feat: Descriptor / 有限 bindless 后端升级

.qoder/repowiki/zh/content/API参考/资源管理API/资源管理API.md

@@ -272,15 +272,14 @@ class RHIWorkGraphPipeline {
 - 静态采样器
   - RHIStaticSamplerDescriptor：按索引绑定的采样器集合，便于在管线布局中固定。
 - 后端实现要点（DX12）
-  - 转换为原生采样器描述，分配 CPU/GPU 描述符对，复制到着色器可见区域。
+  - 转换为原生采样器描述，按完整 descriptor intern 共享 CPU 槽；BindingTable sampler 段才占用 GPU heap。
 
 ```mermaid
 classDiagram
 class RHISampler {
 }
 class Dx12Sampler {
 +NativeCpuDescriptorHandle
-+NativeGpuDescriptorHandle
 }
 RHISampler <|-- Dx12Sampler
 ```

.qoder/repowiki/zh/content/API参考/资源管理API/采样器API.md

@@ -246,7 +246,6 @@ class RHISampler {
 }
 class Dx12Sampler {
 +NativeCpuDescriptorHandle
-+NativeGpuDescriptorHandle
 +Release()
 }
 class MetalSampler {

.qoder/repowiki/zh/content/资源管理/视图管理.md

@@ -192,7 +192,6 @@ class Dx12TextureView {
 +Device
 +DescriptorClass
 +NativeCpuDescriptorHandle
-+NativeGpuDescriptorHandle
 +Release()
 }
 class VulkanTextureView {
@@ -233,7 +232,7 @@ RHITextureView <|-- MetalTextureView
 ## 依赖关系分析
 - 抽象层依赖：无后端特定类型，仅定义描述符与基类
 - DX12 实现依赖：
-  - 设备与描述符堆管理（AllocateCbvSrvUavDescriptorPair、CopyDescriptorToShaderVisible）
+  - 设备与描述符堆管理（view 只占 CPU staging；BindingTable 自有 CPU 镜像 + GPU 段，SetBindingTable flush）
   - 工具函数进行格式/维度转换与描述符填充
 - Vulkan 实现依赖：
   - VkImageView 创建与销毁

