gl_shader_disk_cache: Store and load fast BRX

src/video_core/engines/const_buffer_engine_interface.h

@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <type_traits>
 #include "common/bit_field.h"
 #include "common/common_types.h"
 #include "video_core/engines/shader_bytecode.h"
@@ -29,51 +30,49 @@ struct SamplerDescriptor {
         u32 raw{};
     };
 
+    bool operator==(const SamplerDescriptor& rhs) const noexcept {
+        return raw == rhs.raw;
+    }
+
     static SamplerDescriptor FromTicTexture(Tegra::Texture::TextureType tic_texture_type) {
-        SamplerDescriptor result{};
+        SamplerDescriptor result;
         switch (tic_texture_type) {
-        case Tegra::Texture::TextureType::Texture1D: {
+        case Tegra::Texture::TextureType::Texture1D:
             result.texture_type.Assign(Tegra::Shader::TextureType::Texture1D);
             result.is_array.Assign(0);
             result.is_buffer.Assign(0);
             result.is_shadow.Assign(0);
             return result;
-        }
+        case Tegra::Texture::TextureType::Texture2D:
-        case Tegra::Texture::TextureType::Texture2D: {
             result.texture_type.Assign(Tegra::Shader::TextureType::Texture2D);
             result.is_array.Assign(0);
             result.is_buffer.Assign(0);
             result.is_shadow.Assign(0);
             return result;
-        }
+        case Tegra::Texture::TextureType::Texture3D:
-        case Tegra::Texture::TextureType::Texture3D: {
             result.texture_type.Assign(Tegra::Shader::TextureType::Texture3D);
             result.is_array.Assign(0);
             result.is_buffer.Assign(0);
             result.is_shadow.Assign(0);
             return result;
-        }
+        case Tegra::Texture::TextureType::TextureCubemap:
-        case Tegra::Texture::TextureType::TextureCubemap: {
             result.texture_type.Assign(Tegra::Shader::TextureType::TextureCube);
             result.is_array.Assign(0);
             result.is_buffer.Assign(0);
             result.is_shadow.Assign(0);
             return result;
-        }
+        case Tegra::Texture::TextureType::Texture1DArray:
-        case Tegra::Texture::TextureType::Texture1DArray: {
             result.texture_type.Assign(Tegra::Shader::TextureType::Texture1D);
             result.is_array.Assign(1);
             result.is_buffer.Assign(0);
             result.is_shadow.Assign(0);
             return result;
-        }
+        case Tegra::Texture::TextureType::Texture2DArray:
-        case Tegra::Texture::TextureType::Texture2DArray: {
             result.texture_type.Assign(Tegra::Shader::TextureType::Texture2D);
             result.is_array.Assign(1);
             result.is_buffer.Assign(0);
             result.is_shadow.Assign(0);
             return result;
-        }
         case Tegra::Texture::TextureType::Texture1DBuffer: {
             result.texture_type.Assign(Tegra::Shader::TextureType::Texture1D);
             result.is_array.Assign(0);
@@ -81,21 +80,19 @@ struct SamplerDescriptor {
             result.is_shadow.Assign(0);
             return result;
         }
-        case Tegra::Texture::TextureType::Texture2DNoMipmap: {
+        case Tegra::Texture::TextureType::Texture2DNoMipmap:
             result.texture_type.Assign(Tegra::Shader::TextureType::Texture2D);
             result.is_array.Assign(0);
             result.is_buffer.Assign(0);
             result.is_shadow.Assign(0);
             return result;
-        }
+        case Tegra::Texture::TextureType::TextureCubeArray:
-        case Tegra::Texture::TextureType::TextureCubeArray: {
             result.texture_type.Assign(Tegra::Shader::TextureType::TextureCube);
             result.is_array.Assign(1);
             result.is_buffer.Assign(0);
             result.is_shadow.Assign(0);
             return result;
-        }
+        default:
-        default: {
             result.texture_type.Assign(Tegra::Shader::TextureType::Texture2D);
             result.is_array.Assign(0);
             result.is_buffer.Assign(0);
@@ -103,8 +100,8 @@ struct SamplerDescriptor {
             return result;
         }
     }
-    }
 };
+static_assert(std::is_trivially_copyable_v<SamplerDescriptor>);
 
 class ConstBufferEngineInterface {
 public:

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -392,7 +392,7 @@ std::tuple<GLuint, BaseBindings> CachedShader::GetProgramHandle(const ProgramVar
         ConstBufferLocker locker(GetEnginesShaderType(program_type), *engine);
         program = BuildShader(device, unique_identifier, program_type, program_code, program_code_b,
                               variant, locker);
-        disk_cache.SaveUsage(GetUsage(variant));
+        disk_cache.SaveUsage(GetUsage(variant, locker));
 
