Merge pull request #4923 from jroweboy/diskcachelul

Disk Shader Caching

src/video_core/CMakeLists.txt

@@ -31,6 +31,8 @@ add_library(video_core STATIC
     renderer_opengl/gl_resource_manager.h
     renderer_opengl/gl_shader_decompiler.cpp
     renderer_opengl/gl_shader_decompiler.h
+    renderer_opengl/gl_shader_disk_cache.cpp
+    renderer_opengl/gl_shader_disk_cache.h
     renderer_opengl/gl_shader_gen.cpp
     renderer_opengl/gl_shader_gen.h
     renderer_opengl/gl_shader_manager.cpp

src/video_core/rasterizer_interface.h

@@ -4,6 +4,8 @@
 
 #pragma once
 
+#include <atomic>
+#include <functional>
 #include "common/common_types.h"
 #include "core/hw/gpu.h"
 
@@ -17,6 +19,14 @@ struct OutputVertex;
 
 namespace VideoCore {
 
+enum class LoadCallbackStage {
+    Prepare,
+    Decompile,
+    Build,
+    Complete,
+};
+using DiskResourceLoadCallback = std::function<void(LoadCallbackStage, std::size_t, std::size_t)>;
+
 class RasterizerInterface {
 public:
     virtual ~RasterizerInterface() {}
@@ -71,5 +81,8 @@ public:
     virtual bool AccelerateDrawBatch(bool is_indexed) {
         return false;
     }
+
+    virtual void LoadDiskResources(const std::atomic_bool& stop_loading,
+                                   const DiskResourceLoadCallback& callback) {}
 };
 } // namespace VideoCore

src/video_core/renderer_base.h

@@ -6,8 +6,8 @@
 
 #include <memory>
 #include "common/common_types.h"
-#include "core/core.h"
 #include "video_core/rasterizer_interface.h"
+#include "video_core/video_core.h"
 
 namespace Frontend {
 class EmuWindow;
@@ -23,7 +23,7 @@ public:
     virtual ~RendererBase();
 
     /// Initialize the renderer
-    virtual Core::System::ResultStatus Init() = 0;
+    virtual VideoCore::ResultStatus Init() = 0;
 
     /// Shutdown the renderer
     virtual void ShutDown() = 0;

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -171,6 +171,11 @@ RasterizerOpenGL::RasterizerOpenGL(Frontend::EmuWindow& window)
 
 RasterizerOpenGL::~RasterizerOpenGL() {}
 
+void RasterizerOpenGL::LoadDiskResources(const std::atomic_bool& stop_loading,
+                                         const VideoCore::DiskResourceLoadCallback& callback) {
+    shader_program_manager->LoadDiskCache(stop_loading, callback);
+}
+
 void RasterizerOpenGL::SyncEntireState() {
     // Sync fixed function OpenGL state
     SyncClipEnabled();
@@ -378,16 +383,15 @@ void RasterizerOpenGL::SetupVertexArray(u8* array_ptr, GLintptr buffer_offset,
 
 bool RasterizerOpenGL::SetupVertexShader() {
     MICROPROFILE_SCOPE(OpenGL_VS);
-    PicaVSConfig vs_config(Pica::g_state.regs, Pica::g_state.vs);
-    return shader_program_manager->UseProgrammableVertexShader(vs_config, Pica::g_state.vs);
+    return shader_program_manager->UseProgrammableVertexShader(Pica::g_state.regs,
+                                                               Pica::g_state.vs);
 }
 
 bool RasterizerOpenGL::SetupGeometryShader() {
     MICROPROFILE_SCOPE(OpenGL_GS);
     const auto& regs = Pica::g_state.regs;
     if (regs.pipeline.use_gs == Pica::PipelineRegs::UseGS::No) {
-        PicaFixedGSConfig gs_config(regs);
-        shader_program_manager->UseFixedGeometryShader(gs_config);
+        shader_program_manager->UseFixedGeometryShader(regs);
         return true;
     } else {
         LOG_ERROR(Render_OpenGL, "Accelerate draw doesn't support geometry shader");
@@ -1622,8 +1626,7 @@ void RasterizerOpenGL::SamplerInfo::SyncWithConfig(
 }
 
 void RasterizerOpenGL::SetShader() {
-    auto config = PicaFSConfig::BuildFromRegs(Pica::g_state.regs);
-    shader_program_manager->UseFragmentShader(config);
+    shader_program_manager->UseFragmentShader(Pica::g_state.regs);
 }
 
 void RasterizerOpenGL::SyncClipEnabled() {

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -42,6 +42,9 @@ public:
     explicit RasterizerOpenGL(Frontend::EmuWindow& renderer);
     ~RasterizerOpenGL() override;
 
+    void LoadDiskResources(const std::atomic_bool& stop_loading,
+                           const VideoCore::DiskResourceLoadCallback& callback) override;
+
     void AddTriangle(const Pica::Shader::OutputVertex& v0, const Pica::Shader::OutputVertex& v1,
                      const Pica::Shader::OutputVertex& v2) override;
     void DrawTriangles() override;

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -56,7 +56,7 @@ struct Subroutine {
 /// Analyzes shader code and produces a set of subroutines.
 class ControlFlowAnalyzer {
 public:
-    ControlFlowAnalyzer(const ProgramCode& program_code, u32 main_offset)
+    ControlFlowAnalyzer(const Pica::Shader::ProgramCode& program_code, u32 main_offset)
         : program_code(program_code) {
 
         // Recursively finds all subroutines.
@@ -70,7 +70,7 @@ public:
     }
 
 private:
-    const ProgramCode& program_code;
+    const Pica::Shader::ProgramCode& program_code;
     std::set<Subroutine> subroutines;
     std::map<std::pair<u32, u32>, ExitMethod> exit_method_map;
 
@@ -246,8 +246,9 @@ constexpr auto GetSelectorSrc3 = GetSelectorSrc<&SwizzlePattern::GetSelectorSrc3
 
 class GLSLGenerator {
 public:
-    GLSLGenerator(const std::set<Subroutine>& subroutines, const ProgramCode& program_code,
-                  const SwizzleData& swizzle_data, u32 main_offset,
+    GLSLGenerator(const std::set<Subroutine>& subroutines,
+                  const Pica::Shader::ProgramCode& program_code,
+                  const Pica::Shader::SwizzleData& swizzle_data, u32 main_offset,
                   const RegGetter& inputreg_getter, const RegGetter& outputreg_getter,
                   bool sanitize_mul)
         : subroutines(subroutines), program_code(program_code), swizzle_data(swizzle_data),
@@ -865,8 +866,8 @@ private:
 
 private:
     const std::set<Subroutine>& subroutines;
-    const ProgramCode& program_code;
-    const SwizzleData& swizzle_data;
+    const Pica::Shader::ProgramCode& program_code;
+    const Pica::Shader::SwizzleData& swizzle_data;
     const u32 main_offset;
     const RegGetter& inputreg_getter;
     const RegGetter& outputreg_getter;
@@ -888,16 +889,17 @@ bool exec_shader();
 )";
 }
 
-std::optional<std::string> DecompileProgram(const ProgramCode& program_code,
-                                            const SwizzleData& swizzle_data, u32 main_offset,
-                                            const RegGetter& inputreg_getter,
-                                            const RegGetter& outputreg_getter, bool sanitize_mul) {
+std::optional<ProgramResult> DecompileProgram(const Pica::Shader::ProgramCode& program_code,
+                                              const Pica::Shader::SwizzleData& swizzle_data,
+                                              u32 main_offset, const RegGetter& inputreg_getter,
+                                              const RegGetter& outputreg_getter,
+                                              bool sanitize_mul) {
 
     try {
         auto subroutines = ControlFlowAnalyzer(program_code, main_offset).MoveSubroutines();
         GLSLGenerator generator(subroutines, program_code, swizzle_data, main_offset,
                                 inputreg_getter, outputreg_getter, sanitize_mul);
-        return generator.MoveShaderCode();
+        return {ProgramResult{generator.MoveShaderCode()}};
     } catch (const DecompileFail& exception) {
         LOG_INFO(HW_GPU, "Shader decompilation failed: {}", exception.what());
         return {};

src/video_core/renderer_opengl/gl_shader_decompiler.h

@@ -11,15 +11,17 @@
 
 namespace OpenGL::ShaderDecompiler {
 
-using ProgramCode = std::array<u32, Pica::Shader::MAX_PROGRAM_CODE_LENGTH>;
-using SwizzleData = std::array<u32, Pica::Shader::MAX_SWIZZLE_DATA_LENGTH>;
 using RegGetter = std::function<std::string(u32)>;
 
+struct ProgramResult {
+    std::string code;
+};
+
 std::string GetCommonDeclarations();
 
