Backport non-seperable disk shader from android (#5976)

* VideoCore: Backport Shader cache for non separable shaders

android: disk_shader_cache: remove redundant lookup

* disk_shader_cache: dont compile dumped shaders

* cleanup Seperable shader loading

* gl_shader_disk_cache.cpp: remove redundant checks

* Skip compiling non seperable raws

* gl_shader_manager.cpp: Skip `GLAD_GL_ARB_get_program_binary` check on GLES

src/video_core/renderer_base.cpp

@@ -22,7 +22,7 @@ void RendererBase::RefreshRasterizerSetting() {
         opengl_rasterizer_active = hw_renderer_enabled;
 
         if (hw_renderer_enabled) {
-            rasterizer = std::make_unique<OpenGL::RasterizerOpenGL>();
+            rasterizer = std::make_unique<OpenGL::RasterizerOpenGL>(render_window);
         } else {
             rasterizer = std::make_unique<VideoCore::SWRasterizer>();
         }

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -48,7 +48,7 @@ static bool IsVendorIntel() {
     return gpu_vendor == "Intel Inc.";
 }
 
-RasterizerOpenGL::RasterizerOpenGL()
+RasterizerOpenGL::RasterizerOpenGL(Frontend::EmuWindow& emu_window)
     : is_amd(IsVendorAmd()), vertex_buffer(GL_ARRAY_BUFFER, VERTEX_BUFFER_SIZE, is_amd),
       uniform_buffer(GL_UNIFORM_BUFFER, UNIFORM_BUFFER_SIZE, false),
       index_buffer(GL_ELEMENT_ARRAY_BUFFER, INDEX_BUFFER_SIZE, false),
@@ -172,15 +172,16 @@ RasterizerOpenGL::RasterizerOpenGL()
 #ifdef __APPLE__
     if (IsVendorIntel()) {
         shader_program_manager = std::make_unique<ShaderProgramManager>(
+            emu_window,
             VideoCore::g_separable_shader_enabled ? GLAD_GL_ARB_separate_shader_objects : false,
             is_amd);
     } else {
-        shader_program_manager =
-            std::make_unique<ShaderProgramManager>(GLAD_GL_ARB_separate_shader_objects, is_amd);
+        shader_program_manager = std::make_unique<ShaderProgramManager>(
+            emu_window, GLAD_GL_ARB_separate_shader_objects, is_amd);
     }
 #else
-    shader_program_manager =
-        std::make_unique<ShaderProgramManager>(GLAD_GL_ARB_separate_shader_objects, is_amd);
+    shader_program_manager = std::make_unique<ShaderProgramManager>(
+        emu_window, GLAD_GL_ARB_separate_shader_objects, is_amd);
 #endif
 
     glEnable(GL_BLEND);

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -38,7 +38,7 @@ class ShaderProgramManager;
 
 class RasterizerOpenGL : public VideoCore::RasterizerInterface {
 public:
-    explicit RasterizerOpenGL();
+    explicit RasterizerOpenGL(Frontend::EmuWindow& emu_window);
     ~RasterizerOpenGL() override;
 
     void LoadDiskResources(const std::atomic_bool& stop_loading,

src/video_core/renderer_opengl/gl_shader_disk_cache.cpp

@@ -171,7 +171,7 @@ std::optional<std::vector<ShaderDiskCacheRaw>> ShaderDiskCache::LoadTransferable
 }
 
 std::pair<std::unordered_map<u64, ShaderDiskCacheDecompiled>, ShaderDumpsMap>
-ShaderDiskCache::LoadPrecompiled() {
+ShaderDiskCache::LoadPrecompiled(bool compressed) {
     if (!IsUsable())
         return {};
 
@@ -182,7 +182,7 @@ ShaderDiskCache::LoadPrecompiled() {
         return {};
     }
 
