3 files changed, 46 insertions, 16 deletions

diff --git a/src/core/clip.c b/src/core/clip.c
index 2880448..d5de4fc 100644
--- a/src/core/clip.c
+++ b/src/core/clip.c
@@ -145,7 +145,7 @@ static bool is_inside_plane(Vec4* c, Plane plane) {
 		case NEGATIVE_Y:
 			return c->y >= -c->w;
 		case POSITIVE_Z:
-			return c->z <= +c->w;
+			return c->z <= c->w;
 		case NEGATIVE_Z:
 			return c->z >= -c->w;
 		default:
diff --git a/src/core/device.c b/src/core/device.c
index ac055c9..923a2fa 100644
--- a/src/core/device.c
+++ b/src/core/device.c
@@ -498,18 +498,36 @@ static void render_prims_triangle(uint varying_flag) {
 		c1 = &clip_coords.coords[i1];
 		c2 = &clip_coords.coords[i2];
 
-		/*back face culling*/
+		/*early back face culling*/
+		bool early_culled = FALSE;
 		if (ssr_isenable(ENABLE_BACKFACECULL)) {
 			/*cull in ndc*/
 			float w0 = 1 / c0->w, w1 = 1 / c1->w, w2 = 1 / c2->w;
-			Vec3 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*/
+			bool all_front = w0 > 0 && w1 > 0 && w2 > 0;
+			if (all_front)
+			{
+				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*/
+				}
 			}
+			//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*/
@@ -521,7 +539,7 @@ static void render_prims_triangle(uint varying_flag) {
 				reset_active_reg = FALSE;
 				ssrS_setactiveregr();
 			}
-			ssrR_triangle(c0, c1, c2, i0, i1, i2, state.program, &state.uniforms);
+			ssrR_triangle(c0, c1, c2, i0, i1, i2, state.program, &state.uniforms, early_culled);
 		}
 		else {
 			if (!reset_active_reg) {
@@ -535,7 +553,7 @@ 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, &state.uniforms);
+					ssrR_triangle(c0, c1, c2, i0, i1, i2, state.program, &state.uniforms, early_culled);
 				}
 			}
 		}
diff --git a/src/core/rasterizer.c b/src/core/rasterizer.c
index 28ae639..c5745ba 100644
--- a/src/core/rasterizer.c
+++ b/src/core/rasterizer.c
@@ -86,19 +86,31 @@ void ssrR_triangle(
 	Vec4* CA, Vec4* CB, Vec4* CC, 
 	uint IA, uint IB, uint IC, 
 	Program* program, 
-	UniformCollection* uniforms 
+	UniformCollection* uniforms, 
+	bool early_culled
 ) {
 	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;
+		}
+	}
+
 	Vec4 SA, SB, SC;
 	internal_vec4_dividewnoz(CA, &SA); ssrU_viewport(&SA, &SA);
 	internal_vec4_dividewnoz(CB, &SB); ssrU_viewport(&SB, &SB);
 	internal_vec4_dividewnoz(CC, &SC); ssrU_viewport(&SC, &SC);
 
-/*
-	puttriangle(&SA, &SB, &SC, 0xffff0000);
-	return;
-*/
+	//puttriangle(&SA, &SB, &SC, 0xffff0000);
+	//return;
 
 	Vec4 *sa = &SA, *sb = &SB, *sc = &SC, *tmp; 
 	Vec4 *v4tmp;