-std::optional<std::string> DecompileProgram(const ProgramCode& program_code,
-                                            const SwizzleData& swizzle_data, u32 main_offset,
-                                            const RegGetter& inputreg_getter,
-                                            const RegGetter& outputreg_getter, bool sanitize_mul);
+std::optional<ProgramResult> DecompileProgram(const Pica::Shader::ProgramCode& program_code,
+                                              const Pica::Shader::SwizzleData& swizzle_data,
+                                              u32 main_offset, const RegGetter& inputreg_getter,
+                                              const RegGetter& outputreg_getter, bool sanitize_mul);
 
 } // namespace OpenGL::ShaderDecompiler

src/video_core/renderer_opengl/gl_shader_disk_cache.cpp

@@ -0,0 +1,495 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <cstring>
+#include <fmt/format.h>
+
+#include "common/assert.h"
+#include "common/common_paths.h"
+#include "common/common_types.h"
+#include "common/file_util.h"
+#include "common/logging/log.h"
+#include "common/scm_rev.h"
+#include "common/zstd_compression.h"
+#include "core/core.h"
+#include "core/hle/kernel/process.h"
+#include "core/settings.h"
+#include "video_core/renderer_opengl/gl_shader_disk_cache.h"
+
+namespace OpenGL {
+
+constexpr std::size_t HASH_LENGTH = 64;
+using ShaderCacheVersionHash = std::array<u8, HASH_LENGTH>;
+
+enum class TransferableEntryKind : u32 {
+    Raw,
+};
+
+enum class PrecompiledEntryKind : u32 {
+    Decompiled,
+    Dump,
+};
+
+constexpr u32 NativeVersion = 1;
+
+ShaderCacheVersionHash GetShaderCacheVersionHash() {
+    ShaderCacheVersionHash hash{};
+    const std::size_t length = std::min(std::strlen(Common::g_shader_cache_version), hash.size());
+    std::memcpy(hash.data(), Common::g_shader_cache_version, length);
+    return hash;
+}
+
+ShaderDiskCacheRaw::ShaderDiskCacheRaw(u64 unique_identifier, ProgramType program_type,
+                                       RawShaderConfig config, ProgramCode program_code)
+    : unique_identifier{unique_identifier}, program_type{program_type}, config{config},
+      program_code{std::move(program_code)} {}
+
+bool ShaderDiskCacheRaw::Load(FileUtil::IOFile& file) {
+    if (file.ReadBytes(&unique_identifier, sizeof(u64)) != sizeof(u64) ||
+        file.ReadBytes(&program_type, sizeof(u32)) != sizeof(u32)) {
+        return false;
+    }
+
+    u64 reg_array_len{};
+    if (file.ReadBytes(&reg_array_len, sizeof(u64)) != sizeof(u64)) {
+        return false;
+    }
+
+    if (file.ReadArray(config.reg_array.data(), reg_array_len) != reg_array_len) {
+        return false;
+    }
+
+    // Read in type specific configuration
+    if (program_type == ProgramType::VS) {
+        u64 code_len{};
+        if (file.ReadBytes(&code_len, sizeof(u64)) != sizeof(u64)) {
+            return false;
+        }
+        program_code.resize(code_len);
+        if (file.ReadArray(program_code.data(), code_len) != code_len) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+bool ShaderDiskCacheRaw::Save(FileUtil::IOFile& file) const {
+    if (file.WriteObject(unique_identifier) != 1 ||
+        file.WriteObject(static_cast<u32>(program_type)) != 1) {
+        return false;
+    }
+
+    // Just for future proofing, save the sizes of the array to the file
+    const std::size_t reg_array_len = Pica::Regs::NUM_REGS;
+    if (file.WriteObject(static_cast<u64>(reg_array_len)) != 1) {
+        return false;
+    }
+    if (file.WriteArray(config.reg_array.data(), reg_array_len) != reg_array_len) {
+        return false;
+    }
+
+    if (program_type == ProgramType::VS) {
+        const std::size_t code_len = program_code.size();
+        if (file.WriteObject(static_cast<u64>(code_len)) != 1) {
+            return false;
+        }
+        if (file.WriteArray(program_code.data(), code_len) != code_len) {
+            return false;
+        }
+    }
+    return true;
+}
+
+ShaderDiskCache::ShaderDiskCache(bool separable) : separable{separable} {}
+
+std::optional<std::vector<ShaderDiskCacheRaw>> ShaderDiskCache::LoadTransferable() {
+    const bool has_title_id = GetProgramID() != 0;
+    if (!Settings::values.use_disk_shader_cache || !has_title_id)
+        return {};
+    tried_to_load = true;
+
+    FileUtil::IOFile file(GetTransferablePath(), "rb");
+    if (!file.IsOpen()) {
+        LOG_INFO(Render_OpenGL, "No transferable shader cache found for game with title id={}",
+                 GetTitleID());
+        return {};
+    }
+
+    u32 version{};
+    if (file.ReadBytes(&version, sizeof(version)) != sizeof(version)) {
+        LOG_ERROR(Render_OpenGL,
+                  "Failed to get transferable cache version for title id={} - skipping",
+                  GetTitleID());
+        return {};
+    }
+
+    if (version < NativeVersion) {
+        LOG_INFO(Render_OpenGL, "Transferable shader cache is old - removing");
+        file.Close();
+        InvalidateAll();
+        return {};
+    }
+    if (version > NativeVersion) {
+        LOG_WARNING(Render_OpenGL, "Transferable shader cache was generated with a newer version "
+                                   "of the emulator - skipping");
+        return {};
+    }
+
+    // Version is valid, load the shaders
+    std::vector<ShaderDiskCacheRaw> raws;
+    while (file.Tell() < file.GetSize()) {
+        TransferableEntryKind kind{};
+        if (file.ReadBytes(&kind, sizeof(u32)) != sizeof(u32)) {
+            LOG_ERROR(Render_OpenGL, "Failed to read transferable file - skipping");
+            return {};
+        }
+
+        switch (kind) {
+        case TransferableEntryKind::Raw: {
+            ShaderDiskCacheRaw entry;
+            if (!entry.Load(file)) {
+                LOG_ERROR(Render_OpenGL, "Failed to load transferable raw entry - skipping");
+                return {};
+            }
+            transferable.emplace(entry.GetUniqueIdentifier(), ShaderDiskCacheRaw{});
+            raws.push_back(std::move(entry));
+            break;
+        }
+        default:
+            LOG_ERROR(Render_OpenGL, "Unknown transferable shader cache entry kind={} - skipping",
+                      static_cast<u32>(kind));
+            return {};
+        }
+    }
+
+    LOG_INFO(Render_OpenGL, "Found a transferable disk cache with {} entries", raws.size());
+    return {raws};
+}
+
+std::pair<std::unordered_map<u64, ShaderDiskCacheDecompiled>, ShaderDumpsMap>
+ShaderDiskCache::LoadPrecompiled() {
+    if (!IsUsable())
+        return {};
+
+    FileUtil::IOFile file(GetPrecompiledPath(), "rb");
+    if (!file.IsOpen()) {
+        LOG_INFO(Render_OpenGL, "No precompiled shader cache found for game with title id={}",
+                 GetTitleID());
+        return {};
+    }
+
+    const auto result = LoadPrecompiledFile(file);
+    if (!result) {
+        LOG_INFO(Render_OpenGL,
+                 "Failed to load precompiled cache for game with title id={} - removing",
+                 GetTitleID());
+        file.Close();
+        InvalidatePrecompiled();
+        return {};
+    }
+    return *result;
+}
+
+std::optional<std::pair<std::unordered_map<u64, ShaderDiskCacheDecompiled>, ShaderDumpsMap>>
+ShaderDiskCache::LoadPrecompiledFile(FileUtil::IOFile& file) {
+    // Read compressed file from disk and decompress to virtual precompiled cache file
+    std::vector<u8> compressed(file.GetSize());
+    file.ReadBytes(compressed.data(), compressed.size());
+    const std::vector<u8> decompressed = Common::Compression::DecompressDataZSTD(compressed);
+    SaveArrayToPrecompiled(decompressed.data(), decompressed.size());
+    decompressed_precompiled_cache_offset = 0;
+
+    ShaderCacheVersionHash file_hash{};
+    if (!LoadArrayFromPrecompiled(file_hash.data(), file_hash.size())) {
+        return {};
+    }
+    if (GetShaderCacheVersionHash() != file_hash) {
+        LOG_INFO(Render_OpenGL, "Precompiled cache is from another version of the emulator");
+        return {};
+    }
+
+    std::unordered_map<u64, ShaderDiskCacheDecompiled> decompiled;
+    ShaderDumpsMap dumps;
+    while (decompressed_precompiled_cache_offset < decompressed_precompiled_cache.size()) {
+        PrecompiledEntryKind kind{};
+        if (!LoadObjectFromPrecompiled(kind)) {
+            return {};
+        }
+
+        switch (kind) {
+        case PrecompiledEntryKind::Decompiled: {
+            u64 unique_identifier{};
+            if (!LoadObjectFromPrecompiled(unique_identifier)) {
+                return {};
+            }
+
+            auto entry = LoadDecompiledEntry();
+            if (!entry) {
+                return {};
+            }
+            decompiled.insert({unique_identifier, std::move(*entry)});
