CI: Use separate job for building Windows deps

.github/workflows/release_cut_new.yml

@@ -120,10 +120,16 @@ jobs:
     secrets: inherit
 
   # Windows
+  build_windows_deps:
+    if: github.repository == 'PCSX2/pcsx2'
+    name: "Windows Dependencies"
+    uses: ./.github/workflows/windows_deps_build.yml
+
   build_windows_qt:
     if: github.repository == 'PCSX2/pcsx2'
     needs:
       - cut-release
+      - build_windows_deps
     name: "Windows"
     uses: ./.github/workflows/windows_build_qt.yml
     with:

.github/workflows/windows_build_matrix.yml

@@ -9,6 +9,10 @@ on:
       - master
 
 jobs:
+  build_qt_deps:
+    name: "Dependencies"
+    uses: ./.github/workflows/windows_deps_build.yml
+
   # MSBUILD
   lint_vs_proj_files:
     name: Lint VS Project Files
@@ -21,7 +25,7 @@ jobs:
         run: .github\workflows\scripts\windows\validate-vs-filters.ps1
 
   build_qt_sse4:
-    needs: lint_vs_proj_files
+    needs: [build_qt_deps, lint_vs_proj_files]
     name: "SSE4"
     if: github.repository != 'PCSX2/pcsx2' || github.event_name == 'pull_request'
     uses: ./.github/workflows/windows_build_qt.yml
@@ -33,7 +37,7 @@ jobs:
     secrets: inherit
 
   build_qt_avx2:
-    needs: lint_vs_proj_files
+    needs: [build_qt_deps, lint_vs_proj_files]
     name: "AVX2"
     if: github.repository != 'PCSX2/pcsx2' || github.event_name == 'pull_request'
     uses: ./.github/workflows/windows_build_qt.yml
@@ -44,6 +48,7 @@ jobs:
     secrets: inherit
 
   build_qt_cmake:
+    needs: build_qt_deps
     name: "CMake"
     if: github.repository != 'PCSX2/pcsx2' || github.event_name == 'pull_request'
     uses: ./.github/workflows/windows_build_qt.yml
@@ -55,7 +60,7 @@ jobs:
     secrets: inherit
 
   build_qt_clang_sse4:
-    needs: lint_vs_proj_files
+    needs: [build_qt_deps, lint_vs_proj_files]
     name: "SSE4"
     if: github.repository != 'PCSX2/pcsx2' || github.event_name == 'pull_request'
     uses: ./.github/workflows/windows_build_qt.yml
@@ -67,7 +72,7 @@ jobs:
     secrets: inherit
 
   build_qt_clang_avx2:
-    needs: lint_vs_proj_files
+    needs: [build_qt_deps, lint_vs_proj_files]
     name: "AVX2"
     if: github.repository != 'PCSX2/pcsx2' || github.event_name == 'pull_request'
     uses: ./.github/workflows/windows_build_qt.yml
@@ -78,6 +83,7 @@ jobs:
     secrets: inherit
 
   build_qt_cmake_clang:
+    needs: build_qt_deps
     name: "CMake"
     if: github.repository != 'PCSX2/pcsx2' || github.event_name == 'pull_request'
     uses: ./.github/workflows/windows_build_qt.yml

.github/workflows/windows_build_qt.yml

@@ -50,7 +50,7 @@ jobs:
     name: ${{ inputs.jobName }}
     runs-on: ${{ inputs.os }}
     # Set some sort of timeout in the event of run-away builds.  We are limited on concurrent jobs so, get rid of them.
-    timeout-minutes: 90
+    timeout-minutes: 30
     env:
       POWERSHELL_TELEMETRY_OPTOUT: 1
 
@@ -124,15 +124,10 @@ jobs:
         id: cache-deps
         uses: actions/cache@v6
         with:
+          fail-on-cache-miss: true
           path: deps
           key: ${{ inputs.os }} ${{ inputs.platform }} deps ${{ hashFiles('.github/workflows/scripts/windows/build-dependencies.bat', '.github/workflows/scripts/common/*.patch') }}
 
-      - name: Build Dependencies
-        if: steps.cache-deps.outputs.cache-hit != 'true'
-        env:
-          DEBUG: 0
-          BUILD_FFMPEG: 1
-        run: .github/workflows/scripts/windows/build-dependencies.bat
 
       - name: Generate CMake
         if: inputs.configuration == 'CMake'