-    const auto result = LoadPrecompiledFile(file);
+    const auto result = LoadPrecompiledFile(file, compressed);
     if (!result) {
         LOG_INFO(Render_OpenGL,
                  "Failed to load precompiled cache for game with title id={} - removing",
@@ -195,12 +195,18 @@ ShaderDiskCache::LoadPrecompiled() {
 }
 
 std::optional<std::pair<std::unordered_map<u64, ShaderDiskCacheDecompiled>, ShaderDumpsMap>>
-ShaderDiskCache::LoadPrecompiledFile(FileUtil::IOFile& file) {
+ShaderDiskCache::LoadPrecompiledFile(FileUtil::IOFile& file, bool compressed) {
     // 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);
+    std::vector<u8> precompiled_file(file.GetSize());
+    file.ReadBytes(precompiled_file.data(), precompiled_file.size());
+    if (compressed) {
+        const std::vector<u8> decompressed =
+            Common::Compression::DecompressDataZSTD(precompiled_file);
         SaveArrayToPrecompiled(decompressed.data(), decompressed.size());
+    } else {
+        SaveArrayToPrecompiled(precompiled_file.data(), precompiled_file.size());
+    }
+
     decompressed_precompiled_cache_offset = 0;
 
     ShaderCacheVersionHash file_hash{};
@@ -293,7 +299,23 @@ std::optional<ShaderDiskCacheDecompiled> ShaderDiskCache::LoadDecompiledEntry()
     return entry;
 }
 
-bool ShaderDiskCache::SaveDecompiledFile(u64 unique_identifier,
+void ShaderDiskCache::SaveDecompiledToFile(FileUtil::IOFile& file, u64 unique_identifier,
+                                           const ShaderDecompiler::ProgramResult& result,
+                                           bool sanitize_mul) {
+    if (!IsUsable())
+        return;
+
+    if (file.WriteObject(static_cast<u32>(PrecompiledEntryKind::Decompiled)) != 1 ||
+        file.WriteObject(unique_identifier) != 1 || file.WriteObject(sanitize_mul) != 1 ||
+        file.WriteObject(static_cast<u32>(result.code.size())) != 1 ||
+        file.WriteArray(result.code.data(), result.code.size()) != result.code.size()) {
+        LOG_ERROR(Render_OpenGL, "Failed to save decompiled cache entry - removing");
+        file.Close();
+        InvalidatePrecompiled();
+    }
+}
+
+bool ShaderDiskCache::SaveDecompiledToCache(u64 unique_identifier,
                                             const ShaderDecompiler::ProgramResult& result,
                                             bool sanitize_mul) {
     if (!SaveObjectToPrecompiled(static_cast<u32>(PrecompiledEntryKind::Decompiled)) ||
@@ -315,7 +337,7 @@ void ShaderDiskCache::InvalidateAll() {
 }
 
 void ShaderDiskCache::InvalidatePrecompiled() {
-    // Clear virtaul precompiled cache file
+    // Clear virtual precompiled cache file
     decompressed_precompiled_cache.resize(0);
 
     if (!FileUtil::Delete(GetPrecompiledPath())) {
@@ -351,11 +373,11 @@ void ShaderDiskCache::SaveDecompiled(u64 unique_identifier,
     if (!IsUsable())
         return;
 
-    if (decompressed_precompiled_cache.size() == 0) {
+    if (decompressed_precompiled_cache.empty()) {
         SavePrecompiledHeaderToVirtualPrecompiledCache();
     }
 
-    if (!SaveDecompiledFile(unique_identifier, code, sanitize_mul)) {
+    if (!SaveDecompiledToCache(unique_identifier, code, sanitize_mul)) {
         LOG_ERROR(Render_OpenGL,
                   "Failed to save decompiled entry to the precompiled file - removing");
         InvalidatePrecompiled();
@@ -385,6 +407,37 @@ void ShaderDiskCache::SaveDump(u64 unique_identifier, GLuint program) {
     }
 }
 
