8.3 Metal 与 D3D11 后端
这一节把两个平台后端并在一起看:Apple 的 Metal 与微软的 D3D11。它们都坚持「预编译的状态对象」思路——Metal 的 MTLRenderPipelineState 与 D3D11 的 ID3D11BlendState/RasterizerState/DepthStencilState——但缓存与脏标记策略不同。
本章与站内既有内容的分工
8.1/8.2 的 GLES2 与 Vulkan 各自展示了能力探测与设备管理;本节补齐剩下两个后端的关键机制:Metal 的 PSO 缓存与编码器状态转发、D3D11 的 feature level 分级与描述符延迟应用。lab 把两种机制的最小逻辑分别复现。
Metal:按 psd hash 缓存 MTLRenderPipelineState
Metal 的 MTLRenderPipelineDescriptor(成员 psd)描述管线全部状态;getPipelineState() 先用 [psd hash] 查 mPSOCache,命中直接返回缓存的 id<MTLRenderPipelineState>,未命中才创建——创建失败抛 ERR_RENDERINGAPI_ERROR:
// RenderSystems/Metal/src/OgreMetalRenderSystem.mm:MetalRenderSystem::getPipelineState(v14.6.3 源码摘录,已删减)
id<MTLRenderPipelineState> MetalRenderSystem::getPipelineState()
{
NSUInteger psdHash = [psd hash];
auto it = mPSOCache.find(psdHash);
if(it != mPSOCache.end())
return it->second;
NSError* error = NULL;
id <MTLRenderPipelineState> pso =
[mActiveDevice->mDevice newRenderPipelineStateWithDescriptor:psd error:&error];
if( !pso || error )
{
String errorDesc;
if( error )
errorDesc = [error localizedDescription].UTF8String;
OGRE_EXCEPT( Exception::ERR_RENDERINGAPI_ERROR,
"Failed to created pipeline state for rendering: " + errorDesc);
}
mPSOCache.emplace(psdHash, pso);
return pso;
}
psd 本身是渲染系统的成员,各状态 setter(_setDepthBias、_setAlphaRejectSettings)在绑定前改写它,下次取 PSO 时 hash 自然反映新状态。编码器 mActiveRenderEncoder 则是「当前命令流的接收者」,渲染期状态直接转发:
// RenderSystems/Metal/src/OgreMetalRenderSystem.mm:MetalRenderSystem::_setPolygonMode / _setDepthBias(v14.6.3 源码摘录,已删减)
void MetalRenderSystem::_setDepthBias(float constantBias, float slopeScaleBias)
{
[mActiveRenderEncoder setDepthBias:constantBias
slopeScale:slopeScaleBias
clamp:0.0f];
}
void MetalRenderSystem::_setPolygonMode(PolygonMode level)
{
MTLTriangleFillMode fillMode =
level == PM_SOLID ? MTLTriangleFillModeFill : MTLTriangleFillModeLines;
[mActiveRenderEncoder setTriangleFillMode:fillMode];
}
Metal 后端没有 D3D11 那种「changed 标记」:编码器是即时态对象,每次调用直接改状态。
D3D11:feature level 分级与能力分支
D3D11 的核心机制是 D3D_FEATURE_LEVEL。createD3D11Device 维护一个从高到低的请求数组(11_1 → 9_1),按配置的最小/最大请求裁剪出窗口,由 D3D 运行时从窗口顶部选设备支持的最高 level;如果 minFL 比 maxFL 还高,直接报 ERR_INTERNAL_ERROR:
// RenderSystems/Direct3D11/src/OgreD3D11RenderSystem.cpp:createD3D11Device(v14.6.3 教学示意,参照上游同名实现,已删减)
// determine feature levels
D3D_FEATURE_LEVEL requestedLevels[] = {
#if !__OGRE_WINRT_PHONE
