Merge pull request #10747 from tellowkrinkle/LateUIDFixup

Add a post-cache shader UID fixup pass

Source/Core/VideoBackends/OGL/OGLRender.cpp

@@ -1168,55 +1168,43 @@ void Renderer::ApplyBlendingState(const BlendingState state)
   if (m_current_blend_state == state)
     return;
 
-  bool useDualSource =
-      state.usedualsrc && g_ActiveConfig.backend_info.bSupportsDualSourceBlend &&
-      (!DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DUAL_SOURCE_BLENDING) || state.dstalpha);
-  // Only use shader blend if we need to and we don't support dual-source blending directly
-  bool useShaderBlend = !useDualSource && state.usedualsrc && state.dstalpha &&
-                        g_ActiveConfig.backend_info.bSupportsFramebufferFetch;
-
-  if (useShaderBlend)
-  {
-    glDisable(GL_BLEND);
-  }
+  bool useDualSource = state.usedualsrc;
+
+  const GLenum src_factors[8] = {GL_ZERO,
+                                 GL_ONE,
+                                 GL_DST_COLOR,
+                                 GL_ONE_MINUS_DST_COLOR,
+                                 useDualSource ? GL_SRC1_ALPHA : (GLenum)GL_SRC_ALPHA,
+                                 useDualSource ? GL_ONE_MINUS_SRC1_ALPHA :
+                                                 (GLenum)GL_ONE_MINUS_SRC_ALPHA,
+                                 GL_DST_ALPHA,
+                                 GL_ONE_MINUS_DST_ALPHA};
+  const GLenum dst_factors[8] = {GL_ZERO,
+                                 GL_ONE,
+                                 GL_SRC_COLOR,
+                                 GL_ONE_MINUS_SRC_COLOR,
+                                 useDualSource ? GL_SRC1_ALPHA : (GLenum)GL_SRC_ALPHA,
+                                 useDualSource ? GL_ONE_MINUS_SRC1_ALPHA :
+                                                 (GLenum)GL_ONE_MINUS_SRC_ALPHA,
+                                 GL_DST_ALPHA,
+                                 GL_ONE_MINUS_DST_ALPHA};
+
+  if (state.blendenable)
+    glEnable(GL_BLEND);
   else
-  {
-    const GLenum src_factors[8] = {GL_ZERO,
-                                   GL_ONE,
-                                   GL_DST_COLOR,
-                                   GL_ONE_MINUS_DST_COLOR,
-                                   useDualSource ? GL_SRC1_ALPHA : (GLenum)GL_SRC_ALPHA,
-                                   useDualSource ? GL_ONE_MINUS_SRC1_ALPHA :
-                                                   (GLenum)GL_ONE_MINUS_SRC_ALPHA,
-                                   GL_DST_ALPHA,
-                                   GL_ONE_MINUS_DST_ALPHA};
-    const GLenum dst_factors[8] = {GL_ZERO,
-                                   GL_ONE,
-                                   GL_SRC_COLOR,
-                                   GL_ONE_MINUS_SRC_COLOR,
-                                   useDualSource ? GL_SRC1_ALPHA : (GLenum)GL_SRC_ALPHA,
-                                   useDualSource ? GL_ONE_MINUS_SRC1_ALPHA :
-                                                   (GLenum)GL_ONE_MINUS_SRC_ALPHA,
-                                   GL_DST_ALPHA,
-                                   GL_ONE_MINUS_DST_ALPHA};
-
-    if (state.blendenable)
-      glEnable(GL_BLEND);
-    else
-      glDisable(GL_BLEND);
-
-    // Always call glBlendEquationSeparate and glBlendFuncSeparate, even when
-    // GL_BLEND is disabled, as a workaround for some bugs (possibly graphics
-    // driver issues?). See https://bugs.dolphin-emu.org/issues/10120 : "Sonic
-    // Adventure 2 Battle: graphics crash when loading first Dark level"
-    GLenum equation = state.subtract ? GL_FUNC_REVERSE_SUBTRACT : GL_FUNC_ADD;
-    GLenum equationAlpha = state.subtractAlpha ? GL_FUNC_REVERSE_SUBTRACT : GL_FUNC_ADD;
-    glBlendEquationSeparate(equation, equationAlpha);
-    glBlendFuncSeparate(src_factors[u32(state.srcfactor.Value())],
-                        dst_factors[u32(state.dstfactor.Value())],
-                        src_factors[u32(state.srcfactoralpha.Value())],
-                        dst_factors[u32(state.dstfactoralpha.Value())]);
-  }
+    glDisable(GL_BLEND);
+
+  // Always call glBlendEquationSeparate and glBlendFuncSeparate, even when
+  // GL_BLEND is disabled, as a workaround for some bugs (possibly graphics
+  // driver issues?). See https://bugs.dolphin-emu.org/issues/10120 : "Sonic
+  // Adventure 2 Battle: graphics crash when loading first Dark level"
+  GLenum equation = state.subtract ? GL_FUNC_REVERSE_SUBTRACT : GL_FUNC_ADD;
+  GLenum equationAlpha = state.subtractAlpha ? GL_FUNC_REVERSE_SUBTRACT : GL_FUNC_ADD;
+  glBlendEquationSeparate(equation, equationAlpha);
+  glBlendFuncSeparate(src_factors[u32(state.srcfactor.Value())],
+                      dst_factors[u32(state.dstfactor.Value())],
+                      src_factors[u32(state.srcfactoralpha.Value())],
+                      dst_factors[u32(state.dstfactoralpha.Value())]);
 