+void ShaderDiskCache::SaveDumpToFile(u64 unique_identifier, GLuint program, bool sanitize_mul) {
+    if (!IsUsable())
+        return;
+
+    FileUtil::IOFile file = AppendPrecompiledFile();
+    if (!file.IsOpen())
+        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 (file.WriteObject(static_cast<u32>(PrecompiledEntryKind::Dump)) != 1 ||
+        file.WriteObject(unique_identifier) != 1 ||
+        file.WriteObject(static_cast<u32>(binary_format)) != 1 ||
+        file.WriteObject(static_cast<u32>(binary_length)) != 1 ||
+        file.WriteArray(binary.data(), binary.size()) != binary.size()) {
+        LOG_ERROR(Render_OpenGL, "Failed to save binary program file in shader={:016x} - removing",
+                  unique_identifier);
+        InvalidatePrecompiled();
+        return;
+    }
+
+    // SaveDecompiled is used only to store the accurate multiplication setting, a better way is to
+    // probably change the header in SaveDump
+    SaveDecompiledToFile(file, unique_identifier, {}, sanitize_mul);
+}
+
 bool ShaderDiskCache::IsUsable() const {
     return tried_to_load && Settings::values.use_disk_shader_cache;
 }
@@ -412,6 +465,30 @@ FileUtil::IOFile ShaderDiskCache::AppendTransferableFile() {
     return file;
 }
 
+FileUtil::IOFile ShaderDiskCache::AppendPrecompiledFile() {
+    if (!EnsureDirectories())
+        return {};
+
+    const auto precompiled_path{GetPrecompiledPath()};
+    const bool existed = FileUtil::Exists(precompiled_path);
+
+    FileUtil::IOFile file(precompiled_path, "ab");
+    if (!file.IsOpen()) {
+        LOG_ERROR(Render_OpenGL, "Failed to open precompiled cache in path={}", precompiled_path);
+        return {};
+    }
+    if (!existed || file.GetSize() == 0) {
+        // If the file didn't exist, write its version
+        const auto hash{GetShaderCacheVersionHash()};
+        if (file.WriteArray(hash.data(), hash.size()) != hash.size()) {
+            LOG_ERROR(Render_OpenGL, "Failed to write precompiled cache version in path={}",
+                      precompiled_path);
+            return {};
+        }
+    }
+    return file;
+}
+
 void ShaderDiskCache::SavePrecompiledHeaderToVirtualPrecompiledCache() {
     const auto hash{GetShaderCacheVersionHash()};
     if (!SaveArrayToPrecompiled(hash.data(), hash.size())) {
@@ -451,7 +528,7 @@ bool ShaderDiskCache::EnsureDirectories() const {
 
     return CreateDir(FileUtil::GetUserPath(FileUtil::UserPath::ShaderDir)) &&
            CreateDir(GetBaseDir()) && CreateDir(GetTransferableDir()) &&
-           CreateDir(GetPrecompiledDir());
+           CreateDir(GetPrecompiledDir()) && CreateDir(GetPrecompiledShaderDir());
 }
 
 std::string ShaderDiskCache::GetTransferablePath() {
@@ -459,7 +536,7 @@ std::string ShaderDiskCache::GetTransferablePath() {
 }
 
 std::string ShaderDiskCache::GetPrecompiledPath() {
-    return FileUtil::SanitizePath(GetPrecompiledDir() + DIR_SEP_CHR + GetTitleID() + ".bin");
+    return FileUtil::SanitizePath(GetPrecompiledShaderDir() + DIR_SEP_CHR + GetTitleID() + ".bin");
 }
 
 std::string ShaderDiskCache::GetTransferableDir() const {
@@ -470,6 +547,13 @@ std::string ShaderDiskCache::GetPrecompiledDir() const {
     return GetBaseDir() + DIR_SEP "precompiled";
 }
 
+std::string ShaderDiskCache::GetPrecompiledShaderDir() const {
+    if (separable) {
+        return GetPrecompiledDir() + DIR_SEP "separable";
+    }
+    return GetPrecompiledDir() + DIR_SEP "conventional";
+}
+
 std::string ShaderDiskCache::GetBaseDir() const {
     return FileUtil::GetUserPath(FileUtil::UserPath::ShaderDir) + DIR_SEP "opengl";
 }

src/video_core/renderer_opengl/gl_shader_disk_cache.h

@@ -97,7 +97,7 @@ public:
     std::optional<std::vector<ShaderDiskCacheRaw>> LoadTransferable();
 
