Add example of workgroupsize breaking Enzyme

[vchuravy]

julia> Enzyme.autodiff(Forward, mul_caller_static, Duplicated(A, dA), Duplicated(1.2, 1.0))
C = max(64 ÷ prod(size(A)), 1) = 1
ERROR: 
No forward mode derivative found for jl_get_builtin_fptr
 at context:   %15 = call {} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32)* @jl_get_builtin_fptr({} addrspace(10)* noundef addrspacecast ({}* inttoptr (i64 139935955039648 to {}*) to {} addrspace(10)*)) #20, !dbg !53

Stacktrace:
 [1] construct
   @ ~/.julia/packages/KernelAbstractions/X5fk1/src/KernelAbstractions.jl:779
 [2] _mul!
   @ ~/.julia/packages/KernelAbstractions/X5fk1/src/macros.jl:61
 [3] mul!
   @ ~/.julia/packages/KernelAbstractions/X5fk1/src/macros.jl:68
 [4] mul_caller_static
   @ ./REPL[11]:4


Stacktrace:
  [1] construct
    @ ~/.julia/packages/KernelAbstractions/X5fk1/src/KernelAbstractions.jl:779 [inlined]
  [2] _mul!
    @ ~/.julia/packages/KernelAbstractions/X5fk1/src/macros.jl:61 [inlined]
  [3] mul!
    @ ~/.julia/packages/KernelAbstractions/X5fk1/src/macros.jl:68 [inlined]
  [4] mul_caller_static
    @ ./REPL[11]:4 [inlined]
  [5] fwddiffejulia_mul_caller_static_23728wrap
    @ ./REPL[11]:0
  [6] macro expansion
    @ ~/src/Enzyme/src/compiler.jl:5887 [inlined]
  [7] enzyme_call
    @ ~/src/Enzyme/src/compiler.jl:5421 [inlined]
  [8] ForwardModeThunk
    @ ~/src/Enzyme/src/compiler.jl:5323 [inlined]
  [9] autodiff
    @ ~/src/Enzyme/src/Enzyme.jl:673 [inlined]
 [10] autodiff
    @ ~/src/Enzyme/src/Enzyme.jl:562 [inlined]
 [11] autodiff(::ForwardMode{…}, ::typeof(mul_caller_static), ::Duplicated{…}, ::Duplicated{…})
    @ Enzyme ~/src/Enzyme/src/Enzyme.jl:534
 [12] top-level scope
    @ REPL[12]:1
Some type information was truncated. Use `show(err)` to see complete types.

[wsmoses]

what exactly is jl_get_builtin_fptr

and can we define a relevant custom derivative handler for it?

it's come up a few times

[wsmoses]

per https://github.com/JuliaLang/julia/pull/55803

it looks like its a different way of throwing a method error?

which would explain why the unreachable thing happened?

[vchuravy]

The primal code doesn't throw, so it might be something that is emitted as a fallback.

Now I have a small enough example that I can stare at Monday

[github-actions[bot]]

Benchmark Results

	main	a81b1bc388ff7d...	main / a81b1bc388ff7d...
basics/make_zero/namedtuple	0.053 ± 0.0025 μs	0.0523 ± 0.0027 μs	1.01 ± 0.07
basics/make_zero/struct	0.285 ± 0.0082 μs	0.271 ± 0.0068 μs	1.05 ± 0.04
basics/overhead	4.03 ± 0.001 ns	4.64 ± 0.01 ns	0.868 ± 0.0019
basics/remake_zero!/namedtuple	0.234 ± 0.0072 μs	0.235 ± 0.0061 μs	0.995 ± 0.04
basics/remake_zero!/struct	0.234 ± 0.0066 μs	0.248 ± 0.0073 μs	0.943 ± 0.039
fold_broadcast/multidim_sum_bcast/1D	10.3 ± 0.37 μs	10.3 ± 0.26 μs	1 ± 0.044
fold_broadcast/multidim_sum_bcast/2D	12.2 ± 0.29 μs	12.1 ± 0.25 μs	1 ± 0.032
time_to_load	1.01 ± 0.0046 s	1.02 ± 0.0044 s	0.993 ± 0.0062

Benchmark Plots

A plot of the benchmark results has been uploaded as an artifact at https://github.com/EnzymeAD/Enzyme.jl/actions/runs/19854289811/artifacts/4735610205.