   const GLenum logic_op_codes[16] = {
       GL_CLEAR,         GL_AND,         GL_AND_REVERSE, GL_COPY,  GL_AND_INVERTED, GL_NOOP,

Source/Core/VideoBackends/Software/Rasterizer.cpp

@@ -153,7 +153,7 @@ static void Draw(s32 x, s32 y, s32 xi, s32 yi)
 
   s32 z = (s32)std::clamp<float>(ZSlope.GetValue(x, y), 0.0f, 16777215.0f);
 
-  if (bpmem.UseEarlyDepthTest())
+  if (bpmem.GetEmulatedZ() == EmulatedZ::Early)
   {
     // TODO: Test if perf regs are incremented even if test is disabled
     EfbInterface::IncPerfCounterQuadCount(PQ_ZCOMP_INPUT_ZCOMPLOC);

Source/Core/VideoBackends/Software/Tev.cpp

@@ -840,7 +840,7 @@ void Tev::Draw()
     output[BLU_C] = (output[BLU_C] * invFog + fogInt * bpmem.fog.color.b) >> 8;
   }
 
-  if (bpmem.UseLateDepthTest())
+  if (bpmem.GetEmulatedZ() == EmulatedZ::Late)
   {
     // TODO: Check against hw if these values get incremented even if depth testing is disabled
     EfbInterface::IncPerfCounterQuadCount(PQ_ZCOMP_INPUT);

Source/Core/VideoBackends/Vulkan/VKPipeline.cpp

@@ -137,60 +137,48 @@ GetVulkanAttachmentBlendState(const BlendingState& state, AbstractPipelineUsage
 {
   VkPipelineColorBlendAttachmentState vk_state = {};
 
