Vulkan: Uber shader support

Source/Core/VideoBackends/Vulkan/ObjectCache.cpp

@@ -16,6 +16,7 @@
 
 #include "Core/ConfigManager.h"
 
+#include "VideoBackends/Vulkan/CommandBufferManager.h"
 #include "VideoBackends/Vulkan/ShaderCompiler.h"
 #include "VideoBackends/Vulkan/StreamBuffer.h"
 #include "VideoBackends/Vulkan/Util.h"
@@ -59,6 +60,19 @@ bool ObjectCache::Initialize()
   if (!m_utility_shader_vertex_buffer || !m_utility_shader_uniform_buffer)
     return false;
 
+  m_dummy_texture = Texture2D::Create(1, 1, 1, 1, VK_FORMAT_R8G8B8A8_UNORM, VK_SAMPLE_COUNT_1_BIT,
+                                      VK_IMAGE_VIEW_TYPE_2D_ARRAY, VK_IMAGE_TILING_LINEAR,
+                                      VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT);
+  m_dummy_texture->TransitionToLayout(g_command_buffer_mgr->GetCurrentInitCommandBuffer(),
+                                      VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
+  VkClearColorValue clear_color = {};
+  VkImageSubresourceRange clear_range = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1};
+  vkCmdClearColorImage(g_command_buffer_mgr->GetCurrentInitCommandBuffer(),
+                       m_dummy_texture->GetImage(), VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+                       &clear_color, 1, &clear_range);
+  m_dummy_texture->TransitionToLayout(g_command_buffer_mgr->GetCurrentInitCommandBuffer(),
+                                      VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
+
   return true;
 }
 
@@ -99,17 +113,9 @@ bool ObjectCache::CreateDescriptorSetLayouts()
       {UBO_DESCRIPTOR_SET_BINDING_GS, VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC, 1,
        VK_SHADER_STAGE_GEOMETRY_BIT}};
 
-  // Annoying these have to be split, apparently we can't partially update an array without the
-  // validation layers throwing a warning.
   static const VkDescriptorSetLayoutBinding sampler_set_bindings[] = {
-      {0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 1, VK_SHADER_STAGE_FRAGMENT_BIT},
+      {0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, static_cast<u32>(NUM_PIXEL_SHADER_SAMPLERS),
-      {1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 1, VK_SHADER_STAGE_FRAGMENT_BIT},
+       VK_SHADER_STAGE_FRAGMENT_BIT}};
-      {2, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 1, VK_SHADER_STAGE_FRAGMENT_BIT},
-      {3, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 1, VK_SHADER_STAGE_FRAGMENT_BIT},
-      {4, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 1, VK_SHADER_STAGE_FRAGMENT_BIT},
-      {5, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 1, VK_SHADER_STAGE_FRAGMENT_BIT},
-      {6, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 1, VK_SHADER_STAGE_FRAGMENT_BIT},
-      {7, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 1, VK_SHADER_STAGE_FRAGMENT_BIT}};
 
   static const VkDescriptorSetLayoutBinding ssbo_set_bindings[] = {
       {0, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, 1, VK_SHADER_STAGE_FRAGMENT_BIT}};

Source/Core/VideoBackends/Vulkan/ObjectCache.h

@@ -15,6 +15,7 @@
 #include "Common/LinearDiskCache.h"
 
 #include "VideoBackends/Vulkan/Constants.h"
+#include "VideoBackends/Vulkan/Texture2D.h"
 
 #include "VideoCommon/GeometryShaderGen.h"
 #include "VideoCommon/PixelShaderGen.h"
@@ -62,6 +63,9 @@ public:
   VkSampler GetLinearSampler() const { return m_linear_sampler; }
   VkSampler GetSampler(const SamplerState& info);
 
+  // Dummy image for samplers that are unbound
+  Texture2D* GetDummyImage() const { return m_dummy_texture.get(); }
+  VkImageView GetDummyImageView() const { return m_dummy_texture->GetView(); }
   // Perform at startup, create descriptor layouts, compiles all static shaders.
   bool Initialize();
 
@@ -89,6 +93,9 @@ private:
   VkSampler m_linear_sampler = VK_NULL_HANDLE;
 
   std::map<SamplerState, VkSampler> m_sampler_cache;
+
+  // Dummy image for samplers that are unbound
+  std::unique_ptr<Texture2D> m_dummy_texture;
 };
 
 extern std::unique_ptr<ObjectCache> g_object_cache;

Source/Core/VideoBackends/Vulkan/PostProcessing.cpp

@@ -149,7 +149,7 @@ static const std::string DEFAULT_FRAGMENT_SHADER_SOURCE = R"(
 
 static const std::string POSTPROCESSING_SHADER_HEADER = R"(
   SAMPLER_BINDING(0) uniform sampler2DArray samp0;
-  SAMPLER_BINDING(1) uniform sampler2D samp1;
+  SAMPLER_BINDING(1) uniform sampler2DArray samp1;
 
   layout(location = 0) in float3 uv0;
   layout(location = 1) in float4 col0;
@@ -176,7 +176,7 @@ static const std::string POSTPROCESSING_SHADER_HEADER = R"(
 
   float4 SampleFontLocation(float2 location)
   {
-    return texture(samp1, location);
+    return texture(samp1, float3(location, 0.0));
   }
 
   float2 GetResolution()

Source/Core/VideoBackends/Vulkan/RasterFont.cpp

@@ -150,7 +150,7 @@ layout(std140, push_constant) uniform PCBlock {
   vec4 color;
 } PC;
 
-layout(set = 1, binding = 0) uniform sampler2D samp0;
+layout(set = 1, binding = 0) uniform sampler2DArray samp0;
 
 layout(location = 0) in vec2 uv0;
 
@@ -158,7 +158,7 @@ layout(location = 0) out vec4 ocol0;
 
 void main()
 {
-  ocol0 = texture(samp0, uv0) * PC.color;
+  ocol0 = texture(samp0, float3(uv0, 0.0)) * PC.color;
 }
 
 )";
@@ -209,7 +209,7 @@ bool RasterFont::CreateTexture()
   // create the actual texture object
   m_texture = Texture2D::Create(CHARACTER_WIDTH * CHARACTER_COUNT, CHARACTER_HEIGHT, 1, 1,
                                 VK_FORMAT_R8G8B8A8_UNORM, VK_SAMPLE_COUNT_1_BIT,
-                                VK_IMAGE_VIEW_TYPE_2D, VK_IMAGE_TILING_OPTIMAL,
+                                VK_IMAGE_VIEW_TYPE_2D_ARRAY, VK_IMAGE_TILING_OPTIMAL,
                                 VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT);
   if (!m_texture)
     return false;

Source/Core/VideoBackends/Vulkan/Renderer.cpp

@@ -589,6 +589,9 @@ void Renderer::SwapImpl(u32 xfb_addr, u32 fb_width, u32 fb_stride, u32 fb_height
 
   // Clean up stale textures.
   TextureCache::GetInstance()->Cleanup(frameCount);
+
+  // Pull in now-ready async shaders.
+  g_shader_cache->RetrieveAsyncShaders();
 }
 
 void Renderer::TransitionBuffersForSwap(const TargetRectangle& scaled_rect,
@@ -1132,6 +1135,8 @@ void Renderer::CheckForConfigChanges()
   bool old_force_filtering = g_ActiveConfig.bForceFiltering;
   bool old_use_xfb = g_ActiveConfig.bUseXFB;
   bool old_use_realxfb = g_ActiveConfig.bUseRealXFB;
+  bool old_vertex_ubershaders = g_ActiveConfig.bForceVertexUberShaders;
+  bool old_pixel_ubershaders = g_ActiveConfig.bForcePixelUberShaders;
 
   // Copy g_Config to g_ActiveConfig.
   // NOTE: This can potentially race with the UI thread, however if it does, the changes will be
@@ -1145,6 +1150,8 @@ void Renderer::CheckForConfigChanges()
   bool aspect_changed = old_aspect_ratio != g_ActiveConfig.iAspectRatio;
   bool use_xfb_changed = old_use_xfb != g_ActiveConfig.bUseXFB;
   bool use_realxfb_changed = old_use_realxfb != g_ActiveConfig.bUseRealXFB;
+  bool ubershaders_changed = old_vertex_ubershaders != g_ActiveConfig.bForceVertexUberShaders ||
+                             old_pixel_ubershaders != g_ActiveConfig.bForcePixelUberShaders;
 
   // Update texture cache settings with any changed options.
   TextureCache::GetInstance()->OnConfigChanged(g_ActiveConfig);
@@ -1190,6 +1197,10 @@ void Renderer::CheckForConfigChanges()
   if (anisotropy_changed || force_texture_filtering_changed)
     ResetSamplerStates();
 