     /// Loads current game's precompiled cache. Invalidates on failure.
-    std::pair<ShaderDecompiledMap, ShaderDumpsMap> LoadPrecompiled();
+    std::pair<ShaderDecompiledMap, ShaderDumpsMap> LoadPrecompiled(bool compressed);
 
     /// Removes the transferable (and precompiled) cache file.
     void InvalidateAll();
@@ -115,20 +115,27 @@ public:
     /// Saves a dump entry to the precompiled file. Does not check for collisions.
     void SaveDump(u64 unique_identifier, GLuint program);
 
+    /// Saves a dump entry to the precompiled file. Does not check for collisions.
+    void SaveDumpToFile(u64 unique_identifier, GLuint program, bool sanitize_mul);
+
     /// 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);
+        FileUtil::IOFile& file, bool compressed);
 
     /// 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,
+    /// Saves a decompiled entry to the passed file. Does not check for collisions.
+    void SaveDecompiledToFile(FileUtil::IOFile& file, u64 unique_identifier,
+                              const ShaderDecompiler::ProgramResult& code, bool sanitize_mul);
+
+    /// Saves a decompiled entry to the virtual precompiled cache. Does not check for collisions.
+    bool SaveDecompiledToCache(u64 unique_identifier, const ShaderDecompiler::ProgramResult& code,
                                bool sanitize_mul);
 
     /// Returns if the cache can be used
@@ -155,6 +162,8 @@ private:
     /// Get user's precompiled directory path
     std::string GetPrecompiledDir() const;
 
+    std::string GetPrecompiledShaderDir() const;
+
     /// Get user's shader directory path
     std::string GetBaseDir() const;
 
@@ -197,7 +206,7 @@ private:
         return LoadArrayFromPrecompiled(&object, 1);
     }
 
-    // Stores whole precompiled cache which will be read from or saved to the precompiled chache
+    // Stores whole precompiled cache which will be read from or saved to the precompiled cache
     // file
     std::vector<u8> decompressed_precompiled_cache;
     // Stores the current offset of the precompiled cache file for IO purposes
@@ -213,6 +222,8 @@ private:
 
     u64 program_id{};
     std::string title_id;
+
+    FileUtil::IOFile AppendPrecompiledFile();
 };
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_shader_manager.cpp

@@ -8,8 +8,10 @@
 #include <boost/functional/hash.hpp>
 #include <boost/variant.hpp>
 #include "core/core.h"
+#include "core/frontend/scope_acquire_context.h"
 #include "video_core/renderer_opengl/gl_shader_disk_cache.h"
 #include "video_core/renderer_opengl/gl_shader_manager.h"
+#include "video_core/renderer_opengl/gl_vars.h"
 #include "video_core/video_core.h"
 
 namespace OpenGL {
@@ -26,7 +28,8 @@ static u64 GetUniqueIdentifier(const Pica::Regs& regs, const ProgramCode& code)
 }
 
 static OGLProgram GeneratePrecompiledProgram(const ShaderDiskCacheDump& dump,
-                                             const std::set<GLenum>& supported_formats) {
+                                             const std::set<GLenum>& supported_formats,
+                                             bool separable) {
 
     if (supported_formats.find(dump.binary_format) == supported_formats.end()) {
         LOG_INFO(Render_OpenGL, "Precompiled cache entry with unsupported format - removing");
@@ -35,7 +38,9 @@ static OGLProgram GeneratePrecompiledProgram(const ShaderDiskCacheDump& dump,
 
     auto shader = OGLProgram();
     shader.handle = glCreateProgram();
+    if (separable) {
         glProgramParameteri(shader.handle, GL_PROGRAM_SEPARABLE, GL_TRUE);
+    }
     glProgramBinary(shader.handle, dump.binary_format, dump.binary.data(),
                     static_cast<GLsizei>(dump.binary.size()));
 