+            break;
+        }
+        case PrecompiledEntryKind::Dump: {
+            u64 unique_identifier;
+            if (!LoadObjectFromPrecompiled(unique_identifier)) {
+                return {};
+            }
+
+            ShaderDiskCacheDump dump;
+            if (!LoadObjectFromPrecompiled(dump.binary_format)) {
+                return {};
+            }
+
+            u32 binary_length{};
+            if (!LoadObjectFromPrecompiled(binary_length)) {
+                return {};
+            }
+
+            dump.binary.resize(binary_length);
+            if (!LoadArrayFromPrecompiled(dump.binary.data(), dump.binary.size())) {
+                return {};
+            }
+
+            dumps.insert({unique_identifier, dump});
+            break;
+        }
+        default:
+            return {};
+        }
+    }
+
+    LOG_INFO(Render_OpenGL,
+             "Found a precompiled disk cache with {} decompiled entries and {} binary entries",
+             decompiled.size(), dumps.size());
+    return {{decompiled, dumps}};
+}
+
+std::optional<ShaderDiskCacheDecompiled> ShaderDiskCache::LoadDecompiledEntry() {
+
+    bool sanitize_mul;
+    if (!LoadObjectFromPrecompiled(sanitize_mul)) {
+        return {};
+    }
+
+    u32 code_size{};
+    if (!LoadObjectFromPrecompiled(code_size)) {
+        return {};
+    }
+
+    std::string code(code_size, '\0');
+    if (!LoadArrayFromPrecompiled(code.data(), code.size())) {
+        return {};
+    }
+
+    ShaderDiskCacheDecompiled entry;
+    entry.result.code = std::move(code);
+    entry.sanitize_mul = sanitize_mul;
+
+    return entry;
+}
+
+bool ShaderDiskCache::SaveDecompiledFile(u64 unique_identifier,
+                                         const ShaderDecompiler::ProgramResult& result,
+                                         bool sanitize_mul) {
+    if (!SaveObjectToPrecompiled(static_cast<u32>(PrecompiledEntryKind::Decompiled)) ||
+        !SaveObjectToPrecompiled(unique_identifier) || !SaveObjectToPrecompiled(sanitize_mul) ||
+        !SaveObjectToPrecompiled(static_cast<u32>(result.code.size())) ||
+        !SaveArrayToPrecompiled(result.code.data(), result.code.size())) {
+        return false;
+    }
+
+    return true;
+}
+
+void ShaderDiskCache::InvalidateAll() {
+    if (!FileUtil::Delete(GetTransferablePath())) {
+        LOG_ERROR(Render_OpenGL, "Failed to invalidate transferable file={}",
+                  GetTransferablePath());
+    }
+    InvalidatePrecompiled();
+}
+
+void ShaderDiskCache::InvalidatePrecompiled() {
+    // Clear virtaul precompiled cache file
+    decompressed_precompiled_cache.resize(0);
+
+    if (!FileUtil::Delete(GetPrecompiledPath())) {
+        LOG_ERROR(Render_OpenGL, "Failed to invalidate precompiled file={}", GetPrecompiledPath());
+    }
+}
+
+void ShaderDiskCache::SaveRaw(const ShaderDiskCacheRaw& entry) {
+    if (!IsUsable())
+        return;
+
+    const u64 id = entry.GetUniqueIdentifier();
+    if (transferable.find(id) != transferable.end()) {
+        // The shader already exists
+        return;
+    }
+
+    FileUtil::IOFile file = AppendTransferableFile();
+    if (!file.IsOpen())
+        return;
+    if (file.WriteObject(TransferableEntryKind::Raw) != 1 || !entry.Save(file)) {
+        LOG_ERROR(Render_OpenGL, "Failed to save raw transferable cache entry - removing");
+        file.Close();
+        InvalidateAll();
+        return;
+    }
+    transferable.insert({id, entry});
+}
+
+void ShaderDiskCache::SaveDecompiled(u64 unique_identifier,
+                                     const ShaderDecompiler::ProgramResult& code,
+                                     bool sanitize_mul) {
+    if (!IsUsable())
+        return;
+
+    if (decompressed_precompiled_cache.size() == 0) {
+        SavePrecompiledHeaderToVirtualPrecompiledCache();
+    }
+
+    if (!SaveDecompiledFile(unique_identifier, code, sanitize_mul)) {
+        LOG_ERROR(Render_OpenGL,
+                  "Failed to save decompiled entry to the precompiled file - removing");
+        InvalidatePrecompiled();
+    }
+}
+
+void ShaderDiskCache::SaveDump(u64 unique_identifier, GLuint program) {
+    if (!IsUsable())
+        return;
+
+    GLint binary_length{};
+    glGetProgramiv(program, GL_PROGRAM_BINARY_LENGTH, &binary_length);
+
+    GLenum binary_format{};
+    std::vector<u8> binary(binary_length);
+    glGetProgramBinary(program, binary_length, nullptr, &binary_format, binary.data());
+
+    if (!SaveObjectToPrecompiled(static_cast<u32>(PrecompiledEntryKind::Dump)) ||
+        !SaveObjectToPrecompiled(unique_identifier) ||
+        !SaveObjectToPrecompiled(static_cast<u32>(binary_format)) ||
+        !SaveObjectToPrecompiled(static_cast<u32>(binary_length)) ||
+        !SaveArrayToPrecompiled(binary.data(), binary.size())) {
+        LOG_ERROR(Render_OpenGL, "Failed to save binary program file in shader={:016x} - removing",
+                  unique_identifier);
+        InvalidatePrecompiled();
+        return;
+    }
+}
+
+bool ShaderDiskCache::IsUsable() const {
+    return tried_to_load && Settings::values.use_disk_shader_cache;
+}
+
+FileUtil::IOFile ShaderDiskCache::AppendTransferableFile() {
+    if (!EnsureDirectories())
+        return {};
+
+    const auto transferable_path{GetTransferablePath()};
+    const bool existed = FileUtil::Exists(transferable_path);
+
+    FileUtil::IOFile file(transferable_path, "ab");
+    if (!file.IsOpen()) {
+        LOG_ERROR(Render_OpenGL, "Failed to open transferable cache in path={}", transferable_path);
+        return {};
+    }
+    if (!existed || file.GetSize() == 0) {
+        // If the file didn't exist, write its version
+        if (file.WriteObject(NativeVersion) != 1) {
+            LOG_ERROR(Render_OpenGL, "Failed to write transferable cache version in path={}",
+                      transferable_path);
+            return {};
+        }
+    }
+    return file;
+}
+
+void ShaderDiskCache::SavePrecompiledHeaderToVirtualPrecompiledCache() {
+    const auto hash{GetShaderCacheVersionHash()};
+    if (!SaveArrayToPrecompiled(hash.data(), hash.size())) {
+        LOG_ERROR(
+            Render_OpenGL,
+            "Failed to write precompiled cache version hash to virtual precompiled cache file");
+    }
+}
+
+void ShaderDiskCache::SaveVirtualPrecompiledFile() {
+    decompressed_precompiled_cache_offset = 0;
+    const std::vector<u8>& compressed = Common::Compression::CompressDataZSTDDefault(
+        decompressed_precompiled_cache.data(), decompressed_precompiled_cache.size());
+
+    const auto precompiled_path{GetPrecompiledPath()};
+    FileUtil::IOFile file(precompiled_path, "wb");
+
+    if (!file.IsOpen()) {
+        LOG_ERROR(Render_OpenGL, "Failed to open precompiled cache in path={}", precompiled_path);
+        return;
+    }
+    if (file.WriteBytes(compressed.data(), compressed.size()) != compressed.size()) {
+        LOG_ERROR(Render_OpenGL, "Failed to write precompiled cache version in path={}",
+                  precompiled_path);
+        return;
+    }
+}
+
+bool ShaderDiskCache::EnsureDirectories() const {
+    const auto CreateDir = [](const std::string& dir) {
+        if (!FileUtil::CreateDir(dir)) {
+            LOG_ERROR(Render_OpenGL, "Failed to create directory={}", dir);
+            return false;
+        }
+        return true;
+    };
+
+    return CreateDir(FileUtil::GetUserPath(FileUtil::UserPath::ShaderDir)) &&
+           CreateDir(GetBaseDir()) && CreateDir(GetTransferableDir()) &&
+           CreateDir(GetPrecompiledDir());
+}
+
+std::string ShaderDiskCache::GetTransferablePath() {
+    return FileUtil::SanitizePath(GetTransferableDir() + DIR_SEP_CHR + GetTitleID() + ".bin");
+}
+
+std::string ShaderDiskCache::GetPrecompiledPath() {
+    return FileUtil::SanitizePath(GetPrecompiledDir() + DIR_SEP_CHR + GetTitleID() + ".bin");
+}
+
+std::string ShaderDiskCache::GetTransferableDir() const {
+    return GetBaseDir() + DIR_SEP "transferable";
+}
+
+std::string ShaderDiskCache::GetPrecompiledDir() const {
+    return GetBaseDir() + DIR_SEP "precompiled";
+}
+
+std::string ShaderDiskCache::GetBaseDir() const {
+    return FileUtil::GetUserPath(FileUtil::UserPath::ShaderDir) + DIR_SEP "opengl";
+}
+
+u64 ShaderDiskCache::GetProgramID() {
+    // Skip games without title id
+    if (program_id != 0) {
+        return program_id;
+    }
+    if (Core::System::GetInstance().GetAppLoader().ReadProgramId(program_id) !=
+        Loader::ResultStatus::Success) {
+        return 0;
+    }
+    return program_id;
+}
+
+std::string ShaderDiskCache::GetTitleID() {
+    if (!title_id.empty()) {
+        return title_id;
+    }
+    title_id = fmt::format("{:016X}", GetProgramID());
+    return title_id;
+}
+
+} // namespace OpenGL

