Hyperdrive with RTX Geforce 5090 (compute capability is 12.0)

[syoshiura]

Hi hyperdrive developers,

I encountered an issue when I install hyperdrive(and hyperbeam) with capability of GPU (Geforce 5090, its compute capability is 12.0.). The issue was that hyperbeam and hyperdrive do not accept HYPERDRIVE_CUDA_COMPUTE=120.

To solve the issue, first I downloaded mwa_hyperbeam,

git clone https://github.com/MWATelescope/mwa_hyperbeam.git
cd mwa_hyperbeam

and I edited build.rs as

- const DEFAULT_CUDA_SMS: &[u16] = &[60, 61, 70, 75, 80, 86];
+ const DEFAULT_CUDA_SMS: &[u16] = &[60, 61, 70, 75, 80, 86, 120];
- if compute.len() != 2
+ if compute.len() != 3

.
Similarly for hyperdrive,

git clone https://github.com/MWATelescope/mwa_hyperdrive
cd mwa_hyperdrive

, and I edited build.rs as

- const DEFAULT_CUDA_SMS: &[u16] = &[60, 61, 70, 75, 80, 86];
+ const DEFAULT_CUDA_SMS: &[u16] = &[60, 61, 70, 75, 80, 86, 120];
- if compute.len() != 2
+ if compute.len() != 3

, and also Cargo.toml is edited as

- mwa_hyperbeam = "0.10.2"
+ mwa_hyperbeam = { path = "../mwa_hyperbeam" }

After these edits, I can install hyperdrive with CUDA_COMPUTE=120

export HYPERDRIVE_CUDA_COMPUTE=120
export HYPERBEAM_CUDA_COMPUTE=120
cargo install --path . --locked --features=cuda,gpu-single

It seems that the hyperdrive is working, but could you please let me know if it's dangerous?

[d3v-null]

Thanks for pointing this out Shintaro,

The compute.len() check should be fixed to check values between 2-3 in hyperbeam and hyperdrive.

please let me know if it's dangerous?

I think if you provide only *_CUDA_COMPUTE=120 like this then cuda will compile a nicely optimized binary for your 5090.

I don't have any hardware that can take a three digit cuda compute, but I'll try and make a branch for you to test soon.

Thanks!

[d3v-null]

I think it should be possible to solve this without changing the default cuda compute values, only build.rs.

Would you please be able to test the hyperbeam branch copilot/fix-issue-21-error works for your 5090?

If it works, then I'll apply this change to hyperdrive too.

Thanks!