@@ -239,6 +244,10 @@ public:
         shaders.emplace(key, std::move(stage));
     }
 
+    void Inject(const KeyConfigType& key, OGLShaderStage&& stage) {
+        shaders.emplace(key, std::move(stage));
+    }
+
 private:
     bool separable;
     std::unordered_map<KeyConfigType, OGLShaderStage> shaders;
@@ -294,6 +303,12 @@ public:
         shader_map.insert_or_assign(key, &cached_shader);
     }
 
+    void Inject(const KeyConfigType& key, std::string decomp, OGLShaderStage&& stage) {
+        const auto iter = shader_cache.emplace(std::move(decomp), std::move(stage)).first;
+        OGLShaderStage& cached_shader = iter->second;
+        shader_map.insert_or_assign(key, &cached_shader);
+    }
+
 private:
     bool separable;
     std::unordered_map<KeyConfigType, OGLShaderStage*> shader_map;
@@ -323,6 +338,10 @@ public:
         GLuint gs = 0;
         GLuint fs = 0;
 
+        std::size_t vs_hash = 0;
+        std::size_t gs_hash = 0;
+        std::size_t fs_hash = 0;
+
         bool operator==(const ShaderTuple& rhs) const {
             return std::tie(vs, gs, fs) == std::tie(rhs.vs, rhs.gs, rhs.fs);
         }
@@ -331,16 +350,14 @@ public:
             return std::tie(vs, gs, fs) != std::tie(rhs.vs, rhs.gs, rhs.fs);
         }
 
-        struct Hash {
-            std::size_t operator()(const ShaderTuple& tuple) const {
+        std::size_t GetConfigHash() const {
             std::size_t hash = 0;
-                boost::hash_combine(hash, tuple.vs);
-                boost::hash_combine(hash, tuple.gs);
-                boost::hash_combine(hash, tuple.fs);
+            boost::hash_combine(hash, vs_hash);
+            boost::hash_combine(hash, gs_hash);
+            boost::hash_combine(hash, fs_hash);
             return hash;
         }
     };
-    };
 
     bool is_amd;
     bool separable;
@@ -353,13 +370,14 @@ public:
     FixedGeometryShaders fixed_geometry_shaders;
 
     FragmentShaders fragment_shaders;
-    std::unordered_map<ShaderTuple, OGLProgram, ShaderTuple::Hash> program_cache;
+    std::unordered_map<u64, OGLProgram> program_cache;
     OGLPipeline pipeline;
     ShaderDiskCache disk_cache;
 };
 
-ShaderProgramManager::ShaderProgramManager(bool separable, bool is_amd)
-    : impl(std::make_unique<Impl>(separable, is_amd)) {}
+ShaderProgramManager::ShaderProgramManager(Frontend::EmuWindow& emu_window_, bool separable,
+                                           bool is_amd)
+    : impl(std::make_unique<Impl>(separable, is_amd)), emu_window{emu_window_} {}
 
 ShaderProgramManager::~ShaderProgramManager() = default;
 
@@ -370,6 +388,8 @@ bool ShaderProgramManager::UseProgrammableVertexShader(const Pica::Regs& regs,
     if (handle == 0)
         return false;
     impl->current.vs = handle;
+    impl->current.vs_hash = config.Hash();
+
     // Save VS to the disk cache if its a new shader
     if (result) {
         auto& disk_cache = impl->disk_cache;
@@ -380,28 +400,33 @@ bool ShaderProgramManager::UseProgrammableVertexShader(const Pica::Regs& regs,
         const ShaderDiskCacheRaw raw{unique_identifier, ProgramType::VS, regs,
                                      std::move(program_code)};
         disk_cache.SaveRaw(raw);
+        disk_cache.SaveDecompiled(unique_identifier, *result, VideoCore::g_hw_shader_accurate_mul);
     }
     return true;
 }
 
 void ShaderProgramManager::UseTrivialVertexShader() {
     impl->current.vs = impl->trivial_vertex_shader.Get();
+    impl->current.vs_hash = 0;
 }
 
