aurora: Async pipeline creation

aurora/lib/aurora.cpp

@@ -37,6 +37,7 @@ static void set_window_icon(Icon icon) noexcept {
   );
   if (iconSurface == nullptr) {
     Log.report(logvisor::Fatal, FMT_STRING("Failed to create icon surface: {}"), SDL_GetError());
+    unreachable();
   }
   SDL_SetWindowIcon(g_Window, iconSurface);
   SDL_FreeSurface(iconSurface);
@@ -173,6 +174,7 @@ void app_run(std::unique_ptr<AppDelegate> app, Icon icon, int argc, char** argv)
 
   if (SDL_Init(SDL_INIT_EVERYTHING) < 0) {
     Log.report(logvisor::Fatal, FMT_STRING("Error initializing SDL: {}"), SDL_GetError());
+    unreachable();
   }
 
   Uint32 flags = SDL_WINDOW_SHOWN | SDL_WINDOW_ALLOW_HIGHDPI | SDL_WINDOW_RESIZABLE;
@@ -198,11 +200,12 @@ void app_run(std::unique_ptr<AppDelegate> app, Icon icon, int argc, char** argv)
   g_Window = SDL_CreateWindow("Metaforce", SDL_WINDOWPOS_UNDEFINED, SDL_WINDOWPOS_UNDEFINED, 1280, 720, flags);
   if (g_Window == nullptr) {
     Log.report(logvisor::Fatal, FMT_STRING("Error creating window: {}"), SDL_GetError());
+    unreachable();
   }
   set_window_icon(std::move(icon));
 
   gpu::initialize(g_Window);
-  gfx::construct_state();
+  gfx::initialize();
 
   imgui::create_context();
   g_AppDelegate->onImGuiInit(1.f); // TODO scale
@@ -282,6 +285,7 @@ void app_run(std::unique_ptr<AppDelegate> app, Icon icon, int argc, char** argv)
   g_AppDelegate->onAppExiting();
 
   imgui::shutdown();
+  gfx::shutdown();
   gpu::shutdown();
   SDL_DestroyWindow(g_Window);
   SDL_Quit();

aurora/lib/gfx/common.cpp

@@ -3,7 +3,10 @@
 #include "../gpu.hpp"
 #include "movie_player/shader.hpp"
 
+#include <condition_variable>
+#include <deque>
 #include <logvisor/logvisor.hpp>
+#include <thread>
 #include <unordered_map>
 
 namespace aurora::gfx {
@@ -55,9 +58,13 @@ zeus::CMatrix4f g_mvInv;
 zeus::CMatrix4f g_proj;
 metaforce::CFogState g_fogState;
 
+using NewPipelineCallback = std::function<wgpu::RenderPipeline()>;
 static std::mutex g_pipelineMutex;
-static std::unordered_map<uint64_t, wgpu::RenderPipeline> g_pipelines;
-static std::vector<PipelineCreateCommand> g_queuedPipelines;
+static std::thread g_pipelineThread;
+static bool g_pipelineThreadEnd = false;
+static std::condition_variable g_pipelineCv;
+static std::unordered_map<PipelineRef, wgpu::RenderPipeline> g_pipelines;
+static std::deque<std::pair<PipelineRef, NewPipelineCallback>> g_queuedPipelines;
 static std::unordered_map<BindGroupRef, wgpu::BindGroup> g_cachedBindGroups;
 
 static ByteBuffer g_verts;
@@ -72,22 +79,26 @@ static PipelineRef g_currentPipeline;
 
 static std::vector<Command> g_commands;
 
-using NewPipelineCallback = std::function<wgpu::RenderPipeline()>;
-static PipelineRef find_pipeline(PipelineCreateCommand command, NewPipelineCallback cb) {
+static PipelineRef find_pipeline(PipelineCreateCommand command, NewPipelineCallback&& cb) {
   const auto hash = xxh3_hash(command);
-  bool found;
+  bool found = false;
   {
     std::lock_guard guard{g_pipelineMutex};
-    const auto ref = g_pipelines.find(hash);
-    found = ref != g_pipelines.end();
+    found = g_pipelines.find(hash) != g_pipelines.end();
+    if (!found) {
+      const auto ref =
+          std::find_if(g_queuedPipelines.begin(), g_queuedPipelines.end(), [=](auto v) { return v.first == hash; });
+      if (ref != g_queuedPipelines.end()) {
+        found = true;
+      }
+    }
   }
   if (!found) {
-    // TODO another thread
-    wgpu::RenderPipeline pipeline = cb();
     {
       std::lock_guard guard{g_pipelineMutex};
-      g_pipelines[hash] = std::move(pipeline);
+      g_queuedPipelines.emplace_back(std::pair{hash, std::move(cb)});
     }
+    g_pipelineCv.notify_one();
   }
   return hash;
 }
@@ -157,7 +168,37 @@ PipelineRef pipeline_ref(movie_player::PipelineConfig config) {
                        [=]() { return create_pipeline(g_state.moviePlayer, config); });
 }
 