.qoder/repowiki/zh/content/项目概览/核心特性说明/资源管理/视图管理.md

@@ -140,22 +140,20 @@ RHI-->>App : 可用于命令编码阶段的绑定
 - 作用：将纹理资源按 mip/数组/维度进行切片，并以特定访问语义（如只读采样、写入）暴露给着色器
 - 关键配置：BaseMipLevel、MipCount、BaseArraySlice、ArrayCount、ViewType
 - 后端差异：
-  - D3D12：创建 SRV/UAV 描述符，并复制到可见堆；支持采样反馈特殊路径
+  - D3D12：创建 SRV/UAV 到 CPU staging；BindingTable 在 SetBindingTable 时再 flush 到自有 GPU 段；支持采样反馈特殊路径
   - Vulkan：创建 VkImageView，设置 subresourceRange；通过 GetDescriptorImageInfo 获取绑定信息
-  - Metal：使用池化纹理视图索引，减少频繁创建开销；支持全视图快速路径
+  - Metal：创建期锁定 ViewPool 容量，dispose 清 native 槽后再还软件 index；支持全视图快速路径
 
 ```mermaid
 classDiagram
 class RHITextureView {
 +ViewType
 +Dimension
 +NativeCpuDescriptorHandle()
-+NativeGpuDescriptorHandle()
 }
 class Dx12TextureView {
 +CreateSRV()
 +CreateUAV()
-+CopyToShaderVisible()
 }
 class VulkanTextureView {
 +GetDescriptorImageInfo(layout)

AGENTS.md

@@ -8,7 +8,7 @@
 
 ## Product boundary
 
-Keep RHI and backend boundaries explicit. Preserve the maintained Vortice source and patches; do not replace them with an upstream package that lacks the custom behavior.
+Keep RHI and backend boundaries explicit. Do not expose backend Heap/Pool/View containers on the public RHI surface. Preserve the maintained Vortice source and patches; do not replace them with an upstream package that lacks the custom behavior.
 
 ## Implementation and verification
 

DESIGN.md

@@ -8,6 +8,8 @@ Output and intermediate paths are isolated by project, platform, RID, configurat
 
 Public native-backed operations enforce platform and ownership boundaries. No capability downgrade or compatibility implementation is permitted to conceal unsupported execution. Current extraction acceptance is tracked by InfinityBrowser TASK-20260907-INFINITYSTACK-EXTRACTION; this document is not a claim that migration gates have passed.
 
+Binding and descriptor strategy stays behind the backend. The public surface is `RHIBindingTable` with `Count` and `SetBindElement(..., arrayIndex)` for finite bindless; RHI does not grow Heap/Pool/View containers. DX12 gives every table group its own CPU mirror plus GPU segment and publishes on `SetBindingTable`; views and interned sampler slots occupy CPU staging only. Vulkan keeps a set per table and pages pools on the device. Metal fills argument tables or reference buffers and locks a private ViewPool at device create. Backends do not resize shader-visible GPU heaps or native pools at runtime.
+
 DX12 requires successful Agility device-factory initialization using the application D3D12 directory and UTF-8 paths. Source references and packages both deploy these assets; missing assets fail explicitly. The maintained binding and evidence are described in docs/SharpGPU/VorticeAgilityPathPatch.md.
 
 Backend implementation tests belong to the independent conformance harness. Infinity.Rendering.Tests has no product friend access. Test migration provenance is recorded in docs/provenance/backend-test-migration.json. Native configuration tests exercise actual Configure/Resolve behavior in isolated load contexts; product code does not contain a separate engine-path enumerator solely for tests.

docs/SharpGPU/FeatureAudit-2026.md

@@ -100,7 +100,7 @@ ADR-0064 已完成第一个 clean break：公共 raster attachment 只保留 Inp
 | GPU file I/O | `StorageQueue` | DirectStorage native + qualified | unavailable/fail closed | MTLIO native/runtime-probed，当前无 Metal-specific qualified evidence | DX12 实装；Metal 部分；Vulkan 无同构是正确结果，命名仍应细分 |
 | Sparse 2D/3D | 有 texture API | reserved textures/tile mapping/residency；native Tier4 当前折叠为 RHI Tier3，Tier4-specific 语义/资格未独立公开 | sparse image 2D/3D feature与mapping | sparse/placement texture mapping | 部分：机制存在，2D/3D/MSAA/residency/Tier4 需要拆分 capability 与 GPU evidence |
 | Sparse buffer | capability skeleton | 未形成独立可证明 route | unavailable | unavailable | 不得由 sparse texture 推导 |
-| Bindless/indexing | BindingTable + typed capabilities | descriptor table/indexing | descriptor indexing 部分 probe；无 descriptor buffer/heap lowering | MTL4 argument table strategy | 部分；strategy 与 limits 需要继续分层 |
+| Bindless/indexing | BindingTable + `Count` + `SetBindElement(..., arrayIndex)`；不暴露 Heap/Pool/View 容器 | table 自有 CPU 镜像 + GPU 段；SetBindElement 只写 CPU；SetBindingTable flush `CopyDescriptors` | set + Device pool 页（SetsPerPage 分档 / 超集共页）；`vkUpdateDescriptorSets` + BindSet | argument table / reference buffer + 创建期锁定的私有 ViewPool | 公共面已落地为有限 bindless BindingTable；后端 strategy 不进 RHI，也不热换 GPU heap / native pool |
 | Barriers/sync | state/barrier/fence/semaphore | Enhanced + classic barrier path | synchronization2 + core barrier path | Metal4 command/memory barrier path | 基础实装；timeline/shared event/external ownership 仍需公共查询 |
 | Pipeline cache | 有 factory | native | native | fail closed（当前 blob contract 不同构） | 部分；应引入 typed archive/binary strategy 而非伪 portable blob |