 void ShaderProgramManager::UseFixedGeometryShader(const Pica::Regs& regs) {
     PicaFixedGSConfig gs_config(regs);
     auto [handle, _] = impl->fixed_geometry_shaders.Get(gs_config);
     impl->current.gs = handle;
+    impl->current.gs_hash = gs_config.Hash();
 }
 
 void ShaderProgramManager::UseTrivialGeometryShader() {
     impl->current.gs = 0;
+    impl->current.gs_hash = 0;
 }
 
 void ShaderProgramManager::UseFragmentShader(const Pica::Regs& regs) {
     PicaFSConfig config = PicaFSConfig::BuildFromRegs(regs);
     auto [handle, result] = impl->fragment_shaders.Get(config);
     impl->current.fs = handle;
+    impl->current.fs_hash = config.Hash();
     // Save FS to the disk cache if its a new shader
     if (result) {
         auto& disk_cache = impl->disk_cache;
@@ -429,9 +454,14 @@ void ShaderProgramManager::ApplyTo(OpenGLState& state) {
         state.draw.shader_program = 0;
         state.draw.program_pipeline = impl->pipeline.handle;
     } else {
-        OGLProgram& cached_program = impl->program_cache[impl->current];
+        const u64 unique_identifier = impl->current.GetConfigHash();
+        OGLProgram& cached_program = impl->program_cache[unique_identifier];
         if (cached_program.handle == 0) {
             cached_program.Create(false, {impl->current.vs, impl->current.gs, impl->current.fs});
+            auto& disk_cache = impl->disk_cache;
+            disk_cache.SaveDumpToFile(unique_identifier, cached_program.handle,
+                                      VideoCore::g_hw_shader_accurate_mul);
+
             SetShaderUniformBlockBindings(cached_program.handle);
             SetShaderSamplerBindings(cached_program.handle);
         }
@@ -441,12 +471,7 @@ void ShaderProgramManager::ApplyTo(OpenGLState& state) {
 
 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;
-    }
-    if (!GLAD_GL_ARB_get_program_binary) {
+    if (!GLAD_GL_ARB_get_program_binary && !GLES) {
         LOG_ERROR(Render_OpenGL,
                   "Cannot load disk cache as ARB_get_program_binary is not supported!");
         return;
@@ -459,7 +484,9 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
     }
     const auto& raws = *transferable;
 
-    auto [decompiled, dumps] = disk_cache.LoadPrecompiled();
+    // Load uncompressed precompiled file for non-separable shaders.
+    // Precompiled file for separable shaders is compressed.
+    auto [decompiled, dumps] = disk_cache.LoadPrecompiled(impl->separable);
 
     if (stop_loading) {
         return;
@@ -478,10 +505,11 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
     }
     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>& raw_cache,
-            const ShaderDecompiledMap& decompiled_map, const ShaderDumpsMap& dump_map) {
-            for (std::size_t i = 0; i < end; ++i) {
+    const auto LoadPrecompiledShader = [&](std::size_t begin, std::size_t end,
+                                           const std::vector<ShaderDiskCacheRaw>& raw_cache,
+                                           const ShaderDecompiledMap& decompiled_map,
+                                           const ShaderDumpsMap& dump_map) {
+        for (std::size_t i = begin; i < end; ++i) {
             if (stop_loading || compilation_failed) {
                 return;
             }
@@ -505,14 +533,15 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
             OGLProgram shader;
 
             if (dump != dump_map.end() && decomp != decompiled_map.end()) {
-                    // Only load this shader if its sanitize_mul setting matches
+                // Only load the vertex shader if its sanitize_mul setting matches
                 if (raw.GetProgramType() == ProgramType::VS &&
                     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);
+                shader =
+                    GeneratePrecompiledProgram(dump->second, supported_formats, impl->separable);
                 if (shader.handle == 0) {
                     // If any shader failed, stop trying to compile, delete the cache, and start
                     // loading from raws
@@ -524,7 +553,6 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
                 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) {
@@ -534,7 +562,7 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
                 } else {
                     // Unsupported shader type got stored somehow so nuke the cache
 
