Most but not all samples fail in SwapChain Present() unless I change GetHardwareAdapter() to pass true for third arg - requestHighPerformanceAdapter

[angusgraham]

HelloWindow works, but HelloTriangle fails.

The adapter Intel(R) HD Graphics 4600 is selected because it comes first when DXGI_GPU_PREFERENCE_UNSPECIFIED and it reports D3D_FEATURE_LEVEL_11_0 but then it crashes in Present():

HRESULT is 0x887a0005 : The GPU device instance has been suspended. Use GetDeviceRemovedReason to determine the appropriate action.

Debug output:
The thread 25936 has exited with code 0 (0x0).
D3D12: Removing Device.
D3D12 ERROR: ID3D12Device::RemoveDevice: Device removal has been triggered for the following reason (DXGI_ERROR_DEVICE_HUNG: The Device took an unreasonable amount of time to execute its commands, or the hardware crashed/hung. As a result, the TDR (Timeout Detection and Recovery) mechanism has been triggered. The current Device Context was executing commands when the hang occurred. The application may want to respawn and fallback to less aggressive use of the display hardware). [ EXECUTION ERROR #232: DEVICE_REMOVAL_PROCESS_AT_FAULT]
Exception thrown at 0x00007FFC5F3FAF29 in D3D12HelloTriangle.exe: Microsoft C++ exception: _com_error at memory location 0x000000190E4FC1E0.
Exception thrown at 0x00007FFC5F3FAF29 in D3D12HelloTriangle.exe: Microsoft C++ exception: _com_error at memory location 0x000000190E4FC558.
Exception thrown at 0x00007FFB99778F9D (d3d12SDKLayers.dll) in D3D12HelloTriangle.exe: 0xC0000005: Access violation reading location 0x0000000000000000.

From dxdiag:

Diagnostics

Windows Error Reporting:
+++ WER0 +++:
Fault bucket 1497160328855956209, type 5
Event Name: D3DDRED2
Response: Not available
Cab Id: 0

Problem signature:
P1: VEN_8086
P2: 20.19.15.5126
P3: 0x887A0006
P4: 7
P5: D3D12HelloTriangle.exe
P6: Unknown
P7:
P8:
P9:
P10:

If I change the call to GetHardwareAdapter() to add the requestHighPerformanceAdapter arg then DXGI_GPU_PREFERENCE_HIGH_PERFORMANCE causes NVIDIA Quadro K2100M to be selected which doesn't crash.

[bvanevery]

So do you feel the bug is failure to reject the insufficient feature level of the Intel HD Graphics 4600 chip? I'm nobody, just asking. Happened to be perusing the repo's issues.

[angusgraham]

So do you feel the bug is failure to reject the insufficient feature level of the Intel HD Graphics 4600 chip? I'm nobody, just asking. Happened to be perusing the repo's issues.

I don't think it's that simple since my dedicated graphics (the K2100M) also reports D3D_FEATURE_LEVEL_11_0 (it still works well enough to draw a triangle using DX12 even with feature level 11).
So I'm not sure what the right answer is - the code to select an adapter via DXGI_GPU_PREFERENCE_HIGH_PERFORMANCE instead of DXGI_GPU_PREFERENCE_UNSPECIFIED is there but there must be a reason it's not active.
It does seem like using DXGI_GPU_PREFERENCE_HIGH_PERFORMANCE is no guarantee, after all presumably if the Intel 4600 was the only adapter I had then it would still be selected and the sample would still crash.

So guess the fix should be to

• fix the render to work with the Intel adapter (if that's possible)
• failing that, iterate the adapters and detect if the render will crash on an adapter (if that's possible) and select a different adapter instead
• if there is no adapter that can render without crashing, and that can be detected, then fail friendly and fast instead of crashing out on render with a device removal fatal

[angusgraham]

OK you made me look at this again and one way I've found to make GetHardwareAdapter() reject the crashy Intel adapter without simply using DXGI_GPU_PREFERENCE_HIGH_PERFORMANCE is to change DXSample::GetHardwareAdapter() to require shader model 6.0:

        // Create the device and check it supports shader model 6.0 or higher.
        ComPtr<ID3D12Device> device;
        if (SUCCEEDED(D3D12CreateDevice(adapter.Get(), D3D_FEATURE_LEVEL_11_0, IID_PPV_ARGS(&device))))
        {
            D3D12_FEATURE_DATA_SHADER_MODEL shaderModel = {};
            shaderModel.HighestShaderModel = D3D_SHADER_MODEL_6_0;
            if (SUCCEEDED(device->CheckFeatureSupport(D3D12_FEATURE_SHADER_MODEL, &shaderModel, sizeof(shaderModel))))
            {
                if (shaderModel.HighestShaderModel >= D3D_SHADER_MODEL_6_0)
                {
                    break;
                }
            }
        }