         LabelGLObject(GL_PROGRAM, program->handle, cpu_addr);
     }
@@ -408,10 +408,14 @@ std::tuple<GLuint, BaseBindings> CachedShader::GetProgramHandle(const ProgramVar
     return {program->handle, base_bindings};
 }
 
-ShaderDiskCacheUsage CachedShader::GetUsage(const ProgramVariant& variant) const {
+ShaderDiskCacheUsage CachedShader::GetUsage(const ProgramVariant& variant,
+                                            const ConstBufferLocker& locker) const {
     ShaderDiskCacheUsage usage;
     usage.unique_identifier = unique_identifier;
     usage.variant = variant;
+    usage.keys = locker.GetKeys();
+    usage.bound_samplers = locker.GetBoundSamplers();
+    usage.bindless_samplers = locker.GetBindlessSamplers();
     return usage;
 }
 
@@ -472,6 +476,17 @@ void ShaderCacheOpenGL::LoadDiskCache(const std::atomic_bool& stop_loading,
             }
             if (!shader) {
                 ConstBufferLocker locker(GetEnginesShaderType(unspecialized.program_type));
+                for (const auto& key : usage.keys) {
+                    const auto [buffer, offset] = key.first;
+                    locker.InsertKey(buffer, offset, key.second);
+                }
+                for (const auto& [offset, sampler] : usage.bound_samplers) {
+                    locker.InsertBoundSampler(offset, sampler);
+                }
+                for (const auto& [key, sampler] : usage.bindless_samplers) {
+                    const auto [buffer, offset] = key;
+                    locker.InsertBindlessSampler(buffer, offset, sampler);
+                }
                 shader = BuildShader(device, usage.unique_identifier, unspecialized.program_type,
                                      unspecialized.code, unspecialized.code_b, usage.variant,
                                      locker, true);

src/video_core/renderer_opengl/gl_shader_cache.h

@@ -31,6 +31,10 @@ namespace Core::Frontend {
 class EmuWindow;
 }
 
+namespace VideoCommon::Shader {
+class ConstBufferLocker;
+}
+
 namespace OpenGL {
 
 class CachedShader;
@@ -92,7 +96,8 @@ private:
                           GLShader::ShaderEntries entries, ProgramCode program_code,
                           ProgramCode program_code_b);
 
-    ShaderDiskCacheUsage GetUsage(const ProgramVariant& variant) const;
+    ShaderDiskCacheUsage GetUsage(const ProgramVariant& variant,
+                                  const VideoCommon::Shader::ConstBufferLocker& locker) const;
 
     Core::System& system;
     ShaderDiskCacheOpenGL& disk_cache;

src/video_core/renderer_opengl/gl_shader_disk_cache.cpp

@@ -22,6 +22,29 @@
 
 namespace OpenGL {
 
+using VideoCommon::Shader::BindlessSamplerMap;
+using VideoCommon::Shader::BoundSamplerMap;
+using VideoCommon::Shader::KeyMap;
+
+namespace {
+
+struct ConstBufferKey {
+    u32 cbuf;
+    u32 offset;
+    u32 value;
+};
+
+struct BoundSamplerKey {
+    u32 offset;
+    Tegra::Engines::SamplerDescriptor sampler;
+};
+
+struct BindlessSamplerKey {
+    u32 cbuf;
+    u32 offset;
+    Tegra::Engines::SamplerDescriptor sampler;
+};
+
 using ShaderCacheVersionHash = std::array<u8, 64>;
 
 enum class TransferableEntryKind : u32 {
@@ -33,9 +56,6 @@ constexpr u32 NativeVersion = 5;
 
 // Making sure sizes doesn't change by accident
 static_assert(sizeof(BaseBindings) == 16);
-static_assert(sizeof(ShaderDiskCacheUsage) == 40);
-
-namespace {
 
 ShaderCacheVersionHash GetShaderCacheVersionHash() {
     ShaderCacheVersionHash hash{};
@@ -121,13 +141,13 @@ ShaderDiskCacheOpenGL::LoadTransferable() {
     u32 version{};
     if (file.ReadBytes(&version, sizeof(version)) != sizeof(version)) {
         LOG_ERROR(Render_OpenGL,
-                  "Failed to get transferable cache version for title id={} - skipping",
+                  "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");
+        LOG_INFO(Render_OpenGL, "Transferable shader cache is old, removing");
         file.Close();
         InvalidateTransferable();
         is_usable = true;
@@ -135,17 +155,18 @@ ShaderDiskCacheOpenGL::LoadTransferable() {
     }
     if (version > NativeVersion) {
         LOG_WARNING(Render_OpenGL, "Transferable shader cache was generated with a newer version "
-                                   "of the emulator - skipping");
+                                   "of the emulator, skipping");
         return {};
     }
 