+  // Clear UID state if ubershaders are toggled.
+  if (ubershaders_changed)
+    StateTracker::GetInstance()->ClearShaders();
+
   // Check for a changed post-processing shader and recompile if needed.
   static_cast<VulkanPostProcessing*>(m_post_processor.get())->UpdateConfig();
 }

Source/Core/VideoBackends/Vulkan/ShaderCache.cpp

@@ -15,13 +15,20 @@
 #include "Common/MsgHandler.h"
 
 #include "Core/ConfigManager.h"
+#include "Core/Host.h"
 
+#include "VideoBackends/Vulkan/FramebufferManager.h"
 #include "VideoBackends/Vulkan/ShaderCompiler.h"
 #include "VideoBackends/Vulkan/StreamBuffer.h"
 #include "VideoBackends/Vulkan/Util.h"
 #include "VideoBackends/Vulkan/VertexFormat.h"
 #include "VideoBackends/Vulkan/VulkanContext.h"
+#include "VideoCommon/AsyncShaderCompiler.h"
+#include "VideoCommon/GeometryShaderGen.h"
 #include "VideoCommon/Statistics.h"
+#include "VideoCommon/UberShaderPixel.h"
+#include "VideoCommon/UberShaderVertex.h"
+#include "VideoCommon/VertexLoaderManager.h"
 
 namespace Vulkan
 {
@@ -55,9 +62,22 @@ bool ShaderCache::Initialize()
   if (!CompileSharedShaders())
     return false;
 
+  m_async_shader_compiler = std::make_unique<VideoCommon::AsyncShaderCompiler>();
+  if (g_ActiveConfig.GetShaderCompilerThreads() > 0)
+    m_async_shader_compiler->StartWorkerThreads(g_ActiveConfig.GetShaderCompilerThreads());
+
   return true;
 }
 
+void ShaderCache::Shutdown()
+{
+  if (m_async_shader_compiler)
+  {
+    m_async_shader_compiler->StopWorkerThreads();
+    m_async_shader_compiler->RetrieveWorkItems();
+  }
+}
+
 static bool IsStripPrimitiveTopology(VkPrimitiveTopology topology)
 {
   return topology == VK_PRIMITIVE_TOPOLOGY_LINE_STRIP ||
@@ -365,13 +385,34 @@ std::pair<VkPipeline, bool> ShaderCache::GetPipelineWithCacheResult(const Pipeli
 {
   auto iter = m_pipeline_objects.find(info);
   if (iter != m_pipeline_objects.end())
-    return {iter->second, true};
+  {
+    // If it's background compiling, ignore it, and recompile it synchronously.
+    if (!iter->second.second)
+      return std::make_pair(iter->second.first, true);
+    else
+      m_pipeline_objects.erase(iter);
+  }
 
   VkPipeline pipeline = CreatePipeline(info);
-  m_pipeline_objects.emplace(info, pipeline);
+  m_pipeline_objects.emplace(info, std::make_pair(pipeline, false));
+  _assert_(pipeline != VK_NULL_HANDLE);
   return {pipeline, false};
 }
 
+std::pair<std::pair<VkPipeline, bool>, bool>
+ShaderCache::GetPipelineWithCacheResultAsync(const PipelineInfo& info)
+{
+  auto iter = m_pipeline_objects.find(info);
+  if (iter != m_pipeline_objects.end())
+    return std::make_pair(iter->second, true);
+
+  // Kick a job off.
+  m_async_shader_compiler->QueueWorkItem(
+      m_async_shader_compiler->CreateWorkItem<PipelineCompilerWorkItem>(info));
+  m_pipeline_objects.emplace(info, std::make_pair(static_cast<VkPipeline>(VK_NULL_HANDLE), true));
+  return std::make_pair(std::make_pair(static_cast<VkPipeline>(VK_NULL_HANDLE), true), false);
+}
+
 VkPipeline ShaderCache::CreateComputePipeline(const ComputePipelineInfo& info)
 {
   VkComputePipelineCreateInfo pipeline_info = {VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
@@ -409,10 +450,11 @@ VkPipeline ShaderCache::GetComputePipeline(const ComputePipelineInfo& info)
 
 void ShaderCache::ClearPipelineCache()
 {
+  // TODO: Stop any async compiling happening.
   for (const auto& it : m_pipeline_objects)
   {
-    if (it.second != VK_NULL_HANDLE)
+    if (it.second.first != VK_NULL_HANDLE)
-      vkDestroyPipeline(g_vulkan_context->GetDevice(), it.second, nullptr);
+      vkDestroyPipeline(g_vulkan_context->GetDevice(), it.second.first, nullptr);
   }
   m_pipeline_objects.clear();
 
@@ -620,7 +662,10 @@ void ShaderCache::SavePipelineCache()
 template <typename Uid>
 struct ShaderCacheReader : public LinearDiskCacheReader<Uid, u32>
 {
-  ShaderCacheReader(std::map<Uid, VkShaderModule>& shader_map) : m_shader_map(shader_map) {}
+  ShaderCacheReader(std::map<Uid, std::pair<VkShaderModule, bool>>& shader_map)
+      : m_shader_map(shader_map)
+  {
+  }
   void Read(const Uid& key, const u32* value, u32 value_size) override
   {
     // We don't insert null modules into the shader map since creation could succeed later on.
@@ -630,10 +675,10 @@ struct ShaderCacheReader : public LinearDiskCacheReader<Uid, u32>
     if (module == VK_NULL_HANDLE)
       return;
 
-    m_shader_map.emplace(key, module);
+    m_shader_map.emplace(key, std::make_pair(module, false));
   }
 
-  std::map<Uid, VkShaderModule>& m_shader_map;
+  std::map<Uid, std::pair<VkShaderModule, bool>>& m_shader_map;
 };
 
 void ShaderCache::LoadShaderCaches()
@@ -653,6 +698,13 @@ void ShaderCache::LoadShaderCaches()
                                       gs_reader);
   }
 
+  ShaderCacheReader<UberShader::VertexShaderUid> uber_vs_reader(m_uber_vs_cache.shader_map);
+  m_uber_vs_cache.disk_cache.OpenAndRead(
+      GetDiskShaderCacheFileName(APIType::Vulkan, "UberVS", false, true), uber_vs_reader);
+  ShaderCacheReader<UberShader::PixelShaderUid> uber_ps_reader(m_uber_ps_cache.shader_map);
+  m_uber_ps_cache.disk_cache.OpenAndRead(
+      GetDiskShaderCacheFileName(APIType::Vulkan, "UberPS", false, true), uber_ps_reader);
+
   SETSTAT(stats.numPixelShadersCreated, static_cast<int>(m_ps_cache.shader_map.size()));
   SETSTAT(stats.numPixelShadersAlive, static_cast<int>(m_ps_cache.shader_map.size()));
   SETSTAT(stats.numVertexShadersCreated, static_cast<int>(m_vs_cache.shader_map.size()));
@@ -666,8 +718,8 @@ static void DestroyShaderCache(T& cache)
   cache.disk_cache.Close();
   for (const auto& it : cache.shader_map)
   {
-    if (it.second != VK_NULL_HANDLE)
+    if (it.second.first != VK_NULL_HANDLE)
-      vkDestroyShaderModule(g_vulkan_context->GetDevice(), it.second, nullptr);
+      vkDestroyShaderModule(g_vulkan_context->GetDevice(), it.second.first, nullptr);
   }
   cache.shader_map.clear();
 }
@@ -680,6 +732,9 @@ void ShaderCache::DestroyShaderCaches()
   if (g_vulkan_context->SupportsGeometryShaders())
     DestroyShaderCache(m_gs_cache);
 
+  DestroyShaderCache(m_uber_vs_cache);
+  DestroyShaderCache(m_uber_ps_cache);
+
   SETSTAT(stats.numPixelShadersCreated, 0);
   SETSTAT(stats.numPixelShadersAlive, 0);
   SETSTAT(stats.numVertexShadersCreated, 0);
@@ -690,7 +745,13 @@ VkShaderModule ShaderCache::GetVertexShaderForUid(const VertexShaderUid& uid)
 {
   auto it = m_vs_cache.shader_map.find(uid);
   if (it != m_vs_cache.shader_map.end())
-    return it->second;
+  {
+    // If it's pending, compile it synchronously.
+    if (!it->second.second)
+      return it->second.first;
+    else
+      m_vs_cache.shader_map.erase(it);
+  }
 
   // Not in the cache, so compile the shader.
   ShaderCompiler::SPIRVCodeVector spv;
@@ -712,7 +773,7 @@ VkShaderModule ShaderCache::GetVertexShaderForUid(const VertexShaderUid& uid)
   }
 
   // We still insert null entries to prevent further compilation attempts.
-  m_vs_cache.shader_map.emplace(uid, module);
+  m_vs_cache.shader_map.emplace(uid, std::make_pair(module, false));
   return module;
 }
 