#if defined(_WIN32_WINNT_WIN8) && _WIN32_WINNT >= _WIN32_WINNT_WIN8
D3D_FEATURE_LEVEL_11_1,
#endif
D3D_FEATURE_LEVEL_11_0,
D3D_FEATURE_LEVEL_10_1,
D3D_FEATURE_LEVEL_10_0,
#endif
D3D_FEATURE_LEVEL_9_3,
D3D_FEATURE_LEVEL_9_2,
D3D_FEATURE_LEVEL_9_1
};
D3D_FEATURE_LEVEL *pFirstFL = requestedLevels, *pLastFL = pFirstFL + ARRAYSIZE(requestedLevels) - 1;
for(unsigned int i = 0; i < ARRAYSIZE(requestedLevels); i++)
{
if(minFL == requestedLevels[i])
pLastFL = &requestedLevels[i];
if(maxFL == requestedLevels[i])
pFirstFL = &requestedLevels[i];
}
if(pLastFL < pFirstFL)
{
OGRE_EXCEPT(Exception::ERR_INTERNAL_ERROR,
"Requested min level feature is bigger the requested max level feature.",
"D3D11RenderSystem::initialise");
}
能力探测用 mFeatureLevel >= X 的线性比较:FL 9_2 起支持 32 位索引,FL 10_0 起支持读回纹理与双面模板;MRT 数量随 FL 阶梯上升——9_1 只有 1,9_3 有 4,高于 9_3 用 D3D11 全量 8:
// RenderSystems/Direct3D11/src/OgreD3D11RenderSystem.cpp:createRenderSystemCapabilities(v14.6.3 源码摘录,已删减)
UINT formatSupport;
if(mFeatureLevel >= D3D_FEATURE_LEVEL_9_2
|| SUCCEEDED(mDevice->CheckFormatSupport(DXGI_FORMAT_R32_UINT, &formatSupport)) && 0 != (formatSupport & D3D11_FORMAT_SUPPORT_IA_INDEX_BUFFER))
rsc->setCapability(RSC_32BIT_INDEX);
...
// Cube map
if (mFeatureLevel >= D3D_FEATURE_LEVEL_10_0)
{
rsc->setCapability(RSC_READ_BACK_AS_TEXTURE);
}
...
if(mFeatureLevel >= D3D_FEATURE_LEVEL_10_0)
rsc->setCapability(RSC_TWO_SIDED_STENCIL);
...
#ifdef D3D_FEATURE_LEVEL_9_3
int numMultiRenderTargets = (mFeatureLevel > D3D_FEATURE_LEVEL_9_3) ? D3D11_SIMULTANEOUS_RENDER_TARGET_COUNT : // 8
(mFeatureLevel == D3D_FEATURE_LEVEL_9_3) ? 4/*D3D_FL9_3_SIMULTANEOUS_RENDER_TARGET_COUNT*/ : // 4
1/*D3D_FL9_1_SIMULTANEOUS_RENDER_TARGET_COUNT*/; // 1
#else
int numMultiRenderTargets = D3D11_SIMULTANEOUS_RENDER_TARGET_COUNT; // 8
#endif
rsc->setNumMultiRenderTargets(std::min(numMultiRenderTargets, (int)OGRE_MAX_MULTIPLE_RENDER_TARGETS));
D3D11:描述符 changed 标记
与 Metal 的即时编码器不同,D3D11 的状态对象(blend/rasterizer/depth-stencil)创建成本高,OGRE 用 mXxxDescChanged 布尔标记「描述符已改但未应用」。每个 setter 只改描述符字段并置脏位;到 _render 时若脏位为真才创建新的 ID3D11BlendState 等并绑定:
// RenderSystems/Direct3D11/include/OgreD3D11RenderSystem.h:D3D11RenderSystem(v14.6.3 源码摘录,已删减)
D3D11_BLEND_DESC mBlendDesc;
bool mBlendDescChanged;
D3D11_RASTERIZER_DESC mRasterizerDesc;
bool mRasterizerDescChanged;
UINT mStencilRef;
D3D11_DEPTH_STENCIL_DESC mDepthStencilDesc;
bool mDepthStencilDescChanged;
...
