LV keeps crushing during benchmarking about loopvectorization.jl HOT 5 CLOSED

Julia Version 1.8.2
Commit 36034abf26 (2022-09-29 15:21 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: 64 × Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-13.0.1 (ORCJIT, cascadelake)
  Threads: 32 on 64 virtual cores
Environment:
  JULIA_NUM_THREADS = 32

from loopvectorization.jl.

@turbo broadcasting currently doesn't support dynamically broadcasting a contiguous dimension.

That is, a normally contiguous dimension should either be known to be of size 1 at compile time, or of full size.

It is possible to make this efficient, but I've never had the time.
I'll probably do that at some point in the rewrite.

from loopvectorization.jl.

Thank you for the quick reply! Surprisingly, it gives correct result and won't crush, as long as it is not being benchmarked. And it could be benchmarked without any issues in my MBP...

from loopvectorization.jl.

I'd have to take a closer look.
It doesn't crash for me:

julia> @benchmark fast_inverse(@view($result[:,:]), $A_inv, $b, $c)
BenchmarkTools.Trial: 10000 samples with 1 evaluation.
 Range (min … max):   9.902 μs … 39.703 μs  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     13.838 μs              ┊ GC (median):    0.00%
 Time  (mean ± σ):   14.702 μs ±  3.534 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

    ▂▃ ▂    ▅▆▆▇█▄▄▃▁▁                        ▁        ▅   ▁▂ ▂
  ██████▇█▇▇██████████▇▇▄▄▅▇█▇▆▁▃▆▄▅▄▁▄▁▄▇▆▅▄▃█▇▆▆▃▇▄▄▃█▆▃███ █
  9.9 μs       Histogram: log(frequency) by time      26.5 μs <

 Memory estimate: 1.75 KiB, allocs estimate: 2.

julia> versioninfo()
Julia Version 1.9.0-DEV.1805
Commit 050f21edc4 (2022-11-10 22:44 UTC)
Platform Info:
  OS: Linux (x86_64-redhat-linux)
  CPU: 28 × Intel(R) Core(TM) i9-9940X CPU @ 3.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, skylake-avx512)
  Threads: 28 on 28 virtual cores

Can you try commenting out the @turbo @. result[n+1:n+dt_n, 1:n] = -g*f line to see if it makes a difference?

from loopvectorization.jl.

Stop crushing after changing result[end, end] = g to @turbo @. result[n+1:n+dt_n, n+1:n+dt_n] = g. Don't know why but it just works.

from loopvectorization.jl.