-                        LOG_CRITICAL(Frontend, "failed to load raw programtype {}",
+                    LOG_CRITICAL(Frontend, "failed to load raw ProgramType {}",
                                  raw.GetProgramType());
                     compilation_failed = true;
                     return;
@@ -550,28 +578,79 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
         }
     };
 
-    LoadPrecompiledWorker(0, raws.size(), raws, decompiled, dumps);
+    const auto LoadPrecompiledProgram = [&](const ShaderDecompiledMap& decompiled_map,
+                                            const ShaderDumpsMap& dump_map) {
+        std::size_t i{0};
+        for (const auto& dump : dump_map) {
+            if (stop_loading) {
+                break;
+            }
+            const u64 unique_identifier{dump.first};
+            const auto decomp{decompiled_map.find(unique_identifier)};
 
+            // Only load the program if its sanitize_mul setting matches
+            if (decomp->second.sanitize_mul != VideoCore::g_hw_shader_accurate_mul) {
+                continue;
+            }
+
+            // If the shader program is dumped, attempt to load it
+            OGLProgram shader =
+                GeneratePrecompiledProgram(dump.second, supported_formats, impl->separable);
+            if (shader.handle != 0) {
+                SetShaderUniformBlockBindings(shader.handle);
+                SetShaderSamplerBindings(shader.handle);
+                impl->program_cache.emplace(unique_identifier, std::move(shader));
+            } else {
+                LOG_ERROR(Frontend, "Failed to link Precompiled program!");
+                compilation_failed = true;
+                break;
+            }
+            if (callback) {
+                callback(VideoCore::LoadCallbackStage::Decompile, ++i, dump_map.size());
+            }
+        }
+    };
+
+    if (impl->separable) {
+        LoadPrecompiledShader(0, raws.size(), raws, decompiled, dumps);
+    } else {
+        LoadPrecompiledProgram(decompiled, dumps);
+    }
+
+    bool load_all_raws = false;
     if (compilation_failed) {
         // Invalidate the precompiled cache if a shader dumped shader was rejected
+        impl->program_cache.clear();
         disk_cache.InvalidatePrecompiled();
         dumps.clear();
         precompiled_cache_altered = true;
+        load_all_raws = true;
+    }
+    // TODO(SachinV): Skip loading raws until we implement a proper way to link non-seperable
+    // shaders.
+    if (!impl->separable) {
+        return;
     }
 
+    const std::size_t load_raws_size = load_all_raws ? raws.size() : load_raws_index.size();
+
     if (callback) {
-        callback(VideoCore::LoadCallbackStage::Build, 0, raws.size());
+        callback(VideoCore::LoadCallbackStage::Build, 0, load_raws_size);
     }
 
     compilation_failed = false;
 
-    const auto LoadTransferable = [&](std::size_t begin, std::size_t end,
-                                      const std::vector<ShaderDiskCacheRaw>& raw_cache) {
-        for (std::size_t i = 0; i < end; ++i) {
+    std::size_t built_shaders = 0; // It doesn't have be atomic since it's used behind a mutex
+    const auto LoadRawSepareble = [&](Frontend::GraphicsContext* context, std::size_t begin,
+                                      std::size_t end) {
+        Frontend::ScopeAcquireContext scope(*context);
+        for (std::size_t i = begin; i < end; ++i) {
             if (stop_loading || compilation_failed) {
                 return;
             }
-            const auto& raw{raw_cache[i]};
+
+            const std::size_t raws_index = load_all_raws ? i : load_raws_index[i];
+            const auto& raw{raws[raws_index]};
             const u64 unique_identifier{raw.GetUniqueIdentifier()};
 