ComPtr<ID3D11BlendState> mBoundBlendState;
ComPtr<ID3D11RasterizerState> mBoundRasterizer;
ComPtr<ID3D11DepthStencilState> mBoundDepthStencilState;
mBound* 是已经创建并绑定的状态对象;描述符 + 脏位 + 绑定对象三者构成一套「惰性重建」协议。Metal 侧的 mPSOCache 与 D3D11 侧的脏标记,是同一问题(昂贵状态对象复用)的两种解法:前者按描述符 hash 缓存全部历史,后者只保留「最新一份」。
能力查询对应实际设备与窗口
后端可被构建和加载之后,还需要真正的设备、上下文或窗口才能验证运行行为。Metal 与 D3D11 的设备等级、目标格式和支持特性不同。
同一场景上保留所选渲染后端、capability、驱动和目标格式。命令或截图缺失时,用固定版本源码解释路径并明确证据范围。
源码定位
RenderSystems/Metal/src/OgreMetalRenderSystem.mm:getPipelineState(794-818)、_setDepthBias(762-767)、_setPolygonMode(769-774)、_setAlphaRejectSettings(776-779)RenderSystems/Metal/include/OgreMetalRenderSystem.h:MetalRenderSystem(51-168)、mPSOCache(84)、mActiveRenderEncoder(74)RenderSystems/Direct3D11/src/OgreD3D11RenderSystem.cpp:createD3D11Device的 requestedLevels(272-299)、能力分支(803-822)、MRT 数量(893-901)RenderSystems/Direct3D11/include/OgreD3D11RenderSystem.h:D3D11RenderSystem(58-204)、描述符与脏标记(109-117)、mBound*(128-131)
上游引用固定到 OGRE v14.6.3 源码 。
可运行产物:C++ 最小验证
下面复现两种后端的核心机制:Metal 的 PSO 缓存(命中/未命中/失败)与编码器状态转发(_setPolygonMode 的 Fill/Lines 映射)、D3D11 的 feature level 能力分支(32 位索引、MRT 数量)与描述符脏标记。它与 src/ogre-ext-08-3-metal-d3d11-backend.cpp 逐字节一致:
// file: src/ogre-ext-08-3-metal-d3d11-backend.cpp
// tier: L1
// verify: clang++ -std=c++20 -I. src/ogre-ext-08-3-metal-d3d11-backend.cpp -o bin/ogre-ext-08-3-metal-d3d11-backend && ./bin/ogre-ext-08-3-metal-d3d11-backend
//
// 复现 Metal 与 D3D11 后端的两个核心机制
// (对应 OgreMetalRenderSystem.mm 794-818、OgreMetalRenderSystem.h 51-168、
// OgreD3D11RenderSystem.cpp 272-299、893-901、OgreD3D11RenderSystem.h 58-118):
// Metal:getPipelineState 按 psd hash 查 mPSOCache,未命中创建后缓存,失败抛异常;
// 渲染编码器直接收状态(triangleFillMode/cullMode);
// D3D11:feature level 分级决定 32 位索引、读回纹理、MRT 数量;
// 描述符用 changed 标记延迟应用(mBlendDescChanged 等)。
// 独立最小验证,非 OGRE 源码。
#include "verify.h"
#include <cstdint>
#include <map>
#include <string>
namespace ogre_repro {
// D3D_FEATURE_LEVEL(OgreD3D11RenderSystem.cpp 272-284)
enum FeatureLevel {
D3D_FEATURE_LEVEL_9_1 = 1,
D3D_FEATURE_LEVEL_9_2,
D3D_FEATURE_LEVEL_9_3,
D3D_FEATURE_LEVEL_10_0,
D3D_FEATURE_LEVEL_10_1,
D3D_FEATURE_LEVEL_11_0,
D3D_FEATURE_LEVEL_11_1
};
FeatureLevel parseFeatureLevel(const std::string& s) {
if (s == "11.1") return D3D_FEATURE_LEVEL_11_1;
if (s == "11.0") return D3D_FEATURE_LEVEL_11_0;
if (s == "10.1") return D3D_FEATURE_LEVEL_10_1;
if (s == "10.0") return D3D_FEATURE_LEVEL_10_0;
if (s == "9.3") return D3D_FEATURE_LEVEL_9_3;
if (s == "9.2") return D3D_FEATURE_LEVEL_9_2;
return D3D_FEATURE_LEVEL_9_1;
}