src/video_core/renderer_opengl/gl_shader_disk_cache.h

@@ -0,0 +1,218 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+#include <bitset>
+#include <optional>
+#include <string>
+#include <tuple>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+#include <glad/glad.h>
+
+#include "common/assert.h"
+#include "common/common_types.h"
+#include "video_core/regs.h"
+#include "video_core/renderer_opengl/gl_shader_decompiler.h"
+#include "video_core/renderer_opengl/gl_shader_gen.h"
+
+namespace Core {
+class System;
+}
+
+namespace FileUtil {
+class IOFile;
+}
+
+namespace OpenGL {
+
+struct ShaderDiskCacheDecompiled;
+struct ShaderDiskCacheDump;
+
+using RawShaderConfig = Pica::Regs;
+using ProgramCode = std::vector<u32>;
+using ShaderDecompiledMap = std::unordered_map<u64, ShaderDiskCacheDecompiled>;
+using ShaderDumpsMap = std::unordered_map<u64, ShaderDiskCacheDump>;
+
+/// Describes a shader how it's used by the guest GPU
+class ShaderDiskCacheRaw {
+public:
+    explicit ShaderDiskCacheRaw(u64 unique_identifier, ProgramType program_type,
+                                RawShaderConfig config, ProgramCode program_code);
+    ShaderDiskCacheRaw() = default;
+    ~ShaderDiskCacheRaw() = default;
+
+    bool Load(FileUtil::IOFile& file);
+
+    bool Save(FileUtil::IOFile& file) const;
+
+    u64 GetUniqueIdentifier() const {
+        return unique_identifier;
+    }
+
+    ProgramType GetProgramType() const {
+        return program_type;
+    }
+
+    const ProgramCode& GetProgramCode() const {
+        return program_code;
+    }
+
+    const RawShaderConfig& GetRawShaderConfig() const {
+        return config;
+    }
+
+private:
+    u64 unique_identifier{};
+    ProgramType program_type{};
+    RawShaderConfig config{};
+    ProgramCode program_code{};
+};
+
+/// Contains decompiled data from a shader
+struct ShaderDiskCacheDecompiled {
+    ShaderDecompiler::ProgramResult result;
+    bool sanitize_mul;
+};
+
+/// Contains an OpenGL dumped binary program
+struct ShaderDiskCacheDump {
+    GLenum binary_format;
+    std::vector<u8> binary;
+};
+
+class ShaderDiskCache {
+public:
+    explicit ShaderDiskCache(bool separable);
+    ~ShaderDiskCache() = default;
+
+    /// Loads transferable cache. If file has a old version or on failure, it deletes the file.
+    std::optional<std::vector<ShaderDiskCacheRaw>> LoadTransferable();
+
+    /// Loads current game's precompiled cache. Invalidates on failure.
+    std::pair<ShaderDecompiledMap, ShaderDumpsMap> LoadPrecompiled();
+
+    /// Removes the transferable (and precompiled) cache file.
+    void InvalidateAll();
+
+    /// Removes the precompiled cache file and clears virtual precompiled cache file.
+    void InvalidatePrecompiled();
+
+    /// Saves a raw dump to the transferable file. Checks for collisions.
+    void SaveRaw(const ShaderDiskCacheRaw& entry);
+
+    /// Saves a decompiled entry to the precompiled file. Does not check for collisions.
+    void SaveDecompiled(u64 unique_identifier, const ShaderDecompiler::ProgramResult& code,
+                        bool sanitize_mul);
+
+    /// Saves a dump entry to the precompiled file. Does not check for collisions.
+    void SaveDump(u64 unique_identifier, GLuint program);
+
+    /// Serializes virtual precompiled shader cache file to real file
+    void SaveVirtualPrecompiledFile();
+
+private:
+    /// Loads the transferable cache. Returns empty on failure.
+    std::optional<std::pair<ShaderDecompiledMap, ShaderDumpsMap>> LoadPrecompiledFile(
+        FileUtil::IOFile& file);
+
+    /// Loads a decompiled cache entry from m_precompiled_cache_virtual_file. Returns empty on
+    /// failure.
+    std::optional<ShaderDiskCacheDecompiled> LoadDecompiledEntry();
+
+    /// Saves a decompiled entry to the passed file. Returns true on success.
+    bool SaveDecompiledFile(u64 unique_identifier, const ShaderDecompiler::ProgramResult& code,
+                            bool sanitize_mul);
+
+    /// Returns if the cache can be used
+    bool IsUsable() const;
+
+    /// Opens current game's transferable file and write it's header if it doesn't exist
+    FileUtil::IOFile AppendTransferableFile();
+
+    /// Save precompiled header to precompiled_cache_in_memory
+    void SavePrecompiledHeaderToVirtualPrecompiledCache();
+
+    /// Create shader disk cache directories. Returns true on success.
+    bool EnsureDirectories() const;
+
+    /// Gets current game's transferable file path
+    std::string GetTransferablePath();
+
+    /// Gets current game's precompiled file path
+    std::string GetPrecompiledPath();
+
+    /// Get user's transferable directory path
+    std::string GetTransferableDir() const;
+
+    /// Get user's precompiled directory path
+    std::string GetPrecompiledDir() const;
+
+    /// Get user's shader directory path
+    std::string GetBaseDir() const;
+
+    /// Get current game's title id as u64
+    u64 GetProgramID();
+
+    /// Get current game's title id
+    std::string GetTitleID();
+
+    template <typename T>
+    bool SaveArrayToPrecompiled(const T* data, std::size_t length) {
+        const u8* data_view = reinterpret_cast<const u8*>(data);
+        decompressed_precompiled_cache.insert(decompressed_precompiled_cache.end(), &data_view[0],
+                                              &data_view[length * sizeof(T)]);
+        decompressed_precompiled_cache_offset += length * sizeof(T);
+        return true;
+    }
+
+    template <typename T>
+    bool LoadArrayFromPrecompiled(T* data, std::size_t length) {
+        u8* data_view = reinterpret_cast<u8*>(data);
+        std::copy_n(decompressed_precompiled_cache.data() + decompressed_precompiled_cache_offset,
+                    length * sizeof(T), data_view);
+        decompressed_precompiled_cache_offset += length * sizeof(T);
+        return true;
+    }
+
+    template <typename T>
+    bool SaveObjectToPrecompiled(const T& object) {
+        return SaveArrayToPrecompiled(&object, 1);
+    }
+
+    bool SaveObjectToPrecompiled(bool object) {
+        const auto value = static_cast<u8>(object);
+        return SaveArrayToPrecompiled(&value, 1);
+    }
+
+    template <typename T>
+    bool LoadObjectFromPrecompiled(T& object) {
+        return LoadArrayFromPrecompiled(&object, 1);
+    }
+
+    // Stores whole precompiled cache which will be read from or saved to the precompiled chache
+    // file
+    std::vector<u8> decompressed_precompiled_cache;
+    // Stores the current offset of the precompiled cache file for IO purposes
+    std::size_t decompressed_precompiled_cache_offset = 0;
+
+    // Stored transferable shaders
+    std::unordered_map<u64, ShaderDiskCacheRaw> transferable;
+
+    // The cache has been loaded at boot
+    bool tried_to_load{};
+
+    bool separable{};
+
+    u64 program_id{};
+    std::string title_id;
+};
+
+} // namespace OpenGL

src/video_core/renderer_opengl/gl_shader_gen.cpp

@@ -1231,7 +1231,8 @@ float ProcTexNoiseCoef(vec2 x) {
     }
 }
 
-std::string GenerateFragmentShader(const PicaFSConfig& config, bool separable_shader) {
+ShaderDecompiler::ProgramResult GenerateFragmentShader(const PicaFSConfig& config,
+                                                       bool separable_shader) {
     const auto& state = config.state;
 
     std::string out = R"(
@@ -1482,7 +1483,7 @@ vec4 secondary_fragment_color = vec4(0.0);
     // Do not do any sort of processing if it's obvious we're not going to pass the alpha test
     if (state.alpha_test_func == FramebufferRegs::CompareFunc::Never) {
         out += "discard; }";
-        return out;
+        return {out};
     }
 
