GPU cleanup improvements

[Keluaa]

Fix for #19

• Demangling of LLVM names
• Demangling on Windows?
• Demangling LLVM names in PTX and SASS
  • the names are still very long, it might be preferable to detect the mangled function name and replace it with <current_function> or something else
  • kernel parameters are named with mangled_name followed by _param_<idx>, messing up the demangling, this should be replaced easily
  • extern functions in the form .extern .func (.param .b64 func_retval0) overflow_case are not sorted deterministically
  • There is sometimes a missing newline between the ret; and the closing }
• Demangling LLVM names in GCN
  • update AMDGPU compat
  • same names, no parameters (unlike with CUDA)
• Demangling LLVM names in SPIR-V
• debuginfo alternative for all code types
• Get rid of ; Location comments in SASS (as an option)
• Cleanup LLVM IR generated by the GPU backends
  • with cuda_llvm, the function name is still mangled
• Improve PTX display (colors?)
  • Pygmentize v2.16 supports PTX
  • Pygmentize v2.8 supports AMDGPU
  • GPUCompiler v1.2 now supports highlighting of PTX and GCN
    • Add Pygmentize in CI
    • Test highlighting of PTX (without CUDA.jl)
• Support for Cthulhu.jl, especially for GPU kernels
• Inline LLVM calls in typed IR
  • Having some compound types in arguments messes up the extraction of the LLVM function body
• Nit: fix the alignment of predicates in PTX (@%p1 bra %L1 is aligned with other instruction at @, but doing it at bra would match how nsys does it) (read the PTX spec to match predicates correctly)
• Option to remove loop comments in PTX, e.g. // Parent Loop BB0_8 Depth=1, // => This Loop Header: Depth=4, // Child Loop BB0_265 Depth 5, or // in Loop: Header=BB0_17 Depth=3
• Indenting disparity on PTX between @code_for and @code_diff
• Update docs
  • Docs for cleanup options
  • Docs for stats

Kernel test case:

@kernel function test_copy(a, b)
    i = @index(Global, Linear)
    b[i] = a[i]
end

[codecov]

Codecov Report

Attention: Patch coverage is 83.91720% with 101 lines in your changes missing coverage. Please review.

Project coverage is 81.40%. Comparing base (730b5c9) to head (68a3e3f).
Report is 51 commits behind head on main.

Files with missing lines	Patch %	Lines
ext/CodeDiffsCUDA.jl	0.00%	19 Missing ⚠️
src/cleanup/ptx.jl	80.00%	18 Missing ⚠️
ext/CodeDiffsAMDGPU.jl	0.00%	15 Missing ⚠️
src/stats/ptx.jl	92.75%	10 Missing ⚠️
src/get_code.jl	68.00%	8 Missing ⚠️
src/cleanup/typed_ir.jl	81.57%	7 Missing ⚠️
ext/CodeDiffsMetal.jl	0.00%	4 Missing ⚠️
ext/gpu_common.jl	0.00%	4 Missing ⚠️
src/cleanup/llvm_ir.jl	20.00%	4 Missing ⚠️
ext/CodeDiffsOneAPI.jl	0.00%	3 Missing ⚠️
... and 4 more

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #20      +/-   ##
==========================================
+ Coverage   75.37%   81.40%   +6.02%     
==========================================
  Files          13       25      +12     
  Lines         670     1242     +572     
==========================================
+ Hits          505     1011     +506     
- Misses        165      231      +66     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

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

[Keluaa]

Example code for missing newline between ret; and the last }:

@kernel inbounds=true unsafe_indices=true function sync_shared(x)
   is_main = @index(Local, Linear) == 1
   
   shrd = KernelAbstractions.@localmem Int (1,)
   
   if is_main
       shrd[1] = 42
   end

   KernelAbstractions.@synchronize()
   val = shrd[1]
   
   i = @index(Local, Linear)
   x[i] = val
end