Hi

I'm writing homebrew DX12 engine (mainly for fun & demoscene purposes).

For the rendering part, everything works as expected on both NVidia and AMD hardware, but for DXR part I'm struggling with making it to work on AMD (I have Radeon RX 6800 XT for testing stuff right now, but I'v also tried on different machine with Radeon RX 6700, same effect).

When I enable full debug layer validation:

((ID3D12Debug3 *)D3D12DebugCtrl)->EnableDebugLayer();
((ID3D12Debug3 *)D3D12DebugCtrl)->SetEnableGPUBasedValidation(TRUE);

I'm unable to create ANY root signature with D3D12_ROOT_SIGNATURE_FLAG_LOCAL_ROOT_SIGNATURE flag

It crashes the driver instantly uppon calling ID3D12Device5->CreateRootSignature.

With

((ID3D12Debug3 *)D3D12DebugCtrl)->SetEnableGPUBasedValidation(FALSE);

It creates those signatures just fine, but crashes driver when I want to create ID3D12Device5->CreateStateObject

There is nothing spew to debug output, the callstack is as follows (not very informative):

 amdxc64.dll!00007fffa70dc153() Unknown
 amdxc64.dll!00007fffa70dbf6a() Unknown
 amdxc64.dll!00007fffa70cb08c() Unknown
 amdxc64.dll!00007fffa70cbb1c() Unknown
 amdxc64.dll!00007fffa706d27d() Unknown
 amdxc64.dll!00007fffa70a84d8() Unknown
 D3D12Core.dll!00007fff899907b3() Unknown
 D3D12Core.dll!00007fff8991e33a() Unknown
 D3D12Core.dll!00007fff89901b23() Unknown
 D3D12Core.dll!00007fff898e4f1c() Unknown
 d3d12SDKLayers.dll!00007fffaa38bc55() Unknown
 d3d12SDKLayers.dll!00007fffaa360bb6() Unknown
 d3d12SDKLayers.dll!00007fffaa344f0c() Unknown
 D3D12.dll!00007fff9a5a5737() Unknown
 D3D12Core.dll!00007fff898f59c8() Unknown
 d3d12SDKLayers.dll!00007fffaa358628() Unknown
> SETISequencer.exe!FRHIStateObject::FRHIStateObject(const FComPointer<FRHIDevice> & Dev, const D3D12_STATE_OBJECT_DESC & Desc, unsigned __int64 _Hash) Line 19 C++
 SETISequencer.exe!FRHIDevice::GetCachedStateObject(const D3D12_STATE_OBJECT_DESC & Desc) Line 1854 C++

When I use root signatures defined in DXIL Library, then all works fine, but since my pathtracer materials system is node based, I prefer to have full control over the fragments and generate local root signatures in c++ code (as whole logic & binding is done there).

All this works just fine on NVidia hardware (tested on windows 10 and 11, and few different RTX models and on ranging from 2060 to 4090) with debug runtime & validation on, and off.

Even this very simple StateObject (dump below) crashes driver:

--------------------------------------------------------------------
| D3D12 State Object 0x000000F9423AD688: Raytracing Pipeline
| [0]: DXIL Library 0x000002D59D414980, 6546 bytes
|  [0]: RayGeneration --> RG_3E307212_B6CF74EB_??RayGeneration@@YAXXZ
|  [1]: RayHit --> RH_3E307212_B6CF74EB_??RayHit@@YAXURayPayload@@UBuiltInTriangleIntersectionAttributes@@@Z
|  [2]: RayHitOcclusion --> RH_3E307212_B6CF74EB_??RayHitOcclusion@@YAXURayPayloadOcclusion@@UBuiltInTriangleIntersectionAttributes@@@Z
|  [3]: RayMiss --> RM_3E307212_B6CF74EB_??RayMiss@@YAXURayPayload@@@Z
|  [4]: RayMissOcclusion --> RM_3E307212_B6CF74EB_??RayMissOcclusion@@YAXURayPayloadOcclusion@@@Z
|--------------------------------------------------------------------
| [1]: Hit Group (Hit_3D4802DE_AE444268)
|  [0]: Any Hit Import: [none]
|  [1]: Closest Hit Import: RH_3E307212_B6CF74EB_??RayHit@@YAXURayPayload@@UBuiltInTriangleIntersectionAttributes@@@Z
|  [2]: Intersection Import: [none]
|--------------------------------------------------------------------
| [2]: Hit Group (Occ_3D4802DE_AE444268)
|  [0]: Any Hit Import: [none]
|  [1]: Closest Hit Import: RH_3E307212_B6CF74EB_??RayHitOcclusion@@YAXURayPayloadOcclusion@@UBuiltInTriangleIntersectionAttributes@@@Z
|  [2]: Intersection Import: [none]
|--------------------------------------------------------------------
| [3]: Local Root Signature 0x000002D59D37FB10
|--------------------------------------------------------------------
| [4]: Subobject to Exports Association (Subobject [3])
|  [0]: RG_3E307212_B6CF74EB_??RayGeneration@@YAXXZ
|--------------------------------------------------------------------
| [5]: Raytracing Shader Config
|  [0]: Max Payload Size: 16 bytes
|  [1]: Max Attribute Size: 8 bytes
|--------------------------------------------------------------------
| [6]: Subobject to Exports Association (Subobject [5])
|  [0]: Hit_3D4802DE_AE444268
|  [1]: Occ_3D4802DE_AE444268
|  [2]: RG_3E307212_B6CF74EB_??RayGeneration@@YAXXZ
|  [3]: RM_3E307212_B6CF74EB_??RayMiss@@YAXURayPayload@@@Z
|  [4]: RM_3E307212_B6CF74EB_??RayMissOcclusion@@YAXURayPayloadOcclusion@@@Z
|--------------------------------------------------------------------
| [7]: Raytracing Pipeline Config
|  [0]: Max Recursion Depth: 1
|--------------------------------------------------------------------
| [8]: Global Root Signature 0x000002D5F56009D0
|--------------------------------------------------------------------