 | Video | 无 | 未暴露 D3D12 Video | 未暴露 KHR Video | 未接 VideoToolbox | 缺失但应归 media service，而非 graphics core |

docs/SharpGPU/FeatureMatrix.md

@@ -54,7 +54,7 @@ A Successful native capability probe is never a substitute for a Passed Qualifie
 | OpacityMicromap | `RayTracing.OpacityMicromap` / `OpacityMicromapSerialization`, `RHIOpacityMicromap`, BLAS triangle attachment, instance `ForceOmm2State` / `DisableOmms` | Typed capability + AccelStruct object model | G4-T13 PASSED: DX12 Tier1_2 + factory/build; Vulkan only when `VK_EXT_opacity_micromap` is enabled, AS + sync2 (1.3 or `VK_KHR_synchronization2`) are enabled, and function pointers load; WindowsQualified tiny 2-state OMM + one-triangle BLAS; independent Vulkan OMM 1/1 | Metal and missing native path Unavailable; Create/Build `Require` throw; unknown format fail-closed; OMM input / triangle-array / index buffers must include `ERHIBufferUsage.AccelStruct` (maps to `MICROMAP_BUILD_INPUT_READ_ONLY`); `DisableOmms` requires BLAS `AllowDisableOmms` and sets native `ALLOW_DISABLE_OMMS`; serialization is honest Unavailable (no fake blob) |
 | Motion | `RayTracing.Motion`, motion triangles (`MotionVertexBuffer` / `MotionVertexOffset` / `MotionVertexStride`) and motion instances (`MotionType` / matrix / SRT), `ERHIAccelStructFlag.Motion` | Typed capability + AccelStruct object model | G4-T15 PASSED at SHA `11d5a8c1ba2bb9d131a96458d22e12ff167e3041`: Vulkan Available only when `VK_NV_ray_tracing_motion_blur` is listed, `rayTracingMotionBlur` is enabled, AS deps + function pointers load, and the native motion AS path is used; DX12 Unavailable (no standard D3D12 motion AS); Metal compile-level Available only when `supportsPrimitiveMotionBlur` plus MTL motion triangle / motion instance descriptors bind | Create/Build `Require` throw when Unavailable; `ERHIAccelStructFlag.Motion` must match motion data (XOR fail-closed); failed Update leaves descriptor + native motion mode unchanged; `MotionVertexStride` 0 means `VertexStride`, a different non-zero stride is fail-closed; unknown `MotionType` fail-closed; no silent static path; pipeline ABI **9** is an incompatible revision for the Motion public descriptor (AS is not a pipeline-cache key); SER / HitObject is G4-T14 `BLOCKED_SDK_BINDING` and is not this row |
 | MeshShading | `RHIDeviceCapabilities.Mesh.MeshShader` / `Mesh.TaskShader`, mesh raster pipeline and dispatch | Typed capability | native mesh pipeline plus pixel/readback evidence | mesh factories/encoding throw `NotSupportedException` when unavailable; no silent no-op dispatch |
-| DescriptorIndexing | `RHIDeviceCapabilities.DescriptorIndexing`, `RHIBindingTableLayoutElement.Count`, `RHIBindingTable.SetBindElement` | Typed capability + API contract | required/optional arrays, namespace mapping, cross-device/disposed validation, pool rollback, dispatch/readback | unsupported native descriptor semantics fail at layout/table creation; no typed dummy descriptors |
+| DescriptorIndexing | `RHIDeviceCapabilities.DescriptorIndexing`, `RHIBindingTableLayoutElement.Count`, `RHIBindingTable.SetBindElement(..., arrayIndex)` | Typed capability + API contract | required/optional arrays, `arrayIndex` updates, namespace mapping, cross-device/disposed validation, table-create pool rollback, dispatch/readback | unsupported native descriptor semantics fail at layout/table creation; no typed dummy descriptors |
 | StorageQueue | `RHIDeviceCapabilities.Storage.NativeGpuFileIo` / `GpuDecompression` / `RequestCancellation` / `IoPriority`, `RHIStorageQueue` | Independent typed capabilities + API contract | DX12 DirectStorage file → GPU-local buffer/texture → fence → readback under `SharpGpuDirectStorageQualified`; GDeflate / `CancelRequestsWithTag` / queue `Priority` only when the native queue expresses them | non-native backends and missing DirectStorage support throw `NotSupportedException`; Vulkan and unproven dimensions stay Unavailable; no FileStream/map/staging queue fallback |
 | PipelineCache | `RHIDeviceCapabilities.PipelineCache.NativeCache`, `RHIPipelineCache` | Typed capability + API contract | cold/warm/restart native hit, typed corrupt/incompatible import, full-key non-collision | unavailable native cache strategy throws `NotSupportedException`; caller owns opaque blobs |
 | WorkGraph | `RHIDeviceCapabilities.WorkGraph.Execution`, `RHIWorkGraphPipeline`, `RHIWorkGraphEncoder` | Typed capability | native create/dispatch/readback on each reported strategy | factory/encoding throws `NotSupportedException` when unavailable |

docs/VERIFICATION.md

@@ -39,6 +39,9 @@ dotnet test tests/SharpGPU.Conformance.Tests/SharpGPU.Conformance.Tests.csproj `
   --results-directory "$out/test-results/source-release"
 ```
 