     // Append the scissor test
@@ -1546,7 +1547,7 @@ vec4 secondary_fragment_color = vec4(0.0);
                                                                 "VideoCore_Pica_UseGasMode", true);
         LOG_CRITICAL(Render_OpenGL, "Unimplemented gas mode");
         out += "discard; }";
-        return out;
+        return {out};
     }
 
     if (state.shadow_rendering) {
@@ -1584,10 +1585,10 @@ do {
 
     out += "}";
 
-    return out;
+    return {out};
 }
 
-std::string GenerateTrivialVertexShader(bool separable_shader) {
+ShaderDecompiler::ProgramResult GenerateTrivialVertexShader(bool separable_shader) {
     std::string out = "";
     if (separable_shader) {
         out += "#extension GL_ARB_separate_shader_objects : enable\n";
@@ -1630,11 +1631,11 @@ void main() {
 }
 )";
 
-    return out;
+    return {out};
 }
 
-std::optional<std::string> GenerateVertexShader(const Pica::Shader::ShaderSetup& setup,
-                                                const PicaVSConfig& config, bool separable_shader) {
+std::optional<ShaderDecompiler::ProgramResult> GenerateVertexShader(
+    const Pica::Shader::ShaderSetup& setup, const PicaVSConfig& config, bool separable_shader) {
     std::string out = "";
     if (separable_shader) {
         out += "#extension GL_ARB_separate_shader_objects : enable\n";
@@ -1664,7 +1665,7 @@ std::optional<std::string> GenerateVertexShader(const Pica::Shader::ShaderSetup&
     if (!program_source_opt)
         return {};
 
-    std::string& program_source = *program_source_opt;
+    std::string& program_source = program_source_opt->code;
 
     out += R"(
 #define uniforms vs_uniforms
@@ -1696,7 +1697,7 @@ layout (std140) uniform vs_config {
 
     out += program_source;
 
-    return out;
+    return {{out}};
 }
 
 static std::string GetGSCommonSource(const PicaGSConfigCommonRaw& config, bool separable_shader) {
@@ -1784,7 +1785,8 @@ void EmitPrim(Vertex vtx0, Vertex vtx1, Vertex vtx2) {
     return out;
 };
 
-std::string GenerateFixedGeometryShader(const PicaFixedGSConfig& config, bool separable_shader) {
+ShaderDecompiler::ProgramResult GenerateFixedGeometryShader(const PicaFixedGSConfig& config,
+                                                            bool separable_shader) {
     std::string out = "";
     if (separable_shader) {
         out += "#extension GL_ARB_separate_shader_objects : enable\n\n";
@@ -1814,6 +1816,6 @@ void main() {
     out += "    EmitPrim(prim_buffer[0], prim_buffer[1], prim_buffer[2]);\n";
     out += "}\n";
 
-    return out;
+    return {out};
 }
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_shader_gen.h

@@ -16,6 +16,12 @@
 
 namespace OpenGL {
 
+namespace ShaderDecompiler {
+struct ProgramResult;
+}
+
+enum class ProgramType : u32 { VS, GS, FS };
+
 enum Attributes {
     ATTRIBUTE_POSITION,
     ATTRIBUTE_COLOR,
@@ -161,8 +167,11 @@ struct PicaShaderConfigCommon {
  * shader.
  */
 struct PicaVSConfig : Common::HashableStruct<PicaShaderConfigCommon> {
-    explicit PicaVSConfig(const Pica::Regs& regs, Pica::Shader::ShaderSetup& setup) {
-        state.Init(regs.vs, setup);
+    explicit PicaVSConfig(const Pica::ShaderRegs& regs, Pica::Shader::ShaderSetup& setup) {
+        state.Init(regs, setup);
+    }
+    explicit PicaVSConfig(PicaShaderConfigCommon& conf) {
+        state = conf;
     }
 };
 
@@ -197,20 +206,21 @@ struct PicaFixedGSConfig : Common::HashableStruct<PicaGSConfigCommonRaw> {
  * @param separable_shader generates shader that can be used for separate shader object
  * @returns String of the shader source code
  */
-std::string GenerateTrivialVertexShader(bool separable_shader);
+ShaderDecompiler::ProgramResult GenerateTrivialVertexShader(bool separable_shader);
 
 /**
  * Generates the GLSL vertex shader program source code for the given VS program
  * @returns String of the shader source code; boost::none on failure
  */
-std::optional<std::string> GenerateVertexShader(const Pica::Shader::ShaderSetup& setup,
-                                                const PicaVSConfig& config, bool separable_shader);
+std::optional<ShaderDecompiler::ProgramResult> GenerateVertexShader(
+    const Pica::Shader::ShaderSetup& setup, const PicaVSConfig& config, bool separable_shader);
 
 /*
  * Generates the GLSL fixed geometry shader program source code for non-GS PICA pipeline
  * @returns String of the shader source code
  */
-std::string GenerateFixedGeometryShader(const PicaFixedGSConfig& config, bool separable_shader);
+ShaderDecompiler::ProgramResult GenerateFixedGeometryShader(const PicaFixedGSConfig& config,
+                                                            bool separable_shader);
 
 /**
  * Generates the GLSL fragment shader program source code for the current Pica state
@@ -219,7 +229,8 @@ std::string GenerateFixedGeometryShader(const PicaFixedGSConfig& config, bool se
  * @param separable_shader generates shader that can be used for separate shader object
  * @returns String of the shader source code
  */
-std::string GenerateFragmentShader(const PicaFSConfig& config, bool separable_shader);
+ShaderDecompiler::ProgramResult GenerateFragmentShader(const PicaFSConfig& config,
+                                                       bool separable_shader);
 
 } // namespace OpenGL
 

src/video_core/renderer_opengl/gl_shader_manager.cpp

@@ -3,13 +3,80 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
+#include <thread>
 #include <unordered_map>
 #include <boost/functional/hash.hpp>
 #include <boost/variant.hpp>
+#include "core/core.h"
+#include "video_core/renderer_opengl/gl_shader_disk_cache.h"
 #include "video_core/renderer_opengl/gl_shader_manager.h"
+#include "video_core/video_core.h"
 