-  bool use_dual_source =
-      state.usedualsrc && g_ActiveConfig.backend_info.bSupportsDualSourceBlend &&
-      (!DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DUAL_SOURCE_BLENDING) || state.dstalpha);
-  bool use_shader_blend = !use_dual_source && state.usedualsrc && state.dstalpha &&
-                          g_ActiveConfig.backend_info.bSupportsFramebufferFetch;
-
-  if (use_shader_blend || (usage == AbstractPipelineUsage::GX &&
-                           DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DISCARD_WITH_EARLY_Z)))
+  bool use_dual_source = state.usedualsrc;
+
+  vk_state.blendEnable = static_cast<VkBool32>(state.blendenable);
+  vk_state.colorBlendOp = state.subtract ? VK_BLEND_OP_REVERSE_SUBTRACT : VK_BLEND_OP_ADD;
+  vk_state.alphaBlendOp = state.subtractAlpha ? VK_BLEND_OP_REVERSE_SUBTRACT : VK_BLEND_OP_ADD;
+
+  if (use_dual_source)
   {
-    vk_state.blendEnable = VK_FALSE;
+    static constexpr std::array<VkBlendFactor, 8> src_factors = {
+        {VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_DST_COLOR,
+         VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR, VK_BLEND_FACTOR_SRC1_ALPHA,
+         VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
+         VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
+    static constexpr std::array<VkBlendFactor, 8> dst_factors = {
+        {VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_SRC_COLOR,
+         VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR, VK_BLEND_FACTOR_SRC1_ALPHA,
+         VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
+         VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
+
+    vk_state.srcColorBlendFactor = src_factors[u32(state.srcfactor.Value())];
+    vk_state.srcAlphaBlendFactor = src_factors[u32(state.srcfactoralpha.Value())];
+    vk_state.dstColorBlendFactor = dst_factors[u32(state.dstfactor.Value())];
+    vk_state.dstAlphaBlendFactor = dst_factors[u32(state.dstfactoralpha.Value())];
   }
   else
   {
-    vk_state.blendEnable = static_cast<VkBool32>(state.blendenable);
-    vk_state.colorBlendOp = state.subtract ? VK_BLEND_OP_REVERSE_SUBTRACT : VK_BLEND_OP_ADD;
-    vk_state.alphaBlendOp = state.subtractAlpha ? VK_BLEND_OP_REVERSE_SUBTRACT : VK_BLEND_OP_ADD;
-
-    if (use_dual_source)
-    {
-      static constexpr std::array<VkBlendFactor, 8> src_factors = {
-          {VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_DST_COLOR,
-           VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR, VK_BLEND_FACTOR_SRC1_ALPHA,
-           VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
-           VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
-      static constexpr std::array<VkBlendFactor, 8> dst_factors = {
-          {VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_SRC_COLOR,
-           VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR, VK_BLEND_FACTOR_SRC1_ALPHA,
-           VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
-           VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
-
-      vk_state.srcColorBlendFactor = src_factors[u32(state.srcfactor.Value())];
-      vk_state.srcAlphaBlendFactor = src_factors[u32(state.srcfactoralpha.Value())];
-      vk_state.dstColorBlendFactor = dst_factors[u32(state.dstfactor.Value())];
-      vk_state.dstAlphaBlendFactor = dst_factors[u32(state.dstfactoralpha.Value())];
-    }
-    else
-    {
-      static constexpr std::array<VkBlendFactor, 8> src_factors = {
-          {VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_DST_COLOR,
-           VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR, VK_BLEND_FACTOR_SRC_ALPHA,
-           VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
-           VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
-
-      static constexpr std::array<VkBlendFactor, 8> dst_factors = {
-          {VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_SRC_COLOR,
-           VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR, VK_BLEND_FACTOR_SRC_ALPHA,
-           VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
-           VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
-
-      vk_state.srcColorBlendFactor = src_factors[u32(state.srcfactor.Value())];
-      vk_state.srcAlphaBlendFactor = src_factors[u32(state.srcfactoralpha.Value())];
-      vk_state.dstColorBlendFactor = dst_factors[u32(state.dstfactor.Value())];
-      vk_state.dstAlphaBlendFactor = dst_factors[u32(state.dstfactoralpha.Value())];
-    }
+    static constexpr std::array<VkBlendFactor, 8> src_factors = {
+        {VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_DST_COLOR,
+         VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR, VK_BLEND_FACTOR_SRC_ALPHA,
+         VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
+         VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
+
+    static constexpr std::array<VkBlendFactor, 8> dst_factors = {
+        {VK_BLEND_FACTOR_ZERO, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_SRC_COLOR,
+         VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR, VK_BLEND_FACTOR_SRC_ALPHA,
+         VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA, VK_BLEND_FACTOR_DST_ALPHA,
+         VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA}};
+
+    vk_state.srcColorBlendFactor = src_factors[u32(state.srcfactor.Value())];
+    vk_state.srcAlphaBlendFactor = src_factors[u32(state.srcfactoralpha.Value())];
+    vk_state.dstColorBlendFactor = dst_factors[u32(state.dstfactor.Value())];
+    vk_state.dstAlphaBlendFactor = dst_factors[u32(state.dstfactoralpha.Value())];
   }
 