@@ -721,7 +782,13 @@ VkShaderModule ShaderCache::GetGeometryShaderForUid(const GeometryShaderUid& uid
   _assert_(g_vulkan_context->SupportsGeometryShaders());
   auto it = m_gs_cache.shader_map.find(uid);
   if (it != m_gs_cache.shader_map.end())
-    return it->second;
+  {
+    // If it's pending, compile it synchronously.
+    if (!it->second.second)
+      return it->second.first;
+    else
+      m_gs_cache.shader_map.erase(it);
+  }
 
   // Not in the cache, so compile the shader.
   ShaderCompiler::SPIRVCodeVector spv;
@@ -739,7 +806,7 @@ VkShaderModule ShaderCache::GetGeometryShaderForUid(const GeometryShaderUid& uid
   }
 
   // We still insert null entries to prevent further compilation attempts.
-  m_gs_cache.shader_map.emplace(uid, module);
+  m_gs_cache.shader_map.emplace(uid, std::make_pair(module, false));
   return module;
 }
 
@@ -747,7 +814,13 @@ VkShaderModule ShaderCache::GetPixelShaderForUid(const PixelShaderUid& uid)
 {
   auto it = m_ps_cache.shader_map.find(uid);
   if (it != m_ps_cache.shader_map.end())
-    return it->second;
+  {
+    // If it's pending, compile it synchronously.
+    if (!it->second.second)
+      return it->second.first;
+    else
+      m_ps_cache.shader_map.erase(it);
+  }
 
   // Not in the cache, so compile the shader.
   ShaderCompiler::SPIRVCodeVector spv;
@@ -769,7 +842,79 @@ VkShaderModule ShaderCache::GetPixelShaderForUid(const PixelShaderUid& uid)
   }
 
   // We still insert null entries to prevent further compilation attempts.
-  m_ps_cache.shader_map.emplace(uid, module);
+  m_ps_cache.shader_map.emplace(uid, std::make_pair(module, false));
+  return module;
+}
+
+VkShaderModule ShaderCache::GetVertexUberShaderForUid(const UberShader::VertexShaderUid& uid)
+{
+  auto it = m_uber_vs_cache.shader_map.find(uid);
+  if (it != m_uber_vs_cache.shader_map.end())
+  {
+    // If it's pending, compile it synchronously.
+    if (!it->second.second)
+      return it->second.first;
+    else
+      m_uber_vs_cache.shader_map.erase(it);
+  }
+
+  // Not in the cache, so compile the shader.
+  ShaderCompiler::SPIRVCodeVector spv;
+  VkShaderModule module = VK_NULL_HANDLE;
+  ShaderCode source_code = UberShader::GenVertexShader(
+      APIType::Vulkan, ShaderHostConfig::GetCurrent(), uid.GetUidData());
+  if (ShaderCompiler::CompileVertexShader(&spv, source_code.GetBuffer().c_str(),
+                                          source_code.GetBuffer().length()))
+  {
+    module = Util::CreateShaderModule(spv.data(), spv.size());
+
+    // Append to shader cache if it created successfully.
+    if (module != VK_NULL_HANDLE)
+    {
+      m_uber_vs_cache.disk_cache.Append(uid, spv.data(), static_cast<u32>(spv.size()));
+      INCSTAT(stats.numVertexShadersCreated);
+      INCSTAT(stats.numVertexShadersAlive);
+    }
+  }
+
+  // We still insert null entries to prevent further compilation attempts.
+  m_uber_vs_cache.shader_map.emplace(uid, std::make_pair(module, false));
+  return module;
+}
+
+VkShaderModule ShaderCache::GetPixelUberShaderForUid(const UberShader::PixelShaderUid& uid)
+{
+  auto it = m_uber_ps_cache.shader_map.find(uid);
+  if (it != m_uber_ps_cache.shader_map.end())
+  {
+    // If it's pending, compile it synchronously.
+    if (!it->second.second)
+      return it->second.first;
+    else
+      m_uber_ps_cache.shader_map.erase(it);
+  }
+
+  // Not in the cache, so compile the shader.
+  ShaderCompiler::SPIRVCodeVector spv;
+  VkShaderModule module = VK_NULL_HANDLE;
+  ShaderCode source_code =
+      UberShader::GenPixelShader(APIType::Vulkan, ShaderHostConfig::GetCurrent(), uid.GetUidData());
+  if (ShaderCompiler::CompileFragmentShader(&spv, source_code.GetBuffer().c_str(),
+                                            source_code.GetBuffer().length()))
+  {
+    module = Util::CreateShaderModule(spv.data(), spv.size());
+
+    // Append to shader cache if it created successfully.
+    if (module != VK_NULL_HANDLE)
+    {
+      m_uber_ps_cache.disk_cache.Append(uid, spv.data(), static_cast<u32>(spv.size()));
+      INCSTAT(stats.numPixelShadersCreated);
+      INCSTAT(stats.numPixelShadersAlive);
+    }
+  }
+
+  // We still insert null entries to prevent further compilation attempts.
+  m_uber_ps_cache.shader_map.emplace(uid, std::make_pair(module, false));
   return module;
 }
 
@@ -782,6 +927,9 @@ void ShaderCache::RecompileSharedShaders()
 
 void ShaderCache::ReloadShaderAndPipelineCaches()
 {
+  m_async_shader_compiler->WaitUntilCompletion();
+  m_async_shader_compiler->RetrieveWorkItems();
+
   SavePipelineCache();
   DestroyShaderCaches();
   DestroyPipelineCache();
@@ -795,6 +943,9 @@ void ShaderCache::ReloadShaderAndPipelineCaches()
   {
     CreatePipelineCache();
   }
+
+  if (g_ActiveConfig.CanPrecompileUberShaders())
+    PrecompileUberShaders();
 }
 