+Descriptor-strategy unit tests that do not require a matching GPU host:
+`Dx12DescriptorAllocatorTests`, `VulkanBindingTablePlanTests.PoolPolicy_ShouldTierSetsPerPageAndAllowSupersetShareWithinWasteLimit`, and `MetalBindingTablePlanTests.TextureViewPool_ShouldLockFirstSuccessfulCapacityOnTheLadder`. Metal native slot-clear after dispose remains `TODO(UNVERIFIED)` without a Metal host.
+
 On the current Windows x64 host the Release source build completed with zero
 errors and the full conformance run passed **289/289**. It exercised the
 DirectX 12 and Vulkan compute/draw paths, native memory and synchronization,

src/SharpGPU/Dx12/Dx12AccelStruct.cs

@@ -54,14 +54,13 @@ internal unsafe class Dx12TopLevelAccelStruct : RHITopLevelAccelStruct, IDx12Des
     {
         public Dx12Device Device => m_Dx12Device;
         public Dx12DescriptorClass DescriptorClass => Dx12DescriptorClass.AccelerationStructure;
-        public Vortice.Direct3D12.CpuDescriptorHandle NativeCpuDescriptorHandle => m_Descriptors.Staging.CpuHandle;
-        public Vortice.Direct3D12.GpuDescriptorHandle NativeGpuDescriptorHandle => m_Descriptors.ShaderVisible.GpuHandle;
+        public Vortice.Direct3D12.CpuDescriptorHandle NativeCpuDescriptorHandle => m_Staging.Descriptor.CpuHandle;
         public Vortice.Direct3D12.ID3D12Resource ResultBuffer => m_NativeResultBuffer;
         public Vortice.Direct3D12.BuildRaytracingAccelerationStructureDescription NativeAccelStructDescriptor => m_NativeAccelStructDescriptor;
 