     // Version is valid, load the shaders
+    constexpr const char error_loading[] = "Failed to load transferable raw entry, skipping";
     std::vector<ShaderDiskCacheRaw> raws;
     std::vector<ShaderDiskCacheUsage> usages;
     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");
+            LOG_ERROR(Render_OpenGL, "Failed to read transferable file, skipping");
             return {};
         }
 
@@ -153,7 +174,7 @@ ShaderDiskCacheOpenGL::LoadTransferable() {
         case TransferableEntryKind::Raw: {
             ShaderDiskCacheRaw entry;
             if (!entry.Load(file)) {
-                LOG_ERROR(Render_OpenGL, "Failed to load transferable raw entry - skipping");
+                LOG_ERROR(Render_OpenGL, error_loading);
                 return {};
             }
             transferable.insert({entry.GetUniqueIdentifier(), {}});
@@ -161,16 +182,45 @@ ShaderDiskCacheOpenGL::LoadTransferable() {
             break;
         }
         case TransferableEntryKind::Usage: {
-            ShaderDiskCacheUsage usage{};
+            ShaderDiskCacheUsage usage;
-            if (file.ReadBytes(&usage, sizeof(usage)) != sizeof(usage)) {
+
-                LOG_ERROR(Render_OpenGL, "Failed to load transferable usage entry - skipping");
+            u32 num_keys{};
+            u32 num_bound_samplers{};
+            u32 num_bindless_samplers{};
+            if (file.ReadArray(&usage.unique_identifier, 1) != 1 ||
+                file.ReadArray(&usage.variant, 1) != 1 || file.ReadArray(&num_keys, 1) != 1 ||
+                file.ReadArray(&num_bound_samplers, 1) != 1 ||
+                file.ReadArray(&num_bindless_samplers, 1) != 1) {
+                LOG_ERROR(Render_OpenGL, error_loading);
                 return {};
             }
+
+            std::vector<ConstBufferKey> keys(num_keys);
+            std::vector<BoundSamplerKey> bound_samplers(num_bound_samplers);
+            std::vector<BindlessSamplerKey> bindless_samplers(num_bindless_samplers);
+            if (file.ReadArray(keys.data(), keys.size()) != keys.size() ||
+                file.ReadArray(bound_samplers.data(), bound_samplers.size()) !=
+                    bound_samplers.size() ||
+                file.ReadArray(bindless_samplers.data(), bindless_samplers.size()) !=
+                    bindless_samplers.size()) {
+                LOG_ERROR(Render_OpenGL, error_loading);
+                return {};
+            }
+            for (const auto& key : keys) {
+                usage.keys.insert({{key.cbuf, key.offset}, key.value});
+            }
+            for (const auto& key : bound_samplers) {
+                usage.bound_samplers.emplace(key.offset, key.sampler);
+            }
+            for (const auto& key : bindless_samplers) {
+                usage.bindless_samplers.insert({{key.cbuf, key.offset}, key.sampler});
+            }
+
             usages.push_back(std::move(usage));
             break;
         }
         default:
-            LOG_ERROR(Render_OpenGL, "Unknown transferable shader cache entry kind={} - skipping",
+            LOG_ERROR(Render_OpenGL, "Unknown transferable shader cache entry kind={}, skipping",
                       static_cast<u32>(kind));
             return {};
         }
@@ -197,7 +247,7 @@ ShaderDiskCacheOpenGL::LoadPrecompiled() {
     const auto result = LoadPrecompiledFile(file);
     if (!result) {
         LOG_INFO(Render_OpenGL,
-                 "Failed to load precompiled cache for game with title id={} - removing",
+                 "Failed to load precompiled cache for game with title id={}, removing",
                  GetTitleID());
         file.Close();
         InvalidatePrecompiled();
@@ -228,10 +278,35 @@ ShaderDiskCacheOpenGL::LoadPrecompiledFile(FileUtil::IOFile& file) {
 
