Only load precompiled shaders if their sanitize_mul setting matches

src/common/file_util.cpp

@@ -557,7 +557,7 @@ std::optional<std::string> GetCurrentDir() {
 #endif
     free(dir);
     return strDir;
-}
+} // namespace FileUtil
 
 bool SetCurrentDir(const std::string& directory) {
 #ifdef _WIN32

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -889,16 +889,17 @@ bool exec_shader();
 )";
 }
 
-std::optional<std::string> DecompileProgram(const Pica::Shader::ProgramCode& program_code,
+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) {
+                                              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

@@ -12,7 +12,10 @@
 namespace OpenGL::ShaderDecompiler {
 
 using RegGetter = std::function<std::string(u32)>;
-using ProgramResult = std::string;
+
+struct ProgramResult {
+    std::string code;
+};
 
 std::string GetCommonDeclarations();
 

src/video_core/renderer_opengl/gl_shader_disk_cache.cpp

@@ -270,6 +270,12 @@ ShaderDiskCache::LoadPrecompiledFile(FileUtil::IOFile& file) {
 }
 
 std::optional<ShaderDiskCacheDecompiled> ShaderDiskCache::LoadDecompiledEntry() {
+
+    bool sanitize_mul;
+    if (!LoadObjectFromPrecompiled(sanitize_mul)) {
+        return {};
+    }
+
     u32 code_size{};
     if (!LoadObjectFromPrecompiled(code_size)) {
         return {};
@@ -281,17 +287,19 @@ std::optional<ShaderDiskCacheDecompiled> ShaderDiskCache::LoadDecompiledEntry()
     }
 
     ShaderDiskCacheDecompiled entry;
-    entry.code = std::move(code);
+    entry.result.code = std::move(code);
+    entry.sanitize_mul = sanitize_mul;
 
     return entry;
 }
 
 bool ShaderDiskCache::SaveDecompiledFile(u64 unique_identifier,
-                                         const ShaderDecompiler::ProgramResult& code) {
+                                         const ShaderDecompiler::ProgramResult& result,
+                                         bool sanitize_mul) {
     if (!SaveObjectToPrecompiled(static_cast<u32>(PrecompiledEntryKind::Decompiled)) ||
-        !SaveObjectToPrecompiled(unique_identifier) ||
+        !SaveObjectToPrecompiled(unique_identifier) || !SaveObjectToPrecompiled(sanitize_mul) ||
-        !SaveObjectToPrecompiled(static_cast<u32>(code.size())) ||
+        !SaveObjectToPrecompiled(static_cast<u32>(result.code.size())) ||
-        !SaveArrayToPrecompiled(code.data(), code.size())) {
+        !SaveArrayToPrecompiled(result.code.data(), result.code.size())) {
         return false;
     }
 
@@ -338,7 +346,8 @@ void ShaderDiskCache::SaveRaw(const ShaderDiskCacheRaw& entry) {
 }
 
 void ShaderDiskCache::SaveDecompiled(u64 unique_identifier,
-                                     const ShaderDecompiler::ProgramResult& code) {
+                                     const ShaderDecompiler::ProgramResult& code,
+                                     bool sanitize_mul) {
     if (!IsUsable())
         return;
 
@@ -346,7 +355,7 @@ void ShaderDiskCache::SaveDecompiled(u64 unique_identifier,
         SavePrecompiledHeaderToVirtualPrecompiledCache();
     }
 
-    if (!SaveDecompiledFile(unique_identifier, code)) {
+    if (!SaveDecompiledFile(unique_identifier, code, sanitize_mul)) {
         LOG_ERROR(Render_OpenGL,
                   "Failed to save decompiled entry to the precompiled file - removing");
         InvalidatePrecompiled();

src/video_core/renderer_opengl/gl_shader_disk_cache.h

@@ -78,7 +78,8 @@ private:
 
 /// Contains decompiled data from a shader
 struct ShaderDiskCacheDecompiled {
-    ShaderDecompiler::ProgramResult code;
+    ShaderDecompiler::ProgramResult result;
+    bool sanitize_mul;
 };
 
 /// Contains an OpenGL dumped binary program
@@ -108,7 +109,8 @@ public:
     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);
+    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);
@@ -126,7 +128,8 @@ private:
     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 SaveDecompiledFile(u64 unique_identifier, const ShaderDecompiler::ProgramResult& code,
+                            bool sanitize_mul);
 
