[RFE] Instancing problem in GPU-driven rendering

[vertver]

I'm trying to implement GPU-driven rendering with bindless descriptors and instances, and I have a problem with instance (specifically on D3D12). SV_InstanceID on D3D11/D3D12 starts from 0. However, in Vulkan it starts from firstInstance in draw indexed instanced command. That means that I can't use instance id in D3D12 or D3D11 because it always starts from 0. I can hack this and use additional vertex instance buffer with instance index for each draw, but this method adds complexity. So what do you think, how can we fix this problem?

https://github.com/gpuweb/gpuweb/issues/901 - Problem description

[vertver]

Now it's working. Waiting for your feedback

[dzhdanNV]

"D3D12 + bindless" flickers on RTX 4080. Objects randomly disappear. I'm trying to understand what's going on...

[dzhdanNV]

Works now. The root cause is: BEFORE:

groupshared uint DrawCount = 0;

It's NOP, compiler generates a warning that it's ignored. AFTER:

groupshared uint DrawCount;

[numthreads(THREAD_COUNT, 1, 1)]
void main(uint ThreadID : SV_DispatchThreadId)
{
    if (ThreadID == 0)
        DrawCount = 0;

    GroupMemoryBarrierWithGroupSync();

Polishing, thinking, improving...

[dzhdanNV]

Thanks for your work. Attached is the polished version. Sorry, I'm lazy to mess with GitHub :) Please, copy over all files (assuming you haven't changed anything in the past 16 hours), and arrange two MRs:

• one for NRI
• one for NRISamples

I will accept them, merge in my unrelated changes, update GitHub and make a new release.

NRI changes:

• NRICompatibility.hlsli moved to NRI
• baseAttribute(s) => drawParameters
• added comments to NRIDescs.h
• WrapperD3D12: exposed enableDrawParametersEmulation
• added missing code block to CommandBufferD3D12::Draw
• changed hashing function
• exposed DeviceDesc::isDrawParametersEmulationEnabled, which can be false if emulation is requested but SM6.8 is supported
• fixed sneaked printf
• emulation moved to 999 register space to avoid conflicts

SAMPLE changes:

• added missing counter initialization
• Cmake massaging and cleanup
• cleaned up shaders
• struct definitions moved to a HLSL/C++ shared file
• the checkbox moved to the existing window

Modified.zip

[dzhdanNV]

Forgot to mention: Please, read comments in NRICompatibility.hlsli (on top) and in NRIDescs.h (where command signatures are).

My personal goal is to not lose the default behavior. To reach this:

• added DeviceDesc::isDrawParametersEmulationEnabled, which can be false if SM6.8 is supported
• a Pipeline gets "emulation" on if emulation is requested and "confirmed" by the backend
• on the shader side pipelines with potential "emulation" must have NRI_ENABLE_DRAW_PARAMETERS_EMULATION defined prior NRICompatibility.hlsli inclusion. If SM6.8 is supported, it will be ignored, so we are on the safe side

Open question - RWByteAddressBuffer is undesired... Is it worth switching to RWStructuredBuffer?

[vertver]

Open question - RWByteAddressBuffer is undesired... Is it worth switching to RWStructuredBuffer?

Yes, but it requires using a different structure on each platform. I would prefer more RWBuffer<uint> instead of RWByteAddressBuffer because it doesn't need to provide a stride and can also be cleared by commands.

[dzhdanNV]

RWBuffer<uint> is a good candidate too, I agree. Strictly speaking BaseVertex can be negative, hope casting forth and back won't break anything. Stride is needed in any case (if I understand correctly).

[vertver]

PR is ready - https://github.com/NVIDIAGameWorks/NRI/pull/64

[vertver]

Oups, some problems with Vulkan... fixing them

[vertver]

Fixed, but Vulkan still has 2 allocations on app close

[dzhdanNV]

Fixed, but Vulkan still has 2 allocations on app close

Validation layer enabled? If yes, it's a VK problem.

Sorry about this: MY:

    return ((uint64_t)rootSignature && ((1ull<<52) - 1)) | stride;

FIXED:

    return (stride << 52ull) | ((uint64_t)rootSignature & ((1ull << 52) - 1));

(Will be) fixed in the main branch.

[dzhdanNV]

Merged. Thank you!