GS/DX: Specify shader model directly instead of via feature level

bin/resources/GameIndex.yaml

@@ -31969,6 +31969,9 @@ SLKA-25176:
   name: "펌프 잇 업 익시드"
   name-en: "Pump It Up - Exceed"
   region: "NTSC-K"
+  gsHWFixes:
+    cpuSpriteRenderBW: 4 # Fixes graphical corruption in musical notes.
+    cpuSpriteRenderLevel: 2 # Needed for above.
 SLKA-25180:
   name: "스플린터 셀 - 판도라 투머로우"
   name-en: "Tom Clancy's Splinter Cell - Pandora Tomorrow"
@@ -44039,7 +44042,7 @@ SLPM-65495:
 SLPM-65496:
   name: "ハイパーストリートファイターⅡ アニバーサリーエディション"
   name-sort: "はいぱーすとりーとふぁいたー2 あにばーさりーえでぃしょん"
-  name-en: "Hyper Street Fighter"
+  name-en: "Hyper Street Fighter II - The Anniversary Edition"
   region: "NTSC-J"
 SLPM-65497:
   name: "ワールドサッカーウイニングイレブン7 インターナショナル プレミアムパッケージ supported by アディダス"
@@ -54506,6 +54509,7 @@ SLPM-74405:
   region: "NTSC-J"
   gameFixes:
     - EETimingHack # Stops crash after intro movie.
+    - InstantDMAHack # Stops hang when entering Sword Select screen.
 SLPM-74406:
   name: "WRC ～WORLD RALLY CHAMPIONSHIP～ [PlayStation2 the Best]"
   name-sort: "WRC ～わーるど らりー ちゃんぴおんしっぷ～ [PlayStation2 the Best]"
@@ -70115,6 +70119,9 @@ SLUS-21131:
   name: "Pump It Up - Exceed"
   region: "NTSC-U"
   compat: 5
+  gsHWFixes:
+    cpuSpriteRenderBW: 4 # Fixes graphical corruption in musical notes.
+    cpuSpriteRenderLevel: 2 # Needed for above.
 SLUS-21132:
   name: "Stella Deus - The Gate of Eternity"
   region: "NTSC-U"

bin/resources/game_controller_db.txt

@@ -1538,6 +1538,8 @@ xinput,XInput Controller,a:b0,b:b1,back:b6,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,
 05000000853200000503000000010000,Nacon MGX Pro,a:b0,b:b1,back:b10,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b12,leftshoulder:b6,leftstick:b13,lefttrigger:a5,leftx:a0,lefty:a1,rightshoulder:b7,rightstick:b14,righttrigger:a4,rightx:a2,righty:a3,start:b11,x:b3,y:b4,platform:Linux,
 0300000085320000170d000011010000,Nacon Revolution 5 Pro,a:b1,b:b2,back:b8,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b12,leftshoulder:b4,leftstick:b10,lefttrigger:a3,leftx:a0,lefty:a1,rightshoulder:b5,rightstick:b11,righttrigger:a4,rightx:a2,righty:a5,start:b9,touchpad:b13,x:b0,y:b3,platform:Linux,
 0300000085320000190d000011010000,Nacon Revolution 5 Pro,a:b1,b:b2,back:b8,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b12,leftshoulder:b4,leftstick:b10,lefttrigger:a3,leftx:a0,lefty:a1,rightshoulder:b5,rightstick:b11,righttrigger:a4,rightx:a2,righty:a5,start:b9,touchpad:b13,x:b0,y:b3,platform:Linux,
+0300000085320000180d000011010000,Nacon Revolution 5 Pro,a:b1,b:b2,back:b8,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b12,leftshoulder:b4,leftstick:b10,lefttrigger:a3,leftx:a0,lefty:a1,rightshoulder:b5,rightstick:b11,righttrigger:a4,rightx:a2,righty:a5,start:b9,x:b0,y:b3,platform:Linux,
+0300000085320000160d000011010000,Nacon Revolution 5 Pro,a:b1,b:b2,back:b8,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b12,leftshoulder:b4,leftstick:b10,lefttrigger:a3,leftx:a0,lefty:a1,rightshoulder:b5,rightstick:b11,righttrigger:a4,rightx:a2,righty:a5,start:b9,touchpad:b13,x:b0,y:b3,platform:Linux,
 030000000d0f00000900000010010000,Natec Genesis P44,a:b1,b:b2,back:b8,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b12,leftshoulder:b4,leftstick:b10,lefttrigger:b6,leftx:a0,lefty:a1,rightshoulder:b5,rightstick:b11,righttrigger:b7,rightx:a2,righty:a3,start:b9,x:b0,y:b3,platform:Linux,
 030000004f1f00000800000011010000,NeoGeo PS3,a:b1,b:b2,back:b8,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b12,leftshoulder:b4,lefttrigger:b6,leftx:a0,lefty:a1,rightshoulder:b5,righttrigger:b7,start:b9,x:b0,y:b3,platform:Linux,
 0300000092120000474e000000010000,NeoGeo X Arcade Stick,a:b0,b:b1,back:b8,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,start:b9,x:b3,y:b2,platform:Linux,