[vchuravy]

@wsmoses with Cthulhu:

; │┌ @ /home/vchuravy/.julia/packages/KernelAbstractions/X5fk1/src/macros.jl:61 within `_mul!`
; ││┌ @ /home/vchuravy/.julia/packages/KernelAbstractions/X5fk1/src/KernelAbstractions.jl:779 within `construct`
     %12 = call ptr @jl_get_builtin_fptr(ptr nonnull @"+Core.#_compute_sparams#20514.jit")

builtins.c
2181:JL_CALLABLE(jl_f__compute_sparams)
2183:    JL_NARGSV(_compute_sparams, 1);
2185:    JL_TYPECHK(_compute_sparams, method, (jl_value_t*)m);

But the error that Enzyme gives only the pointer:

  Enzyme: jl_call calling convention not implemented in forward for   %15 = call nonnull "enzyme_type"="{[-1]:Pointer}" {} addrspace(10)* ({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32)*, {} addrspace(10)*, ...) @julia.call({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32)* %14, {} addrspace(10)* noundef addrspacecast ({}* inttoptr (i64 137585066522080 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 137578224249072 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 137578224248752 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 137577969042512 to {}*) to {} addrspace(10)*), {} addrspace(10)* nonnull %13, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 137578224222272 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 137580755331088 to {}*) to {} addrspace(10)*)) #17, !dbg !48

[github-actions[bot]]

Your PR requires formatting changes to meet the project's style guidelines. Please consider running Runic (git runic main) to apply these changes.

Click here to view the suggested changes.

diff --git a/src/compiler/validation.jl b/src/compiler/validation.jl
index aab84faa..a617c0b8 100644
--- a/src/compiler/validation.jl
+++ b/src/compiler/validation.jl
@@ -100,8 +100,8 @@ function __init__()
         "jl_array_typetagdata",
         "ijl_array_typetagdata",
         "jl_idtable_rehash",
-        "jl_f__compute_sparams",
-        "ijl_f__compute_sparams",
+            "jl_f__compute_sparams",
+            "ijl_f__compute_sparams",
     )
     for name in known_names
         sym = LLVM.find_symbol(name)
diff --git a/src/llvm/attributes.jl b/src/llvm/attributes.jl
index c1989bf3..279aea19 100644
--- a/src/llvm/attributes.jl
+++ b/src/llvm/attributes.jl
@@ -130,8 +130,8 @@ const nofreefns = Set{String}((
     "ijl_f__typevar",
     "jl_f_isa",
     "ijl_f_isa",
-    "jl_f__compute_sparams",
-    "ijl_f__compute_sparams",
+        "jl_f__compute_sparams",
+        "ijl_f__compute_sparams",
     "jl_set_task_threadpoolid",
     "ijl_set_task_threadpoolid",
     "jl_types_equal",
@@ -270,8 +270,8 @@ const inactivefns = Set{String}((
     "ijl_f_sizeof",
     "jl_f_isa",
     "ijl_f_isa",
-    "jl_f__compute_sparams",
-    "ijl_f__compute_sparams",
+        "jl_f__compute_sparams",
+        "ijl_f__compute_sparams",
     "jl_set_task_threadpoolid",
     "ijl_set_task_threadpoolid",
     "jl_types_equal",

[vchuravy]

@wsmoses adding jl_f__compute_sparams to the list is not effective... I assume this is because the pointer to name thing failed? Which is why the error message is not helpful?

[codecov[bot]]

Codecov Report

:white_check_mark: All modified and coverable lines are covered by tests. :white_check_mark: Project coverage is 67.77%. Comparing base (3cd3c65) to head (a81b1bc).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2751   +/-   ##
=======================================
  Coverage   67.76%   67.77%           
=======================================
  Files          58       58           
  Lines       20913    20913           
=======================================
+ Hits        14172    14173    +1     
+ Misses       6741     6740    -1     

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

:rocket: New features to boost your workflow:

• :snowflake: Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[wsmoses]

also add it to the list of names to lookup? https://github.com/EnzymeAD/Enzyme.jl/blob/6b30dda03763357b12b338275470a18df97bc7a8/src/compiler/validation.jl#L102

[vchuravy]

None of the other jl_f_ functions are listed there? And it doesn't help.