Support render-to-clut, at least in some cases.

This is pretty much only tested with Brave Story.  See hrydgard#6754.

There may be other cases which are not handled yet.

GPU/Common/FramebufferCommon.cpp

@@ -334,6 +334,7 @@ void FramebufferManagerCommon::DoSetRenderFrameBuffer() {
 		vfb->last_frame_used = 0;
 		vfb->last_frame_attached = 0;
 		vfb->last_frame_displayed = 0;
+		vfb->last_frame_clut = 0;
 		frameLastFramebufUsed_ = gpuStats.numFlips;
 		vfbs_.push_back(vfb);
 		currentRenderVfb_ = vfb;
@@ -491,6 +492,13 @@ bool FramebufferManagerCommon::NotifyFramebufferCopy(u32 src, u32 dst, int size,
 				srcBuffer = vfb;
 				srcY = yOffset;
 				srcH = 1;
+			} else if (yOffset == 0 && yOffset < srcY) {
+				// Okay, last try - it might be a clut.
+				if (vfb->usageFlags & FB_USAGE_CLUT) {
+					srcBuffer = vfb;
+					srcY = yOffset;
+					srcH = 1;
+				}
 			}
 		}
 	}
@@ -584,7 +592,13 @@ void FramebufferManagerCommon::FindTransferFramebuffers(VirtualFramebuffer *&dst
 				// Grand Knights History copies with a mismatching stride but a full line at a time.
 				// Makes it hard to detect the wrong transfers in e.g. God of War.
 				if (width != dstStride || (byteStride * height != vfb_byteStride && byteStride * height != vfb_byteWidth)) {
-					match = false;
+					// However, some other games write cluts to framebuffers.
+					// Let's catch this and upload.  Otherwise reject the match.
+					match = (vfb->usageFlags & FB_USAGE_CLUT) != 0;
+					if (match) {
+						dstWidth = byteStride * height / vfb_bpp;
+						dstHeight = 1;
+					}
 				} else {
 					dstWidth = byteStride * height / vfb_bpp;
 					dstHeight = 1;
@@ -786,4 +800,5 @@ void FramebufferManagerCommon::UpdateFramebufUsage(VirtualFramebuffer *vfb) {
 	checkFlag(FB_USAGE_DISPLAYED_FRAMEBUFFER, vfb->last_frame_displayed);
 	checkFlag(FB_USAGE_TEXTURE, vfb->last_frame_used);
 	checkFlag(FB_USAGE_RENDERTARGET, vfb->last_frame_render);
+	checkFlag(FB_USAGE_CLUT, vfb->last_frame_clut);
 }

GPU/Common/FramebufferCommon.h

@@ -28,6 +28,7 @@ enum {
 	FB_USAGE_DISPLAYED_FRAMEBUFFER = 1,
 	FB_USAGE_RENDERTARGET = 2,
 	FB_USAGE_TEXTURE = 4,
+	FB_USAGE_CLUT = 8,
 };
 
 enum {
@@ -51,6 +52,7 @@ struct VirtualFramebuffer {
 	int last_frame_attached;
 	int last_frame_render;
 	int last_frame_displayed;
+	int last_frame_clut;
 	bool memoryUpdated;
 	bool depthUpdated;
 

GPU/Directx9/TextureCacheDX9.cpp

@@ -823,11 +823,27 @@ inline bool TextureCacheDX9::TexCacheEntry::Matches(u16 dim2, u8 format2, int ma
 }
 
 void TextureCacheDX9::LoadClut() {
-	u32 clutAddr = gstate.getClutAddress();
+	// Clear the uncached bit, etc. to match framebuffers.
+	const u32 clutAddr = gstate.getClutAddress() & 0x3FFFFFFF;
+	bool foundFramebuffer = false;
+
+	for (size_t i = 0, n = fbCache_.size(); i < n; ++i) {
+		auto framebuffer = fbCache_[i];
+		if ((framebuffer->fb_address | 0x04000000) == clutAddr) {
+			framebuffer->last_frame_clut = gpuStats.numFlips;
+			framebuffer->usageFlags |= FB_USAGE_CLUT;
+			foundFramebuffer = true;
+		}
+	}
+
 	clutTotalBytes_ = gstate.getClutLoadBytes();
 	if (Memory::IsValidAddress(clutAddr)) {
 		// It's possible for a game to (successfully) access outside valid memory.
 		u32 bytes = Memory::ValidSize(clutAddr, clutTotalBytes_);
+		if (foundFramebuffer) {
+			gpu->PerformMemoryDownload(clutAddr, bytes);
+		}
+
 #ifdef _M_SSE
 		int numBlocks = bytes / 16;
 		if (bytes == clutTotalBytes_) {
@@ -842,17 +858,17 @@ void TextureCacheDX9::LoadClut() {
 		} else {
 			Memory::MemcpyUnchecked(clutBufRaw_, clutAddr, bytes);
 			if (bytes < clutTotalBytes_) {
-				memset(clutBufRaw_ + bytes, 0x00, clutTotalBytes_ - bytes);
+				memset((u8 *)clutBufRaw_ + bytes, 0x00, clutTotalBytes_ - bytes);
 			}
 		}
 #else
 		Memory::MemcpyUnchecked(clutBufRaw_, clutAddr, bytes);
 		if (bytes < clutTotalBytes_) {
-			memset(clutBufRaw_ + bytes, 0x00, clutTotalBytes_ - bytes);
+			memset((u8 *)clutBufRaw_ + bytes, 0x00, clutTotalBytes_ - bytes);
 		}
 #endif
 	} else {
-		memset(clutBufRaw_, 0x00, clutTotalBytes_);
+		memset((u8 *)clutBufRaw_, 0x00, clutTotalBytes_);
 	}
 	// Reload the clut next time.
 	clutLastFormat_ = 0xFFFFFFFF;

GPU/GLES/TextureCache.cpp

@@ -950,11 +950,27 @@ inline bool TextureCache::TexCacheEntry::Matches(u16 dim2, u8 format2, int maxLe
 }
 
 void TextureCache::LoadClut() {
-	u32 clutAddr = gstate.getClutAddress();
+	// Clear the uncached bit, etc. to match framebuffers.
+	const u32 clutAddr = gstate.getClutAddress() & 0x3FFFFFFF;
+	bool foundFramebuffer = false;
+
+	for (size_t i = 0, n = fbCache_.size(); i < n; ++i) {
+		auto framebuffer = fbCache_[i];
+		if ((framebuffer->fb_address | 0x04000000) == clutAddr) {
+			framebuffer->last_frame_clut = gpuStats.numFlips;
+			framebuffer->usageFlags |= FB_USAGE_CLUT;
+			foundFramebuffer = true;
+		}
+	}
+
 	clutTotalBytes_ = gstate.getClutLoadBytes();
 	if (Memory::IsValidAddress(clutAddr)) {
 		// It's possible for a game to (successfully) access outside valid memory.
 		u32 bytes = Memory::ValidSize(clutAddr, clutTotalBytes_);
+		if (foundFramebuffer) {
+			gpu->PerformMemoryDownload(clutAddr, bytes);
+		}
+
 #ifdef _M_SSE
 		int numBlocks = bytes / 16;
 		if (bytes == clutTotalBytes_) {
@@ -979,7 +995,7 @@ void TextureCache::LoadClut() {
 		}
 #endif
 	} else {
-		memset(clutBufRaw_, 0x00, clutTotalBytes_);
+		memset((u8 *)clutBufRaw_, 0x00, clutTotalBytes_);
 	}
 	// Reload the clut next time.
 	clutLastFormat_ = 0xFFFFFFFF;