pcsx2/GS/GSDrawingContext.cpp

@@ -91,8 +91,8 @@ void GSDrawingContext::UpdateScissor()
 	scissor.in = rscissor + GSVector4i::cxpr(0, 0, 1, 1);
 
 	// Fixed-point scissor min/max, used for rejecting primitives which are entirely outside.
-	// Bottom-right end point should be exclusive for rectangle intersection test to work correctly.
-	scissor.cull = rscissor.sll32<4>() + GSVector4i::cxpr(0, 0, 1, 1);
+	// Add half a pixel around the edges for upscaling and lines/points.
+	scissor.cull = rscissor.sll32<4>() + GSVector4i(-8, -8, 8, 8);
 
 	// Offset applied to vertices for culling.
 	scissor.xyof = GSVector4i::loadl(&XYOFFSET.U64).xyxy();

pcsx2/GS/GSState.cpp

@@ -5930,39 +5930,46 @@ __forceinline void GSState::VertexKick(u32 skip)
 			bbox = v0.runion(v1).runion(v2);
 		}
 
-		if (m_nativeres && (primclass == GS_TRIANGLE_CLASS || primclass == GS_SPRITE_CLASS))
+		if constexpr (primclass == GS_TRIANGLE_CLASS || primclass == GS_SPRITE_CLASS)
 		{
-			// For triangles and sprites at native res take the region strictly within the bounds
-			// omitting the bottom/right.
-			bbox = (bbox + GSVector4i(0xF, 0xF, -1, -1)) & GSVector4i(~0xF);
+			if (m_nativeres)
+			{
+				// For triangles and sprites at native res take the interior pixel centers.
+				const GSVector4i interior = (bbox + GSVector4i(0xF, 0xF, -1, -1)) & GSVector4i(~0xF);
+				bbox = interior + GSVector4i(0, 0, 1, 1); // +1 to bottom/right so empty test works correctly.
+			}
+			else
+			{
+				// For upscaling, remove bottom/right subtexels.
+				bbox -= ((bbox & GSVector4i(0xF)) == GSVector4i(0)) & GSVector4i(0, 0, 1, 1);
+			}
+
+			// For AA1 triangles and lines, expand the bounds by 1 pixel on all sides.
+			// Note: redundant check for the AA1 flag to avoid calling a function if not needed.
+			if (PRIM->AA1 && IsCoverageAlphaSupported())
+			{
+				bbox += GSVector4i(-0x10, -0x10, 0x10, 0x10);
+			}
 		}
 
-		// For AA1 triangles and lines, expand the bounds by 1 pixel on all sides.
-		// Note: redundant check for the AA1 flag to avoid calling a function if not needed.
-		if ((primclass == GS_TRIANGLE_CLASS || primclass == GS_LINE_CLASS) && PRIM->AA1 && IsCoverageAlphaSupported())
+		// Do scissor test.
+		const GSVector4i bbox_ex = bbox + GSVector4i(0, 0, 1, 1); // Exclusive coords for the scissor test.
+		const GSVector4i& scissor = m_context->scissor.cull;
+		u32 test = static_cast<u32>(!bbox_ex.rintersects(scissor));
+
+		// Test for empty bbox.
+		if constexpr (primclass == GS_TRIANGLE_CLASS || primclass == GS_SPRITE_CLASS)
 		{
-			bbox += GSVector4i(-0x10, -0x10, 0x10, 0x10);
+			test |= static_cast<u32>(bbox.rempty());
 		}
 
