Merge pull request #9708 from ameerj/gl-context-flush

gl_shader_cache: Force context flush when loading disk shader cache

src/video_core/renderer_opengl/gl_compute_pipeline.cpp

@@ -30,7 +30,7 @@ bool ComputePipelineKey::operator==(const ComputePipelineKey& rhs) const noexcep
 ComputePipeline::ComputePipeline(const Device& device, TextureCache& texture_cache_,
                                  BufferCache& buffer_cache_, ProgramManager& program_manager_,
                                  const Shader::Info& info_, std::string code,
-                                 std::vector<u32> code_v)
+                                 std::vector<u32> code_v, bool force_context_flush)
     : texture_cache{texture_cache_}, buffer_cache{buffer_cache_},
       program_manager{program_manager_}, info{info_} {
     switch (device.GetShaderBackend()) {
@@ -63,6 +63,15 @@ ComputePipeline::ComputePipeline(const Device& device, TextureCache& texture_cac
     writes_global_memory = !use_storage_buffers &&
                            std::ranges::any_of(info.storage_buffers_descriptors,
                                                [](const auto& desc) { return desc.is_written; });
+    if (force_context_flush) {
+        std::scoped_lock lock{built_mutex};
+        built_fence.Create();
+        // Flush this context to ensure compilation commands and fence are in the GPU pipe.
+        glFlush();
+        built_condvar.notify_one();
+    } else {
+        is_built = true;
+    }
 }
 
 void ComputePipeline::Configure() {
@@ -142,6 +151,9 @@ void ComputePipeline::Configure() {
     }
     texture_cache.FillComputeImageViews(std::span(views.data(), views.size()));
 
+    if (!is_built) {
+        WaitForBuild();
+    }
     if (assembly_program.handle != 0) {
         program_manager.BindComputeAssemblyProgram(assembly_program.handle);
     } else {
@@ -223,4 +235,13 @@ void ComputePipeline::Configure() {
     }
 }
 
+void ComputePipeline::WaitForBuild() {
+    if (built_fence.handle == 0) {
+        std::unique_lock lock{built_mutex};
+        built_condvar.wait(lock, [this] { return built_fence.handle != 0; });
+    }
+    ASSERT(glClientWaitSync(built_fence.handle, 0, GL_TIMEOUT_IGNORED) != GL_WAIT_FAILED);
+    is_built = true;
+}
+
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_compute_pipeline.h

@@ -50,7 +50,8 @@ class ComputePipeline {
 public:
     explicit ComputePipeline(const Device& device, TextureCache& texture_cache_,
                              BufferCache& buffer_cache_, ProgramManager& program_manager_,
-                             const Shader::Info& info_, std::string code, std::vector<u32> code_v);
+                             const Shader::Info& info_, std::string code, std::vector<u32> code_v,
+                             bool force_context_flush = false);
 
     void Configure();
 
@@ -65,6 +66,8 @@ public:
     }
 
 private:
+    void WaitForBuild();
+
     TextureCache& texture_cache;
     BufferCache& buffer_cache;
     Tegra::MemoryManager* gpu_memory;
@@ -81,6 +84,11 @@ private:
 
     bool use_storage_buffers{};
     bool writes_global_memory{};
+
+    std::mutex built_mutex;
+    std::condition_variable built_condvar;
+    OGLSync built_fence{};
+    bool is_built{false};
 };
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_graphics_pipeline.cpp

@@ -176,7 +176,7 @@ GraphicsPipeline::GraphicsPipeline(const Device& device, TextureCache& texture_c
                                    std::array<std::string, 5> sources,
                                    std::array<std::vector<u32>, 5> sources_spirv,
                                    const std::array<const Shader::Info*, 5>& infos,
-                                   const GraphicsPipelineKey& key_)
+                                   const GraphicsPipelineKey& key_, bool force_context_flush)
     : texture_cache{texture_cache_}, buffer_cache{buffer_cache_}, program_manager{program_manager_},
       state_tracker{state_tracker_}, key{key_} {
     if (shader_notify) {
@@ -231,7 +231,8 @@ GraphicsPipeline::GraphicsPipeline(const Device& device, TextureCache& texture_c
     const bool in_parallel = thread_worker != nullptr;
     const auto backend = device.GetShaderBackend();
     auto func{[this, sources = std::move(sources), sources_spirv = std::move(sources_spirv),
-               shader_notify, backend, in_parallel](ShaderContext::Context*) mutable {
+               shader_notify, backend, in_parallel,
+               force_context_flush](ShaderContext::Context*) mutable {
         for (size_t stage = 0; stage < 5; ++stage) {
             switch (backend) {
             case Settings::ShaderBackend::GLSL:
@@ -251,7 +252,7 @@ GraphicsPipeline::GraphicsPipeline(const Device& device, TextureCache& texture_c
                 break;
             }
         }
-        if (in_parallel) {
+        if (force_context_flush || in_parallel) {
             std::scoped_lock lock{built_mutex};
             built_fence.Create();
             // Flush this context to ensure compilation commands and fence are in the GPU pipe.

