feat: 模型异步加载——assimp 解析移后台线程(SunTemple 3s 卡顿消除)

Model 拆两段式:ParseFromDisk(static 纯 CPU,assimp 解析 → 几何 +
材质引用,worker 可调)/ FromParsed(GPU 构建:纹理上传 / mesh 缓冲 /
材质描述符,主线程),同步路径等价两段串行。第二个异步任务类型落地,
线程收割骨架抽成 Core/AsyncTask<T>(thread + atomic + Take,覆盖式
Run;pending 接力语义留调用方)。

实测(SunTemple 1056 mesh):解析 2.5s 全程在 worker,主线程持续
渲染;GPU 构建 0.56s。assets 层经 IAssetUploader/AssetManager 重载
传预解析产物(路径缓存去重,路径已缓存则产物丢弃)。

已知边界:纹理解码(backpack 4k PNG ~1s)仍在 GPU 段主线程——
Texture 需支持预解码像素注入,后续演进。
This commit is contained in:
2026-09-24 19:31:14 +08:00
parent 6df3be2846
commit 99e6712cdf
9 changed files with 260 additions and 26 deletions
@@ -150,6 +150,8 @@ namespace Engine
AssetManager& operator=(const AssetManager&) = delete;
ModelRef LoadModel(const std::string& path);
// 异步路径:解析已在后台线程完成(路径已缓存则产物丢弃去重)
ModelRef LoadModel(const std::string& path, ParsedModel&& parsed);
Model* GetModel(ModelRef h) const { return m_Models.Get(h); }
void UnloadModel(ModelRef h) { m_Models.Unload(h); }
@@ -30,6 +30,10 @@ namespace Engine::Assets {
// 磁盘模型 → 解析 + GPU 上传(失败返回 nullptr)
virtual Scope<Model> LoadModel(const std::string& path) = 0;
// 异步路径:解析已在后台线程完成,此调用只做 GPU 构建(主线程)
virtual Scope<Model> LoadModel(const std::string& path,
ParsedModel&& parsed) = 0;
// 程序化纯色平面(size 为半边长)
virtual Scope<Model> CreatePlane(f32 halfSize, u8 r, u8 g, u8 b) = 0;
+82
View File
@@ -0,0 +1,82 @@
//
// Created by atdunbg on 2026/9/24.
//
// AsyncTask:单 in-flight 异步任务的通用收割工具(header-only)。
//
// 异步加载的公共骨架(已验证两次:环境 DecodeHdr / 模型 ParseFromDisk):
// worker 线程跑纯 CPU 任务 → 原子标志通知 → 主线程每帧轮询收割。
// Vulkan 提交/描述符等 GPU 工作全部留在主线程(设备非线程安全)。
//
// AsyncTask<ParsedModel> task;
// task.Run([] { return Model::ParseFromDisk(path); }); // 发起
// if (task.Done()) { auto p = task.Take(); ... } // 每帧收割
// // 析构自动 join(worker 只做纯 CPU 工作,join 恒安全)
//
// 语义:Run 在忙碌时先 join 并丢弃旧结果(覆盖式);"快速连切只保留
// 最终意图"的 pending 接力由调用方持有(参照 main.cpp 环境加载)。
//
#ifndef HELLOVULKAN_ASYNCTASK_H
#define HELLOVULKAN_ASYNCTASK_H
#include <Engine/Core/Types.h>
#include <atomic>
#include <functional>
#include <thread>
#include <utility>
namespace Engine {
template <typename T>
class AsyncTask {
public:
AsyncTask() = default;
~AsyncTask() { Join(); }
AsyncTask(const AsyncTask&) = delete;
AsyncTask& operator=(const AsyncTask&) = delete;
// 发起任务。忙碌时先等待旧任务完成并放弃其结果。
void Run(std::function<T()> job)
{
Join();
m_Busy = true;
m_Done.store(false);
m_Worker = std::thread([this, j = std::move(job)]() {
m_Result = j();
m_Done.store(true); // release:结果对主线程 acquire 可见
});
}
bool Busy() const { return m_Busy; }
bool Done() const { return m_Busy && m_Done.load(); }
// 收割:join worker 并移交结果。只在 Done() 为 true 时调用。
T Take()
{
m_Worker.join(); // join = 同步点,此后结果归主线程
m_Busy = false;
T out = std::move(m_Result);
m_Result = T{};
return out;
}
// 关闭前兜底等待(析构自动调用)
void Join()
{
if (m_Worker.joinable()) {
m_Worker.join();
}
m_Busy = false;
}
private:
std::thread m_Worker;
std::atomic<bool> m_Done{false};
T m_Result{};
bool m_Busy = false;
};
} // namespace Engine
#endif //HELLOVULKAN_ASYNCTASK_H
+41 -2
View File
@@ -12,6 +12,7 @@
#include <Engine/Renderer/Mesh.h>
#include <vulkan/vulkan.hpp>
#include <array>
