Enzyme Rules, all functioning

[wsmoses]

Requires https://github.com/EnzymeAD/Enzyme.jl/pull/729/checks

I failed to understand the extension work, so please feel free to rework.

[vchuravy]

Also will this only work for the CPU? I would prefer if we could have a generalized implementation that would also work for the GPU backends.

[wsmoses]

@vchuravy forward mode is functioning with the rule in CUDA, etc.

reverse mode, is not because of the previously discussed tape issues. I've actually mostly worked around it, however that would need to be in CUDAKernels.jl as a rule registration to be able to get the parent job (And also because mkcontext, etc is apparently not stable from cpu -> gpu etc).

[vchuravy]

Fails in function verification:

Call parameter type does not match function signature!
{} addrspace(10)* ()* bitcast ({ { {} addrspace(10)*, {} addrspace(10)*, {} addrspace(10)* } ()*, void (i8*)* }* @"_enzyme_reverse_julia_wait_3789'" to {} addrspace(10)* ()*)
 {} addrspace(10)*  %16 = call {} addrspace(10)* @julia_enq_work_3818({} addrspace(10)* ()* bitcast ({ { {} addrspace(10)*, {} addrspace(10)*, {} addrspace(10)* } ()*, void (i8*)* }* @"_enzyme_reverse_julia_wait_3789'" to {} addrspace(10)* ()*)), !dbg !3333

[wsmoses]

This looks lke the rule is no longer triggering?

[vchuravy]

Ah that's just 1.6 being unsupported

[wsmoses]

@vchuravy excludijf nightly which is currently not supported by Enzyme per the swiftcc stuff, this fully builds and runs successfully on CI.

can we merge?

[vchuravy]

@maleadt Reverse CI for CUDA seems to be broken?

Maybe JuliaGPU/CUDA.jl#1970?

[codecov]

Codecov Report

Patch coverage: 35.55% and project coverage change: -3.72 ⚠️

Comparison is base (9526206) 87.50% compared to head (46a1d1b) 83.78%.

❗ Current head 46a1d1b differs from pull request most recent head f07343f. Consider uploading reports for the commit f07343f to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #382      +/-   ##
==========================================
- Coverage   87.50%   83.78%   -3.72%     
==========================================
  Files           8        8              
  Lines         584      623      +39     
==========================================
+ Hits          511      522      +11     
- Misses         73      101      +28     

Impacted Files	Coverage Δ
ext/EnzymeExt.jl	30.00% <30.00%> (ø)
src/KernelAbstractions.jl	83.33% <75.00%> (ø)
src/cpu.jl	79.24% <100.00%> (ø)

... and 1 file with indirect coverage changes

☔ View full report in Codecov by Sentry.
📢 Do you have feedback about the report comment? Let us know in this issue.

[vchuravy]

@wsmoses did you test this with CUDA?

[wsmoses]

I did not, but only forward mode has a rule for all backends. Otherwise the PR intentionally just adds reverse mode for CPU. Reason is the tape nested job stuff we discussed, which will take some thinking to solve.