*early back face culling

4 files changed, 80 insertions, 50 deletions

diff --git a/src/core/device.c b/src/core/device.c
index ce13c72..5ff39ef 100644
--- a/src/core/device.c
+++ b/src/core/device.c
@@ -459,14 +459,28 @@ void ssrU_blend(
 	out->a = src->a + dst->a;
 }
 
-void ssr_bindvertices(Vertex* verts, int vert_count, uint* indices, int prim_count) {
-	ssr_assert(verts && indices);
+void ssr_bindvertices(Vertex* verts, int vert_count) {
+	ssr_assert(verts);
 	state.verts = verts;
 	state.vert_count = vert_count;
+}
+
+void ssr_unbindvertices() {
+	state.verts = NULL; 
+	state.vert_count = 0;  
+}
+
+void ssr_bindindices(uint* indices, int prim_count) {
+	ssr_assert(indices);
 	state.indices = indices;
 	state.prim_count = prim_count;
 }
 
+void ssr_unbindindices() {
+	state.indices = NULL;
+	state.prim_count = 0;
+}
+
 void ssr_useprogram(Program* program) {
 	ssr_assert(program);
 	state.program = program;
@@ -485,11 +499,45 @@ static struct {
 
 extern ClippedBuffer clip_buffer; /*clipping result*/
 
+static bool is_back_face(Vec4* c0, Vec4* c1, Vec4*c2) {
+	/*cull in ndc*/
+	float w0 = 1 / c0->w, w1 = 1 / c1->w, w2 = 1 / c2->w;
+	bool all_front = w0 > 0 && w1 > 0 && w2 > 0;
+	if (all_front)
+	{
+		Vec2 ab, ac;
+		ab.x = c1->x * w1 - c0->x * w0;
+		ab.y = c1->y * w1 - c0->y * w0;
+		ac.x = c2->x * w2 - c0->x * w0;
+		ac.y = c2->y * w2 - c0->y * w0;
+		if (ab.x * ac.y - ab.y * ac.x <= 0) {
+			return TRUE; 
+		}
+	}
+	return FALSE;
+	/*OpenGL algrithm*/
+	/*
+	float w0 = max(1 / c0->w, 0), w1 = 1 / c1->w, w2 = 1 / c2->w;
+	Vec2 a = { c0->x * w0, c0->y * w0 };
+	Vec2 b = { c1->x * w1, c1->y * w1 };
+	Vec2 c = { c2->x * w2, c2->y * w2 };
+	float signed_area = 0;
+	signed_area += a.x * b.y - a.y * b.x;
+	signed_area += b.x * c.y - b.y * c.x;
+	signed_area += c.x * a.y - c.y * a.x;
+	if (signed_area <= 0) {
+		return TRUE;
+	}
+	return FALSE;
+	*/
+}
+
 static void render_prims_triangle(uint varying_flag) {
 	uint i0, i1, i2;
 	Vec4 *c0, *c1, *c2;
 	bool reset_active_reg = FALSE;
 	bool clipped = FALSE;
+	bool is_enable_backface_cull = ssr_isenable(ENABLE_BACKFACECULL);
 	for (int i = 0; i < state.prim_count; ++i) {
 		i0 = state.indices[i * 3];
 		i1 = state.indices[i * 3 + 1];
@@ -501,34 +549,17 @@ static void render_prims_triangle(uint varying_flag) {
 