#include <filesystem>
#include <string>
#include <unordered_map>
#include <vector>
@@ -28,14 +29,42 @@ namespace Engine
class Texture;
class Material;
// ---- CPU 解析产物(异步加载的跨线程交棒数据)----
// 纯数据:几何 + 材质槽位引用(路径 + 因子),无任何 GPU 资源。
// worker 线程产出(ParseFromDisk),主线程消费(FromParsed)。
struct ParsedMesh {
std::vector<Vertex> vertices;
std::vector<u32> indices;
std::array<std::string, Material::kSlotCount> texPaths;
MaterialParams params;
};
struct ParsedModel {
std::string path;
bool isGltf = false;
std::vector<ParsedMesh> meshes;
bool Valid() const { return !meshes.empty(); }
};
class Model
{
public:
// 同步便捷路径(等价 ParseFromDisk + FromParsed,启动期小模型可用)
Model(Device& device, MemoryAllocator& allocator,
DescriptorManager& descriptorManager,
vk::DescriptorSetLayout textureLayout,
const std::string& path);
// ---- 异步加载两段式(assimp 解析占大模型耗时大头,可搬后台线程)----
// 纯 CPU:读盘 + assimp 解析 → 几何 + 材质引用。无线程状态。
static ParsedModel ParseFromDisk(const std::string& path);
// GPU 构建(必须主线程):纹理上传 / mesh 缓冲 / 材质描述符。
// 解析失败返回 nullptr。
static Scope<Model> FromParsed(Device& device, MemoryAllocator& allocator,
DescriptorManager& descriptorManager,
vk::DescriptorSetLayout textureLayout,
ParsedModel&& parsed);
// 程序化模型:单位法线朝上的 XZ 平面(纯色贴图),用于地面等。
// size 为半边长(平面总边长 = 2*size)。
struct ProceduralTag {};
@@ -57,9 +86,19 @@ namespace Engine
vk::DescriptorSetLayout textureLayout,
ProceduralTag, f32 halfSize, const std::array<u8, 3>& color);
// 空模型骨架(FromParsed 用):只建默认纹理,mesh 由调用方填
Model(Device& device, MemoryAllocator& allocator,
DescriptorManager& descriptorManager,
vk::DescriptorSetLayout textureLayout);
void LoadModel(const std::string& path);
void ProcessNode(aiNode* node, const aiScene* scene);
Scope<Mesh> ProcessMesh(aiMesh* mesh, const aiScene* scene);
static void ProcessNode(const aiNode* node, const aiScene* scene,
std::vector<ParsedMesh>& out,
const std::filesystem::path& modelDir,
bool isGltf);
static ParsedMesh ProcessMesh(const aiMesh* mesh, const aiScene* scene,
const std::filesystem::path& modelDir,
bool isGltf);
void InitDefaultTextures();
Texture* GetOrLoadTexture(const std::string& path, bool srgb);
@@ -32,6 +32,8 @@ namespace Engine {
m_Ibl(ibl) {}
Scope<Model> LoadModel(const std::string& path) override;
Scope<Model> LoadModel(const std::string& path,
ParsedModel&& parsed) override;
Scope<Model> CreatePlane(f32 halfSize, u8 r, u8 g, u8 b) override;
Scope<Cubemap> LoadEnvironment(const std::string& path) override;
Scope<Cubemap> LoadEnvironment(const std::string& path,
+11
View File
@@ -20,6 +20,17 @@ namespace Engine::Assets {
[this](const std::string& p) { return m_Uploader.LoadModel(p); });
}
ModelRef AssetManager::LoadModel(const std::string& path,
ParsedModel&& parsed)
{
// 预解析数据移交 loader(Load 同步调用 lambda,引用捕获安全);
// 路径已缓存则 loader 不执行,parsed 自然丢弃(去重)
return m_Models.Load(path,
[this, &parsed](const std::string& p) {
return m_Uploader.LoadModel(p, std::move(parsed));
});
}
ModelRef AssetManager::CreatePlane(f32 halfSize, u8 r, u8 g, u8 b)
{
// 程序化资产复用同一缓存:键 = 伪路径 + 参数
+82 -23
View File
@@ -23,14 +23,22 @@ Model::Model(Device& device, MemoryAllocator& allocator,