src/video_core/renderer_opengl/gl_graphics_pipeline.h

@@ -78,7 +78,7 @@ public:
                               std::array<std::string, 5> sources,
                               std::array<std::vector<u32>, 5> sources_spirv,
                               const std::array<const Shader::Info*, 5>& infos,
-                              const GraphicsPipelineKey& key_);
+                              const GraphicsPipelineKey& key_, bool force_context_flush = false);
 
     void Configure(bool is_indexed) {
         configure_func(this, is_indexed);

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -286,7 +286,7 @@ void ShaderCache::LoadDiskResources(u64 title_id, std::stop_token stop_loading,
         file.read(reinterpret_cast<char*>(&key), sizeof(key));
         queue_work([this, key, env = std::move(env), &state, &callback](Context* ctx) mutable {
             ctx->pools.ReleaseContents();
-            auto pipeline{CreateComputePipeline(ctx->pools, key, env)};
+            auto pipeline{CreateComputePipeline(ctx->pools, key, env, true)};
             std::scoped_lock lock{state.mutex};
             if (pipeline) {
                 compute_cache.emplace(key, std::move(pipeline));
@@ -307,7 +307,7 @@ void ShaderCache::LoadDiskResources(u64 title_id, std::stop_token stop_loading,
                 env_ptrs.push_back(&env);
             }
             ctx->pools.ReleaseContents();
-            auto pipeline{CreateGraphicsPipeline(ctx->pools, key, MakeSpan(env_ptrs), false)};
+            auto pipeline{CreateGraphicsPipeline(ctx->pools, key, MakeSpan(env_ptrs), false, true)};
             std::scoped_lock lock{state.mutex};
             if (pipeline) {
                 graphics_cache.emplace(key, std::move(pipeline));
@@ -439,7 +439,8 @@ std::unique_ptr<GraphicsPipeline> ShaderCache::CreateGraphicsPipeline() {
 
 std::unique_ptr<GraphicsPipeline> ShaderCache::CreateGraphicsPipeline(
     ShaderContext::ShaderPools& pools, const GraphicsPipelineKey& key,
-    std::span<Shader::Environment* const> envs, bool build_in_parallel) try {
+    std::span<Shader::Environment* const> envs, bool use_shader_workers,
+    bool force_context_flush) try {
     LOG_INFO(Render_OpenGL, "0x{:016x}", key.Hash());
     size_t env_index{};
     u32 total_storage_buffers{};
@@ -531,10 +532,10 @@ std::unique_ptr<GraphicsPipeline> ShaderCache::CreateGraphicsPipeline(
         }
         previous_program = &program;
     }
-    auto* const thread_worker{build_in_parallel ? workers.get() : nullptr};
+    auto* const thread_worker{use_shader_workers ? workers.get() : nullptr};
     return std::make_unique<GraphicsPipeline>(device, texture_cache, buffer_cache, program_manager,
                                               state_tracker, thread_worker, &shader_notify, sources,
-                                              sources_spirv, infos, key);
+                                              sources_spirv, infos, key, force_context_flush);
 
 } catch (Shader::Exception& exception) {
     LOG_ERROR(Render_OpenGL, "{}", exception.what());
@@ -559,8 +560,8 @@ std::unique_ptr<ComputePipeline> ShaderCache::CreateComputePipeline(
 }
 
 std::unique_ptr<ComputePipeline> ShaderCache::CreateComputePipeline(
-    ShaderContext::ShaderPools& pools, const ComputePipelineKey& key,
-    Shader::Environment& env) try {
+    ShaderContext::ShaderPools& pools, const ComputePipelineKey& key, Shader::Environment& env,
+    bool force_context_flush) try {
     LOG_INFO(Render_OpenGL, "0x{:016x}", key.Hash());
 
     Shader::Maxwell::Flow::CFG cfg{env, pools.flow_block, env.StartAddress()};
@@ -589,7 +590,7 @@ std::unique_ptr<ComputePipeline> ShaderCache::CreateComputePipeline(
     }
 
     return std::make_unique<ComputePipeline>(device, texture_cache, buffer_cache, program_manager,
-                                             program.info, code, code_spirv);
+                                             program.info, code, code_spirv, force_context_flush);
 } catch (Shader::Exception& exception) {
     LOG_ERROR(Render_OpenGL, "{}", exception.what());
     return nullptr;

src/video_core/renderer_opengl/gl_shader_cache.h

@@ -50,14 +50,16 @@ private:
 
     std::unique_ptr<GraphicsPipeline> CreateGraphicsPipeline(
         ShaderContext::ShaderPools& pools, const GraphicsPipelineKey& key,
-        std::span<Shader::Environment* const> envs, bool build_in_parallel);
+        std::span<Shader::Environment* const> envs, bool use_shader_workers,
+        bool force_context_flush = false);
 
     std::unique_ptr<ComputePipeline> CreateComputePipeline(const ComputePipelineKey& key,
                                                            const VideoCommon::ShaderInfo* shader);
 
     std::unique_ptr<ComputePipeline> CreateComputePipeline(ShaderContext::ShaderPools& pools,
                                                            const ComputePipelineKey& key,
-                                                           Shader::Environment& env);
+                                                           Shader::Environment& env,
+                                                           bool force_context_flush = false);
 
     std::unique_ptr<ShaderWorker> CreateWorkers() const;