 		/*early back face culling*/
 		bool early_culled = FALSE;
-		if (ssr_isenable(ENABLE_BACKFACECULL)) {
-			/*cull in ndc*/
+		if (is_enable_backface_cull) {
 			float w0 = 1 / c0->w, w1 = 1 / c1->w, w2 = 1 / c2->w;
-			bool all_front = w0 > 0 && w1 > 0 && w2 > 0;
-			if (all_front)
-			{
+			if (w0 <= 0 && w1 <= 0 && w2 <= 0) {
+				early_culled = TRUE; 
+				continue;
+			}
+			else if (w0 > 0 && w1 > 0 && w2 > 0) {
 				early_culled = TRUE;
-				Vec2 ab, ac;
-				ab.x = c1->x * w1 - c0->x * w0;
-				ab.y = c1->y * w1 - c0->y * w0;
-				ac.x = c2->x * w2 - c0->x * w0;
-				ac.y = c2->y * w2 - c0->y * w0;
-				if (ab.x * ac.y - ab.y * ac.x <= 0) {
-					continue; /*cull*/
-				}
+				if (is_back_face(c0, c1, c2))
+					continue;
 			}
-			//OpenGL algrithm:
-			//float w0 = max(1 / c0->w, 0), w1 = 1 / c1->w, w2 = 1 / c2->w;
-			//Vec2 a = { c0->x * w0, c0->y * w0 };
-			//Vec2 b = { c1->x * w1, c1->y * w1 };
-			//Vec2 c = { c2->x * w2, c2->y * w2 };
-			//float signed_area = 0;
-			//signed_area += a.x * b.y - a.y * b.x;
-			//signed_area += b.x * c.y - b.y * c.x;
-			//signed_area += c.x * a.y - c.y * a.x;
-			//if (signed_area <= 0) {
-			//	continue; /*cull*/
-			//}
 		}
 
 		/*clipping*/
@@ -540,7 +571,12 @@ static void render_prims_triangle(uint varying_flag) {
 				reset_active_reg = FALSE;
 				ssrS_setactiveregr();
 			}
-			ssrR_triangle(c0, c1, c2, i0, i1, i2, state.program, early_culled);
+			/*late back face culling*/
+			if (!early_culled && is_enable_backface_cull) {
+				if(is_back_face(c0, c1, c2))
+					continue;
+			}	
+			ssrR_triangle(c0, c1, c2, i0, i1, i2, state.program);
 		}
 		else {
 			if (!reset_active_reg) {
@@ -554,7 +590,12 @@ static void render_prims_triangle(uint varying_flag) {
 					vt2 = &clip_buffer.vertices[i + 1];
 					c0 = &vt0->clip_coord; c1 = &vt1->clip_coord; c2 = &vt2->clip_coord;
 					i0 = vt0->index;     i1 = vt1->index;     i2 = vt2->index;
-					ssrR_triangle(c0, c1, c2, i0, i1, i2, state.program, early_culled);
+					/*late back face culling*/
+					if (!early_culled && is_enable_backface_cull) {
+						if (is_back_face(c0, c1, c2))
+							continue;
+					}
+					ssrR_triangle(c0, c1, c2, i0, i1, i2, state.program);
 				}
 			}
 		}
diff --git a/src/core/device.h b/src/core/device.h
index 79ffd4b..6d0b61d 100644
--- a/src/core/device.h
+++ b/src/core/device.h
@@ -118,9 +118,12 @@ void ssr_disable(uint mask);
 bool ssr_isenable(uint mask);
 
 /* 绑定顶点数据 */
-void ssr_bindvertices(Vertex* verts, int vert_count, uint* indices, int prim_count);
+void ssr_bindvertices(Vertex* verts, int vert_count);
 void ssr_unbindvertices();
 
+void ssr_bindindices(uint* indices, int prim_count);
+void ssr_unbindindices();
+
 void ssr_useprogram(Program* program);
 void ssr_unuseprogram();
 
diff --git a/src/core/rasterizer.c b/src/core/rasterizer.c
index 55d4ad3..9276cbd 100644
--- a/src/core/rasterizer.c
+++ b/src/core/rasterizer.c
@@ -44,7 +44,7 @@ float ssrR_area(Vec2* v1, Vec2* v2, Vec2* v3) {
 	return area;
 }
 
