- We're still working on the windows support, for the moment, please use WSL2.
- CUDA support is enable only during installation. If you first installed HVM and then installed CUDA and
nvcc
, please try to install HVM again.
- The current iteration of the
hvm.cu
was written with the RTX 4090 in mind, and won't work on older GPUs, since they contain about half of the newer GPUs shared memory, for better understanding please refer to HVM#283. We are working on support for older GPUs and will release it soon.
- If you are on Unix system (or WSL) then most likely bend did not add itself to the PATH variable in your rc file. To remedy this:
- Determine if you are using bash or zsh (check for presence of
~/.bashrc
or~/.zshrc
) - Make change to the relevant file by running
echo -n 'export PATH=$PATH:$HOME/.cargo/bin' >> ~/.bashrc
- use
~/.zshrc
in place of~/.bashrc
in the case you are on zsh system
- use
- You will then need to run
source ~/.bashrc
orsource ~/.zshrc
for the changes to take effect immediately - Try running
bend
once more - it should work now!
- Determine if you are using bash or zsh (check for presence of
-
If the error happens when compiling the CUDA runtime or contains anything regarding
ccbin
, please refer to HVM#291. -
If the error contains anything regarding
libc
missing, please refer to HVM#355
- We plan on adding support to many other GPUs as soon as the CUDA version is sufficiently stable.
GPUs with >=96KB L1 cache per SM should work. This includes the following:
GPU | Tested? |
---|---|
RTX 4090 | Yes |
All Other Desktop 40 Series | No |
All Desktop 30 Series | No |
All Mobile 40 Series | No |
All Mobile 30 Series | No |
- IO is still being developed and is expected to come soon.
- Not as soon as basic IO, but planned or at least something similar to FFI.
- A package manager will be added soon.
Some possibilities:
- Your program is causing an overflow on 24-bit number values.
- Your program is doing operations on numbers of different types. (e.g.
2.0 + 1
is not allowed, you must use2.0 + 1.0
) - Floating point numbers are currently bugged and are interpreted incorrectly in some cases.
- There's a bug with signed integers numbers that flips that sometimes flips the order of the operations.
- HVM currently has a bug in its conversion of f32 to f24 and it's unable to read the number 0.0. We already have a fix that we're working on.