Merge pull request #7177 from unknownbrackets/spline-minor

Use floats for color values in splines

GPU/Common/SplineCommon.cpp

@@ -18,12 +18,17 @@
 #include <string.h>
 #include <algorithm>
 
+#if defined(_M_SSE)
+#include <emmintrin.h>
+#endif
+
 #include "Core/Config.h"
 
 #include "GPU/Common/SplineCommon.h"
 #include "GPU/ge_constants.h"
 #include "GPU/GPUState.h"
 
+
 #define START_OPEN 1
 #define END_OPEN 2
 
@@ -139,8 +144,8 @@ void _SplinePatchLowQuality(u8 *&dest, u16 *indices, int &count, const SplinePat
 	const int tile_max_u = (spatch.type_u & END_OPEN) ? spatch.count_u - 1 : spatch.count_u - 2;
 	const int tile_max_v = (spatch.type_v & END_OPEN) ? spatch.count_v - 1 : spatch.count_v - 2;
 
-	float tu_width = spatch.count_u - 3;
-	float tv_height = spatch.count_v - 3;
+	float tu_width = (float)spatch.count_u - 3.0f;
+	float tv_height = (float)spatch.count_v - 3.0f;
 	tu_width /= (float)(tile_max_u - tile_min_u);
 	tv_height /= (float)(tile_max_v - tile_min_v);
 
@@ -224,8 +229,8 @@ void  _SplinePatchFullQuality(u8 *&dest, u16 *indices, int &count, const SplineP
 	// First compute all the vertices and put them in an array
 	SimpleVertex *&vertices = (SimpleVertex*&)dest;
 
-	float tu_width = spatch.count_u - 3;
-	float tv_height = spatch.count_v - 3;
+	float tu_width = (float)spatch.count_u - 3.0f;
+	float tv_height = (float)spatch.count_v - 3.0f;
 
 	bool computeNormals = gstate.isLightingEnabled();
 	for (int tile_v = 0; tile_v < patch_div_t + 1; tile_v++) {
@@ -237,25 +242,23 @@ void  _SplinePatchFullQuality(u8 *&dest, u16 *indices, int &count, const SplineP
 			if (u < 0.0f)
 				u = 0.0f;
 			SimpleVertex *vert = &vertices[tile_v * (patch_div_s + 1) + tile_u];
+			Vec4f vert_color;
 			vert->pos.SetZero();
 			if (origVertType & GE_VTYPE_NRM_MASK) {
 				vert->nrm.SetZero();
-			}
-			else {
+			} else {
 				vert->nrm.SetZero();
 				vert->nrm.z = 1.0f;
 			}
 			if (origVertType & GE_VTYPE_COL_MASK) {
-				memset(vert->color, 0, 4);
-			}
-			else {
+				vert_color.SetZero();
+			} else {
 				memcpy(vert->color, spatch.points[0]->color, 4);
 			}
 			if (origVertType & GE_VTYPE_TC_MASK) {
 				vert->uv[0] = 0.0f;
 				vert->uv[1] = 0.0f;
-			}
-			else {
+			} else {
 				vert->uv[0] = tu_width * ((float)tile_u / (float)patch_div_s);
 				vert->uv[1] = tv_height * ((float)tile_v / (float)patch_div_t);
 			}
@@ -288,11 +291,12 @@ void  _SplinePatchFullQuality(u8 *&dest, u16 *indices, int &count, const SplineP
 							vert->uv[1] += a->uv[1] * f;
 						}
 						if (origVertType & GE_VTYPE_COL_MASK) {
-							// TODO: Accumulating values in u8s is crazy. We need floats or something.
-							vert->color[0] += a->color[0] * f;
-							vert->color[1] += a->color[1] * f;
-							vert->color[2] += a->color[2] * f;
-							vert->color[3] += a->color[3] * f;
+							Vec4f a_color = Vec4f::FromRGBA(a->color_32);
+#ifdef _M_SSE
+							vert_color.vec = _mm_add_ps(vert_color.vec, _mm_mul_ps(a_color.vec, _mm_set_ps1(f)));
+#else
+							vert_color += a_color * f;
+#endif
 						}
 						if (origVertType & GE_VTYPE_NRM_MASK) {
 							vert->nrm += a->nrm * f;
@@ -303,6 +307,9 @@ void  _SplinePatchFullQuality(u8 *&dest, u16 *indices, int &count, const SplineP
 			if (origVertType & GE_VTYPE_NRM_MASK) {
 				vert->nrm.Normalize();
 			}
+			if (origVertType & GE_VTYPE_COL_MASK) {
+				vert->color_32 = vert_color.ToRGBA();
+			}
 		}
 	}
 

GPU/Common/SplineCommon.h

@@ -23,7 +23,10 @@
 // PSP compatible format so we can use the end of the pipeline in beziers etc
 struct SimpleVertex {
 	float uv[2];
-	u8 color[4];
+	union {
+		u8 color[4];
+		u32_le color_32;
+	};
 	Vec3Packedf nrm;
 	Vec3Packedf pos;
 };

GPU/Math3D.h

@@ -636,7 +636,7 @@ public:
 
 	void SetZero()
 	{
-		x=0; y=0; z=0;
+		x=0; y=0; z=0; w=0;
 	}
 
 	// Common alias: RGBA (colors)