 std::string ShaderCache::GetUtilityShaderHeader() const
@@ -1026,4 +1177,211 @@ void ShaderCache::DestroySharedShaders()
   DestroyShader(m_screen_quad_geometry_shader);
   DestroyShader(m_passthrough_geometry_shader);
 }
+
+void ShaderCache::CreateDummyPipeline(const UberShader::VertexShaderUid& vuid,
+                                      const GeometryShaderUid& guid,
+                                      const UberShader::PixelShaderUid& puid)
+{
+  PortableVertexDeclaration vertex_decl;
+  std::memset(&vertex_decl, 0, sizeof(vertex_decl));
+
+  PipelineInfo pinfo;
+  pinfo.vertex_format =
+      static_cast<const VertexFormat*>(VertexLoaderManager::GetUberVertexFormat(vertex_decl));
+  pinfo.pipeline_layout = g_object_cache->GetPipelineLayout(
+      g_ActiveConfig.bBBoxEnable && g_ActiveConfig.BBoxUseFragmentShaderImplementation() ?
+          PIPELINE_LAYOUT_BBOX :
+          PIPELINE_LAYOUT_STANDARD);
+  pinfo.vs = GetVertexUberShaderForUid(vuid);
+  pinfo.gs = (!guid.GetUidData()->IsPassthrough() && g_vulkan_context->SupportsGeometryShaders()) ?
+                 GetGeometryShaderForUid(guid) :
+                 VK_NULL_HANDLE;
+  pinfo.ps = GetPixelUberShaderForUid(puid);
+  pinfo.render_pass = FramebufferManager::GetInstance()->GetEFBLoadRenderPass();
+  pinfo.rasterization_state.bits = Util::GetNoCullRasterizationState().bits;
+  pinfo.depth_stencil_state.bits = Util::GetNoDepthTestingDepthStencilState().bits;
+  pinfo.blend_state.hex = Util::GetNoBlendingBlendState().hex;
+  switch (guid.GetUidData()->primitive_type)
+  {
+  case PRIMITIVE_POINTS:
+    pinfo.primitive_topology = VK_PRIMITIVE_TOPOLOGY_POINT_LIST;
+    break;
+  case PRIMITIVE_LINES:
+    pinfo.primitive_topology = VK_PRIMITIVE_TOPOLOGY_LINE_LIST;
+    break;
+  case PRIMITIVE_TRIANGLES:
+    pinfo.primitive_topology = g_ActiveConfig.backend_info.bSupportsPrimitiveRestart ?
+                                   VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP :
+                                   VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST;
+    break;
+  }
+  GetPipelineWithCacheResultAsync(pinfo);
+}
+
+void ShaderCache::PrecompileUberShaders()
+{
+  UberShader::EnumerateVertexShaderUids([&](const UberShader::VertexShaderUid& vuid) {
+    UberShader::EnumeratePixelShaderUids([&](const UberShader::PixelShaderUid& puid) {
+      // UIDs must have compatible texgens, a mismatching combination will never be queried.
+      if (vuid.GetUidData()->num_texgens != puid.GetUidData()->num_texgens)
+        return;
+
+      EnumerateGeometryShaderUids([&](const GeometryShaderUid& guid) {
+        if (guid.GetUidData()->numTexGens != vuid.GetUidData()->num_texgens)
+          return;
+
+        CreateDummyPipeline(vuid, guid, puid);
+      });
+    });
+  });
+
+  WaitForBackgroundCompilesToComplete();
+}
+
+void ShaderCache::WaitForBackgroundCompilesToComplete()
+{
+  m_async_shader_compiler->WaitUntilCompletion([](size_t completed, size_t total) {
+    Host_UpdateProgressDialog(GetStringT("Compiling shaders...").c_str(),
+                              static_cast<int>(completed), static_cast<int>(total));
+  });
+  m_async_shader_compiler->RetrieveWorkItems();
+  Host_UpdateProgressDialog("", -1, -1);
+}
+
+void ShaderCache::RetrieveAsyncShaders()
+{
+  m_async_shader_compiler->RetrieveWorkItems();
+}
+
+std::pair<VkShaderModule, bool> ShaderCache::GetVertexShaderForUidAsync(const VertexShaderUid& uid)
+{
+  auto it = m_vs_cache.shader_map.find(uid);
+  if (it != m_vs_cache.shader_map.end())
+    return it->second;
+
+  // Kick a compile job off.
+  m_async_shader_compiler->QueueWorkItem(
+      m_async_shader_compiler->CreateWorkItem<VertexShaderCompilerWorkItem>(uid));
+  m_vs_cache.shader_map.emplace(uid,
+                                std::make_pair(static_cast<VkShaderModule>(VK_NULL_HANDLE), true));
+  return std::make_pair<VkShaderModule, bool>(VK_NULL_HANDLE, true);
+}
+
+std::pair<VkShaderModule, bool> ShaderCache::GetPixelShaderForUidAsync(const PixelShaderUid& uid)
+{
+  auto it = m_ps_cache.shader_map.find(uid);
+  if (it != m_ps_cache.shader_map.end())
+    return it->second;
+
+  // Kick a compile job off.
+  m_async_shader_compiler->QueueWorkItem(
+      m_async_shader_compiler->CreateWorkItem<PixelShaderCompilerWorkItem>(uid));
+  m_ps_cache.shader_map.emplace(uid,
+                                std::make_pair(static_cast<VkShaderModule>(VK_NULL_HANDLE), true));
+  return std::make_pair<VkShaderModule, bool>(VK_NULL_HANDLE, true);
+}
+
+bool ShaderCache::VertexShaderCompilerWorkItem::Compile()
+{
+  ShaderCode code =
+      GenerateVertexShaderCode(APIType::Vulkan, ShaderHostConfig::GetCurrent(), m_uid.GetUidData());
+  if (!ShaderCompiler::CompileVertexShader(&m_spirv, code.GetBuffer().c_str(),
+                                           code.GetBuffer().length()))
+    return true;
+
+  m_module = Util::CreateShaderModule(m_spirv.data(), m_spirv.size());
+  return true;
+}
+
+void ShaderCache::VertexShaderCompilerWorkItem::Retrieve()
+{
+  auto it = g_shader_cache->m_vs_cache.shader_map.find(m_uid);
+  if (it == g_shader_cache->m_vs_cache.shader_map.end())
+  {
+    g_shader_cache->m_vs_cache.shader_map.emplace(m_uid, std::make_pair(m_module, false));
+    g_shader_cache->m_vs_cache.disk_cache.Append(m_uid, m_spirv.data(),
+                                                 static_cast<u32>(m_spirv.size()));
+    return;
+  }
+
+  // The main thread may have also compiled this shader.
+  if (!it->second.second)
+  {
+    if (m_module != VK_NULL_HANDLE)
+      vkDestroyShaderModule(g_vulkan_context->GetDevice(), m_module, nullptr);
+    return;
+  }
+
+  // No longer pending.
+  it->second.first = m_module;
+  it->second.second = false;
+  g_shader_cache->m_vs_cache.disk_cache.Append(m_uid, m_spirv.data(),
+                                               static_cast<u32>(m_spirv.size()));
+}
+
+bool ShaderCache::PixelShaderCompilerWorkItem::Compile()
+{
+  ShaderCode code =
+      GeneratePixelShaderCode(APIType::Vulkan, ShaderHostConfig::GetCurrent(), m_uid.GetUidData());
+  if (!ShaderCompiler::CompileFragmentShader(&m_spirv, code.GetBuffer().c_str(),
+                                             code.GetBuffer().length()))
+    return true;
+
+  m_module = Util::CreateShaderModule(m_spirv.data(), m_spirv.size());
+  return true;
+}
+
+void ShaderCache::PixelShaderCompilerWorkItem::Retrieve()
+{
+  auto it = g_shader_cache->m_ps_cache.shader_map.find(m_uid);
+  if (it == g_shader_cache->m_ps_cache.shader_map.end())
+  {
+    g_shader_cache->m_ps_cache.shader_map.emplace(m_uid, std::make_pair(m_module, false));
+    g_shader_cache->m_ps_cache.disk_cache.Append(m_uid, m_spirv.data(),
+                                                 static_cast<u32>(m_spirv.size()));
+    return;
+  }
+
+  // The main thread may have also compiled this shader.
+  if (!it->second.second)
+  {
+    if (m_module != VK_NULL_HANDLE)
+      vkDestroyShaderModule(g_vulkan_context->GetDevice(), m_module, nullptr);
+    return;
+  }
+
+  // No longer pending.
+  it->second.first = m_module;
+  it->second.second = false;
+  g_shader_cache->m_ps_cache.disk_cache.Append(m_uid, m_spirv.data(),
+                                               static_cast<u32>(m_spirv.size()));
+}
+
+bool ShaderCache::PipelineCompilerWorkItem::Compile()
+{
+  m_pipeline = g_shader_cache->CreatePipeline(m_info);
+  return true;
+}
+
+void ShaderCache::PipelineCompilerWorkItem::Retrieve()
+{
+  auto it = g_shader_cache->m_pipeline_objects.find(m_info);
+  if (it == g_shader_cache->m_pipeline_objects.end())
+  {
+    g_shader_cache->m_pipeline_objects.emplace(m_info, std::make_pair(m_pipeline, false));
+    return;
+  }
+
+  // The main thread may have also compiled this shader.
+  if (!it->second.second)
+  {
+    if (m_pipeline != VK_NULL_HANDLE)
+      vkDestroyPipeline(g_vulkan_context->GetDevice(), m_pipeline, nullptr);
+    return;
+  }
+
+  // No longer pending.
+  it->second.first = m_pipeline;
+  it->second.second = false;
+}
 }

Source/Core/VideoBackends/Vulkan/ShaderCache.h

@@ -10,16 +10,21 @@
 #include <memory>
 #include <string>
 #include <unordered_map>
+#include <utility>
 
 #include "Common/CommonTypes.h"
 #include "Common/LinearDiskCache.h"
 
 #include "VideoBackends/Vulkan/Constants.h"
 #include "VideoBackends/Vulkan/ObjectCache.h"
+#include "VideoBackends/Vulkan/ShaderCompiler.h"
 
+#include "VideoCommon/AsyncShaderCompiler.h"
 #include "VideoCommon/GeometryShaderGen.h"
 #include "VideoCommon/PixelShaderGen.h"
 #include "VideoCommon/RenderState.h"
+#include "VideoCommon/UberShaderPixel.h"
+#include "VideoCommon/UberShaderVertex.h"
 #include "VideoCommon/VertexShaderGen.h"
 
 namespace Vulkan
@@ -92,8 +97,17 @@ public:
   VkShaderModule GetGeometryShaderForUid(const GeometryShaderUid& uid);
   VkShaderModule GetPixelShaderForUid(const PixelShaderUid& uid);
 
+  // Ubershader caches
+  VkShaderModule GetVertexUberShaderForUid(const UberShader::VertexShaderUid& uid);
+  VkShaderModule GetPixelUberShaderForUid(const UberShader::PixelShaderUid& uid);
+
+  // Accesses ShaderGen shader caches asynchronously
+  std::pair<VkShaderModule, bool> GetVertexShaderForUidAsync(const VertexShaderUid& uid);
+  std::pair<VkShaderModule, bool> GetPixelShaderForUidAsync(const PixelShaderUid& uid);
+
   // Perform at startup, create descriptor layouts, compiles all static shaders.
   bool Initialize();
+  void Shutdown();
 
   // Creates a pipeline for the specified description. The resulting pipeline, if successful
   // is not stored anywhere, this is left up to the caller.
@@ -106,6 +120,8 @@ public:
   // resulted in a pipeline being created, the second field of the return value will be false,
   // otherwise for a cache hit it will be true.
   std::pair<VkPipeline, bool> GetPipelineWithCacheResult(const PipelineInfo& info);
+  std::pair<std::pair<VkPipeline, bool>, bool>
+  GetPipelineWithCacheResultAsync(const PipelineInfo& info);
 
   // Creates a compute pipeline, and does not track the handle.
   VkPipeline CreateComputePipeline(const ComputePipelineInfo& info);
@@ -134,6 +150,10 @@ public:
   VkShaderModule GetPassthroughVertexShader() const { return m_passthrough_vertex_shader; }
   VkShaderModule GetScreenQuadGeometryShader() const { return m_screen_quad_geometry_shader; }
   VkShaderModule GetPassthroughGeometryShader() const { return m_passthrough_geometry_shader; }
+  void PrecompileUberShaders();
+  void WaitForBackgroundCompilesToComplete();
+  void RetrieveAsyncShaders();
+
 private:
   bool CreatePipelineCache();
   bool LoadPipelineCache();
@@ -144,17 +164,26 @@ private:
   bool CompileSharedShaders();
   void DestroySharedShaders();
 
+  // We generate a dummy pipeline with some defaults in the blend/depth states,
+  // that way the driver is forced to compile something (looking at you, NVIDIA).
+  // It can then hopefully re-use part of this pipeline for others in the future.
+  void CreateDummyPipeline(const UberShader::VertexShaderUid& vuid, const GeometryShaderUid& guid,
+                           const UberShader::PixelShaderUid& puid);
+
   template <typename Uid>
   struct ShaderModuleCache
   {
-    std::map<Uid, VkShaderModule> shader_map;
+    std::map<Uid, std::pair<VkShaderModule, bool>> shader_map;
     LinearDiskCache<Uid, u32> disk_cache;
   };
   ShaderModuleCache<VertexShaderUid> m_vs_cache;
   ShaderModuleCache<GeometryShaderUid> m_gs_cache;
   ShaderModuleCache<PixelShaderUid> m_ps_cache;
+  ShaderModuleCache<UberShader::VertexShaderUid> m_uber_vs_cache;
+  ShaderModuleCache<UberShader::PixelShaderUid> m_uber_ps_cache;
 
-  std::unordered_map<PipelineInfo, VkPipeline, PipelineInfoHash> m_pipeline_objects;
+  std::unordered_map<PipelineInfo, std::pair<VkPipeline, bool>, PipelineInfoHash>
+      m_pipeline_objects;
   std::unordered_map<ComputePipelineInfo, VkPipeline, ComputePipelineInfoHash>
       m_compute_pipeline_objects;
   VkPipelineCache m_pipeline_cache = VK_NULL_HANDLE;
@@ -165,6 +194,45 @@ private:
   VkShaderModule m_passthrough_vertex_shader = VK_NULL_HANDLE;
   VkShaderModule m_screen_quad_geometry_shader = VK_NULL_HANDLE;
   VkShaderModule m_passthrough_geometry_shader = VK_NULL_HANDLE;
+
+  std::unique_ptr<VideoCommon::AsyncShaderCompiler> m_async_shader_compiler;
+
+  // TODO: Use templates to reduce the number of these classes.
+  class VertexShaderCompilerWorkItem : public VideoCommon::AsyncShaderCompiler::WorkItem
+  {
+  public:
+    VertexShaderCompilerWorkItem(const VertexShaderUid& uid) : m_uid(uid) {}
+    bool Compile() override;
+    void Retrieve() override;
+
+  private:
+    VertexShaderUid m_uid;
+    ShaderCompiler::SPIRVCodeVector m_spirv;
+    VkShaderModule m_module = VK_NULL_HANDLE;
+  };
+  class PixelShaderCompilerWorkItem : public VideoCommon::AsyncShaderCompiler::WorkItem
+  {
+  public:
+    PixelShaderCompilerWorkItem(const PixelShaderUid& uid) : m_uid(uid) {}
+    bool Compile() override;
+    void Retrieve() override;
+
+  private:
+    PixelShaderUid m_uid;
+    ShaderCompiler::SPIRVCodeVector m_spirv;
+    VkShaderModule m_module = VK_NULL_HANDLE;
+  };
+  class PipelineCompilerWorkItem : public VideoCommon::AsyncShaderCompiler::WorkItem
+  {
+  public:
+    PipelineCompilerWorkItem(const PipelineInfo& info) : m_info(info) {}
+    bool Compile() override;
+    void Retrieve() override;
+
+  private:
+    PipelineInfo m_info;
+    VkPipeline m_pipeline;
+  };
 };
 
 extern std::unique_ptr<ShaderCache> g_shader_cache;

Source/Core/VideoBackends/Vulkan/StateTracker.cpp

@@ -22,6 +22,7 @@
 #include "VideoCommon/GeometryShaderManager.h"
 #include "VideoCommon/PixelShaderManager.h"
 #include "VideoCommon/Statistics.h"
+#include "VideoCommon/VertexLoaderManager.h"
 #include "VideoCommon/VertexShaderManager.h"
 #include "VideoCommon/VideoConfig.h"
 
@@ -77,12 +78,13 @@ bool StateTracker::Initialize()
   m_pipeline_state.pipeline_layout = g_object_cache->GetPipelineLayout(PIPELINE_LAYOUT_STANDARD);
   m_num_active_descriptor_sets = NUM_GX_DRAW_DESCRIPTOR_SETS;
   m_bbox_enabled = false;
+  ClearShaders();
 
   // Initialize all samplers to point by default
   for (size_t i = 0; i < NUM_PIXEL_SHADER_SAMPLERS; i++)
   {
     m_bindings.ps_samplers[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
-    m_bindings.ps_samplers[i].imageView = VK_NULL_HANDLE;
+    m_bindings.ps_samplers[i].imageView = g_object_cache->GetDummyImageView();
     m_bindings.ps_samplers[i].sampler = g_object_cache->GetPointSampler();
   }
 
@@ -178,7 +180,8 @@ bool StateTracker::PrecachePipelineUID(const SerializedPipelineUID& uid)
 
   // Need to create the vertex declaration first, rather than deferring to when a game creates a
   // vertex loader that uses this format, since we need it to create a pipeline.
-  pinfo.vertex_format = VertexFormat::GetOrCreateMatchingFormat(uid.vertex_decl);
+  pinfo.vertex_format =
+      static_cast<VertexFormat*>(VertexLoaderManager::GetOrCreateMatchingFormat(uid.vertex_decl));
   pinfo.pipeline_layout = uid.ps_uid.GetUidData()->bounding_box ?
                               g_object_cache->GetPipelineLayout(PIPELINE_LAYOUT_BBOX) :
                               g_object_cache->GetPipelineLayout(PIPELINE_LAYOUT_STANDARD);
@@ -267,11 +270,11 @@ void StateTracker::SetFramebuffer(VkFramebuffer framebuffer, const VkRect2D& ren
 
 void StateTracker::SetVertexFormat(const VertexFormat* vertex_format)
 {
-  if (m_pipeline_state.vertex_format == vertex_format)
+  if (m_vertex_format == vertex_format)
     return;
 
-  m_pipeline_state.vertex_format = vertex_format;
+  m_vertex_format = vertex_format;
-  m_dirty_flags |= DIRTY_FLAG_PIPELINE;
+  UpdatePipelineVertexFormat();
 }
 
 void StateTracker::SetPrimitiveTopology(VkPrimitiveTopology primitive_topology)
@@ -323,14 +326,94 @@ bool StateTracker::CheckForShaderChanges(u32 gx_primitive_type)
 {
   VertexShaderUid vs_uid = GetVertexShaderUid();
   PixelShaderUid ps_uid = GetPixelShaderUid();
-
   bool changed = false;
 
-  if (vs_uid != m_vs_uid)
+  bool use_ubershaders = g_ActiveConfig.bDisableSpecializedShaders;
+  if (g_ActiveConfig.CanBackgroundCompileShaders() && !g_ActiveConfig.bDisableSpecializedShaders)
   {
-    m_pipeline_state.vs = g_shader_cache->GetVertexShaderForUid(vs_uid);
+    // Look up both VS and PS, and check if we can compile it asynchronously.
-    m_vs_uid = vs_uid;
+    auto vs = g_shader_cache->GetVertexShaderForUidAsync(vs_uid);
-    changed = true;
+    auto ps = g_shader_cache->GetPixelShaderForUidAsync(ps_uid);
+    if (vs.second || ps.second)
+    {
+      // One of the shaders is still pending. Use the ubershader for both.
+      use_ubershaders = true;
+    }
+    else
+    {
+      // Use the standard shaders for both.
+      if (m_pipeline_state.vs != vs.first)
+      {
+        m_pipeline_state.vs = vs.first;
+        m_vs_uid = vs_uid;
+        changed = true;
+      }
+      if (m_pipeline_state.ps != ps.first)
+      {
+        m_pipeline_state.ps = ps.first;
+        m_ps_uid = ps_uid;
+        changed = true;
+      }
+    }
+  }
+  else
+  {
+    // Normal shader path. No ubershaders.
+    if (vs_uid != m_vs_uid)
+    {
+      m_vs_uid = vs_uid;
+      m_pipeline_state.vs = g_shader_cache->GetVertexShaderForUid(vs_uid);
+      changed = true;
+    }
+    if (ps_uid != m_ps_uid)
+    {
+      m_ps_uid = ps_uid;
+      m_pipeline_state.ps = g_shader_cache->GetPixelShaderForUid(ps_uid);
+      changed = true;
+    }
+  }
+
+  // Ubershader fallback?
+  bool uber_vertex_shader = use_ubershaders || g_ActiveConfig.bForceVertexUberShaders;
+  bool uber_pixel_shader = use_ubershaders || g_ActiveConfig.bForcePixelUberShaders;
+  bool using_ubershaders = uber_vertex_shader || uber_pixel_shader;
+  if (!g_ActiveConfig.CanUseUberShaders())
+  {
+    // Per-pixel lighting disables ubershaders.
+    uber_vertex_shader = false;
+    uber_pixel_shader = false;
+    using_ubershaders = false;
+  }
+
+  // Switching to/from ubershaders? Have to adjust the vertex format and pipeline layout.
+  if (using_ubershaders != m_using_ubershaders)
+  {
+    m_using_ubershaders = using_ubershaders;
+    UpdatePipelineLayout();
+    UpdatePipelineVertexFormat();
+  }
+
+  if (uber_vertex_shader)
+  {
+    UberShader::VertexShaderUid uber_vs_uid = UberShader::GetVertexShaderUid();
+    VkShaderModule vs = g_shader_cache->GetVertexUberShaderForUid(uber_vs_uid);
+    if (vs != m_pipeline_state.vs)
+    {
+      m_uber_vs_uid = uber_vs_uid;
+      m_pipeline_state.vs = vs;
+      changed = true;
+    }
+  }
+  if (uber_pixel_shader)
+  {
+    UberShader::PixelShaderUid uber_ps_uid = UberShader::GetPixelShaderUid();
+    VkShaderModule ps = g_shader_cache->GetPixelUberShaderForUid(uber_ps_uid);
+    if (ps != m_pipeline_state.ps)
+    {
+      m_uber_ps_uid = uber_ps_uid;
+      m_pipeline_state.ps = ps;
+      changed = true;
+    }
   }
 
   if (g_vulkan_context->SupportsGeometryShaders())
@@ -338,29 +421,39 @@ bool StateTracker::CheckForShaderChanges(u32 gx_primitive_type)
     GeometryShaderUid gs_uid = GetGeometryShaderUid(gx_primitive_type);
     if (gs_uid != m_gs_uid)
     {
+      m_gs_uid = gs_uid;
       if (gs_uid.GetUidData()->IsPassthrough())
         m_pipeline_state.gs = VK_NULL_HANDLE;
       else
         m_pipeline_state.gs = g_shader_cache->GetGeometryShaderForUid(gs_uid);
 
-      m_gs_uid = gs_uid;
       changed = true;
     }
   }
 
-  if (ps_uid != m_ps_uid)
-  {
-    m_pipeline_state.ps = g_shader_cache->GetPixelShaderForUid(ps_uid);
-    m_ps_uid = ps_uid;
-    changed = true;
-  }
-
   if (changed)
     m_dirty_flags |= DIRTY_FLAG_PIPELINE;
 
   return changed;
 }
 
+void StateTracker::ClearShaders()
+{
+  // Set the UIDs to something that will never match, so on the first access they are checked.
+  std::memset(&m_vs_uid, 0xFF, sizeof(m_vs_uid));
+  std::memset(&m_gs_uid, 0xFF, sizeof(m_gs_uid));
+  std::memset(&m_ps_uid, 0xFF, sizeof(m_ps_uid));
+  std::memset(&m_uber_vs_uid, 0xFF, sizeof(m_uber_vs_uid));
+  std::memset(&m_uber_ps_uid, 0xFF, sizeof(m_uber_ps_uid));
+
+  m_pipeline_state.vs = VK_NULL_HANDLE;
+  m_pipeline_state.gs = VK_NULL_HANDLE;
+  m_pipeline_state.ps = VK_NULL_HANDLE;
+  m_pipeline_state.vertex_format = nullptr;
+
+  m_dirty_flags |= DIRTY_FLAG_PIPELINE;
+}
+
 void StateTracker::UpdateVertexShaderConstants()
 {
   if (!VertexShaderManager::dirty || !ReserveConstantStorage())
@@ -557,24 +650,8 @@ void StateTracker::SetBBoxEnable(bool enable)
   if (m_bbox_enabled == enable)
     return;
 
-  // Change the number of active descriptor sets, as well as the pipeline layout
-  if (enable)
-  {
-    m_pipeline_state.pipeline_layout = g_object_cache->GetPipelineLayout(PIPELINE_LAYOUT_BBOX);
-    m_num_active_descriptor_sets = NUM_GX_DRAW_WITH_BBOX_DESCRIPTOR_SETS;
-
-    // The bbox buffer never changes, so we defer descriptor updates until it is enabled.
-    if (m_descriptor_sets[DESCRIPTOR_SET_BIND_POINT_STORAGE_OR_TEXEL_BUFFER] == VK_NULL_HANDLE)
-      m_dirty_flags |= DIRTY_FLAG_PS_SSBO;
-  }
-  else
-  {
-    m_pipeline_state.pipeline_layout = g_object_cache->GetPipelineLayout(PIPELINE_LAYOUT_STANDARD);
-    m_num_active_descriptor_sets = NUM_GX_DRAW_DESCRIPTOR_SETS;
-  }
-
-  m_dirty_flags |= DIRTY_FLAG_PIPELINE | DIRTY_FLAG_DESCRIPTOR_SET_BINDING;
   m_bbox_enabled = enable;
+  UpdatePipelineLayout();
 }
 
 void StateTracker::SetBBoxBuffer(VkBuffer buffer, VkDeviceSize offset, VkDeviceSize range)
@@ -590,7 +667,7 @@ void StateTracker::SetBBoxBuffer(VkBuffer buffer, VkDeviceSize offset, VkDeviceS
   m_bindings.ps_ssbo.range = range;
 
   // Defer descriptor update until bbox is actually enabled.
-  if (m_bbox_enabled)
+  if (IsSSBODescriptorRequired())
     m_dirty_flags |= DIRTY_FLAG_PS_SSBO;
 }
 
@@ -599,7 +676,7 @@ void StateTracker::UnbindTexture(VkImageView view)
   for (VkDescriptorImageInfo& it : m_bindings.ps_samplers)
   {
     if (it.imageView == view)
-      it.imageView = VK_NULL_HANDLE;
+      it.imageView = g_object_cache->GetDummyImageView();
   }
 }
 
@@ -609,7 +686,7 @@ void StateTracker::InvalidateDescriptorSets()
   m_dirty_flags |= DIRTY_FLAG_ALL_DESCRIPTOR_SETS;
 
   // Defer SSBO descriptor update until bbox is actually enabled.
-  if (!m_bbox_enabled)
+  if (!IsSSBODescriptorRequired())
     m_dirty_flags &= ~DIRTY_FLAG_PS_SSBO;
 }
 
@@ -886,15 +963,49 @@ void StateTracker::EndClearRenderPass()
   EndRenderPass();
 }
 
-VkPipeline StateTracker::GetPipelineAndCacheUID(const PipelineInfo& info)
+VkPipeline StateTracker::GetPipelineAndCacheUID()
 {
-  auto result = g_shader_cache->GetPipelineWithCacheResult(info);
+  // We can't cache ubershader uids, only normal shader uids.
+  if (g_ActiveConfig.CanBackgroundCompileShaders() && !m_using_ubershaders)
+  {
+    // Append to UID cache if it is a new pipeline.
+    auto result = g_shader_cache->GetPipelineWithCacheResultAsync(m_pipeline_state);
+    if (!result.second && g_ActiveConfig.bShaderCache)
+      AppendToPipelineUIDCache(m_pipeline_state);
 
-  // Add to the UID cache if it is a new pipeline.
+    // Still waiting for the pipeline to compile?
-  if (!result.second && g_ActiveConfig.bShaderCache)
+    if (!result.first.second)
-    AppendToPipelineUIDCache(info);
+      return result.first.first;
 
-  return result.first;
+    // Use ubershader instead.
+    m_using_ubershaders = true;
+    UpdatePipelineLayout();
+    UpdatePipelineVertexFormat();
+
+    PipelineInfo uber_info = m_pipeline_state;
+    UberShader::VertexShaderUid uber_vuid = UberShader::GetVertexShaderUid();
+    UberShader::PixelShaderUid uber_puid = UberShader::GetPixelShaderUid();
+    uber_info.vs = g_shader_cache->GetVertexUberShaderForUid(uber_vuid);
+    uber_info.ps = g_shader_cache->GetPixelUberShaderForUid(uber_puid);
+
+    auto uber_result = g_shader_cache->GetPipelineWithCacheResult(uber_info);
+    return uber_result.first;
+  }
+  else
+  {
+    // Add to the UID cache if it is a new pipeline.
+    auto result = g_shader_cache->GetPipelineWithCacheResult(m_pipeline_state);
+    if (!result.second && !m_using_ubershaders && g_ActiveConfig.bShaderCache)
+      AppendToPipelineUIDCache(m_pipeline_state);
+
+    return result.first;
+  }
+}
+
+bool StateTracker::IsSSBODescriptorRequired() const
+{
+  return m_bbox_enabled || (m_using_ubershaders && g_ActiveConfig.bBBoxEnable &&
+                            g_ActiveConfig.BBoxUseFragmentShaderImplementation());
 }
 
 bool StateTracker::UpdatePipeline()
@@ -904,16 +1015,56 @@ bool StateTracker::UpdatePipeline()
     return false;
 
   // Grab a new pipeline object, this can fail.
-  m_pipeline_object = GetPipelineAndCacheUID(m_pipeline_state);
+  m_pipeline_object = GetPipelineAndCacheUID();
 
   m_dirty_flags |= DIRTY_FLAG_PIPELINE_BINDING;
   return m_pipeline_object != VK_NULL_HANDLE;
 }
 
+void StateTracker::UpdatePipelineLayout()
+{
+  const bool use_bbox_pipeline_layout = IsSSBODescriptorRequired();
+  VkPipelineLayout pipeline_layout =
+      use_bbox_pipeline_layout ? g_object_cache->GetPipelineLayout(PIPELINE_LAYOUT_BBOX) :
+                                 g_object_cache->GetPipelineLayout(PIPELINE_LAYOUT_STANDARD);
+  if (m_pipeline_state.pipeline_layout == pipeline_layout)
+    return;
+
+  // Change the number of active descriptor sets, as well as the pipeline layout
+  m_pipeline_state.pipeline_layout = pipeline_layout;
+  if (use_bbox_pipeline_layout)
+  {
+    m_num_active_descriptor_sets = NUM_GX_DRAW_WITH_BBOX_DESCRIPTOR_SETS;
+
+    // The bbox buffer never changes, so we defer descriptor updates until it is enabled.
+    if (m_descriptor_sets[DESCRIPTOR_SET_BIND_POINT_STORAGE_OR_TEXEL_BUFFER] == VK_NULL_HANDLE)
+      m_dirty_flags |= DIRTY_FLAG_PS_SSBO;
+  }
+  else
+  {
+    m_num_active_descriptor_sets = NUM_GX_DRAW_DESCRIPTOR_SETS;
+  }
+
+  m_dirty_flags |= DIRTY_FLAG_PIPELINE | DIRTY_FLAG_DESCRIPTOR_SET_BINDING;
+}
+
+void StateTracker::UpdatePipelineVertexFormat()
+{
+  const NativeVertexFormat* vertex_format =
+      m_using_ubershaders ?
+          VertexLoaderManager::GetUberVertexFormat(m_vertex_format->GetVertexDeclaration()) :
+          m_vertex_format;
+  if (m_pipeline_state.vertex_format == vertex_format)
+    return;
+
+  m_pipeline_state.vertex_format = static_cast<const VertexFormat*>(vertex_format);
+  m_dirty_flags |= DIRTY_FLAG_PIPELINE;
+}
+
 bool StateTracker::UpdateDescriptorSet()
 {
   const size_t MAX_DESCRIPTOR_WRITES = NUM_UBO_DESCRIPTOR_SET_BINDINGS +  // UBO
-                                       NUM_PIXEL_SHADER_SAMPLERS +        // Samplers
+                                       1 +                                // Samplers
                                        1;                                 // SSBO
   std::array<VkWriteDescriptorSet, MAX_DESCRIPTOR_WRITES> writes;
   u32 num_writes = 0;
@@ -954,30 +1105,22 @@ bool StateTracker::UpdateDescriptorSet()
     if (set == VK_NULL_HANDLE)
       return false;
 
-    for (size_t i = 0; i < NUM_PIXEL_SHADER_SAMPLERS; i++)
+    writes[num_writes++] = {VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
-    {
+                            nullptr,
-      const VkDescriptorImageInfo& info = m_bindings.ps_samplers[i];
+                            set,
-      if (info.imageView != VK_NULL_HANDLE && info.sampler != VK_NULL_HANDLE)
+                            0,
-      {
+                            0,
-        writes[num_writes++] = {VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+                            static_cast<u32>(NUM_PIXEL_SHADER_SAMPLERS),
-                                nullptr,
+                            VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
-                                set,
+                            m_bindings.ps_samplers.data(),
-                                static_cast<uint32_t>(i),
+                            nullptr,
-                                0,
+                            nullptr};
-                                1,
-                                VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
-                                &info,
-                                nullptr,
-                                nullptr};
-      }
-    }
 
     m_descriptor_sets[DESCRIPTOR_SET_BIND_POINT_PIXEL_SHADER_SAMPLERS] = set;
     m_dirty_flags |= DIRTY_FLAG_DESCRIPTOR_SET_BINDING;
   }
 
-  if (m_bbox_enabled &&
+  if ((m_dirty_flags & DIRTY_FLAG_PS_SSBO ||
-      (m_dirty_flags & DIRTY_FLAG_PS_SSBO ||
        m_descriptor_sets[DESCRIPTOR_SET_BIND_POINT_STORAGE_OR_TEXEL_BUFFER] == VK_NULL_HANDLE))
   {
     VkDescriptorSetLayout layout =

Source/Core/VideoBackends/Vulkan/StateTracker.h

@@ -16,6 +16,8 @@
 #include "VideoCommon/NativeVertexFormat.h"
 #include "VideoCommon/PixelShaderGen.h"
 #include "VideoCommon/RenderBase.h"
+#include "VideoCommon/UberShaderPixel.h"
+#include "VideoCommon/UberShaderVertex.h"
 #include "VideoCommon/VertexShaderGen.h"
 
 namespace Vulkan
@@ -60,6 +62,7 @@ public:
   void SetBlendState(const BlendingState& state);
 
   bool CheckForShaderChanges(u32 gx_primitive_type);
+  void ClearShaders();
 
   void UpdateVertexShaderConstants();
   void UpdateGeometryShaderConstants();
@@ -159,8 +162,8 @@ private:
     DIRTY_FLAG_DESCRIPTOR_SET_BINDING = (1 << 11),
     DIRTY_FLAG_PIPELINE_BINDING = (1 << 12),
 
-    DIRTY_FLAG_ALL_DESCRIPTOR_SETS =
+    DIRTY_FLAG_ALL_DESCRIPTOR_SETS = DIRTY_FLAG_VS_UBO | DIRTY_FLAG_GS_UBO | DIRTY_FLAG_PS_UBO |
-        DIRTY_FLAG_VS_UBO | DIRTY_FLAG_GS_UBO | DIRTY_FLAG_PS_SAMPLERS | DIRTY_FLAG_PS_SSBO
+                                     DIRTY_FLAG_PS_SAMPLERS | DIRTY_FLAG_PS_SSBO
   };
 
   bool Initialize();
@@ -178,9 +181,15 @@ private:
 
   // Obtains a Vulkan pipeline object for the specified pipeline configuration.
   // Also adds this pipeline configuration to the UID cache if it is not present already.
-  VkPipeline GetPipelineAndCacheUID(const PipelineInfo& info);
+  VkPipeline GetPipelineAndCacheUID();
+
+  // Are bounding box ubershaders enabled? If so, we need to ensure the SSBO is set up,
+  // since the bbox writes are determined by a uniform.
+  bool IsSSBODescriptorRequired() const;
 
   bool UpdatePipeline();
+  void UpdatePipelineLayout();
+  void UpdatePipelineVertexFormat();
   bool UpdateDescriptorSet();
 
   // Allocates storage in the uniform buffer of the specified size. If this storage cannot be
@@ -203,10 +212,14 @@ private:
   VertexShaderUid m_vs_uid = {};
   GeometryShaderUid m_gs_uid = {};
   PixelShaderUid m_ps_uid = {};
+  UberShader::VertexShaderUid m_uber_vs_uid = {};
+  UberShader::PixelShaderUid m_uber_ps_uid = {};
+  bool m_using_ubershaders = false;
 
   // pipeline state
   PipelineInfo m_pipeline_state = {};
   VkPipeline m_pipeline_object = VK_NULL_HANDLE;
+  const VertexFormat* m_vertex_format = nullptr;
 
   // shader bindings
   std::array<VkDescriptorSet, NUM_DESCRIPTOR_SET_BIND_POINTS> m_descriptor_sets = {};