I'm not sure why that would be necessary tho - looking at D3D12HelloTriangle.cpp it seems to be using 5.0 shaders:
ThrowIfFailed(D3DCompileFromFile(GetAssetFullPath(L"shaders.hlsl").c_str(), nullptr, nullptr, "VSMain", "vs_5_0", compileFlags, 0, &vertexShader, nullptr));
ThrowIfFailed(D3DCompileFromFile(GetAssetFullPath(L"shaders.hlsl").c_str(), nullptr, nullptr, "PSMain", "ps_5_0", compileFlags, 0, &pixelShader, nullptr));

So maybe this only works because it's excluding the Intel adapter on the basis of supported shader model even tho the shader model support is not the real issue.

[weltkante]

Note that drivers aren't ever supposed to crash/reset when the calling code is valid, they are supposed to reject the call. Low level APIs like D3D12 aren't always fully validated though, so if the calling code is not valid, a driver reset may be the correct response.

Regardless of which scenario is happening, rejecting the adapter will not be the correct fix, either the example code or the driver need a fix. You already know you can reject the adapter if you need to work around it in your own code, but I feel like an issue in the examples repo isn't going to lead to anything. If you can't identify or suspect a bug in the example then reporting the driver bug in the intel forums (so it can be investigated) would be the correct action.

There's also a Discord linked on the front page of the repo, probably a better way to get quick feedback if you are unsure about how to handle this.

[bvanevery]

I am reading that the Intel HD Graphics 4600 came out in 2013. "Even though it supports DirectX 12, the feature level is only 11_1, which can be problematic with newer DirectX 12 titles." The DirectX Tool Kit also notes some weirdness about determining Shader Model level in old drivers. "Note that a number of Shader Model levels were defined originally as 'experimental' so were not supported by the released DXC compiler or production drivers until later releases of Windows." The code is about Shader Model 6, but it wouldn't surprise me if the comment pertains to older stuff as well.

Trying to get this fixed in the driver would be a blast from the past. Unless Intel has some very long running integrated driver model, and a current profit motive, I would expect the odds of actually fixing the driver to be zero. Certainly, find out what they're capable of if you like, but I wouldn't expect much.

One could try to robustly detect a badly behaved driver. They exist in the real world, and elaborate hocus pocus detection methods for "not very well specified things" have certainly been done in industry. I seem to remember magic Windows version incantations, and magic Intel CPU incantations, sometime in my past.

Requiring Shader Model 6 is certainly not the answer for DX12 sample code. What's the highest Shader Model your driver is reporting as available?

[weltkante]

Trying to get this fixed in the driver would be a blast from the past.

Ye, sorry, I did not check the version and whether the driver it uses is still getting support. I have older hardware with intel chipsets that have bugs which never got fixed before driver support was dropped.

I still don't expect the examples repo to start maintain a blacklist of bad graphics drivers or methods to check for them, but thats just my opinion. If you have to support and work around that hardware I'd personally just maintain a list of known bad chipsets/driver versions instead of trying to figure out algorithmic ways to detect bad hardware/drivers.

[bvanevery]

Nobody's gonna blacklist in DX12 samples. Not the job.

Algorithmically determining a bad driver, that's possibly a legit job for a sample. It might be as simple as verifying Shader Model 5.1 is available or not.

[angusgraham]

Unfortunately checking for shader model 5.1 doesn't do it, since the Intel adapter claims to support it and gets selected and crashes.

[stanard]

Are you using the validation layer to ensure your D3D usage is conformant? If so, and your program is error free, then it's entirely possible that you have found a bug in the Intel driver. If you can isolate what API call hangs the GPU or app, that would be informative.

One trick I've used to select the dGPU instead of an iGPU is to search for the adapter with the most memory. It sounds like your "high performance" flag does the same thing. You can also query the name of the vendor and a description of the GPU. You could create a blacklist of incompatible GPUs.

[angusgraham]

Hi stanard. Too be clear I have no program - I'm just running the samples as they are. The requestHighPerformanceAdapter arg which causes DXGI_GPU_PREFERENCE_HIGH_PERFORMANCE that I mentioned in the op is part of the sample code already - it's just set to false for whatever reason. I don't know what a validation layer is or how I would use it. :-) I'm not actually a rendering programmer I'm a more high-level guy but I thought I'd look at the dx12 samples to see what directx code looks like these days and was kind of surprised to find they don't work on my machine. It's a ten-year-old laptop (Thinkpad W541)

[stanard]

Well, I hope our samples don't have any validation warnings. :-)

I am surprised by this though.