices[2]].w);
-
-		idxtab = &(*ep)->indextab[OBJVNormal];
-		if(idxtab->nindex == 3){
-			t[0][0].n = Vec3(nverts[idxtab->indices[0]].i,
-					 nverts[idxtab->indices[0]].j,
-					 nverts[idxtab->indices[0]].k);
-			t[0][0].n = normvec3(t[0][0].n);
-			t[0][1].n = Vec3(nverts[idxtab->indices[1]].i,
-					 nverts[idxtab->indices[1]].j,
-					 nverts[idxtab->indices[1]].k);
-			t[0][1].n = normvec3(t[0][1].n);
-			t[0][2].n = Vec3(nverts[idxtab->indices[2]].i,
-					 nverts[idxtab->indices[2]].j,
-					 nverts[idxtab->indices[2]].k);
-			t[0][2].n = normvec3(t[0][2].n);
-		}else{
-			/* TODO build a list of per-vertex normals earlier */
-			n = normvec3(crossvec3(subpt3(t[0][1].p, t[0][0].p), subpt3(t[0][2].p, t[0][0].p)));
-			t[0][0].n = t[0][1].n = t[0][2].n = n;
-		}
 
-		idxtab = &(*ep)->indextab[OBJVTexture];
-		if(idxtab->nindex == 3){
-			t[0][0].uv = Pt2(tverts[idxtab->indices[0]].u,
-					 tverts[idxtab->indices[0]].v, 1);
-			t[0][1].uv = Pt2(tverts[idxtab->indices[1]].u,
-					 tverts[idxtab->indices[1]].v, 1);
-			t[0][2].uv = Pt2(tverts[idxtab->indices[2]].u,
-					 tverts[idxtab->indices[2]].v, 1);
-		}else{
-			t[0][0].uv = t[0][1].uv = t[0][2].uv = Vec2(0,0);
-		}
+	while((params = recvp(paramsc)) != nil){
+		vsp.su = params;
+
+		verts = params->entity->mdl->obj->vertdata[OBJVGeometric].verts;
+		tverts = params->entity->mdl->obj->vertdata[OBJVTexture].verts;
+		nverts = params->entity->mdl->obj->vertdata[OBJVNormal].verts;
+
+		eb = params->entity->mdl->elems;
+		ee = eb + params->entity->mdl->nelems;
+
+		for(ep = eb; ep != ee; ep++){
+			nt = 1;	/* start with one. after clipping it might change */
+
+			idxtab = &(*ep)->indextab[OBJVGeometric];
+			t[0][0].p = Pt3(verts[idxtab->indices[0]].x,
+					verts[idxtab->indices[0]].y,
+					verts[idxtab->indices[0]].z,
+					verts[idxtab->indices[0]].w);
+			t[0][1].p = Pt3(verts[idxtab->indices[1]].x,
+					verts[idxtab->indices[1]].y,
+					verts[idxtab->indices[1]].z,
+					verts[idxtab->indices[1]].w);
+			t[0][2].p = Pt3(verts[idxtab->indices[2]].x,
+					verts[idxtab->indices[2]].y,
+					verts[idxtab->indices[2]].z,
+					verts[idxtab->indices[2]].w);
+
+			idxtab = &(*ep)->indextab[OBJVNormal];
+			if(idxtab->nindex == 3){
+				t[0][0].n = Vec3(nverts[idxtab->indices[0]].i,
+						 nverts[idxtab->indices[0]].j,
+						 nverts[idxtab->indices[0]].k);
+				t[0][0].n = normvec3(t[0][0].n);
+				t[0][1].n = Vec3(nverts[idxtab->indices[1]].i,
+						 nverts[idxtab->indices[1]].j,
+						 nverts[idxtab->indices[1]].k);
+				t[0][1].n = normvec3(t[0][1].n);
+				t[0][2].n = Vec3(nverts[idxtab->indices[2]].i,
+						 nverts[idxtab->indices[2]].j,
+						 nverts[idxtab->indices[2]].k);
+				t[0][2].n = normvec3(t[0][2].n);
+			}else{
+				/* TODO build a list of per-vertex normals earlier */
+				n = normvec3(crossvec3(subpt3(t[0][1].p, t[0][0].p), subpt3(t[0][2].p, t[0][0].p)));
+				t[0][0].n = t[0][1].n = t[0][2].n = n;
+			}
 