-		// Make the bottom-right endpoints exclusive for rectangle intersection to work correctly.
-		bbox += GSVector4i::cxpr(0, 0, 1, 1);
-
-		// Do scissor test and empty bbox test.
-		GSVector4i test = GSVector4i(bbox.rintersects(m_context->scissor.cull)) == GSVector4i::cxpr(0);
-
 		// Test for degenerate triangle.
-		if (primclass == GS_TRIANGLE_CLASS)
+		if constexpr (primclass == GS_TRIANGLE_CLASS)
 		{
-			test |= v0.eq64(v1) | v1.eq64(v2) | v0.eq64(v2);
+			test |= static_cast<u32>(v0.eq(v1)) | static_cast<u32>(v1.eq(v2)) | static_cast<u32>(v0.eq(v2));
 		}
 
-#ifndef ARCH_ARM64
-		// We only care about the xy passing the skip test.
-		skip |= test.mask();
-#else
-		// mask() is slow on ARM, so just pull the bits out instead, thankfully we only care about the first 4 bytes.
-		skip |= (static_cast<u64>(test.extract64<0>()) & UINT64_C(0x8080808080808080)) != 0;
-#endif
+		skip |= test;
 	}
 
 	if (skip != 0)

pcsx2/GS/Renderers/DX11/D3D.cpp

@@ -466,28 +466,52 @@ GSRendererType D3D::GetPreferredRenderer()
 	}
 }
 
-wil::com_ptr_nothrow<ID3DBlob> D3D::CompileShader(D3D::ShaderType type, D3D_FEATURE_LEVEL feature_level, bool debug,
+const char* D3D::ShaderModelToCacheString(D3D::ShaderModel shader_model)
+{
+	switch (shader_model)
+	{
+		case ShaderModel::SM40:
+			return "sm40";
+		case ShaderModel::SM41:
+			return "sm41";
+		case ShaderModel::SM50:
+			return "sm50";
+		case ShaderModel::SM51:
+			return "sm51";
+		default:
+			return "unk";
+	}
+}
+
+wil::com_ptr_nothrow<ID3DBlob> D3D::CompileShader(D3D::ShaderType type, D3D::ShaderModel shader_model, bool debug,
 	const std::string_view code, const D3D_SHADER_MACRO* macros /* = nullptr */,
 	const char* entry_point /* = "main" */)
 {
 	const char* target;
-	switch (feature_level)
+	switch (shader_model)
 	{
-		case D3D_FEATURE_LEVEL_10_0:
+		case ShaderModel::SM40:
 		{
 			static constexpr std::array<const char*, 4> targets = {{"vs_4_0", "ps_4_0", "cs_4_0"}};
 			target = targets[static_cast<int>(type)];
 		}
 		break;
 
-		case D3D_FEATURE_LEVEL_11_0:
+		case ShaderModel::SM41:
+		{
+			static constexpr std::array<const char*, 4> targets = {{"vs_4_1", "ps_4_1", "cs_4_1"}};
+			target = targets[static_cast<int>(type)];
+		}
+		break;
+
+		case ShaderModel::SM50:
 		{
 			static constexpr std::array<const char*, 4> targets = {{"vs_5_0", "ps_5_0", "cs_5_0"}};
 			target = targets[static_cast<int>(type)];
 		}
 		break;
 
-		case D3D_FEATURE_LEVEL_11_1:
+		case ShaderModel::SM51:
 		default:
 		{
 			static constexpr std::array<const char*, 4> targets = {{"vs_5_1", "ps_5_1", "cs_5_1"}};

pcsx2/GS/Renderers/DX11/D3D.h

@@ -63,6 +63,16 @@ namespace D3D
 		Compute
 	};
 
