From df94914a63afdfe89cef870caa7d92aca4a57a5a Mon Sep 17 00:00:00 2001
From: Henrik Rydgard <hrydgard@gmail.com>
Date: Sun, 8 Nov 2015 22:50:16 +0100
Subject: [PATCH] D3D: No need to convert blendcolor to float and back

Fix bug with constantAlpha

Remove misleading comment, thanks unknown
---
 GPU/Common/GPUStateUtils.cpp     | 17 ++++++++---------
 GPU/Common/GPUStateUtils.h       |  4 ++--
 GPU/Directx9/StateMappingDX9.cpp | 15 +++------------
 GPU/Directx9/helper/dx_state.h   |  7 +++++++
 4 files changed, 20 insertions(+), 23 deletions(-)

diff --git a/GPU/Common/GPUStateUtils.cpp b/GPU/Common/GPUStateUtils.cpp
index 151f97a567..d5fedcd25b 100644
--- a/GPU/Common/GPUStateUtils.cpp
+++ b/GPU/Common/GPUStateUtils.cpp
@@ -459,7 +459,6 @@ void ConvertViewportAndScissor(bool useBufferedRendering, float renderWidth, flo
 	float offsetY = gstate.getOffsetY();
 
 	if (throughmode) {
-		// No viewport transform here. Let's experiment with using region.
 		out.viewportX = renderX + displayOffsetX;
 		out.viewportY = renderY + displayOffsetY;
 		out.viewportW = curRTWidth * renderWidthFactor;
@@ -800,7 +799,7 @@ void ConvertBlendState(GenericBlendState &blendState) {
 		blendState.applyShaderBlending = true;
 		blendState.resetShaderBlending = false;
 		blendState.replaceAlphaWithStencil = replaceAlphaWithStencil;
-		break;  // Surely this should be return??
+		break;
 
 	case REPLACE_BLEND_PRE_SRC:
 	case REPLACE_BLEND_PRE_SRC_2X_ALPHA:
@@ -827,22 +826,22 @@ void ConvertBlendState(GenericBlendState &blendState) {
 	if (blendFuncB > GE_DSTBLEND_FIXB)
 		blendFuncB = GE_DSTBLEND_FIXB;
 
-	float constantAlpha = 1.0f;
+	int constantAlpha = 255;
 	BlendFactor constantAlphaGL = BlendFactor::ONE;
 	if (gstate.isStencilTestEnabled() && replaceAlphaWithStencil == REPLACE_ALPHA_NO) {
 		switch (ReplaceAlphaWithStencilType()) {
 		case STENCIL_VALUE_UNIFORM:
-			constantAlpha = (float)gstate.getStencilTestRef() * (1.0f / 255.0f);
+			constantAlpha = gstate.getStencilTestRef();
 			break;
 
 		case STENCIL_VALUE_INCR_4:
 		case STENCIL_VALUE_DECR_4:
-			constantAlpha = 1.0f / 15.0f;
+			constantAlpha = 16;
 			break;
 
 		case STENCIL_VALUE_INCR_8:
 		case STENCIL_VALUE_DECR_8:
-			constantAlpha = 1.0f / 255.0f;
+			constantAlpha = 1;
 			break;
 
 		default:
@@ -850,9 +849,9 @@ void ConvertBlendState(GenericBlendState &blendState) {
 		}
 
 		// Otherwise it will stay GL_ONE.
-		if (constantAlpha <= 0.0f) {
+		if (constantAlpha <= 0) {
 			constantAlphaGL = BlendFactor::ZERO;
-		} else if (constantAlpha < 1.0f) {
+		} else if (constantAlpha < 255) {
 			constantAlphaGL = BlendFactor::CONSTANT_ALPHA;
 		}
 	}
@@ -945,7 +944,7 @@ void ConvertBlendState(GenericBlendState &blendState) {
 		}
 	}
 
-	// Some Android devices (especially Mali, it seems) composite badly if there's alpha in the backbuffer.
+	// Some Android devices (especially old Mali, it seems) composite badly if there's alpha in the backbuffer.
 	// So in non-buffered rendering, we will simply consider the dest alpha to be zero in blending equations.
 #ifdef ANDROID
 	if (g_Config.iRenderingMode == FB_NON_BUFFERED_MODE) {
diff --git a/GPU/Common/GPUStateUtils.h b/GPU/Common/GPUStateUtils.h
index 047fef6387..58eda73d96 100644
--- a/GPU/Common/GPUStateUtils.h
+++ b/GPU/Common/GPUStateUtils.h
@@ -126,11 +126,11 @@ struct GenericBlendState {
 		eqAlpha = eqA;
 	}
 	void setBlendColor(uint32_t color, uint8_t alpha) {
-		blendColor = color | (alpha << 24);
+		blendColor = color | ((uint32_t)alpha << 24);
 		useBlendColor = true;
 	}
 	void defaultBlendColor(uint8_t alpha) {
-		blendColor = 0xFFFFFF | (alpha << 24);
+		blendColor = 0xFFFFFF | ((uint32_t)alpha << 24);
 		useBlendColor = true;
 	}
 };
diff --git a/GPU/Directx9/StateMappingDX9.cpp b/GPU/Directx9/StateMappingDX9.cpp
index df901223e0..d55d1e245f 100644
--- a/GPU/Directx9/StateMappingDX9.cpp
+++ b/GPU/Directx9/StateMappingDX9.cpp
@@ -166,30 +166,21 @@ void TransformDrawEngineDX9::ApplyDrawState(int prim) {
 			shaderManager_->DirtyUniform(DIRTY_SHADERBLEND);
 		}
 		if (blendState.useBlendColor) {
-			uint32_t color = blendState.blendColor;
-			const float col[4] = {
-				(float)((color & 0xFF) >> 0) * (1.0f / 255.0f),
-				(float)((color & 0xFF00) >> 8) * (1.0f / 255.0f),
-				(float)((color & 0xFF0000) >> 16) * (1.0f / 255.0f),
-				(float)((color & 0xFF000000) >> 24) * (1.0f / 255.0f),
-			};
-			dxstate.blendColor.set(col);
+			dxstate.blendColor.setDWORD(blendState.blendColor);
 		}
 	} else {
 		dxstate.blend.disable();
-		dxstate.blendSeparate.disable();
 	}
 
 	// Set Dither
 	if (gstate.isDitherEnabled()) {
 		dxstate.dither.enable();
-		dxstate.dither.set(true);
-	} else
+	} else {
 		dxstate.dither.disable();
+	}
 
 	// Set ColorMask/Stencil/Depth
 	if (gstate.isModeClear()) {
-
 		// Set Cull 
 		dxstate.cullMode.set(false, false);
 		
diff --git a/GPU/Directx9/helper/dx_state.h b/GPU/Directx9/helper/dx_state.h
index 3a74f669eb..04a2477228 100644
--- a/GPU/Directx9/helper/dx_state.h
+++ b/GPU/Directx9/helper/dx_state.h
@@ -269,6 +269,13 @@ private:
 				restore();
 			}
 		}
+		void setDWORD(DWORD newc) {
+			newc = ((newc >> 8) & 0xff) | (newc & 0xff00ff00) | ((newc << 16) & 0xff0000);  // ARGB -> ABGR fix
+			if (c != newc) {
+				c = newc;
+				restore();
+			}
+		}
 		void force(const float v[4]) {
 			DWORD old = c;
 			set(v);