Allows "merging" render targets that overlap on the Y access. Fixes #7295 (Juiced 2)

To be safe, gating this behind the related AllowLargeFBTextureOffsets,
which is also required for the effect to work.

Additionally, fixes the offset check for X offsets, which I guess is a
very small risk.

GPU/Common/FramebufferManagerCommon.cpp

@@ -266,13 +266,16 @@ VirtualFramebuffer *FramebufferManagerCommon::DoSetRenderFrameBuffer(const Frame
 	int drawing_width, drawing_height;
 	EstimateDrawingSize(params.fb_address, params.fmt, params.viewportWidth, params.viewportHeight, params.regionWidth, params.regionHeight, params.scissorWidth, params.scissorHeight, std::max(params.fb_stride, 4), drawing_width, drawing_height);
 
-	gstate_c.SetCurRTOffsetX(0);
+	gstate_c.SetCurRTOffset(0, 0);
 	bool vfbFormatChanged = false;
 
 	// Find a matching framebuffer
 	VirtualFramebuffer *vfb = nullptr;
 	for (size_t i = 0; i < vfbs_.size(); ++i) {
 		VirtualFramebuffer *v = vfbs_[i];
+
+		const u32 bpp = v->format == GE_FORMAT_8888 ? 4 : 2;
+
 		if (v->fb_address == params.fb_address) {
 			vfb = v;
 			// Update fb stride in case it changed
@@ -302,18 +305,36 @@ VirtualFramebuffer *FramebufferManagerCommon::DoSetRenderFrameBuffer(const Frame
 				vfb->height = drawing_height;
 			}
 			break;
-		} else if (v->fb_address < params.fb_address && v->fb_address + v->fb_stride * 4 > params.fb_address) {
+		} else if (v->fb_stride == params.fb_stride && v->format == params.fmt) {
-			// Possibly a render-to-offset.
+			u32 v_fb_first_line_end_ptr = v->fb_address + v->fb_stride * bpp;
-			const u32 bpp = v->format == GE_FORMAT_8888 ? 4 : 2;
+			u32 v_fb_end_ptr = v->fb_address + v->fb_stride * v->height * bpp;
-			const int x_offset = (params.fb_address - v->fb_address) / bpp;
+
-			if (v->format == params.fmt && v->fb_stride == params.fb_stride && x_offset < params.fb_stride && v->height >= drawing_height) {
+			if (params.fb_address > v->fb_address && params.fb_address < v_fb_first_line_end_ptr) {
-				WARN_LOG_REPORT_ONCE(renderoffset, HLE, "Rendering to framebuffer offset: %08x +%dx%d", v->fb_address, x_offset, 0);
+				const int x_offset = (params.fb_address - v->fb_address) / bpp;
-				vfb = v;
+				if (x_offset < params.fb_stride && v->height >= drawing_height) {
-				gstate_c.SetCurRTOffsetX(x_offset);
+					// Pretty certainly a pure render-to-X-offset.
-				vfb->width = std::max((int)vfb->width, x_offset + drawing_width);
+					WARN_LOG_REPORT_ONCE(renderoffset, HLE, "Rendering to framebuffer offset: %08x +%dx%d", v->fb_address, x_offset, 0);
-				// To prevent the newSize code from being confused.
+					vfb = v;
-				drawing_width += x_offset;
+					gstate_c.SetCurRTOffset(x_offset, 0);
-				break;
+					vfb->width = std::max((int)vfb->width, x_offset + drawing_width);
+					// To prevent the newSize code from being confused.
+					drawing_width += x_offset;
+					break;
+				}
+			} else if (params.fb_address > v->fb_address && params.fb_address < v_fb_end_ptr && PSP_CoreParameter().compat.flags().AllowLargeFBTextureOffsets) {
+				if (params.fb_address % params.fb_stride == v->fb_address % params.fb_stride) {
+					// Framebuffers are overlapping on the Y axis.
+					const int y_offset = (params.fb_address - v->fb_address) / (bpp * params.fb_stride);
+
+					vfb = v;
+					gstate_c.SetCurRTOffset(0, y_offset);
+					// To prevent the newSize code from being confused.
+					drawing_height += y_offset;
+					break;
+				}
+			} else {
+				// We ignore this match.
+				// TODO: We can allow X/Y overlaps too, but haven't seen any so safer to not.
 			}
 		}
 	}
@@ -386,7 +407,7 @@ VirtualFramebuffer *FramebufferManagerCommon::DoSetRenderFrameBuffer(const Frame
 
 		SetColorUpdated(vfb, skipDrawReason);
 