 namespace OpenGL {
 
+static u64 GetUniqueIdentifier(const Pica::Regs& regs, const ProgramCode& code) {
+    std::size_t hash = 0;
+    u64 regs_uid = Common::ComputeHash64(regs.reg_array.data(), Pica::Regs::NUM_REGS * sizeof(u32));
+    boost::hash_combine(hash, regs_uid);
+    if (code.size() > 0) {
+        u64 code_uid = Common::ComputeHash64(code.data(), code.size() * sizeof(u32));
+        boost::hash_combine(hash, code_uid);
+    }
+    return static_cast<u64>(hash);
+}
+
+static OGLProgram GeneratePrecompiledProgram(const ShaderDiskCacheDump& dump,
+                                             const std::set<GLenum>& supported_formats) {
+
+    if (supported_formats.find(dump.binary_format) == supported_formats.end()) {
+        LOG_INFO(Render_OpenGL, "Precompiled cache entry with unsupported format - removing");
+        return {};
+    }
+
+    auto shader = OGLProgram();
+    shader.handle = glCreateProgram();
+    glProgramParameteri(shader.handle, GL_PROGRAM_SEPARABLE, GL_TRUE);
+    glProgramBinary(shader.handle, dump.binary_format, dump.binary.data(),
+                    static_cast<GLsizei>(dump.binary.size()));
+
+    GLint link_status{};
+    glGetProgramiv(shader.handle, GL_LINK_STATUS, &link_status);
+    if (link_status == GL_FALSE) {
+        LOG_INFO(Render_OpenGL, "Precompiled cache rejected by the driver - removing");
+        return {};
+    }
+
+    return shader;
+}
+
+static std::set<GLenum> GetSupportedFormats() {
+    std::set<GLenum> supported_formats;
+
+    GLint num_formats{};
+    glGetIntegerv(GL_NUM_PROGRAM_BINARY_FORMATS, &num_formats);
+
+    std::vector<GLint> formats(num_formats);
+    glGetIntegerv(GL_PROGRAM_BINARY_FORMATS, formats.data());
+
+    for (const GLint format : formats)
+        supported_formats.insert(static_cast<GLenum>(format));
+    return supported_formats;
+}
+
+static std::tuple<PicaVSConfig, Pica::Shader::ShaderSetup> BuildVSConfigFromRaw(
+    const ShaderDiskCacheRaw& raw) {
+    Pica::Shader::ProgramCode program_code{};
+    Pica::Shader::SwizzleData swizzle_data{};
+    std::copy_n(raw.GetProgramCode().begin(), Pica::Shader::MAX_PROGRAM_CODE_LENGTH,
+                program_code.begin());
+    std::copy_n(raw.GetProgramCode().begin() + Pica::Shader::MAX_PROGRAM_CODE_LENGTH,
+                Pica::Shader::MAX_SWIZZLE_DATA_LENGTH, swizzle_data.begin());
+    Pica::Shader::ShaderSetup setup;
+    setup.program_code = program_code;
+    setup.swizzle_data = swizzle_data;
+    return {PicaVSConfig{raw.GetRawShaderConfig().vs, setup}, setup};
+}
+
 static void SetShaderUniformBlockBinding(GLuint shader, const char* name, UniformBindings binding,
                                          std::size_t expected_size) {
     const GLuint ub_index = glGetUniformBlockIndex(shader, name);
@@ -121,6 +188,12 @@ public:
         }
     }
 
+    void Inject(OGLProgram&& program) {
+        SetShaderUniformBlockBindings(program.handle);
+        SetShaderSamplerBindings(program.handle);
+        shader_or_program = std::move(program);
+    }
+
 private:
     boost::variant<OGLShader, OGLProgram> shader_or_program;
 };
@@ -128,7 +201,7 @@ private:
 class TrivialVertexShader {
 public:
     explicit TrivialVertexShader(bool separable) : program(separable) {
-        program.Create(GenerateTrivialVertexShader(separable).c_str(), GL_VERTEX_SHADER);
+        program.Create(GenerateTrivialVertexShader(separable).code.c_str(), GL_VERTEX_SHADER);
     }
     GLuint Get() const {
         return program.GetHandle();
@@ -138,18 +211,28 @@ private:
     OGLShaderStage program;
 };
 
-template <typename KeyConfigType, std::string (*CodeGenerator)(const KeyConfigType&, bool),
+template <typename KeyConfigType,
+          ShaderDecompiler::ProgramResult (*CodeGenerator)(const KeyConfigType&, bool),
           GLenum ShaderType>
 class ShaderCache {
 public:
     explicit ShaderCache(bool separable) : separable(separable) {}
-    GLuint Get(const KeyConfigType& config) {
+    std::tuple<GLuint, std::optional<ShaderDecompiler::ProgramResult>> Get(
+        const KeyConfigType& config) {
         auto [iter, new_shader] = shaders.emplace(config, OGLShaderStage{separable});
         OGLShaderStage& cached_shader = iter->second;
+        std::optional<ShaderDecompiler::ProgramResult> result{};
         if (new_shader) {
-            cached_shader.Create(CodeGenerator(config, separable).c_str(), ShaderType);
+            result = CodeGenerator(config, separable);
+            cached_shader.Create(result->code.c_str(), ShaderType);
         }
-        return cached_shader.GetHandle();
+        return {cached_shader.GetHandle(), result};
+    }
+
+    void Inject(const KeyConfigType& key, std::string decomp, OGLProgram&& program) {
+        OGLShaderStage stage{separable};
+        stage.Inject(std::move(program));
+        shaders.emplace(key, std::move(stage));
     }
 
 private:
@@ -163,36 +246,47 @@ private:
 // program buffer from the previous shader, which is hashed into the config, resulting several
 // different config values from the same shader program.
 template <typename KeyConfigType,
-          std::optional<std::string> (*CodeGenerator)(const Pica::Shader::ShaderSetup&,
-                                                      const KeyConfigType&, bool),
+          std::optional<ShaderDecompiler::ProgramResult> (*CodeGenerator)(
+              const Pica::Shader::ShaderSetup&, const KeyConfigType&, bool),
           GLenum ShaderType>
 class ShaderDoubleCache {
 public:
     explicit ShaderDoubleCache(bool separable) : separable(separable) {}
-    GLuint Get(const KeyConfigType& key, const Pica::Shader::ShaderSetup& setup) {
+    std::tuple<GLuint, std::optional<ShaderDecompiler::ProgramResult>> Get(
+        const KeyConfigType& key, const Pica::Shader::ShaderSetup& setup) {
+        std::optional<ShaderDecompiler::ProgramResult> result{};
         auto map_it = shader_map.find(key);
         if (map_it == shader_map.end()) {
             auto program_opt = CodeGenerator(setup, key, separable);
             if (!program_opt) {
                 shader_map[key] = nullptr;
-                return 0;
+                return {0, {}};
             }
 
-            std::string& program = *program_opt;
+            std::string& program = program_opt->code;
             auto [iter, new_shader] = shader_cache.emplace(program, OGLShaderStage{separable});
             OGLShaderStage& cached_shader = iter->second;
             if (new_shader) {
+                result->code = program;
                 cached_shader.Create(program.c_str(), ShaderType);
             }
             shader_map[key] = &cached_shader;
-            return cached_shader.GetHandle();
+            return {cached_shader.GetHandle(), result};
         }
 
         if (map_it->second == nullptr) {
-            return 0;
+            return {0, {}};
         }
 
-        return map_it->second->GetHandle();
+        return {map_it->second->GetHandle(), {}};
+    }
+
+    void Inject(const KeyConfigType& key, std::string decomp, OGLProgram&& program) {
+        OGLShaderStage stage{separable};
+        stage.Inject(std::move(program));
+        auto [iter, new_shader] = shader_cache.emplace(decomp, std::move(stage));
+        OGLShaderStage& cached_shader = iter->second;
+        shader_map[key] = &cached_shader;
     }
 
 private:
@@ -214,7 +308,7 @@ public:
     explicit Impl(bool separable, bool is_amd)
         : is_amd(is_amd), separable(separable), programmable_vertex_shaders(separable),
           trivial_vertex_shader(separable), fixed_geometry_shaders(separable),
-          fragment_shaders(separable) {
+          fragment_shaders(separable), disk_cache(separable) {
         if (separable)
             pipeline.Create();
     }
@@ -244,6 +338,7 @@ public:
     };
 
     bool is_amd;
+    bool separable;
 
     ShaderTuple current;
 
@@ -253,10 +348,9 @@ public:
     FixedGeometryShaders fixed_geometry_shaders;
 
     FragmentShaders fragment_shaders;
-
-    bool separable;
     std::unordered_map<ShaderTuple, OGLProgram, ShaderTuple::Hash> program_cache;
     OGLPipeline pipeline;
+    ShaderDiskCache disk_cache;
 };
 
 ShaderProgramManager::ShaderProgramManager(bool separable, bool is_amd)
@@ -264,12 +358,23 @@ ShaderProgramManager::ShaderProgramManager(bool separable, bool is_amd)
 
 ShaderProgramManager::~ShaderProgramManager() = default;
 
-bool ShaderProgramManager::UseProgrammableVertexShader(const PicaVSConfig& config,
-                                                       const Pica::Shader::ShaderSetup setup) {
-    GLuint handle = impl->programmable_vertex_shaders.Get(config, setup);
+bool ShaderProgramManager::UseProgrammableVertexShader(const Pica::Regs& regs,
+                                                       Pica::Shader::ShaderSetup& setup) {
+    PicaVSConfig config{regs.vs, setup};
+    auto [handle, result] = impl->programmable_vertex_shaders.Get(config, setup);
     if (handle == 0)
         return false;
     impl->current.vs = handle;
+    // Save VS to the disk cache if its a new shader
+    if (result) {
+        auto& disk_cache = impl->disk_cache;
+        ProgramCode program_code{setup.program_code.begin(), setup.program_code.end()};
+        program_code.insert(program_code.end(), setup.swizzle_data.begin(),
+                            setup.swizzle_data.end());
+        u64 unique_identifier = GetUniqueIdentifier(regs, program_code);
+        ShaderDiskCacheRaw raw{unique_identifier, ProgramType::VS, regs, program_code};
+        disk_cache.SaveRaw(raw);
+    }
     return true;
 }
 