-		for(i = 0; i < 3; i++){
-			t[0][i].c = Pt3(1,1,1,1);
-			t[0][i].mtl = (*ep)->mtl;
-			t[0][i].attrs = nil;
-			t[0][i].nattrs = 0;
-		}
+			idxtab = &(*ep)->indextab[OBJVTexture];
+			if(idxtab->nindex == 3){
+				t[0][0].uv = Pt2(tverts[idxtab->indices[0]].u,
+						 tverts[idxtab->indices[0]].v, 1);
+				t[0][1].uv = Pt2(tverts[idxtab->indices[1]].u,
+						 tverts[idxtab->indices[1]].v, 1);
+				t[0][2].uv = Pt2(tverts[idxtab->indices[2]].u,
+						 tverts[idxtab->indices[2]].v, 1);
+			}else{
+				t[0][0].uv = t[0][1].uv = t[0][2].uv = Vec2(0,0);
+			}
 
-		vsp.v = &t[0][0];
-		vsp.idx = 0;
-		t[0][0].p = params->vshader(&vsp);
-		vsp.v = &t[0][1];
-		vsp.idx = 1;
-		t[0][1].p = params->vshader(&vsp);
-		vsp.v = &t[0][2];
-		vsp.idx = 2;
-		t[0][2].p = params->vshader(&vsp);
+			for(i = 0; i < 3; i++){
+				t[0][i].c = Pt3(1,1,1,1);
+				t[0][i].mtl = (*ep)->mtl;
+				t[0][i].attrs = nil;
+				t[0][i].nattrs = 0;
+			}
 
-		if(!isvisible(t[0][0].p) || !isvisible(t[0][1].p) || !isvisible(t[0][2].p))
-			nt = cliptriangle(t);
+			vsp.v = &t[0][0];
+			vsp.idx = 0;
+			t[0][0].p = params->vshader(&vsp);
+			vsp.v = &t[0][1];
+			vsp.idx = 1;
+			t[0][1].p = params->vshader(&vsp);
+			vsp.v = &t[0][2];
+			vsp.idx = 2;
+			t[0][2].p = params->vshader(&vsp);
+
+			if(!isvisible(t[0][0].p) || !isvisible(t[0][1].p) || !isvisible(t[0][2].p))
+				nt = cliptriangle(t);
+
+			while(nt--){
+				t[nt][0].p = clip2ndc(t[nt][0].p);
+				t[nt][1].p = clip2ndc(t[nt][1].p);
+				t[nt][2].p = clip2ndc(t[nt][2].p);
+
+				/* culling */
+//				if(isfacingback(t[nt]))
+//					goto skiptri;
+
+				t[nt][0].p = ndc2viewport(params->fb, t[nt][0].p);
+				t[nt][1].p = ndc2viewport(params->fb, t[nt][1].p);
+				t[nt][2].p = ndc2viewport(params->fb, t[nt][2].p);
+
+				rasterize(params, t[nt]);
+//skiptri:
+				delvattrs(&t[nt][0]);
+				delvattrs(&t[nt][1]);
+				delvattrs(&t[nt][2]);
+			}
+		}
+		sendp(params->donec, params);
+	}
+}
 
-		while(nt--){
-			t[nt][0].p = clip2ndc(t[nt][0].p);
-			t[nt][1].p = clip2ndc(t[nt][1].p);
-			t[nt][2].p = clip2ndc(t[nt][2].p);
+static void
+renderer(void *arg)
+{
+	Channel *jobc;
+	Jobqueue jobq;
+	Renderjob *job;
+	Scene *sc;
+	Entity *ent;
+	SUparams *params, *params2;
+	Channel *paramsc, *donec;
 