-/*from https://github.com/ssloy/tinyrenderer */
+/*https://github.com/ssloy/tinyrenderer */
 bool ssrR_barycentric(Vec2* A, Vec2* B, Vec2* C, Vec2* p, Vec3* out) {
 	ssr_assert(A && B && C && p && out);
 	Vec3 s[2], u;
@@ -87,24 +87,10 @@ extern UniformCollection* g_uniforms	;
 void ssrR_triangle( 
 	Vec4* CA, Vec4* CB, Vec4* CC, 
 	uint IA, uint IB, uint IC, 
-	Program* program, 
-	bool early_culled
+	Program* program
 ) {
 	ssr_assert(CA && CB && CC && program);
 
-	/*late back face culling*/
-	if (!early_culled && ssr_isenable(ENABLE_BACKFACECULL)) {
-		float w0 = 1 / CA->w, w1 = 1 / CB->w, w2 = 1 / CC->w;
-		Vec2 ab, ac;
-		ab.x = CB->x * w1 - CA->x * w0;
-		ab.y = CB->y * w1 - CA->y * w0;
-		ac.x = CC->x * w2 - CA->x * w0;
-		ac.y = CC->y * w2 - CA->y * w0;
-		if (ab.x * ac.y - ab.y * ac.x <= 0) {
-			return;
-		}
-	}
-
 	UniformCollection* uniforms = g_uniforms;
 
 	Vec4 SA, SB, SC;
@@ -164,7 +150,7 @@ void ssrR_triangle(
 		for (p.x = from; order * (p.x - to) <= 0; p.x += order) {	          \
 			/*calculate barycentric coordinate*/															\
 			s[0].z = sa->x - p.x;																		          \
-			internal_vec3_cross(&s[0], &s[1], &u);													            \
+			internal_vec3_cross(&s[0], &s[1], &u);													  \
 			discardif(compare(u.z, 0));														            \
 			u.z = 1.f / u.z;																			            \
 			bc.x = 1.f - (u.x + u.y) * u.z;											              \
@@ -173,7 +159,7 @@ void ssrR_triangle(
 			discardif(bc.x < 0 || bc.y < 0 || bc.z < 0);				              \
 			/*perspective correction*/																	      \
 			bc.x *= CAw; bc.y *= CBw; bc.z *= CCw;							              \
-			internal_vec3_scale(&bc, 1.f / (bc.x + bc.y + bc.z), &bc);                 \
+			internal_vec3_scale(&bc, 1.f / (bc.x + bc.y + bc.z), &bc);        \
 			/*early depth testing*/																			      \
 			if(depth_test){											                              \
 				depth = bc.x*sa->z + bc.y*sb->z + bc.z*sc->z;										\
@@ -189,7 +175,7 @@ void ssrR_triangle(
 			/*interpolate varying variables*/															    \
 			ssrS_solveregsbcp(&bc, IA, IB, IC);                               \
 			/*enter fragment shader*/																			    \
-			discard = !frag_shader(out_color[0]);                   \
+			discard = !frag_shader(out_color[0]);                             \
 			discardif(discard);																								\
 			/*put point*/																										  \
 			ssr_blendandputpoint(p.x, p.y, blend);													  \
diff --git a/src/core/rasterizer.h b/src/core/rasterizer.h
index fce8376..c7ccc3c 100644
--- a/src/core/rasterizer.h
+++ b/src/core/rasterizer.h
@@ -13,7 +13,7 @@ bool ssrR_barycentric(Vec2* A, Vec2* B, Vec2* C, Vec2* p, Vec3* out); /*计算重心
 void ssrR_center(Vec2* A, Vec2* B, Vec2* C, Vec2* out); /*获得重心*/
 bool ssrR_ispointintriangle(Vec2* A, Vec2* B, Vec2* C, Vec2* p);
 
-void ssrR_triangle(Vec4* CA, Vec4* CB, Vec4* CC, uint IA, uint IB, uint IC, Program* program, bool early_called);
+void ssrR_triangle(Vec4* CA, Vec4* CB, Vec4* CC, uint IA, uint IB, uint IC, Program* program);
 void ssrR_line(Vec4* CA, Vec4* CB, uint IA, uint IB, Program* program);
 void ssrR_point(Vec4* CA, uint IA, Program* program);