// 对应 D3D11RenderSystem(OgreD3D11RenderSystem.h 58-118)
class D3D11RenderSystem {
public:
explicit D3D11RenderSystem(FeatureLevel fl) : mFeatureLevel(fl) {}
// 能力分支(OgreD3D11RenderSystem.cpp 803-822)
bool has32BitIndex() const { return mFeatureLevel >= D3D_FEATURE_LEVEL_9_2; }
bool hasReadBackAsTexture() const { return mFeatureLevel >= D3D_FEATURE_LEVEL_10_0; }
bool hasTwoSidedStencil() const { return mFeatureLevel >= D3D_FEATURE_LEVEL_10_0; }
// MRT 数量(OgreD3D11RenderSystem.cpp 893-901)
int getNumMultiRenderTargets() const {
return (mFeatureLevel > D3D_FEATURE_LEVEL_9_3) ? 8 // D3D11_SIMULTANEOUS_RENDER_TARGET_COUNT
: (mFeatureLevel == D3D_FEATURE_LEVEL_9_3) ? 4 // FL 9_3
: 1; // FL 9_1
}
// 描述符 changed 标记(OgreD3D11RenderSystem.h 109-117)
void markBlendDescChanged() { mBlendDescChanged = true; }
void applyBlendDesc() { mBlendDescChanged = false; }
bool isBlendDescChanged() const { return mBlendDescChanged; }
private:
FeatureLevel mFeatureLevel;
bool mBlendDescChanged = false; // 对应 mBlendDescChanged
};
// PolygonMode(OgreMetalRenderSystem.mm 769-774)
enum PolygonMode { PM_SOLID = 0, PM_WIREFRAME = 1 };
// 对应 MetalRenderSystem(OgreMetalRenderSystem.h 51-168、.mm 794-818)
class MetalRenderSystem {
public:
// getPipelineState:按 desc hash 查 mPSOCache,未命中创建后缓存
uint32_t getPipelineState(uint32_t descHash) {
auto it = mPSOCache.find(descHash);
if (it != mPSOCache.end()) return it->second; // 命中直接返回
uint32_t pso = createPipelineState(descHash);
if (pso == 0) return 0; // 创建失败(对应 OGRE_EXCEPT)
mPSOCache.emplace(descHash, pso);
return pso;
}
size_t cacheSize() const { return mPSOCache.size(); }
bool createFailed() const { return mCreateFailed; }
// 渲染编码器状态转发(_setPolygonMode / _setCullingMode)
void setPolygonMode(PolygonMode level) {
mTriangleFillMode = (level == PM_SOLID) ? 1 : 0; // Fill / Lines
}
int triangleFillMode() const { return mTriangleFillMode; }
void setCullingMode(int mode) { mCullMode = mode; }
int cullMode() const { return mCullMode; }
private:
uint32_t createPipelineState(uint32_t descHash) {
if (descHash == kBadHash) { mCreateFailed = true; return 0; }
return mNextPsoId++;
}
std::map<uint32_t, uint32_t> mPSOCache; // 对应 mPSOCache
uint32_t mNextPsoId = 1;
int mTriangleFillMode = 0; // 对应 MTLTriangleFillMode
int mCullMode = -1; // 对应编码器 cullMode 状态
bool mCreateFailed = false;
static constexpr uint32_t kBadHash = 0xDEADBEEF;
};
} // namespace ogre_repro
int main() {
using namespace ogre_repro;
// Metal:PSO 缓存命中与创建
{
MetalRenderSystem metal;
uint32_t pso1 = metal.getPipelineState(0x1001);