Anyone have any idea what it might be that the driver just bails out ? (Unreal engine 5.0 uses similar way to bind local root signatures, and it works, so it's not that the driver/card is broken, but there seems to be something that needs to be set and I'm missing it obviously, but the ammout of code UE does in the middle obfuscates all the steps).

I'v also tried the Microsoft D3D12RaytracingLibrarySubobjects example from DXSDK, and just added very simple local root signature to it, and the effect is identical, it just crashes the driver uppon creation of RootSignature (when debug layer is enabled) or during StateObject creation (when debug layer is disabled).

To replicate this behaviour in Microsoft example, one need to add this code:

   std::vector<CD3DX12_ROOT_PARAMETER1> Parameters;

CD3DX12_ROOT_PARAMETER1 ConstantBuffer;  
ConstantBuffer.InitAsConstantBufferView(0, 0, D3D12_ROOT_DESCRIPTOR_FLAG_DATA_STATIC_WHILE_SET_AT_EXECUTE, D3D12_SHADER_VISIBILITY_ALL);
Parameters.push_back(ConstantBuffer);

   CD3DX12_VERSIONED_ROOT_SIGNATURE_DESC RootSignatureDesc;
   RootSignatureDesc.Init_1_1(Parameters.size(), &Parameters[0], 0, NULL, D3D12_ROOT_SIGNATURE_FLAG_LOCAL_ROOT_SIGNATURE);
   ID3DBlob *SignatureData = NULL;
   ID3DBlob *ErrorData     = NULL;

if (FAILED(D3DX12SerializeVersionedRootSignature(&RootSignatureDesc, D3D_ROOT_SIGNATURE_VERSION_1_1, &SignatureData, &ErrorData)))
{
   }
   else
   {
       ID3D12RootSignature *Result = NULL;
       m_dxrDevice->CreateRootSignature(0, SignatureData->GetBufferPointer(), SignatureData->GetBufferSize(), __uuidof(ID3D12RootSignature), (void **)&Result); // << -- here the driver will crash

   }

at the begining of D3D12RaytracingLibrarySubobjects::CreateRaytracingPipelineStateObject function, and enable SetEnableGPUBasedValidation in DeviceResources::InitializeDXGIAdapter function.

Hmm that sure does sound like a AMD driver bug. Are you currently using the latest version of the validation layer from the latest Agility SDK?

@MJP Well, I'v tried various AgilitySDK versions (606 (1.606.3), 600 (1.600.10) and 4 (1.4.10)). I'v also tried without AgilitySDK, plain old DX12. In all cases it works on NVidia, and crashes when trying to create LRS / RS on AMD. I'm targeting SM 6.3, but that should not matter couse creating LRS does not require any compiler related data.

OMG – I'v found what was cousing the crash — the NAMES – when I'v imported library fragment, I'v renamed entries according to shema Export→Hash_MangledName.

(RayGeneration --> RG_3E307212_B6CF74EB_??RayGeneration@@YAXXZ)

It turns out that AMD drivers threads @ in those renames in special way, and thus when it encounters @ character in the rename, IT WILL CRASH (!!!)

replacing the @ with _ solved the issue, so renaming:

(RayGeneration --> RG_3E307212_B6CF74EB_??RayGeneration@@YAXXZ) -→ crash

(RayGeneration --> RG_3E307212_B6CF74EB??RayGeneration__YAXXZ) -→ all is fine and engine works …

no comments … :[ (alnd almost 10h of time wasted on debugging, and searching what is wrong … thank you AMD)

Oh wow, that's wild! I'm glad you at least figured that out though.