-void construct_state() {
+static void pipeline_worker() {
+  bool hasMore = false;
+  while (true) {
+    std::pair<PipelineRef, NewPipelineCallback> cb;
+    {
+      std::unique_lock lock{g_pipelineMutex};
+      if (!hasMore) {
+        g_pipelineCv.wait(lock, [] { return !g_queuedPipelines.empty() || g_pipelineThreadEnd; });
+      }
+      if (g_pipelineThreadEnd) {
+        break;
+      }
+      cb = std::move(g_queuedPipelines.front());
+    }
+    auto result = cb.second();
+    {
+      std::lock_guard lock{g_pipelineMutex};
+      if (g_pipelines.contains(cb.first)) {
+        Log.report(logvisor::Fatal, FMT_STRING("Duplicate pipeline {}"), cb.first);
+        unreachable();
+      }
+      g_pipelines[cb.first] = result;
+      g_queuedPipelines.pop_front();
+      hasMore = !g_queuedPipelines.empty();
+    }
+  }
+}
+
+void initialize() {
+  g_pipelineThread = std::thread(pipeline_worker);
+
   {
     const auto uniformDescriptor = wgpu::BufferDescriptor{
         .label = "Shared Uniform Buffer",
@@ -186,6 +227,12 @@ void construct_state() {
   g_state.moviePlayer = movie_player::construct_state();
 }
 
+void shutdown() {
+  g_pipelineThreadEnd = true;
+  g_pipelineCv.notify_all();
+  g_pipelineThread.join();
+}
+
 void render(const wgpu::RenderPassEncoder& pass) {
   {
     g_queue.WriteBuffer(g_vertexBuffer, 0, g_verts.data(), g_verts.size());
@@ -272,6 +319,7 @@ BindGroupRef bind_group_ref(const wgpu::BindGroupDescriptor& descriptor) {
 const wgpu::BindGroup& find_bind_group(BindGroupRef id) {
   if (!g_cachedBindGroups.contains(id)) {
     Log.report(logvisor::Fatal, FMT_STRING("get_bind_group: failed to locate {}"), id);
+    unreachable();
   }
   return g_cachedBindGroups[id];
 }

aurora/lib/gfx/common.hpp

@@ -126,7 +126,8 @@ enum class ShaderType {
   MoviePlayer,
 };
 
-void construct_state();
+void initialize();
+void shutdown();
 
 void render(const wgpu::RenderPassEncoder& pass);
 

aurora/lib/gfx/movie_player/shader.wgsl

@@ -1,42 +0,0 @@
-struct Uniform {
-    xf: mat4x4<f32>;
-    color: vec4<f32>;
-};
-@group(0) @binding(0)
-var<uniform> ubuf: Uniform;
-@group(0) @binding(1)
-var tex_sampler: sampler;
-@group(1) @binding(0)
-var tex_y: texture_2d<f32>;
-@group(1) @binding(1)
-var tex_u: texture_2d<f32>;
-@group(1) @binding(2)
-var tex_v: texture_2d<f32>;
-
-struct VertexOutput {
-    @builtin(position) pos: vec4<f32>;
-    @location(0) uv: vec2<f32>;
-};
-
-@stage(vertex)
-fn vs_main(@location(0) in_pos: vec3<f32>, @location(1) in_uv: vec2<f32>) -> VertexOutput {
-    var out: VertexOutput;
-    out.pos = ubuf.xf * vec4<f32>(in_pos, 1.0);
-    out.uv = in_uv;
-    return out;
-}
-
-@stage(fragment)
-fn fs_main(in: VertexOutput) -> @location(0) vec4<f32> {
-    var yuv = vec3<f32>(
-        1.1643 * (textureSample(tex_y, tex_sampler, in.uv).x - 0.0625),
-        textureSample(tex_u, tex_sampler, in.uv).x - 0.5,
-        textureSample(tex_v, tex_sampler, in.uv).x - 0.5
-    );
-    return ubuf.color * vec4<f32>(
-        yuv.x + 1.5958 * yuv.z,
-        yuv.x - 0.39173 * yuv.y - 0.8129 * yuv.z,
-        yuv.x + 2.017 * yuv.y,
-        1.0
-    );
-}