             bool sanitize_mul = false;
@@ -580,23 +659,25 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
             // 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 = std::move(r);
+                result = GenerateVertexShader(setup, conf, impl->separable);
+                OGLShaderStage stage{impl->separable};
+                stage.Create(result->code.c_str(), GL_VERTEX_SHADER);
+                handle = stage.GetHandle();
                 sanitize_mul = conf.state.sanitize_mul;
+                std::scoped_lock lock(mutex);
+                impl->programmable_vertex_shaders.Inject(conf, result->code, std::move(stage));
             } else if (raw.GetProgramType() == ProgramType::FS) {
                 PicaFSConfig conf = PicaFSConfig::BuildFromRegs(raw.GetRawShaderConfig());
+                result = GenerateFragmentShader(conf, impl->separable);
+                OGLShaderStage stage{impl->separable};
+                stage.Create(result->code.c_str(), GL_FRAGMENT_SHADER);
+                handle = stage.GetHandle();
                 std::scoped_lock lock(mutex);
-                auto [h, r] = impl->fragment_shaders.Get(conf);
-                handle = h;
-                result = std::move(r);
+                impl->fragment_shaders.Inject(conf, std::move(stage));
             } else {
                 // Unsupported shader type got stored somehow so nuke the cache
-                LOG_ERROR(Frontend, "failed to load raw programtype {}", raw.GetProgramType());
+                LOG_ERROR(Frontend, "failed to load raw ProgramType {}", raw.GetProgramType());
                 compilation_failed = true;
                 return;
             }
@@ -606,7 +687,9 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
                 compilation_failed = true;
                 return;
             }
-            // If this is a new shader, add it the precompiled cache
+
+            std::scoped_lock lock(mutex);
+            // If this is a new separable shader, add it the precompiled cache
             if (result) {
                 disk_cache.SaveDecompiled(unique_identifier, *result, sanitize_mul);
                 disk_cache.SaveDump(unique_identifier, handle);
@@ -614,12 +697,27 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
             }
 
             if (callback) {
-                callback(VideoCore::LoadCallbackStage::Build, i, raw_cache.size());
+                callback(VideoCore::LoadCallbackStage::Build, ++built_shaders, load_raws_size);
             }
         }
     };
 
-    LoadTransferable(0, raws.size(), raws);
+    const std::size_t num_workers{std::max(1U, std::thread::hardware_concurrency())};
+    const std::size_t bucket_size{load_raws_size / num_workers};
+    std::vector<std::unique_ptr<Frontend::GraphicsContext>> contexts(num_workers);
+    std::vector<std::thread> threads(num_workers);
+    for (std::size_t i = 0; i < num_workers; ++i) {
+        const bool is_last_worker = i + 1 == num_workers;
+        const std::size_t start{bucket_size * i};
+        const std::size_t end{is_last_worker ? load_raws_size : start + bucket_size};
+
+        // On some platforms the shared context has to be created from the GUI thread
+        contexts[i] = emu_window.CreateSharedContext();
+        threads[i] = std::thread(LoadRawSepareble, contexts[i].get(), start, end);
+    }
+    for (auto& thread : threads) {
+        thread.join();
+    }
 
     if (compilation_failed) {
         disk_cache.InvalidateAll();

src/video_core/renderer_opengl/gl_shader_manager.h

@@ -99,7 +99,7 @@ static_assert(sizeof(VSUniformData) < 16384,
 /// A class that manage different shader stages and configures them with given config data.
 class ShaderProgramManager {
 public:
-    ShaderProgramManager(bool separable, bool is_amd);
+    ShaderProgramManager(Frontend::EmuWindow& emu_window_, bool separable, bool is_amd);
     ~ShaderProgramManager();
 
     void LoadDiskCache(const std::atomic_bool& stop_loading,
@@ -120,5 +120,7 @@ public:
 private:
     class Impl;
     std::unique_ptr<Impl> impl;
+
+    Frontend::EmuWindow& emu_window;
 };
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_shader_util.cpp

@@ -84,6 +84,7 @@ GLuint LoadProgram(bool separable_program, const std::vector<GLuint>& shaders) {
         glProgramParameteri(program_id, GL_PROGRAM_SEPARABLE, GL_TRUE);
     }
 
+    glProgramParameteri(program_id, GL_PROGRAM_BINARY_RETRIEVABLE_HINT, GL_TRUE);
     glLinkProgram(program_id);
 
     // Check the program