Add CodeInstance support to kernel renaming

[imreddyTeja]

In 1.12, kernel renaming does not function correctly when used interactively in the REPL. For example,
profiling update_jacobian results in 213 calls to kernels named __repl_entry_eval_expanded_with_loc_FILE_src_REPL_jl_L301. It looks like in 1.12 many of the frames that used to be
MethodInstance(s) are now CodeInstance(s). This commit adds support for CodeInstance frames.

Before:

julia> CUDA.@profile CA.update_jacobian!(alg, cache, Y, p, 0.2f0, 1.0f0)
Profiler ran for 97.52 s, capturing 10563 events.

Host-side activity: calling CUDA APIs took 34.89 ms (0.04% of the trace)
┌──────────┬────────────┬───────┬──────────────────────────────────────┬─────────────────────┐
│ Time (%) │ Total time │ Calls │ Time distribution                    │ Name                │
├──────────┼────────────┼───────┼──────────────────────────────────────┼─────────────────────┤
│    0.01% │   13.06 ms │   100 │ 130.61 µs ± 77.73  ( 60.08 ‥ 759.84) │ cuModuleGetFunction │
│    0.01% │    12.4 ms │   100 │ 123.96 µs ± 66.33  (  75.1 ‥ 697.37) │ cuModuleLoadDataEx  │
│    0.00% │     4.5 ms │   213 │  21.14 µs ± 14.89  (  3.81 ‥ 72.96)  │ cuLaunchKernel      │
│    0.00% │     1.4 ms │   100 │  13.96 µs ± 2.02   ( 10.25 ‥ 21.93)  │ cuCtxSynchronize    │
│    0.00% │  428.44 µs │     2 │ 214.22 µs ± 109.41 (136.85 ‥ 291.59) │ cuMemFree           │
│    0.00% │  953.67 ns │     1 │                                      │ cuCtxSetCurrent     │
│    0.00% │  476.84 ns │     1 │                                      │ cuCtxGetDevice      │
│    0.00% │  238.42 ns │     1 │                                      │ cuDeviceGetCount    │
└──────────┴────────────┴───────┴──────────────────────────────────────┴─────────────────────┘

Device-side activity: GPU was busy for 12.0 ms (0.01% of the trace)
┌──────────┬────────────┬───────┬──────────────────────────────────────┬───────────────────────────────────────────────────────────┐
│ Time (%) │ Total time │ Calls │ Time distribution                    │ Name                                                      │
├──────────┼────────────┼───────┼──────────────────────────────────────┼───────────────────────────────────────────────────────────┤
│    0.01% │    12.0 ms │   213 │  56.32 µs ± 30.3   (  24.8 ‥ 184.54) │ __repl_entry_eval_expanded_with_loc_FILE_src_REPL_jl_L301 │
└──────────┴────────────┴───────┴──────────────────────────────────────┴───────────────────────────────────────────────────────────┘

After:

julia> CUDA.@profile CA.update_jacobian!(alg, cache, Y, p, 0.2f0, 1.0f0)
Profiler ran for 76.36 s, capturing 8548 events.

Host-side activity: calling CUDA APIs took 6.71 ms (0.01% of the trace)
┌──────────┬────────────┬───────┬──────────────────────────────────────┬────────────────┐
│ Time (%) │ Total time │ Calls │ Time distribution                    │ Name           │
├──────────┼────────────┼───────┼──────────────────────────────────────┼────────────────┤
│    0.01% │    5.74 ms │   213 │  26.97 µs ± 22.01  (   3.1 ‥ 120.16) │ cuLaunchKernel │
└──────────┴────────────┴───────┴──────────────────────────────────────┴────────────────┘

Device-side activity: GPU was busy for 11.97 ms (0.02% of the trace)
┌──────────┬────────────┬───────┬──────────────────────────────────────┬───────────────────────────────────────────────────────────────────────────────────────────────────────┐
│ Time (%) │ Total time │ Calls │ Time distribution                    │ Name                                                                                                  │
├──────────┼────────────┼───────┼──────────────────────────────────────┼───────────────────────────────────────────────────────────────────────────────────────────────────────┤
│    0.00% │    2.81 ms │    46 │   61.0 µs ± 20.0   ( 25.99 ‥ 112.06) │ update_jacobian__FILE_ClimaAtmos_jl_src_prognostic_equations_implicit_manual_sparse_jacobian_jl_L414  │
│    0.00% │    1.28 ms │     8 │  159.5 µs ± 25.82  (134.71 ‥ 184.54) │ _FILE_ClimaAtmos_jl_src_prognostic_equations_implicit_manual_sparse_jacobian_jl_L1338                 │
│    0.00% │  563.86 µs │     8 │  70.48 µs ± 0.18   ( 70.33 ‥ 70.81)  │ _FILE_ClimaAtmos_jl_src_prognostic_equations_implicit_manual_sparse_jacobian_jl_L1344                 │
│    0.00% │  405.79 µs │     4 │ 101.45 µs ± 0.46   (101.09 ‥ 102.04) │ _FILE_ClimaAtmos_jl_src_prognostic_equations_implicit_manual_sparse_jacobian_jl_L1009                 │
│    0.00% │  387.43 µs │     4 │  96.86 µs ± 0.3    ( 96.56 ‥ 97.27)  │ _FILE_ClimaAtmos_jl_src_prognostic_equations_implicit_manual_sparse_jacobian_jl_L650                  │

• Code follows the style guidelines OR N/A.
• Unit tests are included OR N/A.
• Code is exercised in an integration test OR N/A.
• Documentation has been added/updated OR N/A.

[petebachant]

LGTM. I don't think I ever added a unit test for this, but if it were possible, that would be nice to ensure we can do kernel renaming on Julia 1.10--1.12.

[imreddyTeja]

LGTM. I don't think I ever added a unit test for this, but if it were possible, that would be nice to ensure we can do kernel renaming on Julia 1.10--1.12.

That is a good point. ll add a unit test to our current 1.10 buildkite pipeline. The gpu ci is currently only run with 1.10., so 1.11 and 1.12 won't be covered

[petebachant]

Might want to move this out into the buildkite env vars if that's the typical usage. Not sure how you usually do it, but I typically go through the env vars.