b0nes164/gpuprefixsums

A nearly complete collection of prefix sum algorithms implemented in CUDA, D3D12 and Unity. Theoretically portable to all wave/warp/subgroup sizes.

gpuprefixsums's Issues

Is this a bug that implementation contains something related to timings?

I noticed this, is it a leftover from timing implementation? doesn't seem necessary.

The shader hangs 4090 on large (6M entries) buffer

Hello! So I finished my port and integrated into the main application. I tested out with dummy values and small arrays - it went okay, I readback data on CPU to validate it is working. However when I incorporated the shader into my pipeline, my 4090 just hangs and then the whole system blinks black because of GPU hang.

Do you have an idea of what could be the cause? What is the maximum array size you've tested it on?

The purpose of b_index

Can you explain the purpose of a b_index?

Why Chained Scan with Decoupled Lookback isn't in main?

Hello!

Thank you for this wonderful project and a great write-up, I've learned a lot!

I'm looking into porting the fastest algorithm from your repo into native HLSL/DirectX engine and I wonder why didn't you include Chained Scan with Decoupled Lookback as the main, fastest algo? is this due to this limitation:

Currently the maximum aggregate sum supported in Chained Scan with Decoupled Lookback is 2^30.

? if so, do you think I should port this one, given that it suits my number of objects?

Regarding timing options

I have, on windows, used Nvidia Nsight Graphics to get detailed compute shader timings. Just build a Unity binary and launch it from Nsight. I assume a nvidia GPU is needed though.

Why b_state buffer isn't initialized in CD_Main_Vectorized_Dispatch?

Looking into implementation, I can't understand why b_state buffer never gets initialized and what size is it expected to be?