6 files changed, 32 insertions, 23 deletions

diff --git a/src/core/clip.c b/src/core/clip.c
index 48c376d..eb42fb7 100644
--- a/src/core/clip.c
+++ b/src/core/clip.c
@@ -190,7 +190,7 @@ static bool clip_against_plane(
 	, byte* dst_data[]
 	, uint prim_vcount/*minimum number of primitive vertices*/
 ) {
-	bool varying = varying_flag & VARYING_ANY;
+	bool varying = (varying_flag & VARYING_ANY) != 0;
 	int idx = 0;
 	ClippedVert *prev, *curr;
 	uint previ, curri;
diff --git a/src/core/device.c b/src/core/device.c
index 70ae11f..6b2bb7c 100644
--- a/src/core/device.c
+++ b/src/core/device.c
@@ -22,7 +22,7 @@ typedef struct {
 } VertexAttr;
 
 static struct {
-	Mat4 matrices[3][MATRIXDEPTH];
+	Mat4 matrices[3][MATRIX_DEPTH];
 	uint matrix_top[3];
 	ssr_MatrixMode matrix_mode;
 
@@ -155,13 +155,13 @@ void ssr_loadidentity() {
 }
 
 void ssr_pushmatrix() {
-	ssr_assert(MATRIXTOP < MATRIXDEPTH - 1);
+	ssr_assert(MATRIXTOP < MATRIX_DEPTH - 1);
 	++MATRIXTOP;
 	MATRIX = MATRIXSTACK[MATRIXTOP - 1];
 }
 
 void ssr_popmatrix() {
-	MATRIXTOP = clamp(--MATRIXTOP, 0, MATRIXDEPTH - 1);
+	MATRIXTOP = clamp(--MATRIXTOP, 0, MATRIX_DEPTH - 1);
 }
 
 /*http://warmcat.org/chai/blog/?p=559*/
diff --git a/src/core/limits.h b/src/core/limits.h
index 4ef0b75..229f535 100644
--- a/src/core/limits.h
+++ b/src/core/limits.h
@@ -1,6 +1,6 @@
 #ifndef _SOFTSHADEROOM_LIMITS_H_
 #define _SOFTSHADEROOM_LIMITS_H_
 
-#define MATRIXDEPTH 32
+#define MATRIX_DEPTH 32
 
 #endif
\ No newline at end of file
diff --git a/src/core/rasterizer.c b/src/core/rasterizer.c
index d6a9fc1..d926cb0 100644
--- a/src/core/rasterizer.c
+++ b/src/core/rasterizer.c
@@ -237,15 +237,15 @@ void ssrR_line(
 ) {
 	ssr_assert(CA && CB && program && uniforms);
 
-	Vec3 SA, SB;
-	vec4_dividew(CA, &SA); ssrU_viewport(&SA, &SA);
-	vec4_dividew(CB, &SB); ssrU_viewport(&SB, &SB);
+	Vec4 SA, SB;
+	vec4_dividewnoz(CA, &SA); ssrU_viewport(&SA, &SA);
+	vec4_dividewnoz(CB, &SB); ssrU_viewport(&SB, &SB);
 
 	FragmentShader frag_shader = program->fragmentshader;
 	
 	int x0 = SA.x, y0 = SA.y;
 	int x1 = SB.x, y1 = SB.y;
-	float wA = 1 / CA->w, wB = 1 / CB->w;
+	float wA = SA.w, wB = SB.w;
 	float zA = SA.z, zB = SB.z;
 	
 	bool steep = FALSE;
@@ -262,6 +262,8 @@ void ssrR_line(
 		swapf(zA, zB);
 	}
 
+	float inv_wA = 1 / wA, inv_wB = 1 / wB;
+
 	bool depth_test = ssr_isenable(ENABLE_DEPTHTEST);
 	bool blend = ssr_isenable(ENABLE_BLEND);
 	bool write_depth = ssr_isenable(ENABLE_WRITEDEPTH);
@@ -274,15 +276,15 @@ void ssrR_line(
 