aurora/lib/gfx/texture.cpp

@@ -2,6 +2,7 @@
 
 #include "../gpu.hpp"
 
+#include <logvisor/logvisor.hpp>
 #include <magic_enum.hpp>
 
 namespace aurora::gfx {
@@ -45,7 +46,7 @@ static TextureFormatInfo format_info(wgpu::TextureFormat format) {
     return {4, 4, 8, true};
   default:
     Log.report(logvisor::Fatal, FMT_STRING("format_info: unimplemented format {}"), magic_enum::enum_name(format));
-    return {0, 0, 0, false};
+    unreachable();
   }
 }
 static wgpu::Extent3D physical_size(wgpu::Extent3D size, TextureFormatInfo info) {
@@ -74,7 +75,7 @@ TextureHandle new_static_texture_2d(uint32_t width, uint32_t height, uint32_t mi
     if (offset + dataSize > data.size()) {
       Log.report(logvisor::Fatal, FMT_STRING("new_static_texture_2d[{}]: expected at least {} bytes, got {}"), label,
                  offset + dataSize, data.size());
-      return {};
+      unreachable();
     }
     const auto dstView = wgpu::ImageCopyTexture{
         .texture = ref.texture,
@@ -133,7 +134,7 @@ void write_texture(const TextureHandle& handle, ArrayRef<uint8_t> data) noexcept
   };
   if (ref.size.depthOrArrayLayers != 1) {
     Log.report(logvisor::Fatal, FMT_STRING("write_texture: unsupported depth {}"), ref.size.depthOrArrayLayers);
-    return;
+    unreachable();
   }
   const auto info = format_info(ref.format);
   const auto physicalSize = physical_size(ref.size, info);
@@ -143,7 +144,7 @@ void write_texture(const TextureHandle& handle, ArrayRef<uint8_t> data) noexcept
   const uint32_t dataSize = bytesPerRow * heightBlocks * ref.size.depthOrArrayLayers;
   if (dataSize > data.size()) {
     Log.report(logvisor::Fatal, FMT_STRING("write_texture: expected at least {} bytes, got {}"), dataSize, data.size());
-    return;
+    unreachable();
   }
   const auto dataLayout = wgpu::TextureDataLayout{
       .bytesPerRow = bytesPerRow,

aurora/lib/gpu.cpp

@@ -129,6 +129,7 @@ void initialize(SDL_Window* window) {
     });
     if (adapterIt == adapters.end()) {
       Log.report(logvisor::Fatal, FMT_STRING("Failed to find usable graphics backend"));
+      unreachable();
     }
     g_Adapter = *adapterIt;
   }
@@ -157,6 +158,7 @@ void initialize(SDL_Window* window) {
       std::unique_ptr<utils::BackendBinding>(utils::CreateBinding(g_backendType, window, g_device.Get()));
   if (!g_BackendBinding) {
     Log.report(logvisor::Fatal, FMT_STRING("Unsupported backend {}"), backendName);
+    unreachable();
   }
 
   auto swapChainFormat = static_cast<wgpu::TextureFormat>(g_BackendBinding->GetPreferredSwapChainTextureFormat());

aurora/lib/gpu.hpp

@@ -4,6 +4,14 @@
 #include <cstdint>
 #include <dawn/webgpu_cpp.h>
 
+#ifdef __GNUC__
+[[noreturn]] inline __attribute__((always_inline)) void unreachable() { __builtin_unreachable(); }
+#elif defined(_MSC_VER)
+[[noreturn]] __forceinline void unreachable() { __assume(false); }
+#else
+#error Unknown compiler
+#endif
+
 struct SDL_Window;
 
 namespace aurora::gpu {