This repository was archived by the owner on Mar 21, 2024. It is now read-only.
cub::DeviceSpmv needs attention #285
Comments
|
CuPy currently disables CUB SpMV if it's from CUDA's bundle, so it should not affect us if you decide to remove 🙂 |
alliepiper
added a commit
to alliepiper/cub
that referenced
this issue
Apr 15, 2021
`cub::DeviceSpmv` has not been ported to use cub::detail::ptx_dispatch and uses an older dispatch mechanism that is incompatible with nvc++. It was not ported because there are no tests and thus no way to verify whether or not the ported version works. Rather than risk breaking user code, we'll port this later, after we add tests. NV HPC doesn't require this algorithm so it's low priority for now. Tracked by NVIDIA#285.
|
I'd like to use CUB's SPMV as an analysis tool for the memory access patterns of merge-path-based SPMV, since original repo from Duane's paper is no longer compatible with the latest versions of CUDA. |
alliepiper
added a commit
to alliepiper/cub
that referenced
this issue
May 27, 2021
`cub::DeviceSpmv` has not been ported to use cub::detail::ptx_dispatch and uses an older dispatch mechanism that is incompatible with nvc++. It was not ported because there are no tests and thus no way to verify whether or not the ported version works. Rather than risk breaking user code, we'll port this later, after we add tests. NV HPC doesn't require this algorithm so it's low priority for now. Tracked by NVIDIA#285.
alliepiper
added a commit
to alliepiper/cub
that referenced
this issue
Jun 9, 2021
`cub::DeviceSpmv` has not been ported to use cub::detail::ptx_dispatch and uses an older dispatch mechanism that is incompatible with nvc++. It was not ported because there are no tests and thus no way to verify whether or not the ported version works. Rather than risk breaking user code, we'll port this later, after we add tests. NV HPC doesn't require this algorithm so it's low priority for now. Tracked by NVIDIA#285.
alliepiper
added a commit
to alliepiper/cub
that referenced
this issue
Jun 30, 2021
`cub::DeviceSpmv` has not been ported to use cub::detail::ptx_dispatch and uses an older dispatch mechanism that is incompatible with nvc++. It was not ported because there are no tests and thus no way to verify whether or not the ported version works. Rather than risk breaking user code, we'll port this later, after we add tests. NV HPC doesn't require this algorithm so it's low priority for now. Tracked by NVIDIA#285.
alliepiper
added a commit
to alliepiper/cub
that referenced
this issue
Jul 7, 2021
`cub::DeviceSpmv` has not been ported to use cub::detail::ptx_dispatch and uses an older dispatch mechanism that is incompatible with nvc++. It was not ported because there are no tests and thus no way to verify whether or not the ported version works. Rather than risk breaking user code, we'll port this later, after we add tests. NV HPC doesn't require this algorithm so it's low priority for now. Tracked by NVIDIA#285.
alliepiper
added a commit
to alliepiper/cub
that referenced
this issue
Jul 20, 2021
`cub::DeviceSpmv` has not been ported to use cub::detail::ptx_dispatch and uses an older dispatch mechanism that is incompatible with nvc++. It was not ported because there are no tests and thus no way to verify whether or not the ported version works. Rather than risk breaking user code, we'll port this later, after we add tests. NV HPC doesn't require this algorithm so it's low priority for now. Tracked by NVIDIA#285.
alliepiper
added a commit
to alliepiper/cub
that referenced
this issue
Jul 27, 2021
`cub::DeviceSpmv` has not been ported to use cub::detail::ptx_dispatch and uses an older dispatch mechanism that is incompatible with nvc++. It was not ported because there are no tests and thus no way to verify whether or not the ported version works. Rather than risk breaking user code, we'll port this later, after we add tests. NV HPC doesn't require this algorithm so it's low priority for now. Tracked by NVIDIA#285.
|
Just to close the loop on this -- we're going to continue supporting it for now, but will likely deprecate it in the future. #352 fixes some bugs and adds tests. |
alliepiper
added a commit
to alliepiper/cub
that referenced
this issue
Nov 3, 2021
`cub::DeviceSpmv` has not been ported to use cub::detail::ptx_dispatch and uses an older dispatch mechanism that is incompatible with nvc++. It was not ported because there are no tests and thus no way to verify whether or not the ported version works. Rather than risk breaking user code, we'll port this later, after we add tests. NV HPC doesn't require this algorithm so it's low priority for now. Tracked by NVIDIA#285.
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
This algorithm has no tests or examples, and some of the names look like they're from a work-in-progress (eg:
dispatch_spmv_orig.cuh). We'll need to either deprecate and remove this code, or fix it up (there are a few bugs filed for it) and add tests/examples.If anyone is actively using this and would be impacted by its removal, leave a comment and let us know.
The text was updated successfully, but these errors were encountered: