Merge pull request #7999 from stenzek/pipeline-cache-data

Implement pipeline data cache for OpenGL and D3D12

Source/Core/VideoBackends/D3D/Render.cpp

@@ -127,7 +127,9 @@ std::unique_ptr<AbstractShader> Renderer::CreateShaderFromBinary(ShaderStage sta
   return DXShader::CreateFromBytecode(stage, DXShader::CreateByteCode(data, length));
 }
 
-std::unique_ptr<AbstractPipeline> Renderer::CreatePipeline(const AbstractPipelineConfig& config)
+std::unique_ptr<AbstractPipeline> Renderer::CreatePipeline(const AbstractPipelineConfig& config,
+                                                           const void* cache_data,
+                                                           size_t cache_data_length)
 {
   return DXPipeline::Create(config);
 }

Source/Core/VideoBackends/D3D/Render.h

@@ -34,7 +34,9 @@ public:
                                                          size_t length) override;
   std::unique_ptr<NativeVertexFormat>
   CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl) override;
-  std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config) override;
+  std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config,
+                                                   const void* cache_data = nullptr,
+                                                   size_t cache_data_length = 0) override;
   std::unique_ptr<AbstractFramebuffer>
   CreateFramebuffer(AbstractTexture* color_attachment, AbstractTexture* depth_attachment) override;
 

Source/Core/VideoBackends/D3D/main.cpp

@@ -82,6 +82,8 @@ void VideoBackend::FillBackendInfo()
   g_Config.backend_info.bSupportsFragmentStoresAndAtomics = true;
   g_Config.backend_info.bSupportsGSInstancing = true;
   g_Config.backend_info.bSupportsSSAA = true;
+  g_Config.backend_info.bSupportsShaderBinaries = true;
+  g_Config.backend_info.bSupportsPipelineCacheData = false;
 
   g_Config.backend_info.Adapters = D3DCommon::GetAdapterNames();
   g_Config.backend_info.AAModes = D3D::GetAAModes(g_Config.iAdapter);

Source/Core/VideoBackends/D3D12/DXPipeline.cpp

@@ -156,7 +156,8 @@ static void GetD3DBlendDesc(D3D12_BLEND_DESC* desc, const BlendingState& state)
   }
 }
 
-std::unique_ptr<DXPipeline> DXPipeline::Create(const AbstractPipelineConfig& config)
+std::unique_ptr<DXPipeline> DXPipeline::Create(const AbstractPipelineConfig& config,
+                                               const void* cache_data, size_t cache_data_size)
 {
   DEBUG_ASSERT(config.vertex_shader && config.pixel_shader);
 
@@ -202,16 +203,36 @@ std::unique_ptr<DXPipeline> DXPipeline::Create(const AbstractPipelineConfig& con
         D3DCommon::GetDSVFormatForAbstractFormat(config.framebuffer_state.depth_texture_format);
   desc.SampleDesc.Count = config.framebuffer_state.samples;
   desc.NodeMask = 1;
+  desc.CachedPSO.pCachedBlob = cache_data;
+  desc.CachedPSO.CachedBlobSizeInBytes = cache_data_size;
 
   ID3D12PipelineState* pso;
   HRESULT hr = g_dx_context->GetDevice()->CreateGraphicsPipelineState(&desc, IID_PPV_ARGS(&pso));
-  CHECK(SUCCEEDED(hr), "Create PSO");
   if (FAILED(hr))
+  {
+    WARN_LOG(VIDEO, "CreateGraphicsPipelineState() %sfailed with HRESULT %08X",
+             cache_data ? "with cache data " : "", hr);
     return nullptr;
+  }
 
   const bool use_integer_rtv =
       !config.blending_state.blendenable && config.blending_state.logicopenable;
   return std::make_unique<DXPipeline>(pso, desc.pRootSignature, config.usage,
                                       GetD3DTopology(config.rasterization_state), use_integer_rtv);
 }
+
+AbstractPipeline::CacheData DXPipeline::GetCacheData() const
+{
+  ComPtr<ID3DBlob> blob;
+  HRESULT hr = m_pipeline->GetCachedBlob(&blob);
+  if (FAILED(hr))
+  {
+    WARN_LOG(VIDEO, "ID3D12Pipeline::GetCachedBlob() failed with HRESULT %08X", hr);
+    return {};
+  }
+
+  CacheData data(blob->GetBufferSize());
+  std::memcpy(data.data(), blob->GetBufferPointer(), blob->GetBufferSize());
+  return data;
+}
 }  // namespace DX12

Source/Core/VideoBackends/D3D12/DXPipeline.h

@@ -19,7 +19,8 @@ public:
              bool use_integer_rtv);
   ~DXPipeline() override;
 
-  static std::unique_ptr<DXPipeline> Create(const AbstractPipelineConfig& config);
+  static std::unique_ptr<DXPipeline> Create(const AbstractPipelineConfig& config,
+                                            const void* cache_data, size_t cache_data_size);
 
   ID3D12PipelineState* GetPipeline() const { return m_pipeline; }
   ID3D12RootSignature* GetRootSignature() const { return m_root_signature; }
@@ -27,6 +28,8 @@ public:
   D3D12_PRIMITIVE_TOPOLOGY GetPrimitiveTopology() const { return m_primitive_topology; }
   bool UseIntegerRTV() const { return m_use_integer_rtv; }
 
+  CacheData GetCacheData() const override;
+
 private:
   ID3D12PipelineState* m_pipeline;
   ID3D12RootSignature* m_root_signature;

Source/Core/VideoBackends/D3D12/Renderer.cpp

@@ -99,9 +99,11 @@ Renderer::CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl)
   return std::make_unique<DXVertexFormat>(vtx_decl);
 }
 
-std::unique_ptr<AbstractPipeline> Renderer::CreatePipeline(const AbstractPipelineConfig& config)
+std::unique_ptr<AbstractPipeline> Renderer::CreatePipeline(const AbstractPipelineConfig& config,
+                                                           const void* cache_data,
+                                                           size_t cache_data_length)
 {
-  return DXPipeline::Create(config);
+  return DXPipeline::Create(config, cache_data, cache_data_length);
 }
 
 u16 Renderer::BBoxRead(int index)

Source/Core/VideoBackends/D3D12/Renderer.h

@@ -41,7 +41,9 @@ public:
                                                          size_t length) override;
   std::unique_ptr<NativeVertexFormat>
   CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl) override;
-  std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config) override;
+  std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config,
+                                                   const void* cache_data = nullptr,
+                                                   size_t cache_data_length = 0) override;
 
   u16 BBoxRead(int index) override;
   void BBoxWrite(int index, u16 value) override;

Source/Core/VideoBackends/D3D12/VideoBackend.cpp

@@ -80,6 +80,8 @@ void VideoBackend::FillBackendInfo()
   g_Config.backend_info.bSupportsPartialDepthCopies = false;
   g_Config.backend_info.Adapters = D3DCommon::GetAdapterNames();
   g_Config.backend_info.AAModes = DXContext::GetAAModes(g_Config.iAdapter);
+  g_Config.backend_info.bSupportsShaderBinaries = true;
+  g_Config.backend_info.bSupportsPipelineCacheData = true;
 
   // We can only check texture support once we have a device.
   if (g_dx_context)

Source/Core/VideoBackends/D3DCommon/Shader.cpp

@@ -23,11 +23,6 @@ Shader::Shader(ShaderStage stage, BinaryData bytecode)
 
 Shader::~Shader() = default;
 
-bool Shader::HasBinary() const
-{
-  return true;
-}
-
 AbstractShader::BinaryData Shader::GetBinary() const
 {
   return m_bytecode;

Source/Core/VideoBackends/D3DCommon/Shader.h

@@ -16,7 +16,6 @@ public:
 
   const BinaryData& GetByteCode() const { return m_bytecode; }
 
-  bool HasBinary() const override;
   BinaryData GetBinary() const override;
 
   static bool CompileShader(D3D_FEATURE_LEVEL feature_level, BinaryData* out_bytecode,

Source/Core/VideoBackends/Null/NullBackend.cpp

@@ -52,6 +52,8 @@ void VideoBackend::InitBackendInfo()
   g_Config.backend_info.bSupportsLogicOp = false;
   g_Config.backend_info.bSupportsLargePoints = false;
   g_Config.backend_info.bSupportsPartialDepthCopies = false;
+  g_Config.backend_info.bSupportsShaderBinaries = false;
+  g_Config.backend_info.bSupportsPipelineCacheData = false;
 
   // aamodes: We only support 1 sample, so no MSAA
   g_Config.backend_info.Adapters.clear();

Source/Core/VideoBackends/Null/Render.cpp

@@ -46,9 +46,6 @@ class NullShader final : public AbstractShader
 public:
   explicit NullShader(ShaderStage stage) : AbstractShader(stage) {}
   ~NullShader() = default;
-
-  bool HasBinary() const override { return false; }
-  BinaryData GetBinary() const override { return {}; }
 };
 
 std::unique_ptr<AbstractShader> Renderer::CreateShaderFromSource(ShaderStage stage,
@@ -70,7 +67,9 @@ public:
   ~NullPipeline() override = default;
 };
 
-std::unique_ptr<AbstractPipeline> Renderer::CreatePipeline(const AbstractPipelineConfig& config)
+std::unique_ptr<AbstractPipeline> Renderer::CreatePipeline(const AbstractPipelineConfig& config,
+                                                           const void* cache_data,
+                                                           size_t cache_data_length)
 {
   return std::make_unique<NullPipeline>();
 }

Source/Core/VideoBackends/Null/Render.h

@@ -28,7 +28,9 @@ public:
                                                          size_t length) override;
   std::unique_ptr<NativeVertexFormat>
   CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl) override;
-  std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config) override;
+  std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config,
+                                                   const void* cache_data = nullptr,
+                                                   size_t cache_data_length = 0) override;
 
   u32 AccessEFB(EFBAccessType type, u32 x, u32 y, u32 poke_data) override { return 0; }
   void PokeEFB(EFBAccessType type, const EfbPokeData* points, size_t num_points) override {}

Source/Core/VideoBackends/OGL/OGLPipeline.cpp

@@ -9,6 +9,7 @@
 #include "VideoBackends/OGL/ProgramShaderCache.h"
 #include "VideoBackends/OGL/Render.h"
 #include "VideoBackends/OGL/VertexManager.h"
+#include "VideoCommon/VideoConfig.h"
 
 namespace OGL
 {
@@ -31,7 +32,7 @@ static GLenum MapToGLPrimitive(PrimitiveType primitive_type)
 OGLPipeline::OGLPipeline(const GLVertexFormat* vertex_format,
                          const RasterizationState& rasterization_state,
                          const DepthState& depth_state, const BlendingState& blending_state,
-                         const PipelineProgram* program, GLuint gl_primitive)
+                         PipelineProgram* program, GLuint gl_primitive)
     : m_vertex_format(vertex_format), m_rasterization_state(rasterization_state),
       m_depth_state(depth_state), m_blending_state(blending_state), m_program(program),
       m_gl_primitive(gl_primitive)
@@ -44,13 +45,47 @@ OGLPipeline::~OGLPipeline()
   ProgramShaderCache::ReleasePipelineProgram(m_program);
 }
 
-std::unique_ptr<OGLPipeline> OGLPipeline::Create(const AbstractPipelineConfig& config)
+AbstractPipeline::CacheData OGLPipeline::GetCacheData() const
 {
-  const PipelineProgram* program = ProgramShaderCache::GetPipelineProgram(
+  // More than one pipeline can share the same shaders. To avoid bloating the cache with multiple
+  // copies of the same program combination, we set a flag on the program object so that it can't
+  // be retrieved again. When booting, the pipeline cache is loaded in-order, so the additional
+  // pipelines which use the program combination will re-use the already-created object.
+  if (!g_ActiveConfig.backend_info.bSupportsPipelineCacheData || m_program->binary_retrieved)
+    return {};
+
+  GLint program_size = 0;
+  glGetProgramiv(m_program->shader.glprogid, GL_PROGRAM_BINARY_LENGTH, &program_size);
+  if (program_size == 0)
+    return {};
+
+  // Clear any existing error.
+  glGetError();
+
+  // We pack the format at the start of the buffer.
+  CacheData data(program_size + sizeof(u32));
+  GLsizei data_size = 0;
+  GLenum program_format = 0;
+  glGetProgramBinary(m_program->shader.glprogid, program_size, &data_size, &program_format,
+                     &data[sizeof(u32)]);
+  if (glGetError() != GL_NO_ERROR || data_size == 0)
+    return {};
+
+  u32 program_format_u32 = static_cast<u32>(program_format);
+  std::memcpy(&data[0], &program_format_u32, sizeof(u32));
+  data.resize(data_size + sizeof(u32));
+  m_program->binary_retrieved = true;
+  return data;
+}
+
+std::unique_ptr<OGLPipeline> OGLPipeline::Create(const AbstractPipelineConfig& config,
+                                                 const void* cache_data, size_t cache_data_size)
+{
+  PipelineProgram* program = ProgramShaderCache::GetPipelineProgram(
       static_cast<const GLVertexFormat*>(config.vertex_format),
       static_cast<const OGLShader*>(config.vertex_shader),
       static_cast<const OGLShader*>(config.geometry_shader),
-      static_cast<const OGLShader*>(config.pixel_shader));
+      static_cast<const OGLShader*>(config.pixel_shader), cache_data, cache_data_size);
   if (!program)
     return nullptr;
 

Source/Core/VideoBackends/OGL/OGLPipeline.h

@@ -18,7 +18,7 @@ class OGLPipeline final : public AbstractPipeline
 public:
   explicit OGLPipeline(const GLVertexFormat* vertex_format,
                        const RasterizationState& rasterization_state, const DepthState& depth_state,
-                       const BlendingState& blending_state, const PipelineProgram* program,
+                       const BlendingState& blending_state, PipelineProgram* program,
                        GLenum gl_primitive);
   ~OGLPipeline() override;
 
@@ -29,14 +29,16 @@ public:
   const PipelineProgram* GetProgram() const { return m_program; }
   bool HasVertexInput() const { return m_vertex_format != nullptr; }
   GLenum GetGLPrimitive() const { return m_gl_primitive; }
-  static std::unique_ptr<OGLPipeline> Create(const AbstractPipelineConfig& config);
+  CacheData GetCacheData() const override;
+  static std::unique_ptr<OGLPipeline> Create(const AbstractPipelineConfig& config,
+                                             const void* cache_data, size_t cache_data_size);
 
 private:
   const GLVertexFormat* m_vertex_format;
   RasterizationState m_rasterization_state;
   DepthState m_depth_state;
   BlendingState m_blending_state;
-  const PipelineProgram* m_program;
+  PipelineProgram* m_program;
   GLenum m_gl_primitive;
 };
 

Source/Core/VideoBackends/OGL/OGLShader.cpp

@@ -44,17 +44,6 @@ OGLShader::~OGLShader()
     glDeleteProgram(m_gl_compute_program_id);
 }
 
-bool OGLShader::HasBinary() const
-{
-  // NOTE: GL shaders do not have binaries, programs do.
-  return false;
-}
-
-AbstractShader::BinaryData OGLShader::GetBinary() const
-{
-  return {};
-}
-
 std::unique_ptr<OGLShader> OGLShader::CreateFromSource(ShaderStage stage, const char* source,
                                                        size_t length)
 {

Source/Core/VideoBackends/OGL/OGLShader.h

@@ -24,8 +24,6 @@ public:
   GLenum GetGLShaderType() const { return m_type; }
   GLuint GetGLShaderID() const { return m_gl_id; }
   GLuint GetGLComputeProgramID() const { return m_gl_compute_program_id; }
-  bool HasBinary() const override;
-  BinaryData GetBinary() const override;
 
   static std::unique_ptr<OGLShader> CreateFromSource(ShaderStage stage, const char* source,
                                                      size_t length);

Source/Core/VideoBackends/OGL/ProgramShaderCache.cpp

@@ -312,7 +312,7 @@ bool ProgramShaderCache::CompileShader(SHADER& shader, const std::string& vcode,
   if (shader.gsid)
     glAttachShader(shader.glprogid, shader.gsid);
 
-  if (g_ogl_config.bSupportsGLSLCache)
+  if (g_ActiveConfig.backend_info.bSupportsPipelineCacheData)
     glProgramParameteri(shader.glprogid, GL_PROGRAM_BINARY_RETRIEVABLE_HINT, GL_TRUE);
 
   shader.SetProgramBindings(false);
@@ -562,10 +562,12 @@ void ProgramShaderCache::InvalidateLastProgram()
   CurrentProgram = 0;
 }
 
-const PipelineProgram* ProgramShaderCache::GetPipelineProgram(const GLVertexFormat* vertex_format,
+PipelineProgram* ProgramShaderCache::GetPipelineProgram(const GLVertexFormat* vertex_format,
-                                                              const OGLShader* vertex_shader,
+                                                        const OGLShader* vertex_shader,
-                                                              const OGLShader* geometry_shader,
+                                                        const OGLShader* geometry_shader,
-                                                              const OGLShader* pixel_shader)
+                                                        const OGLShader* pixel_shader,
+                                                        const void* cache_data,
+                                                        size_t cache_data_size)
 {
   PipelineProgramKey key = {vertex_shader ? vertex_shader->GetID() : 0,
                             geometry_shader ? geometry_shader->GetID() : 0,
@@ -580,39 +582,69 @@ const PipelineProgram* ProgramShaderCache::GetPipelineProgram(const GLVertexForm
     }
   }
 
-  // We temporarily change the vertex array to the pipeline's vertex format.
-  // This can prevent the NVIDIA OpenGL driver from recompiling on first use.
-  GLuint vao = vertex_format ? vertex_format->VAO : s_attributeless_VAO;
-  if (s_is_shared_context || vao != s_last_VAO)
-    glBindVertexArray(vao);
-
   std::unique_ptr<PipelineProgram> prog = std::make_unique<PipelineProgram>();
   prog->key = key;
-
-  // Attach shaders.
-  ASSERT(vertex_shader && vertex_shader->GetStage() == ShaderStage::Vertex);
-  ASSERT(pixel_shader && pixel_shader->GetStage() == ShaderStage::Pixel);
   prog->shader.glprogid = glCreateProgram();
-  glAttachShader(prog->shader.glprogid, vertex_shader->GetGLShaderID());
+
-  glAttachShader(prog->shader.glprogid, pixel_shader->GetGLShaderID());
+  // Use the cache data, if present. If this fails, we want to return an error, so the shader cache
-  if (geometry_shader)
+  // doesn't attempt to use the same binary data in the future.
+  if (cache_data_size >= sizeof(u32))
   {
-    ASSERT(geometry_shader->GetStage() == ShaderStage::Geometry);
+    u32 program_binary_type;
-    glAttachShader(prog->shader.glprogid, geometry_shader->GetGLShaderID());
+    std::memcpy(&program_binary_type, cache_data, sizeof(u32));
+    glProgramBinary(prog->shader.glprogid, static_cast<GLenum>(program_binary_type),
+                    static_cast<const u8*>(cache_data) + sizeof(u32),
+                    static_cast<GLsizei>(cache_data_size - sizeof(u32)));
+
+    // Check the link status. If this fails, it means the binary was invalid.
+    GLint link_status;
+    glGetProgramiv(prog->shader.glprogid, GL_LINK_STATUS, &link_status);
+    if (link_status != GL_TRUE)
+    {
+      WARN_LOG(VIDEO, "Failed to create GL program from program binary.");
+      prog->shader.Destroy();
+      return nullptr;
+    }
+
+    // We don't want to retrieve this binary and duplicate entries in the cache again.
+    // See the explanation in OGLPipeline.cpp.
+    prog->binary_retrieved = true;
   }
-
+  else
-  // Link program.
-  prog->shader.SetProgramBindings(false);
-  glLinkProgram(prog->shader.glprogid);
-
-  // Restore VAO binding after linking.
-  if (!s_is_shared_context && vao != s_last_VAO)
-    glBindVertexArray(s_last_VAO);
-
-  if (!ProgramShaderCache::CheckProgramLinkResult(prog->shader.glprogid, {}, {}, {}))
   {
-    prog->shader.Destroy();
+    // We temporarily change the vertex array to the pipeline's vertex format.
-    return nullptr;
+    // This can prevent the NVIDIA OpenGL driver from recompiling on first use.
+    GLuint vao = vertex_format ? vertex_format->VAO : s_attributeless_VAO;
+    if (s_is_shared_context || vao != s_last_VAO)
+      glBindVertexArray(vao);
+
+    // Attach shaders.
+    ASSERT(vertex_shader && vertex_shader->GetStage() == ShaderStage::Vertex);
+    ASSERT(pixel_shader && pixel_shader->GetStage() == ShaderStage::Pixel);
+    glAttachShader(prog->shader.glprogid, vertex_shader->GetGLShaderID());
+    glAttachShader(prog->shader.glprogid, pixel_shader->GetGLShaderID());
+    if (geometry_shader)
+    {
+      ASSERT(geometry_shader->GetStage() == ShaderStage::Geometry);
+      glAttachShader(prog->shader.glprogid, geometry_shader->GetGLShaderID());
+    }
+
+    if (g_ActiveConfig.backend_info.bSupportsPipelineCacheData)
+      glProgramParameteri(prog->shader.glprogid, GL_PROGRAM_BINARY_RETRIEVABLE_HINT, GL_TRUE);
+
+    // Link program.
+    prog->shader.SetProgramBindings(false);
+    glLinkProgram(prog->shader.glprogid);
+
+    // Restore VAO binding after linking.
+    if (!s_is_shared_context && vao != s_last_VAO)
+      glBindVertexArray(s_last_VAO);
+
+    if (!ProgramShaderCache::CheckProgramLinkResult(prog->shader.glprogid, {}, {}, {}))
+    {
+      prog->shader.Destroy();
+      return nullptr;
+    }
   }
 
   // Lock to insert. A duplicate program may have been created in the meantime.
@@ -639,16 +671,17 @@ const PipelineProgram* ProgramShaderCache::GetPipelineProgram(const GLVertexForm
   return ip.first->second.get();
 }
 
-void ProgramShaderCache::ReleasePipelineProgram(const PipelineProgram* prog)
+void ProgramShaderCache::ReleasePipelineProgram(PipelineProgram* prog)
 {
+  if (--prog->reference_count > 0)
+    return;
+
+  prog->shader.Destroy();
+
+  std::lock_guard<std::mutex> guard(s_pipeline_program_lock);
   auto iter = s_pipeline_programs.find(prog->key);
   ASSERT(iter != s_pipeline_programs.end() && prog == iter->second.get());
-
+  s_pipeline_programs.erase(iter);
-  if (--iter->second->reference_count == 0)
-  {
-    iter->second->shader.Destroy();
-    s_pipeline_programs.erase(iter);
-  }
 }
 
 void ProgramShaderCache::CreateHeader()

Source/Core/VideoBackends/OGL/ProgramShaderCache.h

@@ -63,6 +63,7 @@ struct PipelineProgram
   PipelineProgramKey key;
   SHADER shader;
   std::atomic_size_t reference_count{1};
+  bool binary_retrieved = false;
 };
 
 class ProgramShaderCache
@@ -97,11 +98,12 @@ public:
   // pipeline do not match the pipeline configuration.
   static u64 GenerateShaderID();
 
-  static const PipelineProgram* GetPipelineProgram(const GLVertexFormat* vertex_format,
+  static PipelineProgram* GetPipelineProgram(const GLVertexFormat* vertex_format,
-                                                   const OGLShader* vertex_shader,
+                                             const OGLShader* vertex_shader,
-                                                   const OGLShader* geometry_shader,
+                                             const OGLShader* geometry_shader,
-                                                   const OGLShader* pixel_shader);
+                                             const OGLShader* pixel_shader, const void* cache_data,
-  static void ReleasePipelineProgram(const PipelineProgram* prog);
+                                             size_t cache_data_size);
+  static void ReleasePipelineProgram(PipelineProgram* prog);
 
 private:
   typedef std::unordered_map<PipelineProgramKey, std::unique_ptr<PipelineProgram>,

Source/Core/VideoBackends/OGL/Render.cpp

@@ -350,6 +350,7 @@ Renderer::Renderer(std::unique_ptr<GLContext> main_gl_context, float backbuffer_
   }
 
   bool bSuccess = true;
+  bool supports_glsl_cache = false;
 
   g_ogl_config.gl_vendor = (const char*)glGetString(GL_VENDOR);
   g_ogl_config.gl_renderer = (const char*)glGetString(GL_RENDERER);
@@ -466,7 +467,7 @@ Renderer::Renderer(std::unique_ptr<GLContext> main_gl_context, float backbuffer_
       GLExtensions::Supports("GL_ARB_gpu_shader5");
 
   g_ogl_config.bIsES = m_main_gl_context->IsGLES();
-  g_ogl_config.bSupportsGLSLCache = GLExtensions::Supports("GL_ARB_get_program_binary");
+  supports_glsl_cache = GLExtensions::Supports("GL_ARB_get_program_binary");
   g_ogl_config.bSupportsGLPinnedMemory = GLExtensions::Supports("GL_AMD_pinned_memory");
   g_ogl_config.bSupportsGLSync = GLExtensions::Supports("GL_ARB_sync");
   g_ogl_config.bSupportsGLBaseVertex = GLExtensions::Supports("GL_ARB_draw_elements_base_vertex") ||
@@ -507,7 +508,7 @@ Renderer::Renderer(std::unique_ptr<GLContext> main_gl_context, float backbuffer_
                                                 EsTexbufType::TexbufExt :
                                                 EsTexbufType::TexbufNone;
 
-    g_ogl_config.bSupportsGLSLCache = true;
+    supports_glsl_cache = true;
     g_ogl_config.bSupportsGLSync = true;
 
     // TODO: Implement support for GL_EXT_clip_cull_distance when there is an extension for
@@ -675,6 +676,16 @@ Renderer::Renderer(std::unique_ptr<GLContext> main_gl_context, float backbuffer_
   g_Config.backend_info.bSupportsBackgroundCompiling =
       !DriverDetails::HasBug(DriverDetails::BUG_SHARED_CONTEXT_SHADER_COMPILATION);
 
+  // Program binaries are supported on GL4.1+, ARB_get_program_binary, or ES3.
+  if (supports_glsl_cache)
+  {
+    // We need to check the number of formats supported. If zero, don't bother getting the binaries.
+    GLint num_formats = 0;
+    glGetIntegerv(GL_NUM_PROGRAM_BINARY_FORMATS, &num_formats);
+    supports_glsl_cache = num_formats > 0;
+  }
+  g_Config.backend_info.bSupportsPipelineCacheData = supports_glsl_cache;
+
   if (g_ogl_config.bSupportsDebug)
   {
     if (GLExtensions::Supports("GL_KHR_debug"))
@@ -739,7 +750,7 @@ Renderer::Renderer(std::unique_ptr<GLContext> main_gl_context, float backbuffer_
            g_ActiveConfig.backend_info.bSupportsPrimitiveRestart ? "" : "PrimitiveRestart ",
            g_ActiveConfig.backend_info.bSupportsEarlyZ ? "" : "EarlyZ ",
            g_ogl_config.bSupportsGLPinnedMemory ? "" : "PinnedMemory ",
-           g_ogl_config.bSupportsGLSLCache ? "" : "ShaderCache ",
+           supports_glsl_cache ? "" : "ShaderCache ",
            g_ogl_config.bSupportsGLBaseVertex ? "" : "BaseVertex ",
            g_ogl_config.bSupportsGLBufferStorage ? "" : "BufferStorage ",
            g_ogl_config.bSupportsGLSync ? "" : "Sync ", g_ogl_config.bSupportsMSAA ? "" : "MSAA ",
@@ -828,9 +839,11 @@ std::unique_ptr<AbstractShader> Renderer::CreateShaderFromBinary(ShaderStage sta
   return nullptr;
 }
 
-std::unique_ptr<AbstractPipeline> Renderer::CreatePipeline(const AbstractPipelineConfig& config)
+std::unique_ptr<AbstractPipeline> Renderer::CreatePipeline(const AbstractPipelineConfig& config,
+                                                           const void* cache_data,
+                                                           size_t cache_data_length)
 {
-  return OGLPipeline::Create(config);
+  return OGLPipeline::Create(config, cache_data, cache_data_length);
 }
 
 void Renderer::SetScissorRect(const MathUtil::Rectangle<int>& rc)

Source/Core/VideoBackends/OGL/Render.h

@@ -48,7 +48,6 @@ enum class EsFbFetchType
 struct VideoConfig
 {
   bool bIsES;
-  bool bSupportsGLSLCache;
   bool bSupportsGLPinnedMemory;
   bool bSupportsGLSync;
   bool bSupportsGLBaseVertex;
@@ -102,7 +101,9 @@ public:
                                                          size_t length) override;
   std::unique_ptr<NativeVertexFormat>
   CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl) override;
-  std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config) override;
+  std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config,
+                                                   const void* cache_data = nullptr,
+                                                   size_t cache_data_length = 0) override;
   std::unique_ptr<AbstractFramebuffer>
   CreateFramebuffer(AbstractTexture* color_attachment, AbstractTexture* depth_attachment) override;
 

Source/Core/VideoBackends/OGL/main.cpp

@@ -90,6 +90,8 @@ void VideoBackend::InitBackendInfo()
   g_Config.backend_info.bSupportsCopyToVram = true;
   g_Config.backend_info.bSupportsLargePoints = true;
   g_Config.backend_info.bSupportsPartialDepthCopies = true;
+  g_Config.backend_info.bSupportsShaderBinaries = false;
+  g_Config.backend_info.bSupportsPipelineCacheData = false;
 
   // TODO: There is a bug here, if texel buffers or SSBOs/atomics are not supported the graphics
   // options will show the option when it is not supported. The only way around this would be

Source/Core/VideoBackends/Software/SWRenderer.cpp

@@ -64,7 +64,6 @@ public:
   explicit SWShader(ShaderStage stage) : AbstractShader(stage) {}
   ~SWShader() = default;
 
-  bool HasBinary() const override { return false; }
   BinaryData GetBinary() const override { return {}; }
 };
 
@@ -87,7 +86,9 @@ public:
   ~SWPipeline() override = default;
 };
 
-std::unique_ptr<AbstractPipeline> SWRenderer::CreatePipeline(const AbstractPipelineConfig& config)
+std::unique_ptr<AbstractPipeline> SWRenderer::CreatePipeline(const AbstractPipelineConfig& config,
+                                                             const void* cache_data,
+                                                             size_t cache_data_length)
 {
   return std::make_unique<SWPipeline>();
 }

Source/Core/VideoBackends/Software/SWRenderer.h

@@ -33,7 +33,9 @@ public:
                                                          size_t length) override;
   std::unique_ptr<NativeVertexFormat>
   CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl) override;
-  std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config) override;
+  std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config,
+                                                   const void* cache_data = nullptr,
+                                                   size_t cache_data_length = 0) override;
 
   u32 AccessEFB(EFBAccessType type, u32 x, u32 y, u32 poke_data) override;
   void PokeEFB(EFBAccessType type, const EfbPokeData* points, size_t num_points) override {}

Source/Core/VideoBackends/Software/SWmain.cpp

@@ -74,6 +74,8 @@ void VideoSoftware::InitBackendInfo()
   g_Config.backend_info.bSupportsFramebufferFetch = false;
   g_Config.backend_info.bSupportsBackgroundCompiling = false;
   g_Config.backend_info.bSupportsLogicOp = true;
+  g_Config.backend_info.bSupportsShaderBinaries = false;
+  g_Config.backend_info.bSupportsPipelineCacheData = false;
 
   // aamodes
   g_Config.backend_info.AAModes = {1};

Source/Core/VideoBackends/Vulkan/Renderer.cpp

@@ -112,7 +112,9 @@ Renderer::CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl)
   return std::make_unique<VertexFormat>(vtx_decl);
 }
 
-std::unique_ptr<AbstractPipeline> Renderer::CreatePipeline(const AbstractPipelineConfig& config)
+std::unique_ptr<AbstractPipeline> Renderer::CreatePipeline(const AbstractPipelineConfig& config,
+                                                           const void* cache_data,
+                                                           size_t cache_data_length)
 {
   return VKPipeline::Create(config);
 }

Source/Core/VideoBackends/Vulkan/Renderer.h

@@ -48,7 +48,9 @@ public:
                                                          size_t length) override;
   std::unique_ptr<NativeVertexFormat>
   CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl) override;
-  std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config) override;
+  std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config,
+                                                   const void* cache_data = nullptr,
+                                                   size_t cache_data_length = 0) override;
 
   SwapChain* GetSwapChain() const { return m_swap_chain.get(); }
   BoundingBox* GetBoundingBox() const { return m_bounding_box.get(); }

Source/Core/VideoBackends/Vulkan/VKShader.cpp

@@ -32,12 +32,6 @@ VKShader::~VKShader()
     vkDestroyPipeline(g_vulkan_context->GetDevice(), m_compute_pipeline, nullptr);
 }
 
-bool VKShader::HasBinary() const
-{
-  ASSERT(!m_spv.empty());
-  return true;
-}
-
 AbstractShader::BinaryData VKShader::GetBinary() const
 {
   BinaryData ret(sizeof(u32) * m_spv.size());

Source/Core/VideoBackends/Vulkan/VKShader.h

@@ -23,7 +23,6 @@ public:
 
   VkShaderModule GetShaderModule() const { return m_module; }
   VkPipeline GetComputePipeline() const { return m_compute_pipeline; }
-  bool HasBinary() const override;
   BinaryData GetBinary() const override;
 
   static std::unique_ptr<VKShader> CreateFromSource(ShaderStage stage, const char* source,

Source/Core/VideoBackends/Vulkan/VulkanContext.cpp

@@ -266,6 +266,8 @@ void VulkanContext::PopulateBackendInfo(VideoConfig* config)
   config->backend_info.bSupportsGPUTextureDecoding = true;    // Assumed support.
   config->backend_info.bSupportsBitfield = true;              // Assumed support.
   config->backend_info.bSupportsPartialDepthCopies = true;    // Assumed support.
+  config->backend_info.bSupportsShaderBinaries = true;        // Assumed support.
+  config->backend_info.bSupportsPipelineCacheData = false;    // Handled via pipeline caches.
   config->backend_info.bSupportsDynamicSamplerIndexing = true;     // Assumed support.
   config->backend_info.bSupportsPostProcessing = true;             // Assumed support.
   config->backend_info.bSupportsBackgroundCompiling = true;        // Assumed support.

Source/Core/VideoCommon/AbstractPipeline.h

@@ -75,4 +75,10 @@ class AbstractPipeline
 public:
   AbstractPipeline() = default;
   virtual ~AbstractPipeline() = default;
+
+  // "Cache data" can be used to assist a driver with creating pipelines by using previously
+  // compiled shader ISA. The abstract shaders and creation struct are still required to create
+  // pipeline objects, the cache is optionally used by the driver to speed up compilation.
+  using CacheData = std::vector<u8>;
+  virtual CacheData GetCacheData() const { return {}; }
 };

Source/Core/VideoCommon/AbstractShader.h

@@ -25,9 +25,11 @@ public:
   virtual ~AbstractShader() = default;
 
   ShaderStage GetStage() const { return m_stage; }
+
+  // Shader binaries represent the input source code in a lower-level form. e.g. SPIR-V or DXBC.
+  // The shader source code is not required to create a shader object from the binary.
   using BinaryData = std::vector<u8>;
-  virtual bool HasBinary() const = 0;
+  virtual BinaryData GetBinary() const { return {}; }
-  virtual BinaryData GetBinary() const = 0;
 
 protected:
   ShaderStage m_stage;

Source/Core/VideoCommon/GXPipelineTypes.h

@@ -99,6 +99,16 @@ struct SerializedGXPipelineUid
   u32 depth_state_bits;
   u32 blending_state_bits;
 };
+struct SerializedGXUberPipelineUid
+{
+  PortableVertexDeclaration vertex_decl;
+  UberShader::VertexShaderUid vs_uid;
+  GeometryShaderUid gs_uid;
+  UberShader::PixelShaderUid ps_uid;
+  u32 rasterization_state_bits;
+  u32 depth_state_bits;
+  u32 blending_state_bits;
+};
 #pragma pack(pop)
 
 }  // namespace VideoCommon

Source/Core/VideoCommon/RenderBase.h

@@ -130,8 +130,9 @@ public:
   CreateShaderFromBinary(ShaderStage stage, const void* data, size_t length) = 0;
   virtual std::unique_ptr<NativeVertexFormat>
   CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl) = 0;
-  virtual std::unique_ptr<AbstractPipeline>
+  virtual std::unique_ptr<AbstractPipeline> CreatePipeline(const AbstractPipelineConfig& config,
-  CreatePipeline(const AbstractPipelineConfig& config) = 0;
+                                                           const void* cache_data = nullptr,
+                                                           size_t cache_data_length = 0) = 0;
   std::unique_ptr<AbstractShader> CreateShaderFromSource(ShaderStage stage,
                                                          const std::string& source);
 

Source/Core/VideoCommon/ShaderCache.cpp

@@ -25,8 +25,7 @@ namespace VideoCommon
 ShaderCache::ShaderCache() = default;
 ShaderCache::~ShaderCache()
 {
-  ClearShaderCaches();
+  ClearCaches();
-  ClearPipelineCaches();
 }
 
 bool ShaderCache::Initialize()
@@ -48,7 +47,7 @@ void ShaderCache::InitializeShaderCache()
   // Load shader and UID caches.
   if (g_ActiveConfig.bShaderCache && m_api_type != APIType::Nothing)
   {
-    LoadShaderCaches();
+    LoadCaches();
     LoadPipelineUIDCache();
   }
 
@@ -69,11 +68,10 @@ void ShaderCache::Reload()
 {
   WaitForAsyncCompiler();
   ClosePipelineUIDCache();
-  InvalidateCachedPipelines();
+  ClearCaches();
-  ClearShaderCaches();
 
   if (g_ActiveConfig.bShaderCache)
-    LoadShaderCaches();
+    LoadCaches();
 
   // Switch to the precompiling shader configuration while we rebuild.
   m_async_shader_compiler->ResizeWorkerThreads(g_ActiveConfig.GetShaderPrecompilerThreads());
@@ -140,7 +138,7 @@ const AbstractPipeline* ShaderCache::GetUberPipelineForUid(const GXUberPipelineU
     return it->second.first.get();
 
   std::unique_ptr<AbstractPipeline> pipeline;
-  std::optional<AbstractPipelineConfig> pipeline_config = GetGXUberPipelineConfig(uid);
+  std::optional<AbstractPipelineConfig> pipeline_config = GetGXPipelineConfig(uid);
   if (pipeline_config)
     pipeline = g_renderer->CreatePipeline(*pipeline_config);
   return InsertGXUberPipeline(uid, std::move(pipeline));
@@ -176,8 +174,34 @@ void ShaderCache::WaitForAsyncCompiler()
   }
 }
 
+template <typename SerializedUidType, typename UidType>
+static void SerializePipelineUid(const UidType& uid, SerializedUidType& serialized_uid)
+{
+  // Convert to disk format. Ensure all padding bytes are zero.
+  std::memset(&serialized_uid, 0, sizeof(serialized_uid));
+  serialized_uid.vertex_decl = uid.vertex_format->GetVertexDeclaration();
+  serialized_uid.vs_uid = uid.vs_uid;
+  serialized_uid.gs_uid = uid.gs_uid;
+  serialized_uid.ps_uid = uid.ps_uid;
+  serialized_uid.rasterization_state_bits = uid.rasterization_state.hex;
+  serialized_uid.depth_state_bits = uid.depth_state.hex;
+  serialized_uid.blending_state_bits = uid.blending_state.hex;
+}
+
+template <typename UidType, typename SerializedUidType>
+static void UnserializePipelineUid(const SerializedUidType& uid, UidType& real_uid)
+{
+  real_uid.vertex_format = VertexLoaderManager::GetOrCreateMatchingFormat(uid.vertex_decl);
+  real_uid.vs_uid = uid.vs_uid;
+  real_uid.gs_uid = uid.gs_uid;
+  real_uid.ps_uid = uid.ps_uid;
+  real_uid.rasterization_state.hex = uid.rasterization_state_bits;
+  real_uid.depth_state.hex = uid.depth_state_bits;
+  real_uid.blending_state.hex = uid.blending_state_bits;
+}
+
 template <ShaderStage stage, typename K, typename T>
-static void LoadShaderCache(T& cache, APIType api_type, const char* type, bool include_gameid)
+void ShaderCache::LoadShaderCache(T& cache, APIType api_type, const char* type, bool include_gameid)
 {
   class CacheReader : public LinearDiskCacheReader<K, u8>
   {
@@ -219,38 +243,127 @@ static void LoadShaderCache(T& cache, APIType api_type, const char* type, bool i
 }
 
 template <typename T>
-static void ClearShaderCache(T& cache)
+void ShaderCache::ClearShaderCache(T& cache)
 {
   cache.disk_cache.Sync();
   cache.disk_cache.Close();
   cache.shader_map.clear();
 }
 
-void ShaderCache::LoadShaderCaches()
+template <typename KeyType, typename DiskKeyType, typename T>
+void ShaderCache::LoadPipelineCache(T& cache, LinearDiskCache<DiskKeyType, u8>& disk_cache,
+                                    APIType api_type, const char* type, bool include_gameid)
 {
-  // Ubershader caches, if present.
+  class CacheReader : public LinearDiskCacheReader<DiskKeyType, u8>
-  LoadShaderCache<ShaderStage::Vertex, UberShader::VertexShaderUid>(m_uber_vs_cache, m_api_type,
+  {
-                                                                    "uber-vs", false);
+  public:
-  LoadShaderCache<ShaderStage::Pixel, UberShader::PixelShaderUid>(m_uber_ps_cache, m_api_type,
+    CacheReader(ShaderCache* this_ptr_, T& cache_) : this_ptr(this_ptr_), cache(cache_) {}
-                                                                  "uber-ps", false);
+    bool AnyFailed() const { return failed; }
+    void Read(const DiskKeyType& key, const u8* value, u32 value_size)
+    {
+      KeyType real_uid;
+      UnserializePipelineUid(key, real_uid);
 
-  // We also share geometry shaders, as there aren't many variants.
+      // Skip those which are already compiled.
-  if (m_host_config.backend_geometry_shaders)
+      if (failed || cache.find(real_uid) != cache.end())
-    LoadShaderCache<ShaderStage::Geometry, GeometryShaderUid>(m_gs_cache, m_api_type, "gs", false);
+        return;
 
-  // Specialized shaders, gameid-specific.
+      auto config = this_ptr->GetGXPipelineConfig(real_uid);
-  LoadShaderCache<ShaderStage::Vertex, VertexShaderUid>(m_vs_cache, m_api_type, "specialized-vs",
+      if (!config)
-                                                        true);
+        return;
-  LoadShaderCache<ShaderStage::Pixel, PixelShaderUid>(m_ps_cache, m_api_type, "specialized-ps",
+
-                                                      true);
+      auto pipeline = g_renderer->CreatePipeline(*config, value, value_size);
+      if (!pipeline)
+      {
+        // If any of the pipelines fail to create, consider the cache stale.
+        failed = true;
+        return;
+      }
+
+      auto& entry = cache[real_uid];
+      entry.first = std::move(pipeline);
+      entry.second = false;
+    }
+
+  private:
+    ShaderCache* this_ptr;
+    T& cache;
+    bool failed = false;
+  };
+
+  std::string filename = GetDiskShaderCacheFileName(api_type, type, include_gameid, true);
+  CacheReader reader(this, cache);
+  u32 count = disk_cache.OpenAndRead(filename, reader);
+  INFO_LOG(VIDEO, "Loaded %u cached pipelines from %s", count, filename.c_str());
+
+  // If any of the pipelines in the cache failed to create, it's likely because of a change of
+  // driver version, or system configuration. In this case, when the UID cache picks up the pipeline
+  // later on, we'll write a duplicate entry to the pipeline cache. There's also no point in keeping
+  // the old cache data around, so discard and recreate the disk cache.
+  if (reader.AnyFailed())
+  {
+    WARN_LOG(VIDEO, "Failed to load one or more pipelines from cache '%s'. Discarding.",
+             filename.c_str());
+    disk_cache.Close();
+    File::Delete(filename);
+    disk_cache.OpenAndRead(filename, reader);
+  }
 }
 
