Improve SSE usage in software transform.

It's actually already pretty decent (unlike the softgpu), but there were a
few places it could use a bit of help.  Speeds up things with hardware
transform off, or areas that need to use software transform.

GPU/Common/VertexDecoderCommon.h

@@ -67,8 +67,14 @@ struct TransformedVertex
 {
 	float x, y, z, fog;     // in case of morph, preblend during decode
 	float u; float v; float w;   // scaled by uscale, vscale, if there
-	u8 color0[4];   // prelit
+	union {
-	u8 color1[4];   // prelit
+		u8 color0[4];   // prelit
+		u32 color0_32;
+	};
+	union {
+		u8 color1[4];   // prelit
+		u32 color1_32;
+	};
 };
 
 void GetIndexBounds(const void *inds, int count, u32 vertType, u16 *indexLowerBound, u16 *indexUpperBound);

GPU/GLES/SoftwareTransform.cpp

@@ -318,8 +318,8 @@ void TransformDrawEngine::SoftwareTransformAndDraw(
 		reader.Goto(index);
 
 		float v[3] = {0, 0, 0};
-		float c0[4] = {1, 1, 1, 1};
+		Vec4f c0 = Vec4f(1, 1, 1, 1);
-		float c1[4] = {0, 0, 0, 0};
+		Vec4f c1 = Vec4f(0, 0, 0, 0);
 		float uv[3] = {0, 0, 1};
 		float fogCoef = 1.0f;
 
@@ -327,15 +327,10 @@ void TransformDrawEngine::SoftwareTransformAndDraw(
 			// Do not touch the coordinates or the colors. No lighting.
 			reader.ReadPos(v);
 			if (reader.hasColor0()) {
-				reader.ReadColor0(c0);
+				reader.ReadColor0(&c0.x);
-				for (int j = 0; j < 4; j++) {
+				// c1 is already 0.
-					c1[j] = 0.0f;
-				}
 			} else {
-				c0[0] = gstate.getMaterialAmbientR() / 255.f;
+				c0 = Vec4f::FromRGBA(gstate.getMaterialAmbientRGBA());
-				c0[1] = gstate.getMaterialAmbientG() / 255.f;
-				c0[2] = gstate.getMaterialAmbientB() / 255.f;
-				c0[3] = gstate.getMaterialAmbientA() / 255.f;
 			}
 
 			if (reader.hasUV()) {
@@ -389,18 +384,15 @@ void TransformDrawEngine::SoftwareTransformAndDraw(
 			}
 
 			// Perform lighting here if enabled. don't need to check through, it's checked above.
-			float unlitColor[4] = {1, 1, 1, 1};
+			Vec4f unlitColor = Vec4f(1, 1, 1, 1);
 			if (reader.hasColor0()) {
-				reader.ReadColor0(unlitColor);
+				reader.ReadColor0(&unlitColor.x);
 			} else {
-				unlitColor[0] = gstate.getMaterialAmbientR() / 255.f;
+				unlitColor = Vec4f::FromRGBA(gstate.getMaterialAmbientRGBA());
-				unlitColor[1] = gstate.getMaterialAmbientG() / 255.f;
-				unlitColor[2] = gstate.getMaterialAmbientB() / 255.f;
-				unlitColor[3] = gstate.getMaterialAmbientA() / 255.f;
 			}
 			float litColor0[4];
 			float litColor1[4];
-			lighter.Light(litColor0, litColor1, unlitColor, out, normal);
+			lighter.Light(litColor0, litColor1, unlitColor.AsArray(), out, normal);
 
 			if (gstate.isLightingEnabled()) {
 				// Don't ignore gstate.lmode - we should send two colors in that case
@@ -424,15 +416,10 @@ void TransformDrawEngine::SoftwareTransformAndDraw(
 						c0[j] = unlitColor[j];
 					}
 				} else {
-					c0[0] = gstate.getMaterialAmbientR() / 255.f;
+					c0 = Vec4f::FromRGBA(gstate.getMaterialAmbientRGBA());
-					c0[1] = gstate.getMaterialAmbientG() / 255.f;
-					c0[2] = gstate.getMaterialAmbientB() / 255.f;
-					c0[3] = gstate.getMaterialAmbientA() / 255.f;
 				}
 				if (lmode) {
-					for (int j = 0; j < 4; j++) {
+					// c1 is already 0.
-						c1[j] = 0.0f;
-					}
 				}
 			}
 
@@ -528,12 +515,8 @@ void TransformDrawEngine::SoftwareTransformAndDraw(
 		if (gstate_c.flipTexture) {
 			transformed[index].v = 1.0f - transformed[index].v;
 		}
-		for (int i = 0; i < 4; i++) {
+		transformed[index].color0_32 = c0.ToRGBA();
-			transformed[index].color0[i] = c0[i] * 255.0f;
+		transformed[index].color1_32 = c1.ToRGBA();
-		}
-		for (int i = 0; i < 3; i++) {
-			transformed[index].color1[i] = c1[i] * 255.0f;
-		}
 	}
 
 	// Here's the best opportunity to try to detect rectangles used to clear the screen, and

GPU/GPUState.h

@@ -331,6 +331,7 @@ struct GPUgstate
 	unsigned int getMaterialAmbientG() const { return (materialambient>>8)&0xFF; }
 	unsigned int getMaterialAmbientB() const { return (materialambient>>16)&0xFF; }
 	unsigned int getMaterialAmbientA() const { return materialalpha&0xFF; }
+	unsigned int getMaterialAmbientRGBA() const { return (materialambient & 0x00FFFFFF) | (materialalpha << 24); }
 	unsigned int getMaterialDiffuseR() const { return materialdiffuse&0xFF; }
 	unsigned int getMaterialDiffuseG() const { return (materialdiffuse>>8)&0xFF; }
 	unsigned int getMaterialDiffuseB() const { return (materialdiffuse>>16)&0xFF; }