     ShaderDumpsMap dumps;
     while (precompiled_cache_virtual_file_offset < precompiled_cache_virtual_file.GetSize()) {
+        u32 num_keys{};
+        u32 num_bound_samplers{};
+        u32 num_bindless_samplers{};
         ShaderDiskCacheUsage usage;
-        if (!LoadObjectFromPrecompiled(usage)) {
+        if (!LoadObjectFromPrecompiled(usage.unique_identifier) ||
+            !LoadObjectFromPrecompiled(usage.variant) || !LoadObjectFromPrecompiled(num_keys) ||
+            !LoadObjectFromPrecompiled(num_bound_samplers) ||
+            !LoadObjectFromPrecompiled(num_bindless_samplers)) {
             return {};
         }
+        std::vector<ConstBufferKey> keys(num_keys);
+        std::vector<BoundSamplerKey> bound_samplers(num_bound_samplers);
+        std::vector<BindlessSamplerKey> bindless_samplers(num_bindless_samplers);
+        if (!LoadArrayFromPrecompiled(keys.data(), keys.size()) ||
+            !LoadArrayFromPrecompiled(bound_samplers.data(), bound_samplers.size()) !=
+                bound_samplers.size() ||
+            !LoadArrayFromPrecompiled(bindless_samplers.data(), bindless_samplers.size()) !=
+                bindless_samplers.size()) {
+            return {};
+        }
+        for (const auto& key : keys) {
+            usage.keys.insert({{key.cbuf, key.offset}, key.value});
+        }
+        for (const auto& key : bound_samplers) {
+            usage.bound_samplers.emplace(key.offset, key.sampler);
+        }
+        for (const auto& key : bindless_samplers) {
+            usage.bindless_samplers.insert({{key.cbuf, key.offset}, key.sampler});
+        }
 
         ShaderDiskCacheDump dump;
         if (!LoadObjectFromPrecompiled(dump.binary_format)) {
@@ -248,7 +323,7 @@ ShaderDiskCacheOpenGL::LoadPrecompiledFile(FileUtil::IOFile& file) {
             return {};
         }
 
-        dumps.emplace(usage, dump);
+        dumps.emplace(std::move(usage), dump);
     }
     return dumps;
 }
@@ -282,10 +357,11 @@ void ShaderDiskCacheOpenGL::SaveRaw(const ShaderDiskCacheRaw& entry) {
     }
 
     FileUtil::IOFile file = AppendTransferableFile();
-    if (!file.IsOpen())
+    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");
+        LOG_ERROR(Render_OpenGL, "Failed to save raw transferable cache entry, removing");
         file.Close();
         InvalidateTransferable();
         return;
@@ -311,13 +387,40 @@ void ShaderDiskCacheOpenGL::SaveUsage(const ShaderDiskCacheUsage& usage) {
     FileUtil::IOFile file = AppendTransferableFile();
     if (!file.IsOpen())
         return;
-
+    const auto Close = [&] {
-    if (file.WriteObject(TransferableEntryKind::Usage) != 1 || file.WriteObject(usage) != 1) {
+        LOG_ERROR(Render_OpenGL, "Failed to save usage transferable cache entry, removing");
-        LOG_ERROR(Render_OpenGL, "Failed to save usage transferable cache entry - removing");
         file.Close();
         InvalidateTransferable();
+    };
+
+    if (file.WriteObject(TransferableEntryKind::Usage) != 1 ||
+        file.WriteObject(usage.unique_identifier) != 1 || file.WriteObject(usage.variant) != 1 ||
+        file.WriteObject(static_cast<u32>(usage.keys.size())) != 1 ||
+        file.WriteObject(static_cast<u32>(usage.bound_samplers.size())) != 1 ||
+        file.WriteObject(static_cast<u32>(usage.bindless_samplers.size())) != 1) {
+        Close();
         return;
     }
+    for (const auto& [pair, value] : usage.keys) {
+        const auto [cbuf, offset] = pair;
+        if (file.WriteObject(ConstBufferKey{cbuf, offset, value}) != 1) {
+            Close();
+            return;
+        }
+    }
+    for (const auto& [offset, sampler] : usage.bound_samplers) {
+        if (file.WriteObject(BoundSamplerKey{offset, sampler}) != 1) {
+            Close();
+            return;
+        }
+    }
+    for (const auto& [pair, sampler] : usage.bindless_samplers) {
+        const auto [cbuf, offset] = pair;
+        if (file.WriteObject(BindlessSamplerKey{cbuf, offset, sampler}) != 1) {
+            Close();
+            return;
+        }
+    }
 }
 
 void ShaderDiskCacheOpenGL::SaveDump(const ShaderDiskCacheUsage& usage, GLuint program) {
@@ -339,15 +442,45 @@ void ShaderDiskCacheOpenGL::SaveDump(const ShaderDiskCacheUsage& usage, GLuint p
     std::vector<u8> binary(binary_length);
     glGetProgramBinary(program, binary_length, nullptr, &binary_format, binary.data());
 