-void ShaderCache::ClearShaderCaches()
+template <typename T, typename Y>
+void ShaderCache::ClearPipelineCache(T& cache, Y& disk_cache)
 {
+  disk_cache.Sync();
+  disk_cache.Close();
+
+  // Set the pending flag to false, and destroy the pipeline.
+  for (auto& it : cache)
+  {
+    it.second.first.reset();
+    it.second.second = false;
+  }
+}
+
+void ShaderCache::LoadCaches()
+{
+  // Ubershader caches, if present.
+  if (g_ActiveConfig.backend_info.bSupportsShaderBinaries)
+  {
+    LoadShaderCache<ShaderStage::Vertex, UberShader::VertexShaderUid>(m_uber_vs_cache, m_api_type,
+                                                                      "uber-vs", false);
+    LoadShaderCache<ShaderStage::Pixel, UberShader::PixelShaderUid>(m_uber_ps_cache, m_api_type,
+                                                                    "uber-ps", false);
+
+    // We also share geometry shaders, as there aren't many variants.
+    if (m_host_config.backend_geometry_shaders)
+      LoadShaderCache<ShaderStage::Geometry, GeometryShaderUid>(m_gs_cache, m_api_type, "gs",
+                                                                false);
+
+    // Specialized shaders, gameid-specific.
+    LoadShaderCache<ShaderStage::Vertex, VertexShaderUid>(m_vs_cache, m_api_type, "specialized-vs",
+                                                          true);
+    LoadShaderCache<ShaderStage::Pixel, PixelShaderUid>(m_ps_cache, m_api_type, "specialized-ps",
+                                                        true);
+  }
+
+  if (g_ActiveConfig.backend_info.bSupportsPipelineCacheData)
+  {
+    LoadPipelineCache<GXPipelineUid, SerializedGXPipelineUid>(
+        m_gx_pipeline_cache, m_gx_pipeline_disk_cache, m_api_type, "specialized-pipeline", true);
+    LoadPipelineCache<GXUberPipelineUid, SerializedGXUberPipelineUid>(
+        m_gx_uber_pipeline_cache, m_gx_uber_pipeline_disk_cache, m_api_type, "uber-pipeline",
+        false);
+  }
+}
+
+void ShaderCache::ClearCaches()
+{
+  ClearPipelineCache(m_gx_pipeline_cache, m_gx_pipeline_disk_cache);
   ClearShaderCache(m_vs_cache);
   ClearShaderCache(m_gs_cache);
   ClearShaderCache(m_ps_cache);
 
+  ClearPipelineCache(m_gx_uber_pipeline_cache, m_gx_uber_pipeline_disk_cache);
   ClearShaderCache(m_uber_vs_cache);
   ClearShaderCache(m_uber_ps_cache);
 
@@ -265,37 +378,16 @@ void ShaderCache::CompileMissingPipelines()
   // Queue all uids with a null pipeline for compilation.
   for (auto& it : m_gx_pipeline_cache)
   {
-    if (!it.second.second)
+    if (!it.second.first)
       QueuePipelineCompile(it.first, COMPILE_PRIORITY_SHADERCACHE_PIPELINE);
   }
   for (auto& it : m_gx_uber_pipeline_cache)
   {
-    if (!it.second.second)
+    if (!it.second.first)
       QueueUberPipelineCompile(it.first, COMPILE_PRIORITY_UBERSHADER_PIPELINE);
   }
 }
 
-void ShaderCache::InvalidateCachedPipelines()
-{
-  // Set the pending flag to false, and destroy the pipeline.
-  for (auto& it : m_gx_pipeline_cache)
-  {
-    it.second.first.reset();
-    it.second.second = false;
-  }
-  for (auto& it : m_gx_uber_pipeline_cache)
-  {
-    it.second.first.reset();
-    it.second.second = false;
-  }
-}
-
-void ShaderCache::ClearPipelineCaches()
-{
-  m_gx_pipeline_cache.clear();
-  m_gx_uber_pipeline_cache.clear();
-}
-
 std::unique_ptr<AbstractShader> ShaderCache::CompileVertexShader(const VertexShaderUid& uid) const
 {
   ShaderCode source_code = GenerateVertexShaderCode(m_api_type, m_host_config, uid.GetUidData());
@@ -334,7 +426,7 @@ const AbstractShader* ShaderCache::InsertVertexShader(const VertexShaderUid& uid
 
   if (shader && !entry.shader)
   {
-    if (g_ActiveConfig.bShaderCache && shader->HasBinary())
+    if (g_ActiveConfig.bShaderCache && g_ActiveConfig.backend_info.bSupportsShaderBinaries)
     {
       auto binary = shader->GetBinary();
       if (!binary.empty())
@@ -356,7 +448,7 @@ const AbstractShader* ShaderCache::InsertVertexUberShader(const UberShader::Vert
 
   if (shader && !entry.shader)
   {
-    if (g_ActiveConfig.bShaderCache && shader->HasBinary())
+    if (g_ActiveConfig.bShaderCache && g_ActiveConfig.backend_info.bSupportsShaderBinaries)
     {
       auto binary = shader->GetBinary();
       if (!binary.empty())
@@ -378,7 +470,7 @@ const AbstractShader* ShaderCache::InsertPixelShader(const PixelShaderUid& uid,
 
   if (shader && !entry.shader)
   {
-    if (g_ActiveConfig.bShaderCache && shader->HasBinary())
+    if (g_ActiveConfig.bShaderCache && g_ActiveConfig.backend_info.bSupportsShaderBinaries)
     {
       auto binary = shader->GetBinary();
       if (!binary.empty())
@@ -400,7 +492,7 @@ const AbstractShader* ShaderCache::InsertPixelUberShader(const UberShader::Pixel
 
   if (shader && !entry.shader)
   {
-    if (g_ActiveConfig.bShaderCache && shader->HasBinary())
+    if (g_ActiveConfig.bShaderCache && g_ActiveConfig.backend_info.bSupportsShaderBinaries)
     {
       auto binary = shader->GetBinary();
       if (!binary.empty())
@@ -425,7 +517,7 @@ const AbstractShader* ShaderCache::CreateGeometryShader(const GeometryShaderUid&
 
   if (shader && !entry.shader)
   {
-    if (g_ActiveConfig.bShaderCache && shader->HasBinary())
+    if (g_ActiveConfig.bShaderCache && g_ActiveConfig.backend_info.bSupportsShaderBinaries)
     {
       auto binary = shader->GetBinary();
       if (!binary.empty())
@@ -505,7 +597,7 @@ std::optional<AbstractPipelineConfig> ShaderCache::GetGXPipelineConfig(const GXP
 }
 
 std::optional<AbstractPipelineConfig>
-ShaderCache::GetGXUberPipelineConfig(const GXUberPipelineUid& config)
+ShaderCache::GetGXPipelineConfig(const GXUberPipelineUid& config)
 {
   const AbstractShader* vs;
   auto vs_iter = m_uber_vs_cache.shader_map.find(config.vs_uid);
@@ -551,6 +643,18 @@ const AbstractPipeline* ShaderCache::InsertGXPipeline(const GXPipelineUid& confi
   if (!entry.first && pipeline)
     entry.first = std::move(pipeline);
 
