Allow JIT Compile to, and Link from, LTOIR for cuda source input #48
Conversation
Co-authored-by: Graham Markall <535640+gmarkall@users.noreply.github.com>
isVoid
changed the title
| # TODO: when linker is configured to generate LTOIR, assembly is not | ||
| # directly visible via the Linker API. We should provide `-ptx` as | ||
| # an additional flag to nvjitlink to generate PTX. However, this | ||
| # could break the linking pipeline. We need to investigate this further. |
There was a problem hiding this comment.
To get PTX when doing LTO, get_cubin() will need to run the linker twice, once without "-ptx" (as it does now) and once again with "-ptx" (additionally) to get the PTX - it can then dump it in a similar way to how it's dumped here.
There was a problem hiding this comment.
As a set of changes on top of #23 this looks good - if you can add the logic to dump assembly correctly when using LTO (and some tests) I think that should be enough for this PR. Do you see anything else that might be needed?
Without looking very deep into the code base, where is
Ok sounds good. |
numba_cuda/numba/cuda/codegen.py
Outdated
| ptx = linker.get_linked_ptx().decode('utf-8') | ||
|
|
||
| if config.DUMP_ASSEMBLY: | ||
| print(("ASSEMBLY (AFTER LTO) %s" % self._name).center(80, '-')) |
There was a problem hiding this comment.
If dump_assembly=1, the pipeline will dump PTX twice: once before LTO, and once after LTO. I think it might be helpful to keep both results. So I added (AFTER LTO) here for clarity.
Good point, I think it's missing - I think when #23 adds tests that would have become apparent, and it will be required. |
gmarkall
added
4 - Waiting on reviewer
In 159e77e I added the line to make lto enabled by default. |
gmarkall
added
the
develop
|
Closing in favour of #62, which is based on a recent |
Depends on #23
This PR adds the functionality supporting kernel and FFI functions being JIT Compile to and link from LTOIR, allowing better optimization when foreign function is used in Numba-cuda.