CHECK(pso1 == 1); // 首次创建
CHECK(metal.cacheSize() == 1);
CHECK(metal.getPipelineState(0x1001) == pso1); // 缓存命中
CHECK(metal.cacheSize() == 1); // 未新增
uint32_t pso2 = metal.getPipelineState(0x1002);
CHECK(pso2 == 2); // 不同描述符 -> 新 PSO
CHECK(metal.cacheSize() == 2);
}
// Metal:创建失败路径(对应 OGRE_EXCEPT)
{
MetalRenderSystem metal;
CHECK(metal.getPipelineState(0xDEADBEEF) == 0);
CHECK(metal.createFailed());
CHECK(metal.cacheSize() == 0); // 失败不缓存
}
// Metal:渲染编码器状态转发
{
MetalRenderSystem metal;
metal.setPolygonMode(PM_SOLID);
CHECK(metal.triangleFillMode() == 1); // MTLTriangleFillModeFill
metal.setPolygonMode(PM_WIREFRAME);
CHECK(metal.triangleFillMode() == 0); // MTLTriangleFillModeLines
metal.setCullingMode(2);
CHECK(metal.cullMode() == 2);
}
// D3D11:feature level 解析
CHECK(parseFeatureLevel("9.3") == D3D_FEATURE_LEVEL_9_3);
CHECK(parseFeatureLevel("11.0") == D3D_FEATURE_LEVEL_11_0);
CHECK(parseFeatureLevel("10.1") == D3D_FEATURE_LEVEL_10_1);
// D3D11:能力分支
{
D3D11RenderSystem d3d9_1(D3D_FEATURE_LEVEL_9_1);
CHECK(d3d9_1.getNumMultiRenderTargets() == 1);
CHECK(!d3d9_1.has32BitIndex());
CHECK(!d3d9_1.hasReadBackAsTexture());
CHECK(!d3d9_1.hasTwoSidedStencil());
D3D11RenderSystem d3d9_2(D3D_FEATURE_LEVEL_9_2);
CHECK(d3d9_2.has32BitIndex()); // >= 9_2
CHECK(d3d9_2.getNumMultiRenderTargets() == 1);
D3D11RenderSystem d3d9_3(D3D_FEATURE_LEVEL_9_3);
CHECK(d3d9_3.getNumMultiRenderTargets() == 4);
D3D11RenderSystem d3d10(D3D_FEATURE_LEVEL_10_0);
CHECK(d3d10.hasReadBackAsTexture()); // >= 10_0
CHECK(d3d10.hasTwoSidedStencil());
CHECK(d3d10.getNumMultiRenderTargets() == 8);
D3D11RenderSystem d3d11(D3D_FEATURE_LEVEL_11_0);
CHECK(d3d11.getNumMultiRenderTargets() == 8);
}
// D3D11:描述符 changed 标记延迟应用
{
D3D11RenderSystem d3d(D3D_FEATURE_LEVEL_11_0);
CHECK(!d3d.isBlendDescChanged());
d3d.markBlendDescChanged();
CHECK(d3d.isBlendDescChanged());
d3d.applyBlendDesc();
CHECK(!d3d.isBlendDescChanged());
}
return lab::report("08-3 Metal 与 D3D11 后端");
}
本机实测(Apple clang,clang++ -std=c++20 -Wall):编译通过(零警告),断言 29/29 通过。
小结
- Metal 用
[psd hash]查mPSOCache:命中复用、未命中创建后缓存、失败抛ERR_RENDERINGAPI_ERROR。 - Metal 渲染编码器是即时态对象,
_setPolygonMode等 setter 直接转发到mActiveRenderEncoder。 - D3D11 的
requestedLevels数组从 11_1 排到 9_1,按 min/max 请求裁剪窗口,min > max 报内部错误。 - D3D11 能力按
mFeatureLevel >= X线性比较:9_2 起 32 位索引,10_0 起读回纹理/双面模板,MRT 数量 1/4/8 阶梯上升。 - D3D11 描述符用
mXxxDescChanged脏标记惰性重建ID3D11*State;Metal 用描述符 hash 缓存全部历史——同一问题的两种解法。
阅读导航:上一节:8.2 Vulkan 后端 · 下一节:9.1 性能剖析器 。
继续阅读
探索更多技术文章
浏览归档,发现更多关于系统设计、工具链和工程实践的内容。