-	wil::com_ptr_nothrow<ID3DBlob> CompileShader(ShaderType type, D3D_FEATURE_LEVEL feature_level, bool debug,
+	enum class ShaderModel
+	{
+		SM40 = 0x40, // DX11 FL 10_0
+		SM41 = 0x41, // DX11 FL 10_1
+		SM50 = 0x50, // DX11 FL 11_0
+		SM51 = 0x51, // DX12
+	};
+
+	const char* ShaderModelToCacheString(ShaderModel shader_model);
+
+	wil::com_ptr_nothrow<ID3DBlob> CompileShader(ShaderType type, ShaderModel shader_model, bool debug,
 		const std::string_view code, const D3D_SHADER_MACRO* macros = nullptr, const char* entry_point = "main");
 }; // namespace D3D

pcsx2/GS/Renderers/DX11/D3D11ShaderCache.cpp

@@ -56,12 +56,27 @@ bool D3D11ShaderCache::CacheIndexKey::operator!=(const CacheIndexKey& key) const
 
 bool D3D11ShaderCache::Open(D3D_FEATURE_LEVEL feature_level, bool debug)
 {
-	m_feature_level = feature_level;
+	switch (feature_level)
+	{
+		case D3D_FEATURE_LEVEL_10_0:
+			m_shader_model = D3D::ShaderModel::SM40;
+			break;
+		case D3D_FEATURE_LEVEL_10_1:
+			m_shader_model = D3D::ShaderModel::SM41;
+			break;
+		case D3D_FEATURE_LEVEL_11_0:
+			m_shader_model = D3D::ShaderModel::SM50;
+			break;
+		default:
+			pxAssert(false);
+			return false;
+	}
+
 	m_debug = debug;
 
 	if (!GSConfig.DisableShaderCache)
 	{
-		const std::string base_filename = GetCacheBaseFileName(feature_level, debug);
+		const std::string base_filename = GetCacheBaseFileName(m_shader_model, debug);
 		const std::string index_filename = base_filename + ".idx";
 		const std::string blob_filename = base_filename + ".bin";
 
@@ -198,19 +213,10 @@ bool D3D11ShaderCache::ReadExisting(const std::string& index_filename, const std
 	return true;
 }
 
-std::string D3D11ShaderCache::GetCacheBaseFileName(D3D_FEATURE_LEVEL feature_level, bool debug)
+std::string D3D11ShaderCache::GetCacheBaseFileName(D3D::ShaderModel shader_model, bool debug)
 {
 	std::string base_filename = "d3d_shaders_";
-
-	switch (feature_level)
-	{
-		case D3D_FEATURE_LEVEL_11_0:
-			base_filename += "sm50";
-			break;
-		default:
-			base_filename += "unk";
-			break;
-	}
+	base_filename += D3D::ShaderModelToCacheString(shader_model);
 
 	if (debug)
 		base_filename += "_debug";
@@ -400,7 +406,7 @@ wil::com_ptr_nothrow<ID3DBlob> D3D11ShaderCache::CompileAndAddShaderBlob(const C
 	const std::string_view shader_code, const D3D_SHADER_MACRO* macros, const char* entry_point)
 {
 	wil::com_ptr_nothrow<ID3DBlob> blob =
-		D3D::CompileShader(key.shader_type, m_feature_level, m_debug, shader_code, macros, entry_point);
+		D3D::CompileShader(key.shader_type, m_shader_model, m_debug, shader_code, macros, entry_point);
 	if (!blob)
 		return {};
 

pcsx2/GS/Renderers/DX11/D3D11ShaderCache.h

@@ -16,7 +16,6 @@ public:
 	D3D11ShaderCache();
 	~D3D11ShaderCache();
 
-	D3D_FEATURE_LEVEL GetFeatureLevel() const { return m_feature_level; }
 	bool UsingDebugShaders() const { return m_debug; }
 
 	bool Open(D3D_FEATURE_LEVEL feature_level, bool debug);
@@ -74,7 +73,7 @@ private:
 
 	using CacheIndex = std::unordered_map<CacheIndexKey, CacheIndexData, CacheIndexEntryHasher>;
 