Source/Core/VideoBackends/Vulkan/Util.cpp

@@ -575,8 +575,7 @@ void UtilityShaderDraw::BindDescriptors()
 {
   // TODO: This method is a mess, clean it up
   std::array<VkDescriptorSet, NUM_DESCRIPTOR_SET_BIND_POINTS> bind_descriptor_sets = {};
-  std::array<VkWriteDescriptorSet, NUM_UBO_DESCRIPTOR_SET_BINDINGS + NUM_PIXEL_SHADER_SAMPLERS>
+  std::array<VkWriteDescriptorSet, NUM_UBO_DESCRIPTOR_SET_BINDINGS + 1> set_writes = {};
-      set_writes = {};
   uint32_t num_set_writes = 0;
 
   VkDescriptorBufferInfo dummy_uniform_buffer = {
@@ -633,29 +632,32 @@ void UtilityShaderDraw::BindDescriptors()
   // Check if we have any at all, skip the binding process entirely if we don't
   if (first_active_sampler != NUM_PIXEL_SHADER_SAMPLERS)
   {
+    // We need to fill it with non-empty images.
+    for (size_t i = 0; i < NUM_PIXEL_SHADER_SAMPLERS; i++)
+    {
+      if (m_ps_samplers[i].imageView == VK_NULL_HANDLE)
+      {
+        m_ps_samplers[i].imageView = g_object_cache->GetDummyImageView();
+        m_ps_samplers[i].sampler = g_object_cache->GetPointSampler();
+      }
+    }
+
     // Allocate a new descriptor set
     VkDescriptorSet set = g_command_buffer_mgr->AllocateDescriptorSet(
         g_object_cache->GetDescriptorSetLayout(DESCRIPTOR_SET_LAYOUT_PIXEL_SHADER_SAMPLERS));
     if (set == VK_NULL_HANDLE)
       PanicAlert("Failed to allocate descriptor set for utility draw");
 
-    for (size_t i = 0; i < NUM_PIXEL_SHADER_SAMPLERS; i++)
+    set_writes[num_set_writes++] = {VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
-    {
+                                    nullptr,
-      const VkDescriptorImageInfo& info = m_ps_samplers[i];
+                                    set,
-      if (info.imageView != VK_NULL_HANDLE && info.sampler != VK_NULL_HANDLE)
+                                    0,
-      {
+                                    0,
-        set_writes[num_set_writes++] = {VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+                                    static_cast<u32>(NUM_PIXEL_SHADER_SAMPLERS),
-                                        nullptr,
+                                    VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
-                                        set,
+                                    m_ps_samplers.data(),
-                                        static_cast<uint32_t>(i),
+                                    nullptr,
-                                        0,
+                                    nullptr};
-                                        1,
-                                        VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
-                                        &info,
-                                        nullptr,
-                                        nullptr};
-      }
-    }
 
     bind_descriptor_sets[DESCRIPTOR_SET_BIND_POINT_PIXEL_SHADER_SAMPLERS] = set;
   }

Source/Core/VideoBackends/Vulkan/VertexFormat.cpp

@@ -53,17 +53,9 @@ VertexFormat::VertexFormat(const PortableVertexDeclaration& in_vtx_decl)
   SetupInputState();
 }
 
-VertexFormat* VertexFormat::GetOrCreateMatchingFormat(const PortableVertexDeclaration& decl)
+const VkPipelineVertexInputStateCreateInfo& VertexFormat::GetVertexInputStateInfo() const
 {
-  auto vertex_format_map = VertexLoaderManager::GetNativeVertexFormatMap();
+  return m_input_state_info;
-  auto iter = vertex_format_map->find(decl);
-  if (iter == vertex_format_map->end())
-  {
-    auto ipair = vertex_format_map->emplace(decl, std::make_unique<VertexFormat>(decl));
-    iter = ipair.first;
-  }
-
-  return static_cast<VertexFormat*>(iter->second.get());
 }
 
 void VertexFormat::MapAttributes()

Source/Core/VideoBackends/Vulkan/VertexFormat.h

@@ -16,16 +16,8 @@ class VertexFormat : public ::NativeVertexFormat
 public:
   VertexFormat(const PortableVertexDeclaration& in_vtx_decl);
 
-  // Creates or obtains a pointer to a VertexFormat representing decl.
-  // If this results in a VertexFormat being created, if the game later uses a matching vertex
-  // declaration, the one that was previously created will be used.
-  static VertexFormat* GetOrCreateMatchingFormat(const PortableVertexDeclaration& decl);
-
   // Passed to pipeline state creation
-  const VkPipelineVertexInputStateCreateInfo& GetVertexInputStateInfo() const
+  const VkPipelineVertexInputStateCreateInfo& GetVertexInputStateInfo() const;
-  {
-    return m_input_state_info;
-  }
 
   // Converting PortableVertexDeclaration -> Vulkan types
   void MapAttributes();

Source/Core/VideoBackends/Vulkan/main.cpp

@@ -253,6 +253,7 @@ bool VideoBackend::Initialize(void* window_handle)
     g_renderer.reset();
     StateTracker::DestroyInstance();
     g_framebuffer_manager.reset();
+    g_shader_cache->Shutdown();
     g_shader_cache.reset();
     g_object_cache.reset();
     g_command_buffer_mgr.reset();
@@ -262,6 +263,11 @@ bool VideoBackend::Initialize(void* window_handle)
     return false;
   }
 
+  // Lastly, precompile ubershaders, if requested.
+  // This has to be done after the texture cache and shader cache are initialized.
+  if (g_ActiveConfig.CanPrecompileUberShaders())
+    g_shader_cache->PrecompileUberShaders();
+
   return true;
 }
 
@@ -293,6 +299,7 @@ void VideoBackend::Shutdown()
 void VideoBackend::Video_Cleanup()
 {
   g_command_buffer_mgr->WaitForGPUIdle();
+  g_shader_cache->Shutdown();
 
   // Save all cached pipelines out to disk for next time.
   if (g_ActiveConfig.bShaderCache)

Source/Core/VideoCommon/PixelShaderGen.cpp

@@ -354,21 +354,10 @@ void WritePixelShaderCommonHeader(ShaderCode& out, APIType ApiType, bool boundin
             "int3 iround(float3 x) { return int3(round(x)); }\n"
             "int4 iround(float4 x) { return int4(round(x)); }\n\n");
 
-  if (ApiType == APIType::OpenGL)
+  if (ApiType == APIType::OpenGL || ApiType == APIType::Vulkan)
   {
     out.Write("SAMPLER_BINDING(0) uniform sampler2DArray samp[8];\n");
   }
-  else if (ApiType == APIType::Vulkan)
-  {
-    out.Write("SAMPLER_BINDING(0) uniform sampler2DArray samp0;\n");
-    out.Write("SAMPLER_BINDING(1) uniform sampler2DArray samp1;\n");
-    out.Write("SAMPLER_BINDING(2) uniform sampler2DArray samp2;\n");
-    out.Write("SAMPLER_BINDING(3) uniform sampler2DArray samp3;\n");
-    out.Write("SAMPLER_BINDING(4) uniform sampler2DArray samp4;\n");
-    out.Write("SAMPLER_BINDING(5) uniform sampler2DArray samp5;\n");
-    out.Write("SAMPLER_BINDING(6) uniform sampler2DArray samp6;\n");
-    out.Write("SAMPLER_BINDING(7) uniform sampler2DArray samp7;\n");
-  }
   else  // D3D
   {
     // Declare samplers
@@ -1191,11 +1180,6 @@ static void SampleTexture(ShaderCode& out, const char* texcoords, const char* te
               "[%d].xy, %s))).%s;\n",
               texmap, texmap, texcoords, texmap, stereo ? "layer" : "0.0", texswap);
   }
-  else if (ApiType == APIType::Vulkan)
-  {
-    out.Write("iround(255.0 * texture(samp%d, float3(%s.xy * " I_TEXDIMS "[%d].xy, %s))).%s;\n",
-              texmap, texcoords, texmap, stereo ? "layer" : "0.0", texswap);
-  }
   else
   {
     out.Write("iround(255.0 * texture(samp[%d], float3(%s.xy * " I_TEXDIMS "[%d].xy, %s))).%s;\n",

Source/Core/VideoCommon/VertexLoaderManager.cpp

@@ -44,13 +44,6 @@ static VertexLoaderMap s_vertex_loader_map;
 
 u8* cached_arraybases[12];
 
-// Used in the Vulkan backend
-
-NativeVertexFormatMap* GetNativeVertexFormatMap()
-{
-  return &s_native_vertex_map;
-}
-
 void Init()
 {
   MarkAllDirty();

Source/Core/VideoCommon/VertexLoaderManager.h

@@ -24,8 +24,6 @@ void Clear();
 
 void MarkAllDirty();
 
-NativeVertexFormatMap* GetNativeVertexFormatMap();
-
 // Creates or obtains a pointer to a VertexFormat representing decl.
 // If this results in a VertexFormat being created, if the game later uses a matching vertex
 // declaration, the one that was previously created will be used.