Merge pull request #21151 from hrydgard/brave-story-performance-hack

Brave Story: Hack to make the bloom effect run much more efficiently

Core/Compatibility.cpp

@@ -155,6 +155,7 @@ void Compatibility::CheckSettings(IniFile &iniFile, const std::string &gameID) {
 	CheckSetting(iniFile, gameID, "IgnoreEnqueue", &flags_.IgnoreEnqueue);
 	CheckSetting(iniFile, gameID, "MsgDialogAutoStatus", &flags_.MsgDialogAutoStatus);
 	CheckSetting(iniFile, gameID, "NullPageValid", &flags_.NullPageValid);
+	CheckSetting(iniFile, gameID, "DetectDestBlendSquared", &flags_.DetectDestBlendSquared);
 }
 
 void Compatibility::CheckVRSettings(IniFile &iniFile, const std::string &gameID) {

Core/Compatibility.h

@@ -118,6 +118,7 @@ struct CompatFlags {
 	bool IgnoreEnqueue;
 	bool MsgDialogAutoStatus;
 	bool NullPageValid;
+	bool DetectDestBlendSquared;
 };
 
 struct VRCompat {

Core/MemMap.cpp

@@ -149,7 +149,7 @@ static bool Memory_TryBase(MemMapSetupFlags flags) {
 		}
 	}
 
-	int i;  // the final values is used in the next loop
+	int i;  // the final value of i is used in the next loop
 	for (i = 0; i < ARRAY_SIZE(views); i++) {
 		const MemoryView &view = views[i];
 		if (view.size == 0)

GPU/Common/FramebufferManagerCommon.cpp

@@ -1273,6 +1273,21 @@ bool FramebufferManagerCommon::BindFramebufferAsColorTexture(int stage, VirtualF
 			return true;
 		}
 
+		// There's a special case we can handle here, where the game is texturing from the same pixels being read, in order to
+		// implement a DST*DST blending function, which the PSP can't do. However on the PC we can absolutely do this!
+		// TODO: Add more checks here.
+		if (PSP_CoreParameter().compat.flags().DetectDestBlendSquared &&
+			gstate.isAlphaBlendEnabled() && gstate.getBlendEq() == GE_BLENDMODE_MUL_AND_ADD && gstate.getBlendFuncA() == GE_SRCBLEND_DSTCOLOR && gstate.getBlendFuncB() == GE_DSTBLEND_FIXB && gstate.getFixB() == 0x0 &&
+			gstate.getMaterialAmbientRGBA() == 0xFFFFFFFF) {
+			// This is the pure DST*DST case, the SRC color is ignored.
+			// Used by Brave Story - New Traveller. This assumes that texture coordinates are set to match the framebuffer pixels - and to
+			// be able to make that assumption reasonably safely we use a compat flag to restrict it to that game.
+			// We can just override the blend mode. Let's set a state variable.
+			// We also just leave the last texture bound, ideally we should bind a placeholder here.
+			gstate_c.dstSquared = true;
+			return true;
+		}
+
 		Draw::Framebuffer *renderCopy = GetTempFBO(TempFBO::COPY, framebuffer->renderWidth, framebuffer->renderHeight);
 		if (renderCopy) {
 			VirtualFramebuffer copyInfo = *framebuffer;

GPU/Common/GPUStateUtils.cpp

@@ -1154,6 +1154,19 @@ static void ConvertBlendState(GenericBlendState &blendState, FBReadSetting useFB
 	//
 	// If we can't apply blending, we make a copy of the framebuffer and do it manually.
 
+	if (gstate_c.dstSquared && useFBRead != FBReadSetting::Forced) {
+		blendState.blendEnabled = true;
+		blendState.applyFramebufferRead = false;
+		blendState.dirtyShaderBlendFixValues = false;
+		blendState.useBlendColor = false;
+		blendState.replaceBlend = REPLACE_BLEND_NO;
+		blendState.simulateLogicOpType = SimulateLogicOpShaderTypeIfNeeded();
+		blendState.replaceAlphaWithStencil = REPLACE_ALPHA_NO;
+		blendState.setEquation(BlendEq::ADD, BlendEq::ADD);
+		blendState.setFactors(BlendFactor::ZERO, BlendFactor::DST_COLOR, BlendFactor::ZERO, BlendFactor::ONE);
+		return;
+	}
+
 	blendState.applyFramebufferRead = false;
 	blendState.dirtyShaderBlendFixValues = false;
 	blendState.useBlendColor = false;

GPU/GPUState.h

@@ -656,6 +656,9 @@ public:
 	// We detect this case and go into a special drawing mode.
 	bool blueToAlpha;
 
+	// DST squared, used in Brave Story
+	bool dstSquared;
+
 	// U/V is 1:1 to pixels. Can influence texture sampling.
 	bool pixelMapped;
 

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -470,6 +470,7 @@ void DrawEngineVulkan::Flush() {
 		if (result.action == SW_DRAW_INDEXED) {
 			if (textureNeedsApply) {
 				gstate_c.pixelMapped = result.pixelMapped;
+				gstate_c.dstSquared = false;
 				textureCache_->ApplyTexture();
 				gstate_c.pixelMapped = false;
 				textureCache_->GetVulkanHandles(imageView, sampler);
@@ -477,6 +478,9 @@ void DrawEngineVulkan::Flush() {
 					imageView = (VkImageView)draw_->GetNativeObject(gstate_c.textureIsArray ? Draw::NativeObject::NULL_IMAGEVIEW_ARRAY : Draw::NativeObject::NULL_IMAGEVIEW);
 				if (sampler == VK_NULL_HANDLE)
 					sampler = nullSampler_;
+				if (gstate_c.dstSquared) {
+					gstate_c.Dirty(DIRTY_BLEND_STATE);
+				}
 			}
 			if (!lastPipeline_ || gstate_c.IsDirty(DIRTY_BLEND_STATE | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_RASTER_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_VERTEXSHADER_STATE | DIRTY_FRAGMENTSHADER_STATE | DIRTY_GEOMETRYSHADER_STATE) || prim != lastPrim_) {
 				if (prim != lastPrim_ || gstate_c.IsDirty(DIRTY_BLEND_STATE | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_RASTER_STATE | DIRTY_DEPTHSTENCIL_STATE)) {

assets/compat.ini

@@ -2106,3 +2106,10 @@ UCES01473 = true
 NPJG90095 = true
 NPEG90035 = true
 NPUG70125 = true
+
+[DetectDestBlendSquared]
+# Fixes the problem in Brave Story where we had to copy the framebuffer to blend it with itself.
+# We can just do that blend natively on PC. See #19820
+ULUS10279 = true
+UCJS10024 = true
+UCAS40096 = true