From 370e480c8c6eee1ffdc2b718eb824112e1710f52 Mon Sep 17 00:00:00 2001
From: Ameer J <52414509+ameerj@users.noreply.github.com>
Date: Sun, 6 Mar 2022 10:46:49 -0500
Subject: [PATCH] gl_graphics_pipeline: Improve shader builder synchronization
 using fences (#7969)

* gl_graphics_pipeline: Improve shader builder synchronization

Make use of GLsync objects to ensure better synchronization between shader builder threads and the main context

* gl_graphics_pipeline: Make built_fence access threadsafe

* gl_graphics_pipeline: Use GLsync objects only when building in parallel

* gl_graphics_pipeline: Replace GetSync calls with non-blocking waits

The spec states that a ClientWait on a Fence object ensures the changes propagate to the calling context
---
 .../renderer_opengl/gl_graphics_pipeline.cpp  | 46 ++++++++++++-------
 .../renderer_opengl/gl_graphics_pipeline.h    |  7 ++-
 2 files changed, 32 insertions(+), 21 deletions(-)

diff --git a/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp b/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
index f8495896c8..9e6732abd4 100644
--- a/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
+++ b/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
@@ -243,10 +243,6 @@ GraphicsPipeline::GraphicsPipeline(
             case Settings::ShaderBackend::GLASM:
                 if (!sources[stage].empty()) {
                     assembly_programs[stage] = CompileProgram(sources[stage], AssemblyStage(stage));
-                    if (in_parallel) {
-                        // Make sure program is built before continuing when building in parallel
-                        glGetString(GL_PROGRAM_ERROR_STRING_NV);
-                    }
                 }
                 break;
             case Settings::ShaderBackend::SPIRV:
@@ -256,20 +252,18 @@ GraphicsPipeline::GraphicsPipeline(
                 break;
             }
         }
-        if (in_parallel && backend != Settings::ShaderBackend::GLASM) {
-            // Make sure programs have built if we are building shaders in parallel
-            for (OGLProgram& program : source_programs) {
-                if (program.handle != 0) {
-                    GLint status{};
-                    glGetProgramiv(program.handle, GL_LINK_STATUS, &status);
-                }
-            }
+        if (in_parallel) {
+            std::lock_guard 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;
         }
         if (shader_notify) {
             shader_notify->MarkShaderComplete();
         }
-        is_built = true;
-        built_condvar.notify_one();
     }};
     if (thread_worker) {
         thread_worker->QueueWork(std::move(func));
@@ -440,7 +434,7 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
     buffer_cache.UpdateGraphicsBuffers(is_indexed);
     buffer_cache.BindHostGeometryBuffers(is_indexed);
 
-    if (!is_built.load(std::memory_order::relaxed)) {
+    if (!IsBuilt()) {
         WaitForBuild();
     }
     const bool use_assembly{assembly_programs[0].handle != 0};
@@ -585,8 +579,26 @@ void GraphicsPipeline::GenerateTransformFeedbackState() {
 }
 
 void GraphicsPipeline::WaitForBuild() {
-    std::unique_lock lock{built_mutex};
-    built_condvar.wait(lock, [this] { return is_built.load(std::memory_order::relaxed); });
+    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;
+}
+
+bool GraphicsPipeline::IsBuilt() noexcept {
+    if (is_built) {
+        return true;
+    }
+    if (built_fence.handle == 0) {
+        return false;
+    }
+    // Timeout of zero means this is non-blocking
+    const auto sync_status = glClientWaitSync(built_fence.handle, 0, 0);
+    ASSERT(sync_status != GL_WAIT_FAILED);
+    is_built = sync_status != GL_TIMEOUT_EXPIRED;
+    return is_built;
 }
 
 } // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_graphics_pipeline.h b/src/video_core/renderer_opengl/gl_graphics_pipeline.h
index 4e28d9a423..311d49f3ff 100644
--- a/src/video_core/renderer_opengl/gl_graphics_pipeline.h
+++ b/src/video_core/renderer_opengl/gl_graphics_pipeline.h
@@ -100,9 +100,7 @@ public:
         return writes_global_memory;
     }
 
-    [[nodiscard]] bool IsBuilt() const noexcept {
-        return is_built.load(std::memory_order::relaxed);
-    }
+    [[nodiscard]] bool IsBuilt() noexcept;
 
     template <typename Spec>
     static auto MakeConfigureSpecFunc() {
@@ -154,7 +152,8 @@ private:
 
     std::mutex built_mutex;
     std::condition_variable built_condvar;
-    std::atomic_bool is_built{false};
+    OGLSync built_fence{};
+    bool is_built{false};
 };
 
 } // namespace OpenGL