@@ -277,25 +382,36 @@ void ShaderProgramManager::UseTrivialVertexShader() {
     impl->current.vs = impl->trivial_vertex_shader.Get();
 }
 
-void ShaderProgramManager::UseFixedGeometryShader(const PicaFixedGSConfig& config) {
-    impl->current.gs = impl->fixed_geometry_shaders.Get(config);
+void ShaderProgramManager::UseFixedGeometryShader(const Pica::Regs& regs) {
+    PicaFixedGSConfig gs_config(regs);
+    auto [handle, _] = impl->fixed_geometry_shaders.Get(gs_config);
+    impl->current.gs = handle;
 }
 
 void ShaderProgramManager::UseTrivialGeometryShader() {
     impl->current.gs = 0;
 }
 
-void ShaderProgramManager::UseFragmentShader(const PicaFSConfig& config) {
-    impl->current.fs = impl->fragment_shaders.Get(config);
+void ShaderProgramManager::UseFragmentShader(const Pica::Regs& regs) {
+    PicaFSConfig config = PicaFSConfig::BuildFromRegs(regs);
+    auto [handle, result] = impl->fragment_shaders.Get(config);
+    impl->current.fs = handle;
+    // Save FS to the disk cache if its a new shader
+    if (result) {
+        auto& disk_cache = impl->disk_cache;
+        u64 unique_identifier = GetUniqueIdentifier(regs, {});
+        ShaderDiskCacheRaw raw{unique_identifier, ProgramType::FS, regs, {}};
+        disk_cache.SaveRaw(raw);
+        disk_cache.SaveDecompiled(unique_identifier, *result, false);
+    }
 }
 
 void ShaderProgramManager::ApplyTo(OpenGLState& state) {
     if (impl->separable) {
         if (impl->is_amd) {
-            // Without this reseting, AMD sometimes freezes when one stage is changed but not for
-            // the others.
-            // On the other hand, including this reset seems to introduce memory leak in Intel
-            // Graphics.
+            // Without this reseting, AMD sometimes freezes when one stage is changed but not
+            // for the others. On the other hand, including this reset seems to introduce memory
+            // leak in Intel Graphics.
             glUseProgramStages(
                 impl->pipeline.handle,
                 GL_VERTEX_SHADER_BIT | GL_GEOMETRY_SHADER_BIT | GL_FRAGMENT_SHADER_BIT, 0);
@@ -316,4 +432,192 @@ void ShaderProgramManager::ApplyTo(OpenGLState& state) {
         state.draw.shader_program = cached_program.handle;
     }
 }
+
+void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
+                                         const VideoCore::DiskResourceLoadCallback& callback) {
+    if (!impl->separable) {
+        LOG_ERROR(Render_OpenGL,
+                  "Cannot load disk cache as separate shader programs are unsupported!");
+        return;
+    }
+    auto& disk_cache = impl->disk_cache;
+    const auto transferable = disk_cache.LoadTransferable();
+    if (!transferable) {
+        return;
+    }
+    const auto raws = *transferable;
+
+    auto [decompiled, dumps] = disk_cache.LoadPrecompiled();
+
+    if (stop_loading) {
+        return;
+    }
+
+    std::set<GLenum> supported_formats = GetSupportedFormats();
+
+    // Track if precompiled cache was altered during loading to know if we have to serialize the
+    // virtual precompiled cache file back to the hard drive
+    bool precompiled_cache_altered = false;
+
+    std::mutex mutex;
+    std::size_t built_shaders = 0; // It doesn't have be atomic since it's used behind a mutex
+    std::atomic_bool compilation_failed = false;
+    if (callback) {
+        callback(VideoCore::LoadCallbackStage::Decompile, 0, raws.size());
+    }
+    std::vector<std::size_t> load_raws_index;
+    // Loads both decompiled and precompiled shaders from the cache. If either one is missing for
+    const auto LoadPrecompiledWorker =
+        [&](std::size_t begin, std::size_t end, const std::vector<ShaderDiskCacheRaw>& raws,
+            const ShaderDecompiledMap& decompiled, const ShaderDumpsMap& dumps) {
+            for (std::size_t i = 0; i < end; ++i) {
+                if (stop_loading || compilation_failed) {
+                    return;
+                }
+                const auto& raw{raws[i]};
+                const u64 unique_identifier{raw.GetUniqueIdentifier()};
+
+                const u64 calculated_hash =
+                    GetUniqueIdentifier(raw.GetRawShaderConfig(), raw.GetProgramCode());
+                if (unique_identifier != calculated_hash) {
+                    LOG_ERROR(Render_OpenGL,
+                              "Invalid hash in entry={:016x} (obtained hash={:016x}) - removing "
+                              "shader cache",
+                              raw.GetUniqueIdentifier(), calculated_hash);
+                    disk_cache.InvalidateAll();
+                    return;
+                }
+
+                const auto dump{dumps.find(unique_identifier)};
+                const auto decomp{decompiled.find(unique_identifier)};
+
+                OGLProgram shader;
+
+                if (dump != dumps.end() && decomp != decompiled.end()) {
+                    // Only load this shader if its sanitize_mul setting matches
+                    if (decomp->second.sanitize_mul == VideoCore::g_hw_shader_accurate_mul) {
+                        continue;
+                    }
+
+                    // If the shader is dumped, attempt to load it
+                    shader = GeneratePrecompiledProgram(dump->second, supported_formats);
+                    if (shader.handle == 0) {
+                        // If any shader failed, stop trying to compile, delete the cache, and start
+                        // loading from raws
+                        compilation_failed = true;
+                        return;
+                    }
+                    // we have both the binary shader and the decompiled, so inject it into the
+                    // cache
+                    if (raw.GetProgramType() == ProgramType::VS) {
+                        auto [conf, setup] = BuildVSConfigFromRaw(raw);
+                        std::scoped_lock lock(mutex);
+
+                        impl->programmable_vertex_shaders.Inject(conf, decomp->second.result.code,
+                                                                 std::move(shader));
+                    } else if (raw.GetProgramType() == ProgramType::FS) {
+                        PicaFSConfig conf = PicaFSConfig::BuildFromRegs(raw.GetRawShaderConfig());
+                        std::scoped_lock lock(mutex);
+                        impl->fragment_shaders.Inject(conf, decomp->second.result.code,
+                                                      std::move(shader));
+                    } else {
+                        // Unsupported shader type got stored somehow so nuke the cache
+
+                        LOG_CRITICAL(Frontend, "failed to load raw programtype {}",
+                                     static_cast<u32>(raw.GetProgramType()));
+                        compilation_failed = true;
+                        return;
+                    }
+                } else {
+                    // Since precompiled didn't have the dump, we'll load them in the next phase
+                    std::scoped_lock lock(mutex);
+                    load_raws_index.push_back(i);
+                }
+                if (callback) {
+                    callback(VideoCore::LoadCallbackStage::Decompile, i, raws.size());
+                }
+            }
+        };
+
+    LoadPrecompiledWorker(0, raws.size(), raws, decompiled, dumps);
+
+    if (compilation_failed) {
+        // Invalidate the precompiled cache if a shader dumped shader was rejected
+        disk_cache.InvalidatePrecompiled();
+        dumps.clear();
+        precompiled_cache_altered = true;
+    }
+
+    if (callback) {
+        callback(VideoCore::LoadCallbackStage::Build, 0, raws.size());
+    }
+
+    compilation_failed = false;
+
+    const auto LoadTransferable = [&](std::size_t begin, std::size_t end,
+                                      const std::vector<ShaderDiskCacheRaw>& raws) {
+        for (std::size_t i = 0; i < end; ++i) {
+            if (stop_loading || compilation_failed) {
+                return;
+            }
+            const auto& raw{raws[i]};
+            const u64 unique_identifier{raw.GetUniqueIdentifier()};
+
+            bool sanitize_mul = false;
+            GLuint handle{0};
+            std::optional<ShaderDecompiler::ProgramResult> result;
+            // Otherwise decompile and build the shader at boot and save the result to the
+            // precompiled file
+            if (raw.GetProgramType() == ProgramType::VS) {
+                // TODO: This isn't the ideal place to lock, since we really only want to
+                // lock access to the shared cache
+                auto [conf, setup] = BuildVSConfigFromRaw(raw);
+                std::scoped_lock lock(mutex);
+                auto [h, r] = impl->programmable_vertex_shaders.Get(conf, setup);
+                handle = h;
+                result = r;
+                sanitize_mul = conf.state.sanitize_mul;
+            } else if (raw.GetProgramType() == ProgramType::FS) {
+                PicaFSConfig conf = PicaFSConfig::BuildFromRegs(raw.GetRawShaderConfig());
+                std::scoped_lock lock(mutex);
+                auto [h, r] = impl->fragment_shaders.Get(conf);
+                handle = h;
+                result = r;
+            } else {
+                // Unsupported shader type got stored somehow so nuke the cache
+                LOG_ERROR(Frontend, "failed to load raw programtype {}",
+                          static_cast<u32>(raw.GetProgramType()));
+                compilation_failed = true;
+                return;
+            }
+            if (handle == 0) {
+                LOG_ERROR(Frontend, "compilation from raw failed {:x} {:x}",
+                          raw.GetProgramCode().at(0), raw.GetProgramCode().at(1));
+                compilation_failed = true;
+                return;
+            }
+            // If this is a new shader, add it the precompiled cache
+            if (result) {
+                disk_cache.SaveDecompiled(unique_identifier, *result, sanitize_mul);
+                disk_cache.SaveDump(unique_identifier, handle);
+                precompiled_cache_altered = true;
+            }
+
+            if (callback) {
+                callback(VideoCore::LoadCallbackStage::Build, i, raws.size());
+            }
+        }
+    };
+
+    LoadTransferable(0, raws.size(), raws);
+
+    if (compilation_failed) {
+        disk_cache.InvalidateAll();
+    }
+
+    if (precompiled_cache_altered) {
+        disk_cache.SaveVirtualPrecompiledFile();
+    }
+} // namespace OpenGL
+
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_shader_manager.h