     /// Returns if the cache can be used
     bool IsUsable() const;

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,
+std::optional<ShaderDecompiler::ProgramResult> GenerateVertexShader(
-                                                const PicaVSConfig& config, bool separable_shader) {
+    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 +16,10 @@
 
 namespace OpenGL {
 
+namespace ShaderDecompiler {
+struct ProgramResult;
+}
+
 enum class ProgramType : u32 { VS, GS, FS };
 
 enum Attributes {
@@ -202,20 +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,
+std::optional<ShaderDecompiler::ProgramResult> GenerateVertexShader(
-                                                const PicaVSConfig& config, bool separable_shader);
+    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
@@ -224,7 +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

@@ -10,18 +10,19 @@
 #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) {
-    u64 hash = 0;
+    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 hash;
+    return static_cast<u64>(hash);
 }
 
 static OGLProgram GeneratePrecompiledProgram(const ShaderDiskCacheDump& dump,
@@ -200,7 +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();
@@ -210,7 +211,8 @@ 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:
@@ -222,7 +224,7 @@ public:
         std::optional<ShaderDecompiler::ProgramResult> result{};
         if (new_shader) {
             result = CodeGenerator(config, separable);
-            cached_shader.Create(result->c_str(), ShaderType);
+            cached_shader.Create(result->code.c_str(), ShaderType);
         }
         return {cached_shader.GetHandle(), result};
     }
@@ -244,8 +246,8 @@ 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&,
+          std::optional<ShaderDecompiler::ProgramResult> (*CodeGenerator)(
-                                                      const KeyConfigType&, bool),
+              const Pica::Shader::ShaderSetup&, const KeyConfigType&, bool),
           GLenum ShaderType>
 class ShaderDoubleCache {
 public:
@@ -261,11 +263,11 @@ public:
                 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 = program;
+                result->code = program;
                 cached_shader.Create(program.c_str(), ShaderType);
             }
             shader_map[key] = &cached_shader;
@@ -336,6 +338,7 @@ public:
     };
 
     bool is_amd;
+    bool separable;
 
     ShaderTuple current;
 
@@ -345,8 +348,6 @@ public:
     FixedGeometryShaders fixed_geometry_shaders;
 
     FragmentShaders fragment_shaders;
-
-    bool separable;
     std::unordered_map<ShaderTuple, OGLProgram, ShaderTuple::Hash> program_cache;
     OGLPipeline pipeline;
     ShaderDiskCache disk_cache;
@@ -401,7 +402,7 @@ void ShaderProgramManager::UseFragmentShader(const Pica::Regs& regs) {
         u64 unique_identifier = GetUniqueIdentifier(regs, {});
         ShaderDiskCacheRaw raw{unique_identifier, ProgramType::FS, regs, {}};
         disk_cache.SaveRaw(raw);
-        disk_cache.SaveDecompiled(unique_identifier, *result);
+        disk_cache.SaveDecompiled(unique_identifier, *result, false);
     }
 }
 
@@ -489,6 +490,12 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
 
                 const auto dump{dumps.find(unique_identifier)};
                 const auto decomp{decompiled.find(unique_identifier)};
+
+                // Only load this shader if its sanitize_mul setting matches
+                if (decomp->second.sanitize_mul == VideoCore::g_hw_shader_accurate_mul) {
+                    continue;
+                }
+
                 OGLProgram shader;
 
                 if (dump != dumps.end() && decomp != decompiled.end()) {
@@ -505,12 +512,14 @@ 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.code,
+
+                        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.code, std::move(shader));
+                        impl->fragment_shaders.Inject(conf, decomp->second.result.code,
+                                                      std::move(shader));
                     } else {
                         // Unsupported shader type got stored somehow so nuke the cache
 
@@ -554,6 +563,7 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
             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
@@ -566,6 +576,7 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
                 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);
@@ -587,7 +598,7 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
             }
             // If this is a new shader, add it the precompiled cache
             if (result) {
-                disk_cache.SaveDecompiled(unique_identifier, *result);
+                disk_cache.SaveDecompiled(unique_identifier, *result, sanitize_mul);
                 disk_cache.SaveDump(unique_identifier, handle);
                 precompiled_cache_altered = true;
             }
@@ -607,6 +618,6 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
     if (precompiled_cache_altered) {
         disk_cache.SaveVirtualPrecompiledFile();
     }
-}
+} // namespace OpenGL
 
 } // namespace OpenGL