jaycedowell
jaycedowell
The `test_matmul_ab_beamformer_kernel_large` and `test_matmul_ab_beamformer_kernel_small` failures are related to the shared memory size (16K just isn't enough). With CUDA 12.6 supporting down to compute capability 50 I think we are safe...
Everything looks good and I'm going to sidestep the issue of "docker builds as a CI test" for now.
Sounds like it is similar to this: https://github.com/ledatelescope/bifrost/issues/130
For what it's worth I also ran into this on Red Hat 8.8 and `dnf install ctags` worked for me. I was able to build/install Bifrost and run a pipeline....
I'll try to find some time this week to look into this.
I spent some time looking into this and I *think* I know what is happening for this particular case. I think the underlying problem is how Bifrost computes array sizes....
I'm not sure that `A_nelement_total` should include elements from other batches (that last term in the sum). For the batched execution you are only really concerned with those elements that...
According to `memory.py` it looks like we should just drop `memcpy`/`memcpy2D` and `memset`/`memset2D`. https://github.com/ledatelescope/bifrost/blob/7dc1c5cd6919501b39757cc6db5d05d9049285b0/python/bifrost/memory.py#L63
I've created a test PR at #230 to see what this would even look like.
This could just be user error on my part.