@@ -6,12 +6,17 @@
 
 #include <memory>
 #include <glad/glad.h>
+#include "video_core/rasterizer_interface.h"
 #include "video_core/regs_lighting.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 #include "video_core/renderer_opengl/gl_shader_gen.h"
 #include "video_core/renderer_opengl/gl_state.h"
 #include "video_core/renderer_opengl/pica_to_gl.h"
 
+namespace Core {
+class System;
+}
+
 namespace OpenGL {
 
 enum class UniformBindings : GLuint { Common, VS, GS };
@@ -97,16 +102,18 @@ public:
     ShaderProgramManager(bool separable, bool is_amd);
     ~ShaderProgramManager();
 
-    bool UseProgrammableVertexShader(const PicaVSConfig& config,
-                                     const Pica::Shader::ShaderSetup setup);
+    void LoadDiskCache(const std::atomic_bool& stop_loading,
+                       const VideoCore::DiskResourceLoadCallback& callback);
+
+    bool UseProgrammableVertexShader(const Pica::Regs& config, Pica::Shader::ShaderSetup& setup);
 
     void UseTrivialVertexShader();
 
-    void UseFixedGeometryShader(const PicaFixedGSConfig& config);
+    void UseFixedGeometryShader(const Pica::Regs& regs);
 
     void UseTrivialGeometryShader();
 
-    void UseFragmentShader(const PicaFSConfig& config);
+    void UseFragmentShader(const Pica::Regs& config);
 
     void ApplyTo(OpenGLState& state);
 

src/video_core/renderer_opengl/renderer_opengl.cpp

@@ -1039,9 +1039,9 @@ static void APIENTRY DebugHandler(GLenum source, GLenum type, GLuint id, GLenum
 }
 
 /// Initialize the renderer
-Core::System::ResultStatus RendererOpenGL::Init() {
+VideoCore::ResultStatus RendererOpenGL::Init() {
     if (!gladLoadGL()) {
-        return Core::System::ResultStatus::ErrorVideoCore_ErrorBelowGL33;
+        return VideoCore::ResultStatus::ErrorBelowGL33;
     }
 
     if (GLAD_GL_KHR_debug) {
@@ -1063,18 +1063,18 @@ Core::System::ResultStatus RendererOpenGL::Init() {
     telemetry_session.AddField(Telemetry::FieldType::UserSystem, "GPU_OpenGL_Version", gl_version);
 
     if (!strcmp(gpu_vendor, "GDI Generic")) {
-        return Core::System::ResultStatus::ErrorVideoCore_ErrorGenericDrivers;
+        return VideoCore::ResultStatus::ErrorGenericDrivers;
     }
 
     if (!(GLAD_GL_VERSION_3_3 || GLAD_GL_ES_VERSION_3_1)) {
-        return Core::System::ResultStatus::ErrorVideoCore_ErrorBelowGL33;
+        return VideoCore::ResultStatus::ErrorBelowGL33;
     }
 
     InitOpenGLObjects();
 
     RefreshRasterizerSetting();
 
-    return Core::System::ResultStatus::Success;
+    return VideoCore::ResultStatus::Success;
 }
 
 /// Shutdown the renderer

src/video_core/renderer_opengl/renderer_opengl.h

@@ -48,7 +48,7 @@ public:
     ~RendererOpenGL() override;
 
     /// Initialize the renderer
-    Core::System::ResultStatus Init() override;
+    VideoCore::ResultStatus Init() override;
 
     /// Shutdown the renderer
     void ShutDown() override;

src/video_core/shader/shader.h

@@ -26,6 +26,8 @@ namespace Pica::Shader {
 
 constexpr unsigned MAX_PROGRAM_CODE_LENGTH = 4096;
 constexpr unsigned MAX_SWIZZLE_DATA_LENGTH = 4096;
+using ProgramCode = std::array<u32, MAX_PROGRAM_CODE_LENGTH>;
+using SwizzleData = std::array<u32, MAX_SWIZZLE_DATA_LENGTH>;
 
 struct AttributeBuffer {
     alignas(16) Common::Vec4<float24> attr[16];
@@ -196,8 +198,8 @@ struct Uniforms {
 struct ShaderSetup {
     Uniforms uniforms;
 
-    std::array<u32, MAX_PROGRAM_CODE_LENGTH> program_code;
-    std::array<u32, MAX_SWIZZLE_DATA_LENGTH> swizzle_data;
+    ProgramCode program_code;
+    SwizzleData swizzle_data;
 
     /// Data private to ShaderEngines
     struct EngineData {

src/video_core/video_core.cpp

@@ -22,6 +22,7 @@ std::atomic<bool> g_hw_renderer_enabled;
 std::atomic<bool> g_shader_jit_enabled;
 std::atomic<bool> g_hw_shader_enabled;
 std::atomic<bool> g_hw_shader_accurate_mul;
+std::atomic<bool> g_use_disk_shader_cache;
 std::atomic<bool> g_renderer_bg_color_update_requested;
 std::atomic<bool> g_renderer_sampler_update_requested;
 std::atomic<bool> g_renderer_shader_update_requested;
@@ -34,16 +35,16 @@ Layout::FramebufferLayout g_screenshot_framebuffer_layout;
 Memory::MemorySystem* g_memory;
 
 /// Initialize the video core
-Core::System::ResultStatus Init(Frontend::EmuWindow& emu_window, Memory::MemorySystem& memory) {
+ResultStatus Init(Frontend::EmuWindow& emu_window, Memory::MemorySystem& memory) {
     g_memory = &memory;
     Pica::Init();
 
     OpenGL::GLES = Settings::values.use_gles;
 
     g_renderer = std::make_unique<OpenGL::RendererOpenGL>(emu_window);
-    Core::System::ResultStatus result = g_renderer->Init();
+    ResultStatus result = g_renderer->Init();
 
-    if (result != Core::System::ResultStatus::Success) {
+    if (result != ResultStatus::Success) {
         LOG_ERROR(Render, "initialization failed !");
     } else {
         LOG_DEBUG(Render, "initialized OK");

src/video_core/video_core.h

@@ -6,7 +6,6 @@
 
 #include <atomic>
 #include <memory>
-#include "core/core.h"
 #include "core/frontend/emu_window.h"
 
 namespace Frontend {
@@ -32,6 +31,7 @@ extern std::atomic<bool> g_hw_renderer_enabled;
 extern std::atomic<bool> g_shader_jit_enabled;
 extern std::atomic<bool> g_hw_shader_enabled;
 extern std::atomic<bool> g_hw_shader_accurate_mul;
+extern std::atomic<bool> g_use_disk_shader_cache;
 extern std::atomic<bool> g_renderer_bg_color_update_requested;
 extern std::atomic<bool> g_renderer_sampler_update_requested;
 extern std::atomic<bool> g_renderer_shader_update_requested;
@@ -43,8 +43,14 @@ extern Layout::FramebufferLayout g_screenshot_framebuffer_layout;
 
 extern Memory::MemorySystem* g_memory;
 
+enum class ResultStatus {
+    Success,
+    ErrorGenericDrivers,
+    ErrorBelowGL33,
+};
+
 /// Initialize the video core
-Core::System::ResultStatus Init(Frontend::EmuWindow& emu_window, Memory::MemorySystem& memory);
+ResultStatus Init(Frontend::EmuWindow& emu_window, Memory::MemorySystem& memory);
 
 /// Shutdown the video core
 void Shutdown();