 #define discardif(condition) if(condition) continue
 
-	float t, depth;
+	float t, z, w, depth;
 	Vec2 c;
 	int x, y, px, py;
 	for (x = x0; x <= x1; ++x) {
 		t = (x - x0) / (float)(x1 - x0);
 		y = y0 + (y1 - y0)*t;
 		/*caculate center*/
-		c.x = (1 - t) * wA;
-		c.y = t * wB;
+		c.x = (1 - t) * inv_wA;
+		c.y = t * inv_wB;
 		discardif(compare(c.x+c.y, 0));
 		c.x /= (c.x + c.y);
 		t = 1 - c.x;
@@ -295,7 +297,9 @@ void ssrR_line(
 			py = y; 
 		}
 		if (depth_test) {
-			ssrS_lerpnum(t, &zA, &zB, &depth);
+			ssrS_lerpnum(t, &zA, &zB, &z);
+			ssrS_lerpnum(t, &wA, &wB, &w);
+			depth = z / w;
 			pass_depth_test = ssr_testdepth(px, py, depth);
 		}
 		if (stencil_test) {
diff --git a/src/core/shader.c b/src/core/shader.c
index 899a160..b523df7 100644
--- a/src/core/shader.c
+++ b/src/core/shader.c
@@ -267,35 +267,40 @@ Vec4* ssrS_bcpvec4(Vec3* bc, Vec4* A, Vec4* B, Vec4* C, Vec4* out) {
 	return out;
 }
 
-void ssrS_lerpnum(float t, float* A, float* B, float* out) {
+float* ssrS_lerpnum(float t, float* A, float* B, float* out) {
 	*out = lerp(*A, *B, t);
+	return out;
 }
 
-void ssrS_lerpcolor(float t, Color A, Color B, Color* out) {
+Color* ssrS_lerpcolor(float t, Color A, Color B, Color* out) {
 	*out = ssr_color(
 		lerp(COLOR_R(A), COLOR_R(B), t),
 		lerp(COLOR_G(A), COLOR_G(B), t),
 		lerp(COLOR_B(A), COLOR_B(B), t),
 		lerp(COLOR_A(A), COLOR_A(B), t)
 	);
+	return out;
 }
 
-void ssrS_lerpvec2(float t, Vec2* A, Vec2* B, Vec2* out) {
+Vec2* ssrS_lerpvec2(float t, Vec2* A, Vec2* B, Vec2* out) {
 	out->x = lerp(A->x, B->x, t);
 	out->y = lerp(A->y, B->y, t);
+	return out;
 }
 
-void ssrS_lerpvec3(float t, Vec3* A, Vec3* B, Vec3* out) {
+Vec3* ssrS_lerpvec3(float t, Vec3* A, Vec3* B, Vec3* out) {
 	out->x = lerp(A->x, B->x, t);
 	out->y = lerp(A->y, B->y, t);
 	out->z = lerp(A->z, B->z, t);
+	return out;
 }
 
-void ssrS_lerpvec4(float t, Vec4* A, Vec4* B, Vec4* out) {
+Vec4* ssrS_lerpvec4(float t, Vec4* A, Vec4* B, Vec4* out) {
 	out->x = lerp(A->x, B->x, t);
 	out->y = lerp(A->y, B->y, t);
 	out->z = lerp(A->z, B->z, t);
 	out->w = lerp(A->w, B->w, t);
+	return out;
 }
 
 static Color32 _out_color[RENDER_TARGET_COUNT];
diff --git a/src/core/shader.h b/src/core/shader.h
index 26606b8..70b4ce9 100644
--- a/src/core/shader.h
+++ b/src/core/shader.h
@@ -113,11 +113,11 @@ Vec2*  ssrS_bcpvec2(Vec3* bc, Vec2* A, Vec2* B, Vec2* C, Vec2* out);
 Vec3* ssrS_bcpvec3(Vec3* bc, Vec3* A, Vec3* B, Vec3* C, Vec3* out);
 Vec4* ssrS_bcpvec4(Vec3* bc, Vec4* A, Vec4* B, Vec4* C, Vec4* out);
 
-void ssrS_lerpcolor(float t, Color A, Color B, Color* out);
-void ssrS_lerpnum(float t, float* A, float* B, float* out);
-void ssrS_lerpvec2(float t, Vec2* A, Vec2* B, Vec2* out);
-void ssrS_lerpvec3(float t, Vec3* A, Vec3* B, Vec3* out);
-void ssrS_lerpvec4(float t, Vec4* A, Vec4* B, Vec4* out);
+Color* ssrS_lerpcolor(float t, Color A, Color B, Color* out);
+float* ssrS_lerpnum(float t, float* A, float* B, float* out);
+Vec2* ssrS_lerpvec2(float t, Vec2* A, Vec2* B, Vec2* out);
+Vec3* ssrS_lerpvec3(float t, Vec3* A, Vec3* B, Vec3* out);
+Vec4* ssrS_lerpvec4(float t, Vec4* A, Vec4* B, Vec4* out);
 
 typedef enum {
 	REGTYPE_NUM = 1,