-	static std::string GetCacheBaseFileName(D3D_FEATURE_LEVEL feature_level, bool debug);
+	static std::string GetCacheBaseFileName(D3D::ShaderModel shader_mode, bool debug);
 	static CacheIndexKey GetCacheKey(D3D::ShaderType type, const std::string_view shader_code,
 		const D3D_SHADER_MACRO* macros, const char* entry_point);
 
@@ -89,6 +88,6 @@ private:
 
 	CacheIndex m_index;
 
-	D3D_FEATURE_LEVEL m_feature_level = D3D_FEATURE_LEVEL_11_0;
+	D3D::ShaderModel m_shader_model = D3D::ShaderModel::SM50;
 	bool m_debug = false;
 };

pcsx2/GS/Renderers/DX12/D3D12ShaderCache.cpp

@@ -2,7 +2,6 @@
 // SPDX-License-Identifier: GPL-3.0+
 
 #include "GS/Renderers/DX12/D3D12ShaderCache.h"
-#include "GS/Renderers/DX11/D3D.h"
 #include "GS/GS.h"
 
 #include "Config.h"
@@ -54,16 +53,18 @@ bool D3D12ShaderCache::CacheIndexKey::operator!=(const CacheIndexKey& key) const
 			source_length != key.source_length);
 }
 
-bool D3D12ShaderCache::Open(D3D_FEATURE_LEVEL feature_level, bool debug)
+bool D3D12ShaderCache::Open(D3D::ShaderModel shader_model, bool debug)
 {
-	m_feature_level = feature_level;
+	// Only support SM5.1 for now, which is the minimum for D3D12.
+	pxAssert(shader_model >= D3D::ShaderModel::SM51);
+	m_shader_model = shader_model;
 	m_debug = debug;
 
 	bool result = true;
 
 	if (!GSConfig.DisableShaderCache)
 	{
-		const std::string base_shader_filename = GetCacheBaseFileName("shaders", feature_level, debug);
+		const std::string base_shader_filename = GetCacheBaseFileName("shaders", m_shader_model, debug);
 		const std::string shader_index_filename = base_shader_filename + ".idx";
 		const std::string shader_blob_filename = base_shader_filename + ".bin";
 
@@ -75,7 +76,7 @@ bool D3D12ShaderCache::Open(D3D_FEATURE_LEVEL feature_level, bool debug)
 
 		if (result)
 		{
-			const std::string base_pipelines_filename = GetCacheBaseFileName("pipelines", feature_level, debug);
+			const std::string base_pipelines_filename = GetCacheBaseFileName("pipelines", m_shader_model, debug);
 			const std::string pipelines_index_filename = base_pipelines_filename + ".idx";
 			const std::string pipelines_blob_filename = base_pipelines_filename + ".bin";
 
@@ -133,7 +134,7 @@ void D3D12ShaderCache::InvalidatePipelineCache()
 	if (GSConfig.DisableShaderCache)
 		return;
 
-	const std::string base_pipelines_filename = GetCacheBaseFileName("pipelines", m_feature_level, m_debug);
+	const std::string base_pipelines_filename = GetCacheBaseFileName("pipelines", m_shader_model, m_debug);
 	const std::string pipelines_index_filename = base_pipelines_filename + ".idx";
 	const std::string pipelines_blob_filename = base_pipelines_filename + ".bin";
 	CreateNew(pipelines_index_filename, pipelines_blob_filename, m_pipeline_index_file, m_pipeline_blob_file);
@@ -253,21 +254,12 @@ bool D3D12ShaderCache::ReadExisting(const std::string& index_filename, const std
 	return true;
 }
 
-std::string D3D12ShaderCache::GetCacheBaseFileName(const std::string_view type, D3D_FEATURE_LEVEL feature_level, bool debug)
+std::string D3D12ShaderCache::GetCacheBaseFileName(const std::string_view type, D3D::ShaderModel shader_model, bool debug)
 {
 	std::string base_filename = "d3d12_";
 	base_filename += type;
 	base_filename += "_";
-
-	switch (feature_level)
-	{
-		case D3D_FEATURE_LEVEL_11_0:
-			base_filename += "sm50";
-			break;
-		default:
-			base_filename += "unk";
-			break;
-	}
+	base_filename += D3D::ShaderModelToCacheString(shader_model);
 