         private Dx12Device m_Dx12Device;
-        private int m_DescriptionHeapIndex;
-        private Dx12DescriptorPair m_Descriptors;
+        private bool m_HasDescriptors;
+        private Dx12CpuDescriptorAllocation m_Staging;
         private Vortice.Direct3D12.ID3D12Resource m_NativeResultBuffer;
         private Vortice.Direct3D12.ID3D12Resource m_NativeScratchBuffer;
         private Vortice.Direct3D12.ID3D12Resource m_NativeInstancesBuffer;
@@ -71,7 +70,7 @@ public Dx12TopLevelAccelStruct(Dx12Device device, in RHITopLevelAccelStructDescr
         {
             m_Dx12Device = device;
             m_Descriptor = descriptor;
-            m_DescriptionHeapIndex = -1;
+            m_HasDescriptors = false;
             RHIOpacityMicromapContract.ValidateTlasDescriptor(device, in descriptor);
             RHIAccelStructMotionContract.ValidateTlasDescriptor(device, in descriptor);
             Span<RHIAccelStructInstance> asInstances = descriptor.Instances.Span;
@@ -118,16 +117,15 @@ public Dx12TopLevelAccelStruct(Dx12Device device, in RHITopLevelAccelStructDescr
             m_NativeScratchBuffer = Dx12RaytracingHelper.CreateBuffer(m_Dx12Device.NativeDevice, (uint)nativeAccelStructPrebuildInfo.ScratchDataSizeInBytes, Vortice.Direct3D12.ResourceFlags.AllowUnorderedAccess, Vortice.Direct3D12.ResourceStates.Common | Vortice.Direct3D12.ResourceStates.UnorderedAccess, Dx12RaytracingHelper.kDefaultHeapProps);
             m_NativeResultBuffer = Dx12RaytracingHelper.CreateBuffer(m_Dx12Device.NativeDevice, (uint)nativeAccelStructPrebuildInfo.ResultDataMaxSizeInBytes, Vortice.Direct3D12.ResourceFlags.AllowUnorderedAccess, Vortice.Direct3D12.ResourceStates.Common | Vortice.Direct3D12.ResourceStates.RaytracingAccelerationStructure, Dx12RaytracingHelper.kDefaultHeapProps);
 
-            m_Descriptors = m_Dx12Device.AllocateCbvSrvUavDescriptorPair();
-            m_DescriptionHeapIndex = m_Descriptors.ShaderVisible.Index;
+            m_Staging = m_Dx12Device.AllocateStagingCbvSrvUavDescriptor(1);
+            m_HasDescriptors = true;
 
             Vortice.Direct3D12.ShaderResourceViewDescription accelStructSrvDesc = new Vortice.Direct3D12.ShaderResourceViewDescription();
             accelStructSrvDesc.Format = Vortice.DXGI.Format.Unknown;
             accelStructSrvDesc.ViewDimension = Vortice.Direct3D12.ShaderResourceViewDimension.RaytracingAccelerationStructure;
             accelStructSrvDesc.Shader4ComponentMapping = 5768;
             accelStructSrvDesc.RaytracingAccelerationStructure.Location = m_NativeResultBuffer.GPUVirtualAddress;
-            m_Dx12Device.NativeDevice.CreateShaderResourceView(null, accelStructSrvDesc, m_Descriptors.Staging.CpuHandle);
-            m_Dx12Device.CopyDescriptorToShaderVisible(m_Descriptors);
+            m_Dx12Device.NativeDevice.CreateShaderResourceView(null, accelStructSrvDesc, m_Staging.Descriptor.CpuHandle);
 
             m_NativeAccelStructDescriptor.Inputs = nativeAccelStructDescriptor;
             m_NativeAccelStructDescriptor.DestinationAccelerationStructureData = m_NativeResultBuffer.GPUVirtualAddress;
@@ -191,10 +189,10 @@ public override void UpdateAccelerationStructure(in RHITopLevelAccelStructDescri
 
         protected override void Release()
         {
-            if (m_DescriptionHeapIndex >= 0)
+            if (m_HasDescriptors)
             {
-                m_Dx12Device.FreeDescriptorPair(m_Descriptors);
-                m_DescriptionHeapIndex = -1;
+                m_Dx12Device.FreeStagingCbvSrvUavDescriptor(m_Staging);
+                m_HasDescriptors = false;
             }
             m_NativeResultBuffer.Release();
             m_NativeScratchBuffer.Release();