GS-HW: Expand target on EE write and improve overlap check

pcsx2/GS/GSState.cpp

@@ -1764,7 +1764,7 @@ void GSState::FlushWrite()
 	r.top = m_env.TRXPOS.DSAY;
 	r.right = r.left + m_env.TRXREG.RRW;
 	r.bottom = r.top + m_env.TRXREG.RRH;
-
+	ExpandTarget(m_env.BITBLTBUF, r);
 	InvalidateVideoMem(m_env.BITBLTBUF, r);
 
 	const GSLocalMemory::writeImage wi = GSLocalMemory::m_psm[m_env.BITBLTBUF.DPSM].wi;
@@ -2026,7 +2026,7 @@ void GSState::Write(const u8* mem, int len)
 		r.top = m_env.TRXPOS.DSAY;
 		r.right = r.left + m_env.TRXREG.RRW;
 		r.bottom = r.top + m_env.TRXREG.RRH;
-
+		ExpandTarget(m_env.BITBLTBUF, r);
 		InvalidateVideoMem(blit, r);
 
 		(m_mem.*psm.wi)(m_tr.x, m_tr.y, mem, m_tr.total, blit, m_env.TRXPOS, m_env.TRXREG);

pcsx2/GS/GSState.h

@@ -373,6 +373,7 @@ class GSState : public GSAlignedClass<32>
 	virtual void PurgePool() = 0;
 	virtual void InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r) {}
 	virtual void InvalidateLocalMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r, bool clut = false) {}
+	virtual void ExpandTarget(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r) {}
 
 	virtual void Move();
 

pcsx2/GS/Renderers/HW/GSRendererHW.cpp

@@ -853,6 +853,11 @@ GSVector2i GSRendererHW::GetTargetSize(GSVector2i* unscaled_size)
 		static_cast<int>(static_cast<float>(height) * GSConfig.UpscaleMultiplier));
 }
 
+void GSRendererHW::ExpandTarget(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r)
+{
+	m_tc->ExpandTarget(BITBLTBUF, r);
+}
+
 void GSRendererHW::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r)
 {
 	// printf("[%d] InvalidateVideoMem %d,%d - %d,%d %05x (%d)\n", (int)m_perfmon.GetFrame(), r.left, r.top, r.right, r.bottom, (int)BITBLTBUF.DBP, (int)BITBLTBUF.DPSM);

pcsx2/GS/Renderers/HW/GSRendererHW.h

@@ -198,6 +198,7 @@ class GSRendererHW : public GSRenderer
 
 	GSTexture* GetOutput(int i, int& y_offset) override;
 	GSTexture* GetFeedbackOutput() override;
+	void ExpandTarget(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r);
 	void InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r) override;
 	void InvalidateLocalMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r, bool clut = false) override;
 	void Move() override;

pcsx2/GS/Renderers/HW/GSTextureCache.cpp

@@ -700,6 +700,43 @@ void GSTextureCache::ScaleTargetForDisplay(Target* t, const GIFRegTEX0& dispfb,
 	GetTargetHeight(t->m_TEX0.TBP0, t->m_TEX0.TBW, t->m_TEX0.PSM, static_cast<u32>(needed_height));
 }
 
+// Expands targets where the write from the EE overlaps the edge of a render target and uses the same base pointer.
+void GSTextureCache::ExpandTarget(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r)
+{
+	GIFRegTEX0 TEX0;
+	TEX0.TBP0 = BITBLTBUF.DBP;
+	TEX0.TBW = BITBLTBUF.DBW;
+	TEX0.PSM = BITBLTBUF.DPSM;
+	Target* dst = nullptr;
+	auto& list = m_dst[RenderTarget];
+
+	for (auto i = list.begin(); i != list.end(); ++i)
+	{
+		Target* t = *i;
+
+		if (TEX0.TBP0 == t->m_TEX0.TBP0 && t->Overlaps(TEX0.TBP0, TEX0.TBW, TEX0.PSM, r))
+		{
+			list.MoveFront(i.Index());
+
+			dst = t;
+			break;
+		}
+	}
+
+	if (dst)
+	{
+		const GSVector2i rect_scaled = GSVector2i(r.z * g_gs_renderer->GetUpscaleMultiplier(), r.w * g_gs_renderer->GetUpscaleMultiplier());
+		const int upsc_width = std::max(rect_scaled.x, dst->m_texture->GetWidth());
+		const int upsc_height = std::max(rect_scaled.y, dst->m_texture->GetHeight());
+		if (dst->m_texture->GetWidth() < upsc_width || dst->m_texture->GetHeight() < upsc_height)
+		{
+			dst->ResizeTexture(upsc_width, upsc_height);
+			dst->m_dirty.push_back(GSDirtyRect(r, TEX0.PSM, TEX0.TBW));
+			GetTargetHeight(TEX0.TBP0, TEX0.TBW, TEX0.PSM, r.w);
+			dst->Update();
+		}
+	}
+}
 // Goal: Depth And Target at the same address is not possible. On GS it is
 // the same memory but not on the Dx/GL. Therefore a write to the Depth/Target
 // must invalidate the Target/Depth respectively
@@ -2283,7 +2320,8 @@ bool GSTextureCache::Surface::Overlaps(u32 bp, u32 bw, u32 psm, const GSVector4i
 {
 	// Valid only for color formats.
 	const u32 end_block = GSLocalMemory::m_psm[psm].info.bn(rect.z - 1, rect.w - 1, bp, bw);
-	const bool overlap = GSTextureCache::CheckOverlap(m_TEX0.TBP0, m_end_block, bp, end_block);
+	const u32 start_block = GSLocalMemory::m_psm[psm].info.bn(rect.x, rect.y, bp, bw);
+	const bool overlap = GSTextureCache::CheckOverlap(m_TEX0.TBP0, m_end_block, start_block, end_block);
 	return overlap;
 }
 

pcsx2/GS/Renderers/HW/GSTextureCache.h

@@ -345,6 +345,7 @@ class GSTextureCache
 
 	u32 GetTargetHeight(u32 fbp, u32 fbw, u32 psm, u32 min_height);
 
+	void ExpandTarget(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r);
 	void InvalidateVideoMemType(int type, u32 bp);
 	void InvalidateVideoMemSubTarget(GSTextureCache::Target* rt);
 	void InvalidateVideoMem(const GSOffset& off, const GSVector4i& r, bool target = true);

pcsx2/GS/Renderers/SW/GSRendererSW.cpp

@@ -616,6 +616,8 @@ void GSRendererSW::Sync(int reason)
 	g_perfmon.Put(GSPerfMon::Fillrate, pixels);
 }
 
+void  GSRendererSW::ExpandTarget(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r) {}
+
 void GSRendererSW::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r)
 {
 	if (LOG)

pcsx2/GS/Renderers/SW/GSRendererSW.h

@@ -75,6 +75,7 @@ class GSRendererSW final : public GSRenderer
 	void Draw() override;
 	void Queue(GSRingHeap::SharedPtr<GSRasterizerData>& item);
 	void Sync(int reason);
+	void ExpandTarget(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r);
 	void InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r) override;
 	void InvalidateLocalMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r, bool clut = false) override;