 	if (debug)
 		base_filename += "_debug";
@@ -497,15 +489,15 @@ D3D12ShaderCache::ComPtr<ID3DBlob> D3D12ShaderCache::CompileAndAddShaderBlob(
 	{
 		case EntryType::VertexShader:
 			blob =
-				D3D::CompileShader(D3D::ShaderType::Vertex, m_feature_level, m_debug, shader_code, macros, entry_point);
+				D3D::CompileShader(D3D::ShaderType::Vertex, m_shader_model, m_debug, shader_code, macros, entry_point);
 			break;
 		case EntryType::PixelShader:
 			blob =
-				D3D::CompileShader(D3D::ShaderType::Pixel, m_feature_level, m_debug, shader_code, macros, entry_point);
+				D3D::CompileShader(D3D::ShaderType::Pixel, m_shader_model, m_debug, shader_code, macros, entry_point);
 			break;
 		case EntryType::ComputeShader:
-			blob = D3D::CompileShader(
-				D3D::ShaderType::Compute, m_feature_level, m_debug, shader_code, macros, entry_point);
+			blob =
+				D3D::CompileShader(D3D::ShaderType::Compute, m_shader_model, m_debug, shader_code, macros, entry_point);
 			break;
 		default:
 			break;

pcsx2/GS/Renderers/DX12/D3D12ShaderCache.h

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: GPL-3.0+
 
 #pragma once
+#include "GS/Renderers/DX11/D3D.h"
 
 #include "common/Pcsx2Defs.h"
 #include "common/HashCombine.h"
@@ -32,10 +33,9 @@ public:
 	D3D12ShaderCache();
 	~D3D12ShaderCache();
 
-	__fi D3D_FEATURE_LEVEL GetFeatureLevel() const { return m_feature_level; }
 	__fi bool UsingDebugShaders() const { return m_debug; }
 
-	bool Open(D3D_FEATURE_LEVEL feature_level, bool debug);
+	bool Open(D3D::ShaderModel shader_model, bool debug);
 	void Close();
 
 	__fi ComPtr<ID3DBlob> GetVertexShader(
@@ -95,7 +95,7 @@ private:
 
 	using CacheIndex = std::unordered_map<CacheIndexKey, CacheIndexData, CacheIndexEntryHasher>;
 
-	static std::string GetCacheBaseFileName(const std::string_view type, D3D_FEATURE_LEVEL feature_level, bool debug);
+	static std::string GetCacheBaseFileName(const std::string_view type, D3D::ShaderModel shader_model, bool debug);
 	static CacheIndexKey GetShaderCacheKey(
 		EntryType type, const std::string_view shader_code, const D3D_SHADER_MACRO* macros, const char* entry_point);
 	static CacheIndexKey GetPipelineCacheKey(const D3D12_GRAPHICS_PIPELINE_STATE_DESC& gpdesc);
@@ -123,6 +123,6 @@ private:
 	std::FILE* m_pipeline_blob_file = nullptr;
 	CacheIndex m_pipeline_index;
 
-	D3D_FEATURE_LEVEL m_feature_level = D3D_FEATURE_LEVEL_11_0;
+	D3D::ShaderModel m_shader_model = D3D::ShaderModel::SM51;
 	bool m_debug = false;
 };

pcsx2/GS/Renderers/DX12/GSDevice12.cpp

@@ -892,7 +892,7 @@ bool GSDevice12::Create(GSVSyncMode vsync_mode, bool allow_present_throttle)
 		m_tfx_source = std::move(*shader);
 	}
 
-	if (!m_shader_cache.Open(m_feature_level, GSConfig.UseDebugDevice))
+	if (!m_shader_cache.Open(D3D::ShaderModel::SM51, GSConfig.UseDebugDevice))
 		Console.Warning("D3D12: Shader cache failed to open.");
 
 	if (!CreateRootSignatures())