   if (state.colorupdate)

Source/Core/VideoBackends/Vulkan/VulkanContext.cpp

@@ -371,13 +371,6 @@ void VulkanContext::PopulateBackendInfoFeatures(VideoConfig* config, VkPhysicalD
   if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_REVERSED_DEPTH_RANGE))
     config->backend_info.bSupportsReversedDepthRange = false;
 
-  // Calling discard when early depth test is enabled can break on some Apple Silicon GPU drivers.
-  if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DISCARD_WITH_EARLY_Z))
-  {
-    // We will use shader blending, so disable hardware dual source blending.
-    config->backend_info.bSupportsDualSourceBlend = false;
-  }
-
   // Dynamic sampler indexing locks up Intel GPUs on MoltenVK/Metal
   if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DYNAMIC_SAMPLER_INDEXING))
     config->backend_info.bSupportsDynamicSamplerIndexing = false;

Source/Core/VideoCommon/BPMemory.h

@@ -2338,6 +2338,16 @@ struct BPCmd
   int newvalue;
 };
 
+enum class EmulatedZ : u32
+{
+  Disabled = 0,
+  Early = 1,
+  Late = 2,
+  ForcedEarly = 3,
+  EarlyWithFBFetch = 4,
+  EarlyWithZComplocHack = 5,
+};
+
 struct BPMemory
 {
   GenMode genMode;
@@ -2405,8 +2415,15 @@ struct BPMemory
   u32 bpMask;          // 0xFE
   u32 unknown18;       // ff
 
-  bool UseEarlyDepthTest() const { return zcontrol.early_ztest && zmode.testenable; }
-  bool UseLateDepthTest() const { return !zcontrol.early_ztest && zmode.testenable; }
+  EmulatedZ GetEmulatedZ() const
+  {
+    if (!zmode.testenable)
+      return EmulatedZ::Disabled;
+    if (zcontrol.early_ztest)
+      return EmulatedZ::Early;
+    else
+      return EmulatedZ::Late;
+  }
 };
 
 #pragma pack()

Source/Core/VideoCommon/DriverDetails.h

@@ -237,7 +237,8 @@ enum Bug
   // crash. Sometimes this happens in the kernel mode part of the driver, resulting in a BSOD.
   // These shaders are also particularly problematic on macOS's Intel drivers. On OpenGL, they can
   // cause depth issues. On Metal, they can cause the driver to not write a primitive to the depth
-  // buffer whenever a fragment is discarded. Disable dual-source blending support on these drivers.
+  // buffer if dual source blending is output in the shader but not subsequently used in blending.
+  // Compile separate shaders for DSB on vs off for these drivers.
   BUG_BROKEN_DUAL_SOURCE_BLENDING,
 
   // BUG: ImgTec GLSL shader compiler fails when negating the input to a bitwise operation

Source/Core/VideoCommon/GXPipelineTypes.h

@@ -19,7 +19,7 @@ namespace VideoCommon
 // As pipelines encompass both shader UIDs and render states, changes to either of these should
 // also increment the pipeline UID version. Incrementing the UID version will cause all UID
 // caches to be invalidated.
-constexpr u32 GX_PIPELINE_UID_VERSION = 4;  // Last changed in PR 10215
+constexpr u32 GX_PIPELINE_UID_VERSION = 5;  // Last changed in PR 10747
 
 struct GXPipelineUid
 {