-    if (!SaveObjectToPrecompiled(usage) ||
+    const auto Close = [&] {
-        !SaveObjectToPrecompiled(static_cast<u32>(binary_format)) ||
+        LOG_ERROR(Render_OpenGL, "Failed to save binary program file in shader={:016X}, removing",
-        !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",
                   usage.unique_identifier);
         InvalidatePrecompiled();
+    };
+
+    if (!SaveObjectToPrecompiled(usage.unique_identifier) ||
+        !SaveObjectToPrecompiled(usage.variant) ||
+        !SaveObjectToPrecompiled(static_cast<u32>(usage.keys.size())) ||
+        !SaveObjectToPrecompiled(static_cast<u32>(usage.bound_samplers.size())) ||
+        !SaveObjectToPrecompiled(static_cast<u32>(usage.bindless_samplers.size()))) {
+        Close();
         return;
     }
+    for (const auto& [pair, value] : usage.keys) {
+        const auto [cbuf, offset] = pair;
+        if (SaveObjectToPrecompiled(ConstBufferKey{cbuf, offset, value}) != 1) {
+            Close();
+            return;
+        }
+    }
+    for (const auto& [offset, sampler] : usage.bound_samplers) {
+        if (SaveObjectToPrecompiled(BoundSamplerKey{offset, sampler}) != 1) {
+            Close();
+            return;
+        }
+    }
+    for (const auto& [pair, sampler] : usage.bindless_samplers) {
+        const auto [cbuf, offset] = pair;
+        if (SaveObjectToPrecompiled(BindlessSamplerKey{cbuf, offset, sampler}) != 1) {
+            Close();
+            return;
+        }
+    }
+    if (!SaveObjectToPrecompiled(static_cast<u32>(binary_format)) ||
+        !SaveObjectToPrecompiled(static_cast<u32>(binary_length)) ||
+        !SaveArrayToPrecompiled(binary.data(), binary.size())) {
+        Close();
+    }
 }
 
 FileUtil::IOFile ShaderDiskCacheOpenGL::AppendTransferableFile() const {

src/video_core/renderer_opengl/gl_shader_disk_cache.h

@@ -8,6 +8,7 @@
 #include <optional>
 #include <string>
 #include <tuple>
+#include <type_traits>
 #include <unordered_map>
 #include <unordered_set>
 #include <utility>
@@ -19,6 +20,7 @@
 #include "common/common_types.h"
 #include "core/file_sys/vfs_vector.h"
 #include "video_core/renderer_opengl/gl_shader_gen.h"
+#include "video_core/shader/const_buffer_locker.h"
 
 namespace Core {
 class System;
@@ -53,6 +55,7 @@ struct BaseBindings {
         return !operator==(rhs);
     }
 };
+static_assert(std::is_trivially_copyable_v<BaseBindings>);
 
 /// Describes the different variants a single program can be compiled.
 struct ProgramVariant {
@@ -70,13 +73,20 @@ struct ProgramVariant {
     }
 };
 
+static_assert(std::is_trivially_copyable_v<ProgramVariant>);
+
 /// Describes how a shader is used.
 struct ShaderDiskCacheUsage {
     u64 unique_identifier{};
     ProgramVariant variant;
+    VideoCommon::Shader::KeyMap keys;
+    VideoCommon::Shader::BoundSamplerMap bound_samplers;
+    VideoCommon::Shader::BindlessSamplerMap bindless_samplers;
 
     bool operator==(const ShaderDiskCacheUsage& rhs) const {
-        return std::tie(unique_identifier, variant) == std::tie(rhs.unique_identifier, rhs.variant);
+        return std::tie(unique_identifier, variant, keys, bound_samplers, bindless_samplers) ==
+               std::tie(rhs.unique_identifier, rhs.variant, rhs.keys, rhs.bound_samplers,
+                        rhs.bindless_samplers);
     }
 
     bool operator!=(const ShaderDiskCacheUsage& rhs) const {

src/video_core/shader/const_buffer_locker.cpp

@@ -90,14 +90,14 @@ bool ConstBufferLocker::IsConsistent() const {
                        [this](const auto& sampler) {
                            const auto [key, value] = sampler;
                            const auto other_value = engine->AccessBoundSampler(stage, key);
-                           return value.raw == other_value.raw;
+                           return value == other_value;
                        }) &&
            std::all_of(
                bindless_samplers.begin(), bindless_samplers.end(), [this](const auto& sampler) {
                    const auto [cbuf, offset] = sampler.first;
                    const auto value = sampler.second;
                    const auto other_value = engine->AccessBindlessSampler(stage, cbuf, offset);
-                   return value.raw == other_value.raw;
+                   return value == other_value;
                });
 }