-			/* culling */
-//			if(isfacingback(t[nt]))
-//				goto skiptri;
+	threadsetname("renderer");
 
-			t[nt][0].p = ndc2viewport(params->fb, t[nt][0].p);
-			t[nt][1].p = ndc2viewport(params->fb, t[nt][1].p);
-			t[nt][2].p = ndc2viewport(params->fb, t[nt][2].p);
+	jobc = arg;
+	jobq.tl = jobq.hd = nil;
+	ent = nil;
+	paramsc = chancreate(sizeof(SUparams*), 8);
+	donec = chancreate(sizeof(SUparams*), 0);
 
-			rasterize(params, t[nt]);
+	proccreate(entityproc, paramsc, mainstacksize);
 
-//skiptri:
-			delvattrs(&t[nt][0]);
-			delvattrs(&t[nt][1]);
-			delvattrs(&t[nt][2]);
+	enum { JOB, PARM, DONE };
+	Alt a[] = {
+	 [JOB]	{jobc, &job, CHANRCV},
+	 [PARM]	{paramsc, &params, CHANNOP},
+	 [DONE]	{donec, &params2, CHANRCV},
+		{nil, nil, CHANEND}
+	};
+	for(;;)
+		switch(alt(a)){
+		case JOB:
+			sc = job->scene;
+			job->nrem = sc->nents;
+			job->lastid = 0;
+			job->time0 = nanosec();
+
+			if(jobq.tl == nil){
+				jobq.tl = jobq.hd = job;
+				ent = sc->ents.next;
+				a[PARM].op = CHANSND;
+				goto sendparams;
+			}else
+				jobq.tl = jobq.tl->next = job;
+			break;
+		case PARM:
+sendparams:
+			job = jobq.hd;
+			sc = job->scene;
+
+			if(ent != nil && ent != &sc->ents){
+				params = emalloc(sizeof *params);
+				memset(params, 0, sizeof *params);
+				params->fb = job->fb;
+				params->id = job->lastid++;
+				params->frag = rgb(DBlack);
+				params->donec = donec;
+				params->job = job;
+				params->entity = ent;
+				params->uni_time = job->time0;
+				params->vshader = job->shaders->vshader;
+				params->fshader = job->shaders->fshader;
+				ent = ent->next;
+			}else{
+				jobq.hd = job->next;
+				if((job = jobq.hd) != nil){
+					ent = job->scene->ents.next;
+					goto sendparams;
+				}
+
+				jobq.tl = jobq.hd;
+				a[PARM].op = CHANNOP;
+			}
+			break;
+		case DONE:
+			if(--params2->job->nrem < 1)
+				send(params2->job->donec, nil);
+
+			freememimage(params2->frag);
+			free(params2);
+			break;
 		}
-	}
-
-	free(t);
-	sendp(params->donec, params);
-	threadexits(nil);
 }
 
-void
-shade(Framebuf *fb, Scene *sc, Shader *s)
+Renderer *
+initgraphics(void)
 {
-	int i;
-	uvlong time;
-	Entity *ent;
-	SUparams *params;
-	Channel *donec;
-
-	time = nanosec();
-	donec = chancreate(sizeof(void*), 0);
-
-	/* TODO come up with an actual concurrent architecture */
-	for(i = 0, ent = sc->ents.next; i < sc->nents; i++, ent = ent->next){
-		params = emalloc(sizeof *params);
-		params->fb = fb;
-		params->id = i;
-		params->frag = rgb(DBlack);
-		params->donec = donec;
-		params->entity = ent;
-		params->uni_time = time;
-		params->vshader = s->vshader;
-		params->fshader = s->fshader;
-		proccreate(shaderunit, params, mainstacksize);
-	}
+	Renderer *r;
 
-	while(i--){
-		params = recvp(donec);
-		freememimage(params->frag);
-		free(params);
-	}
-	chanfree(donec);
+	r = emalloc(sizeof *r);
+	r->c = chancreate(sizeof(Renderjob*), 8);
+	proccreate(renderer, r->c, mainstacksize);
+	return r;
 }