DescriptorManager& descriptorManager,
vk::DescriptorSetLayout textureLayout,
const std::string& path)
: Model(device, allocator, descriptorManager, textureLayout)
{
// 同步便捷路径 = 两段式串行(等价异步收割后的主线程消费)
LoadModel(path);
}
// 空模型骨架:只建默认纹理;mesh/材质由 FromParsed 填充
Model::Model(Device& device, MemoryAllocator& allocator,
DescriptorManager& descriptorManager,
vk::DescriptorSetLayout textureLayout)
: m_Device(device)
, m_Allocator(allocator)
, m_DescriptorManager(descriptorManager)
, m_TextureLayout(textureLayout)
{
InitDefaultTextures();
LoadModel(path);
}
Model::~Model() = default;
@@ -92,9 +100,10 @@ Scope<Model> Model::CreatePlane(Device& device, MemoryAllocator& allocator,
halfSize, std::array<u8, 3>{r, g, b}));
}
void Model::LoadModel(const std::string& path)
ParsedModel Model::ParseFromDisk(const std::string& path)
{
m_ModelPath = path;
ParsedModel out;
out.path = path;
// UV 翻转按格式区分:
// OBJ/FBX —— v 原点在底部,需翻到左上惯例(配合"v=0=图像顶"采样)
@@ -106,36 +115,84 @@ void Model::LoadModel(const std::string& path)
[](unsigned char c) { return std::tolower(c); });
return s;
}();
m_IsGltf = lower.size() >= 5 &&
out.isGltf = lower.size() >= 5 &&
(lower.rfind(".gltf") == lower.size() - 5 ||
lower.rfind(".glb") == lower.size() - 4);
Assimp::Importer importer;
const aiScene* scene = importer.ReadFile(path,
aiProcess_Triangulate |
(m_IsGltf ? 0 : aiProcess_FlipUVs) |
(out.isGltf ? 0 : aiProcess_FlipUVs) |
aiProcess_GenSmoothNormals |
aiProcess_CalcTangentSpace | // 法线贴图 TBN
aiProcess_JoinIdenticalVertices);
if (!scene || (scene->mFlags & AI_SCENE_FLAGS_INCOMPLETE) || !scene->mRootNode) {
ENGINE_ERROR("Assimp error: {}", importer.GetErrorString());
return;
return out; // meshes 空 = 无效
}
ProcessNode(scene->mRootNode, scene);
const std::filesystem::path modelDir =
std::filesystem::path(path).parent_path();
ProcessNode(scene->mRootNode, scene, out.meshes, modelDir, out.isGltf);
ENGINE_INFO("Model parsed: {} ({} meshes)", path, out.meshes.size());
return out;
}
Scope<Model> Model::FromParsed(Device& device, MemoryAllocator& allocator,
DescriptorManager& descriptorManager,
vk::DescriptorSetLayout textureLayout,
ParsedModel&& parsed)
{
if (!parsed.Valid()) {
ENGINE_ERROR("Model::FromParsed: invalid parsed model ({})", parsed.path);
return nullptr;
}
auto model = Scope<Model>(new Model(device, allocator,
descriptorManager, textureLayout));
model->m_ModelPath = parsed.path;
model->m_IsGltf = parsed.isGltf;
// 材质缓存去重 + 纹理上传 + mesh 缓冲(全部 GPU,主线程)
for (auto& pm : parsed.meshes) {
Material* mat = model->GetOrCreateMaterial(pm.texPaths.data(), pm.params);
model->m_Meshes.push_back(CreateScope<Mesh>(
device, allocator, std::move(pm.vertices), std::move(pm.indices), mat));
}
ENGINE_INFO("Model loaded: {} ({} meshes, {} textures)",
parsed.path, model->m_Meshes.size(), model->m_TextureCache.size());
return model;
}
void Model::LoadModel(const std::string& path)
{
// 同步路径:解析 + GPU 构建串行(与 FromParsed 相同的消费逻辑)
ParsedModel parsed = ParseFromDisk(path);
if (!parsed.Valid()) {
return;
}
m_ModelPath = parsed.path;
m_IsGltf = parsed.isGltf;
for (auto& pm : parsed.meshes) {
Material* mat = GetOrCreateMaterial(pm.texPaths.data(), pm.params);
