DX12 Performance

[alpine schooner]

I think I need a different approach for how I've set up rendering. I'm getting serious bottlenecks from having a shit load of GameObjects
About 500-600 simple cubes sharing the same mat/shader/etc will bring the fps down to <30

void Batch::Render(ID3D12GraphicsCommandList2* commandList, D3D12_VIEWPORT viewPort, D3D12_RECT scissorRect, D3D12_CPU_DESCRIPTOR_HANDLE rtv, D3D12_CPU_DESCRIPTOR_HANDLE dsv, XMMATRIX viewProj, Frustum& frustum)
{
    for (int i = 0; i < m_gameObjectList.size(); i++) 
    {
        XMFLOAT3 pos;
        float radius;
        m_gameObjectList[i]->GetBoundingSphere(pos, radius);

        if (!FRUSTUM_CULLING_ENABLED || frustum.CheckSphere(pos, radius))
            m_gameObjectList[i]->Render(commandList, m_rootSignature.Get(), viewPort, scissorRect, rtv, dsv, viewProj);
    }

    for (int i = 0; i < m_gameObjectListTransparent.size(); i++)
    {
        XMFLOAT3 pos;
        float radius;
        m_gameObjectListTransparent[i]->GetBoundingSphere(pos, radius);

        if (!FRUSTUM_CULLING_ENABLED || frustum.CheckSphere(pos, radius))
            m_gameObjectListTransparent[i]->Render(commandList, m_rootSignature.Get(), viewPort, scissorRect, rtv, dsv, viewProj);
    }
}```

void GameObject::Render(ID3D12GraphicsCommandList2* commandListDirect, ID3D12RootSignature* rootSig, D3D12_VIEWPORT viewPort, D3D12_RECT scissorRect, D3D12_CPU_DESCRIPTOR_HANDLE rtv, D3D12_CPU_DESCRIPTOR_HANDLE dsv, XMMATRIX viewProj)
{
    if (!m_model || !m_shader)
        return;

    commandListDirect->SetPipelineState(m_shader->GetPSO().Get());
    commandListDirect->SetGraphicsRootSignature(rootSig);
    commandListDirect->IASetPrimitiveTopology(D3D_PRIMITIVE_TOPOLOGY_TRIANGLELIST);

    commandListDirect->RSSetViewports(1, &viewPort);
    commandListDirect->RSSetScissorRects(1, &scissorRect);

    UINT numRenderTargets = 1;
    commandListDirect->OMSetRenderTargets(numRenderTargets, &rtv, FALSE, &dsv);

    if (m_material)
    {
        auto texHeap = m_material->GetTextureHeap();
        ID3D12DescriptorHeap* ppHeaps[] = { texHeap };
        commandListDirect->SetDescriptorHeaps(_countof(ppHeaps), ppHeaps);

        commandListDirect->SetGraphicsRootDescriptorTable(1, texHeap->GetGPUDescriptorHandleForHeapStart());
    }   

    MatricesCB matricesCB;
    matricesCB.M = m_worldMatrix;
    matricesCB.InverseTransposeM = XMMatrixTranspose(XMMatrixInverse(nullptr, m_worldMatrix));
    matricesCB.VP = viewProj;
    commandListDirect->SetGraphicsRoot32BitConstants(0, sizeof(MatricesCB) / 4, &matricesCB, 0);

    m_model->Render(commandListDirect);
}```

[thick trellis]

Use a profiler. Visual Studio has a pretty good one built-in, and PIX can profile your application as well

[alpine schooner]

I heard something once that using a vector of pointers can cause performance issues due to needing to refill the cache or something. Should I try refactor to avoid this?

[thick trellis]

Nothing stands out to me in the code you've shown

A vector of pointers is usually not the fastest possible thing. However, I strongly doubt it'd hurt your framerate this much

[alpine schooner]

Anyone know what this is about?

Both

if (m_material)```
and especially
```cpp
Model* sphereModel = nullptr```

I feel like shouldn't take more than a few cycles each and yet they're eating loads

[thick sedge]

At 30 fps that’s 33ms per frame, which would be about 65 microseconds per object. That’s definitely very high.have you confirmed that the time spent is even in the function you listed?

For microoptimization you would have to look at the disassembly for optimized code, since there not a simple one-to-one translation from source to instructions. Likely you just have a cache miss on a load that’s causing a long stall.

[alpine schooner]

Anyone know how I can debug something like this? How do I know what this d3d12sdklayers represents?
Apparently when there's a massive performance hit on the first variable declaration line, what it's really showing is how long it took to enter the function (Aka copying arguments and stuff). But every argument I use is either a pointer or lvalue so I'm not sure why else it might be taking long to enter

I have ideas for how to refactor lots of stuff to avoid cache misses but I'm not sure if that's what's actually causing an issue or not

[thick trellis]

d3d12sdklayers is part of the D3D12 driver

That data is telling you that your program is spending most of its time in the driver - your code isn't doing very much

Do you know if the GPU is taking longer to render a frame than the CPU? How much time if your CPU taking each frame?

Hmm, looks like VS doens't really show you the time spent waiting for the GPU. Its sampling profiler just tells you what percentage of your time is spend in which function - not what your total time is

If your GPU is taking longer than your CPU, focusing on CPU optimizations won't help you

[alpine schooner]

Yeah the GPU profiler in VS isn't working for me for some reason, not really sure why it just errors.
Might try Nsight Graphics. Haven't used it before but will that give me some info on if the GPU is a bottleneck?

[thick trellis]

I've never used the GPU profiler in VS, idk how good it is

Nsight is really good, though

It'll tell you if you're CPU-bound of GPU-bound for sure

You'll want to use the GPU Trace Profiler

[thick sedge]

d3d12sdklayers is the debug layer