pcsx2/GS/Renderers/HW/GSRendererHW.cpp

@@ -2995,7 +2995,7 @@ void GSRendererHW::Draw()
 	// I hate that I have to do this, but some games (like Pac-Man World Rally) troll us by causing a flush with degenerate triangles, so we don't have all available information about the next draw.
 	// So we have to check when the next draw happens if our frame has changed or if it's become recursive.
 	const bool has_colclip_texture = g_gs_device->GetColorClipTexture() != nullptr;
-	if (!no_rt && has_colclip_texture && (m_conf.colclip_frame.FBP != m_cached_ctx.FRAME.FBP || m_conf.colclip_frame.Block() == m_cached_ctx.TEX0.TBP0))
+	if (!no_rt && has_colclip_texture && (m_conf.colclip_frame.FBP != m_cached_ctx.FRAME.FBP || (PRIM->TME && m_conf.colclip_frame.Block() == m_cached_ctx.TEX0.TBP0)))
 	{
 		GIFRegTEX0 FRAME;
 		FRAME.TBP0 = m_conf.colclip_frame.Block();

pcsx2/GS/Renderers/HW/GSTextureCache.cpp

@@ -2261,20 +2261,34 @@ GSTextureCache::Source* GSTextureCache::LookupSource(const bool is_color, const
 			}
 		}
 
-		if (src->m_from_target && src->m_target_direct && src->m_region.HasEither())
+		if (src->m_from_target && src->m_target_direct)
 		{
-			if (src->m_from_target->m_TEX0.TBP0 == src->m_TEX0.TBP0)
+			GSVector4i req_rect = r;
+
+			// The read area might be offset but the start of the texture is at the beginning of the space.
+			req_rect.x = region.HasX() ? region.GetMinX() : 0;
+			req_rect.y = region.HasY() ? region.GetMinY() : 0;
+
+			const bool dirty_overlap = !src->m_from_target->m_dirty.GetTotalRect(src->m_from_target->m_TEX0, src->m_from_target->m_unscaled_size).rintersect(req_rect).rempty();
+
+			if (dirty_overlap)
+				src->m_from_target->Update();
+
+			if (src->m_region.HasEither())
 			{
-				src->m_region.bits = 0;
-				src->m_region.SetX(0, region.HasX() ? region.GetMaxX() : (1 << TEX0.TW));
-				src->m_region.SetY(0, region.HasY() ? region.GetMaxY() : (1 << TEX0.TH));
-			}
-			else if (src->m_TEX0.TBP0 > src->m_from_target->m_TEX0.TBP0)
-			{
-				GSVector4i dst_offset = TranslateAlignedRectByPage(src->m_from_target, src->m_TEX0.TBP0, src->m_TEX0.PSM, src->m_TEX0.TBW, GSVector4i(0, 0, 1, 1), false);
-				src->m_region.bits = 0;
-				src->m_region.SetX(dst_offset.x, dst_offset.x + (region.HasX() ? std::min(region.GetMaxX(), (1 << TEX0.TW)) : (1 << TEX0.TW)));
-				src->m_region.SetY(dst_offset.y, dst_offset.y + (region.HasY() ? std::min(region.GetMaxY(), (1 << TEX0.TH)) : (1 << TEX0.TH)));
+				if (src->m_from_target->m_TEX0.TBP0 == src->m_TEX0.TBP0)
+				{
+					src->m_region.bits = 0;
+					src->m_region.SetX(0, region.HasX() ? region.GetMaxX() : (1 << TEX0.TW));
+					src->m_region.SetY(0, region.HasY() ? region.GetMaxY() : (1 << TEX0.TH));
+				}
+				else if (src->m_TEX0.TBP0 > src->m_from_target->m_TEX0.TBP0)
+				{
+					GSVector4i dst_offset = TranslateAlignedRectByPage(src->m_from_target, src->m_TEX0.TBP0, src->m_TEX0.PSM, src->m_TEX0.TBW, GSVector4i(0, 0, 1, 1), false);
+					src->m_region.bits = 0;
+					src->m_region.SetX(dst_offset.x, dst_offset.x + (region.HasX() ? std::min(region.GetMaxX(), (1 << TEX0.TW)) : (1 << TEX0.TW)));
+					src->m_region.SetY(dst_offset.y, dst_offset.y + (region.HasY() ? std::min(region.GetMaxY(), (1 << TEX0.TH)) : (1 << TEX0.TH)));
+				}
 			}
 		}