m_Meshes.push_back(CreateScope<Mesh>(
m_Device, m_Allocator, std::move(pm.vertices), std::move(pm.indices), mat));
}
ENGINE_INFO("Model loaded: {} ({} meshes, {} textures)",
path, m_Meshes.size(), m_TextureCache.size());
}
void Model::ProcessNode(aiNode* node, const aiScene* scene)
void Model::ProcessNode(const aiNode* node, const aiScene* scene,
std::vector<ParsedMesh>& out,
const std::filesystem::path& modelDir, bool isGltf)
{
for (u32 i = 0; i < node->mNumMeshes; ++i) {
aiMesh* mesh = scene->mMeshes[node->mMeshes[i]];
m_Meshes.push_back(ProcessMesh(mesh, scene));
out.push_back(ProcessMesh(mesh, scene, modelDir, isGltf));
}
for (u32 i = 0; i < node->mNumChildren; ++i) {
ProcessNode(node->mChildren[i], scene);
ProcessNode(node->mChildren[i], scene, out, modelDir, isGltf);
}
}
@@ -193,10 +250,13 @@ Material* Model::GetOrCreateMaterial(const std::string paths[Material::kSlotCoun
return ptr;
}
Scope<Mesh> Model::ProcessMesh(aiMesh* mesh, const aiScene* scene)
ParsedMesh Model::ProcessMesh(const aiMesh* mesh, const aiScene* scene,
const std::filesystem::path& modelDir,
bool isGltf)
{
std::vector<Vertex> vertices;
std::vector<u32> indices;
ParsedMesh out;
std::vector<Vertex>& vertices = out.vertices;
std::vector<u32>& indices = out.indices;
vertices.reserve(mesh->mNumVertices);
@@ -248,9 +308,7 @@ Scope<Mesh> Model::ProcessMesh(aiMesh* mesh, const aiScene* scene)
MaterialParams params;
if (mesh->mMaterialIndex >= 0) {
aiMaterial* material = scene->mMaterials[mesh->mMaterialIndex];
const std::filesystem::path modelDir =
std::filesystem::path(m_ModelPath).parent_path();
const aiMaterial* material = scene->mMaterials[mesh->mMaterialIndex];
// glTF2 贴图在 assimp 5.x 的落点:MR 合图 → METALNESS(15)/
// DIFFUSE_ROUGHNESS(16)(同图双槽);baseColor → BASE_COLOR(12)。
@@ -304,7 +362,7 @@ Scope<Mesh> Model::ProcessMesh(aiMesh* mesh, const aiScene* scene)
// roughness 有独立贴图 → 因子 1(G 通道直达);
// 无贴图 → Blinn-Phong Ns 换算 R=sqrt(2/(Ns+2)),
// 补偿默认 MR 贴图 G=0.5 → 因子 = 2R
if (m_IsGltf) {
if (isGltf) {
ai_real metallic = 1.0f, roughness = 1.0f;
material->Get(AI_MATKEY_METALLIC_FACTOR, metallic);
material->Get(AI_MATKEY_ROUGHNESS_FACTOR, roughness);
@@ -323,11 +381,12 @@ Scope<Mesh> Model::ProcessMesh(aiMesh* mesh, const aiScene* scene)
}
}
// 材质缓存:同贴图组合+因子 → 同一 Material(一个描述符集)
Material* mat = GetOrCreateMaterial(texPaths, params);
return CreateScope<Mesh>(
m_Device, m_Allocator, vertices, indices, mat);
// 材质创建延迟到 GPU 构建段(FromParsed):解析期只留引用
for (u32 slot = 0; slot < Material::kSlotCount; ++slot) {
out.texPaths[slot] = std::move(texPaths[slot]);
}
out.params = params;
return out;
}
} // namespace Engine
@@ -15,6 +15,14 @@ namespace Engine {
m_Device, m_Allocator, m_Descriptors, m_TextureLayout, path);
}
Scope<Model> VulkanAssetUploader::LoadModel(const std::string& path,
ParsedModel&& parsed)
{
// 异步路径:assimp 解析已在 worker 线程完成,此处只剩 GPU 构建
return Model::FromParsed(m_Device, m_Allocator, m_Descriptors,
m_TextureLayout, std::move(parsed));
}
Scope<Model> VulkanAssetUploader::CreatePlane(f32 halfSize, u8 r, u8 g, u8 b)
{
return Model::CreatePlane(m_Device, m_Allocator, m_Descriptors,