.github/workflows/windows_deps_build.yml

@@ -0,0 +1,41 @@
+name: Windows Build Steps - Deps
+
+on:
+  workflow_call:
+    inputs:
+      os:
+        required: false
+        type: string
+        default: windows-2025-vs2026
+      platform:
+        required: false
+        type: string
+        default: x64
+
+jobs:
+  build_windows_deps:
+    name: "Cache Dependencies"
+    runs-on: ${{ inputs.os }}
+    # Set some sort of timeout in the event of run-away builds.  We are limited on concurrent jobs so, get rid of them.
+    timeout-minutes: 90
+    env:
+      POWERSHELL_TELEMETRY_OPTOUT: 1
+
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v7
+
+      - name: Check Dependencies
+        id: cache-deps
+        uses: actions/cache@v6
+        with:
+          lookup-only: true
+          path: deps
+          key: ${{ inputs.os }} ${{ inputs.platform }} deps ${{ hashFiles('.github/workflows/scripts/windows/build-dependencies.bat', '.github/workflows/scripts/common/*.patch') }}
+
+      - name: Build Dependencies
+        if: steps.cache-deps.outputs.cache-hit != 'true'
+        env:
+          DEBUG: 0
+          BUILD_FFMPEG: 1
+        run: .github/workflows/scripts/windows/build-dependencies.bat

bin/resources/GameIndex.yaml

@@ -7336,6 +7336,44 @@ SCKA-20027:
     halfPixelOffset: 2 # Aligns post effects.
     nativeScaling: 1 # Fixes post effects.
     textureInsideRT: 1 # Fixes post shuffles.
+  patches:
+    A3643EB1:
+      content: |-
+        comment=Rearranging COP2 ops to avoid macflag bad stuff.
+        // Solves door problems, possibly other issues later on.
+        patch=0,EE,0010bc88,word,48468800
+        patch=0,EE,0010bc8c,word,4BEC682C
+        patch=0,EE,0010bc94,word,4B8D617D
+        patch=0,EE,0010bc98,word,4A0002FF
+        patch=0,EE,0012a358,word,48468800
+        patch=0,EE,0012a35c,word,4BEC682C
+        patch=0,EE,0012a364,word,4B8D617D
+        patch=0,EE,0012a368,word,4A0002FF
+        patch=0,EE,0012a680,word,48468800
+        patch=0,EE,0012a684,word,4BEC682C
+        patch=0,EE,0012a68c,word,4B8D617D
+        patch=0,EE,0012a690,word,4A0002FF
+        patch=0,EE,0012aafc,word,48468800
+        patch=0,EE,0012ab00,word,4BEC682C
+        patch=0,EE,0012ab04,word,4B8D617D
+        patch=0,EE,0012ab08,word,4A6D617C
+        patch=0,EE,0012ab0c,word,4A0002FF
+        patch=0,EE,001644cc,word,48468800
+        patch=0,EE,001644d0,word,4BEC682C
+        patch=0,EE,001644d8,word,4B8D617D
+        patch=0,EE,001644dc,word,4A0002FF
+        patch=0,EE,001ae958,word,48468800
+        patch=0,EE,001ae95c,word,4BEC682C
+        patch=0,EE,001ae960,word,4BED617D
+        patch=0,EE,001ae968,word,4A0002FF
+        patch=0,EE,001bd410,word,48588800
+        patch=0,EE,001bd41c,word,4A5DEF40
+        patch=0,EE,001bd62c,word,48588800
+        patch=0,EE,001bd638,word,4A5DEF40
+        patch=0,EE,001bd9cc,word,48588800
+        patch=0,EE,001bd9d8,word,4A5DEF40
+        patch=0,EE,001bdf90,word,48588800
+        patch=0,EE,001bdf9c,word,4A5DEF40
 SCKA-20028:
   name: "이코 [PlayStation 2 BigHit Series]"
   name-en: "Ico [PlayStation 2 BigHit Series]"

pcsx2/GS/Renderers/Metal/GSDeviceMTL.mm

@@ -1073,9 +1073,10 @@ bool GSDeviceMTL::Create(GSVSyncMode vsync_mode, bool allow_present_throttle)
 
 	// Init metal stuff
 	m_fn_constants = MRCTransfer([MTLFunctionConstantValues new]);
-	setFnConstantB(m_fn_constants, m_features.framebuffer_fetch,    GSMTLConstantIndex_FRAMEBUFFER_FETCH);
-	setFnConstantB(m_fn_constants, m_features.depth_feedback,       GSMTLConstantIndex_DEPTH_FEEDBACK);
-	setFnConstantB(m_fn_constants, m_dev.features.rov_requires_r32, GSMTLConstantIndex_ROV_NEEDS_R32);
+	setFnConstantB(m_fn_constants, m_features.framebuffer_fetch,       GSMTLConstantIndex_FRAMEBUFFER_FETCH);
+	setFnConstantB(m_fn_constants, m_features.depth_feedback,          GSMTLConstantIndex_DEPTH_FEEDBACK);
+	setFnConstantB(m_fn_constants, m_dev.features.rov_requires_r32,    GSMTLConstantIndex_ROV_NEEDS_R32);
+	setFnConstantB(m_fn_constants, m_dev.features.broken_shader_depth, GSMTLConstantIndex_BROKEN_SHADER_DEPTH);
 
 	m_draw_sync_fence = MRCTransfer([m_dev.dev newFence]);
 	[m_draw_sync_fence setLabel:@"Draw Sync Fence"];
@@ -2350,6 +2351,9 @@ void GSDeviceMTL::RenderHW(GSHWDrawConfig& config)
 	if (config.tex && (config.ds == config.tex || config.rt == config.tex))
 		EndRenderPass(); // Barrier
 