+  if (g_ActiveConfig.bShaderCache)
+  {
+    auto cache_data = entry.first->GetCacheData();
+    if (!cache_data.empty())
+    {
+      SerializedGXPipelineUid disk_uid;
+      SerializePipelineUid(config, disk_uid);
+      m_gx_pipeline_disk_cache.Append(disk_uid, cache_data.data(),
+                                      static_cast<u32>(cache_data.size()));
+    }
+  }
+
   return entry.first.get();
 }
 
@@ -563,6 +667,18 @@ ShaderCache::InsertGXUberPipeline(const GXUberPipelineUid& config,
   if (!entry.first && pipeline)
     entry.first = std::move(pipeline);
 
+  if (g_ActiveConfig.bShaderCache)
+  {
+    auto cache_data = entry.first->GetCacheData();
+    if (!cache_data.empty())
+    {
+      SerializedGXUberPipelineUid disk_uid;
+      SerializePipelineUid(config, disk_uid);
+      m_gx_uber_pipeline_disk_cache.Append(disk_uid, cache_data.data(),
+                                           static_cast<u32>(cache_data.size()));
+    }
+  }
+
   return entry.first.get();
 }
 
@@ -648,14 +764,8 @@ void ShaderCache::ClosePipelineUIDCache()
 
 void ShaderCache::AddSerializedGXPipelineUID(const SerializedGXPipelineUid& uid)
 {
-  GXPipelineUid real_uid = {};
+  GXPipelineUid real_uid;
-  real_uid.vertex_format = VertexLoaderManager::GetOrCreateMatchingFormat(uid.vertex_decl);
+  UnserializePipelineUid(uid, real_uid);
-  real_uid.vs_uid = uid.vs_uid;
-  real_uid.gs_uid = uid.gs_uid;
-  real_uid.ps_uid = uid.ps_uid;
-  real_uid.rasterization_state.hex = uid.rasterization_state_bits;
-  real_uid.depth_state.hex = uid.depth_state_bits;
-  real_uid.blending_state.hex = uid.blending_state_bits;
 
   auto iter = m_gx_pipeline_cache.find(real_uid);
   if (iter != m_gx_pipeline_cache.end())
@@ -671,16 +781,8 @@ void ShaderCache::AppendGXPipelineUID(const GXPipelineUid& config)
   if (!m_gx_pipeline_uid_cache_file.IsOpen())
     return;
 
-  // Convert to disk format. Ensure all padding bytes are zero.
   SerializedGXPipelineUid disk_uid;
-  std::memset(&disk_uid, 0, sizeof(disk_uid));
+  SerializePipelineUid(config, disk_uid);
-  disk_uid.vertex_decl = config.vertex_format->GetVertexDeclaration();
-  disk_uid.vs_uid = config.vs_uid;
-  disk_uid.gs_uid = config.gs_uid;
-  disk_uid.ps_uid = config.ps_uid;
-  disk_uid.rasterization_state_bits = config.rasterization_state.hex;
-  disk_uid.depth_state_bits = config.depth_state.hex;
-  disk_uid.blending_state_bits = config.blending_state.hex;
   if (!m_gx_pipeline_uid_cache_file.WriteBytes(&disk_uid, sizeof(disk_uid)))
   {
     WARN_LOG(VIDEO, "Writing pipeline UID to cache failed, closing file.");
@@ -885,7 +987,7 @@ void ShaderCache::QueueUberPipelineCompile(const GXUberPipelineUid& uid, u32 pri
       // Check if all the stages required for this UberPipeline have been compiled.
       // If not, this work item becomes a no-op, and re-queues the UberPipeline for the next frame.
       if (SetStagesReady())
-        config = shader_cache->GetGXUberPipelineConfig(uid);
+        config = shader_cache->GetGXPipelineConfig(uid);
     }
 
     bool SetStagesReady()

Source/Core/VideoCommon/ShaderCache.h

@@ -111,13 +111,11 @@ private:
   static constexpr size_t NUM_PALETTE_CONVERSION_SHADERS = 3;
 
   void WaitForAsyncCompiler();
-  void LoadShaderCaches();
+  void LoadCaches();
-  void ClearShaderCaches();
+  void ClearCaches();
   void LoadPipelineUIDCache();
   void ClosePipelineUIDCache();
   void CompileMissingPipelines();
-  void InvalidateCachedPipelines();
-  void ClearPipelineCaches();
   void QueueUberShaderPipelines();
   bool CompileSharedPipelines();
 
@@ -149,7 +147,7 @@ private:
                       const RasterizationState& rasterization_state, const DepthState& depth_state,
                       const BlendingState& blending_state);
   std::optional<AbstractPipelineConfig> GetGXPipelineConfig(const GXPipelineUid& uid);
-  std::optional<AbstractPipelineConfig> GetGXUberPipelineConfig(const GXUberPipelineUid& uid);
+  std::optional<AbstractPipelineConfig> GetGXPipelineConfig(const GXUberPipelineUid& uid);
   const AbstractPipeline* InsertGXPipeline(const GXPipelineUid& config,
                                            std::unique_ptr<AbstractPipeline> pipeline);
   const AbstractPipeline* InsertGXUberPipeline(const GXUberPipelineUid& config,
@@ -165,6 +163,17 @@ private:
   void QueuePipelineCompile(const GXPipelineUid& uid, u32 priority);
   void QueueUberPipelineCompile(const GXUberPipelineUid& uid, u32 priority);
 
+  // Populating various caches.
+  template <ShaderStage stage, typename K, typename T>
+  void LoadShaderCache(T& cache, APIType api_type, const char* type, bool include_gameid);
+  template <typename T>
+  void ClearShaderCache(T& cache);
+  template <typename KeyType, typename DiskKeyType, typename T>
+  void LoadPipelineCache(T& cache, LinearDiskCache<DiskKeyType, u8>& disk_cache, APIType api_type,
+                         const char* type, bool include_gameid);
+  template <typename T, typename Y>
+  void ClearPipelineCache(T& cache, Y& disk_cache);
+
   // Priorities for compiling. The lower the value, the sooner the pipeline is compiled.
   // The shader cache is compiled last, as it is the least likely to be required. On demand
   // shaders are always compiled before pending ubershaders, as we want to use the ubershader
@@ -213,6 +222,8 @@ private:
   std::map<GXUberPipelineUid, std::pair<std::unique_ptr<AbstractPipeline>, bool>>
       m_gx_uber_pipeline_cache;
   File::IOFile m_gx_pipeline_uid_cache_file;
+  LinearDiskCache<SerializedGXPipelineUid, u8> m_gx_pipeline_disk_cache;
+  LinearDiskCache<SerializedGXUberPipelineUid, u8> m_gx_uber_pipeline_disk_cache;
 
   // EFB copy to VRAM/RAM pipelines
   std::map<TextureConversionShaderGen::TCShaderUid, std::unique_ptr<AbstractPipeline>>

Source/Core/VideoCommon/VideoConfig.h

@@ -219,6 +219,8 @@ struct VideoConfig final
     bool bSupportsBackgroundCompiling;
     bool bSupportsLargePoints;
     bool bSupportsPartialDepthCopies;
+    bool bSupportsShaderBinaries;
+    bool bSupportsPipelineCacheData;
   } backend_info;
 
   // Utility