3,485 karma · joined August 10, 2017
There is no way to invest in the company, and the only way of becoming a "customer" is to engage in trading.
HRT trades their own money so if it didn't beat VOO then they'd just buy VOO. There are no external investors to scam.
It definitely sucks to be led astray and have time wasted by a bug inherited from the original repo though, sorry to hear that :/
You can look at the generated sass on godbolt: https://cuda.godbolt.org/z/19excTxM3
Note that there are 1024 FFMA instructions in the loop but you would expect 16*8*BK = 2048. This would suggest half the operations are skipped, which lines up with the half of writes that are out of bounds being omitted.
After the compute loop when you're calculating the final result and storing it, you can see that the FFMAs referencing out of bounds indices write QNAN instead of any real results.
Is it possible that the NANs are what are messing with your tests? Those are notoriously hard to deal with correctly, but you should assert that the result doesn't have any NANs whatsoever.
E.g. consider the line:
sum[y2*8 + x2] += ...
In the final loop iteration when y2=15 and x2=7, the index is 127.In fact cuBLAS and CUDA are kinda orthogonal in that you're either calling a pre-built cuBLAS kernel or writing your own CUDA kernel but not really combining the two.
I'd say CUDA shines more because of stability, documentation, community support + examples, and ability to use modern C++ features in GPU code.
Tbh I don't really know, I've done extensive ML and graphics work and I'd just pick the right tool for the job (PyTorch and just drop down to Triton/CUDA/whatever when you need custom GPU code).
You can use any two sources of infrared light instead!
wgmma.mma_async.sync.aligned.m64n256k16.f32.bf16.bf16
Do you reverse it back into C++ that does the corresponding FMAs manually instead of using tensor hardware? Or are you able to convert it into a series of __builtin_amdgcn_mfma_CDFmt_MxNxKABFmt instructions that emulate the same behavior?Presumably that stuff doesn't "just work" but they don't want to mention it?