+	if (m_dev.features.broken_shader_depth && (config.depth.ztst >= ZTST_GEQUAL || config.depth.zwe))
+		config.ps.zfloor = true; // Depth must always go through shader (see tfx vs for comment with details)
+
 	size_t vertsize = config.nverts * sizeof(*config.verts);
 	size_t idxsize = config.vs.UseFixedExpandIndexBuffer() ? 0 : (config.nindices * sizeof(*config.indices));
 	Map allocation = Allocate(m_vertex_upload_buf, vertsize + idxsize);

pcsx2/GS/Renderers/Metal/GSMTLDeviceInfo.h

@@ -37,6 +37,7 @@ struct GSMTLDevice
 		bool rov                    : 1;
 		bool rov_requires_rt        : 1;
 		bool rov_requires_r32       : 1;
+		bool broken_shader_depth    : 1;
 		MetalVersion shader_version;
 		int max_texsize;
 	};

pcsx2/GS/Renderers/Metal/GSMTLDeviceInfo.mm

@@ -207,7 +207,7 @@ GSMTLDevice::GSMTLDevice(MRCOwned<id<MTLDevice>> dev)
 	}
 	else if ([name containsString:@"Apple"])
 	{
-		// No special settings
+		features.broken_shader_depth = true;
 	}
 	else
 	{
@@ -228,6 +228,9 @@ GSMTLDevice::GSMTLDevice(MRCOwned<id<MTLDevice>> dev)
 	if (char* env = getenv("MTL_ROV_WITH_RT"))
 		features.rov_requires_rt = env[0] == '1' || env[0] == 'y' || env[0] == 'Y';
 
+	if (char* env = getenv("MTL_SHADER_DEPTH_WORKAROUND"))
+		features.broken_shader_depth = env[0] == '1' || env[0] == 'y' || env[0] == 'Y';
+
 	features.max_texsize = GetMaxTextureSize(dev);
 
 	this->dev = std::move(dev);

pcsx2/GS/Renderers/Metal/GSMTLSharedHeader.h

@@ -166,6 +166,7 @@ enum GSMTLFnConstants
 	GSMTLConstantIndex_FRAMEBUFFER_FETCH,
 	GSMTLConstantIndex_DEPTH_FEEDBACK,
 	GSMTLConstantIndex_ROV_NEEDS_R32,
+	GSMTLConstantIndex_BROKEN_SHADER_DEPTH,
 	GSMTLConstantIndex_FST,
 	GSMTLConstantIndex_IIP,
 	GSMTLConstantIndex_VS_POINT_SIZE,

pcsx2/GS/Renderers/Metal/tfx.metal

@@ -14,6 +14,7 @@ constant uint SHUFFLE_READWRITE = 3;
 constant bool HAS_FBFETCH           [[function_constant(GSMTLConstantIndex_FRAMEBUFFER_FETCH)]];
 constant bool DEPTH_FEEDBACK        [[function_constant(GSMTLConstantIndex_DEPTH_FEEDBACK)]];
 constant bool ROV_NEEDS_R32         [[function_constant(GSMTLConstantIndex_ROV_NEEDS_R32)]];
+constant bool BROKEN_SHADER_DEPTH   [[function_constant(GSMTLConstantIndex_BROKEN_SHADER_DEPTH)]];
 constant bool FST                   [[function_constant(GSMTLConstantIndex_FST)]];
 constant bool IIP                   [[function_constant(GSMTLConstantIndex_IIP)]];
 constant bool VS_POINT_SIZE         [[function_constant(GSMTLConstantIndex_VS_POINT_SIZE)]];
@@ -248,6 +249,14 @@ static MainVSOut vs_main_run(thread const MainVSIn& v, constant GSMTLMainVSUnifo
 	if (VS_POINT_SIZE)
 		out.point_size = cb.point_size.x;
 
+	// Apple GPUs use slightly different algorithms to calculate the Z they send to the shader vs the Z they use in hardware.
+	// This breaks a lot of things (the most common is conservative depth rejecting pixels that should have depth equal to current depth but now don't).
+	// Work around by always routing depth through the shader, and never using hardware depth values.
+	// To allow us to continue to use [[depth(less)]] optimizations, add a bit in the VS and subtract it off in the FS,
+	// so that "equal" depth doesn't ever fail a hardware depth test.
+	if (BROKEN_SHADER_DEPTH)
+		out.p.z += exp_min32;
+
 	return out;
 }
 
@@ -1494,6 +1503,8 @@ struct PSMain
 	{
 		MainResult out = {};
 		float input_z = in.p.z;
+		if (BROKEN_SHADER_DEPTH)
+			input_z -= 0x1p-32;
 		if (PS_ZFLOOR)
 			input_z = floor(input_z * 0x1p32) * 0x1p-32;