-		INFO_LOG(FRAMEBUF, "Creating FBO for %08x (z: %08x) : %i x %i x %i", vfb->fb_address, vfb->z_address, vfb->width, vfb->height, vfb->format);
+		INFO_LOG(FRAMEBUF, "Creating FBO for %08x (z: %08x) : %d x %d x %s", vfb->fb_address, vfb->z_address, vfb->width, vfb->height, GeBufferFormatToString(vfb->format));
 
 		vfb->last_frame_render = gpuStats.numFlips;
 		frameLastFramebufUsed_ = gpuStats.numFlips;

GPU/Common/GPUStateUtils.cpp

@@ -584,6 +584,7 @@ void ConvertViewportAndScissor(bool useBufferedRendering, float renderWidth, flo
 	}
 
 	renderX = gstate_c.curRTOffsetX;
+	renderY = gstate_c.curRTOffsetY;
 
 	// Scissor
 	int scissorX1 = gstate.getScissorX1();

GPU/Common/TextureCacheCommon.cpp

@@ -633,8 +633,17 @@ std::vector<AttachCandidate> TextureCacheCommon::GetFramebufferCandidates(const
 
 	if (candidates.size() > 1) {
 		bool depth = channel == FramebufferNotificationChannel::NOTIFY_FB_DEPTH;
-		WARN_LOG_REPORT_ONCE(multifbcandidate, G3D, "GetFramebufferCandidates(%s): Multiple (%d) candidate framebuffers. texaddr: %08x offset: %d (%dx%d stride %d, %s)",
+
-			depth ? "DEPTH" : "COLOR", (int)candidates.size(), entry.addr, texAddrOffset, dimWidth(entry.dim), dimHeight(entry.dim), entry.bufw, GeTextureFormatToString(entry.format));
+		std::string cands;
+		for (auto &candidate : candidates) {
+			cands += candidate.ToString() + " ";
+		}
+
+		WARN_LOG_REPORT_ONCE(multifbcandidate, G3D, "GetFramebufferCandidates(%s): Multiple (%d) candidate framebuffers. First will be chosen. texaddr: %08x offset: %d (%dx%d stride %d, %s):\n%s",
+			depth ? "DEPTH" : "COLOR", (int)candidates.size(),
+			entry.addr, texAddrOffset, dimWidth(entry.dim), dimHeight(entry.dim), entry.bufw, GeTextureFormatToString(entry.format),
+			cands.c_str()
+		);
 	}
 
 	return candidates;
@@ -1996,3 +2005,7 @@ void TextureCacheCommon::InvalidateAll(GPUInvalidationType /*unused*/) {
 void TextureCacheCommon::ClearNextFrame() {
 	clearCacheNextFrame_ = true;
 }
+
+std::string AttachCandidate::ToString() {
+	return StringFromFormat("[C:%08x/%d Z:%08x/%d X:%d Y:%d reint: %s]", this->fb->fb_address, this->fb->fb_stride, this->fb->z_address, this->fb->z_stride, this->match.xOffset, this->match.yOffset, this->match.reinterpret ? "true" : "false");
+}

GPU/Common/TextureCacheCommon.h

@@ -203,7 +203,6 @@ typedef std::map<u64, std::unique_ptr<TexCacheEntry>> TexCache;
 #undef IGNORE
 #endif
 
-// TODO: Try to get rid of IGNORE, it doesn't match what we want to do
 enum class FramebufferMatch {
 	// Valid, exact match.
 	VALID = 0,
@@ -224,6 +223,8 @@ struct AttachCandidate {
 	TextureDefinition entry;
 	VirtualFramebuffer *fb;
 	FramebufferNotificationChannel channel;
+
+	std::string ToString();
 };
 
 class FramebufferManagerCommon;

GPU/GPUState.h

@@ -606,13 +606,15 @@ struct GPUStateCache {
 	u32 curRTRenderWidth;
 	u32 curRTRenderHeight;
 
-	void SetCurRTOffsetX(int off) {
+	void SetCurRTOffset(u32 xoff, u32 yoff) {
-		if (off != curRTOffsetX) {
+		if (xoff != curRTOffsetX || yoff != curRTOffsetY) {
-			curRTOffsetX = off;
+			curRTOffsetX = xoff;
+			curRTOffsetY = yoff;
 			Dirty(DIRTY_VIEWPORTSCISSOR_STATE);
 		}
 	}
 	u32 curRTOffsetX;
+	u32 curRTOffsetY;
 
 	// Set if we are doing hardware bezier/spline.
 	SubmitType submitType;

assets/compat.ini

@@ -1252,3 +1252,8 @@ ULKS46143 = true
 ULES00981 = true
 ULES00982 = true
 LBSW10345 = true  # Some modded version found in our report logs
+
+# Juiced 2 bloom effect (see #7295)
+ULES00928 = true
+ULUS10312 = true
+ULKS46154 = true