Switch to PrecompileTools.jl.

[maleadt]

Testing examples/kernel.jl without a precompile workload at all:

❯ hyperfine -w 1 " julia --project examples/kernel.jl"
Benchmark 1:  julia --project examples/kernel.jl
  Time (mean ± σ):      8.131 s ±  0.241 s    [User: 8.139 s, System: 1.156 s]
  Range (min … max):    7.873 s …  8.375 s    10 runs

Current master:

Benchmark 1:  julia --project examples/kernel.jl
  Time (mean ± σ):      7.732 s ±  0.085 s    [User: 7.723 s, System: 1.173 s]
  Range (min … max):    7.560 s …  7.818 s    10 runs

Before #600:

Benchmark 1:  julia --project examples/kernel.jl
  Time (mean ± σ):      1.782 s ±  0.022 s    [User: 1.825 s, System: 1.130 s]
  Range (min … max):    1.744 s …  1.815 s    10 runs

This PR:

Benchmark 1:  julia --project examples/kernel.jl
  Time (mean ± σ):     991.3 ms ±  14.8 ms    [User: 1045.8 ms, System: 1120.8 ms]
  Range (min … max):   968.7 ms … 1013.1 ms    10 runs

So still an improvement. Probably doesn't matter too much for downstream users, as the whole @nospecialize(job::CompilerJob) doesn't really work and I think we end up respecializing the GPUCompiler code anyway (i.e., downstream packages require their own precompile workload).

[maleadt]

@vchuravy GPU methods leaking into the package image on Windows/1.11?

lld: error: undefined symbol: gpu_malloc
>>> referenced by D:\a\GPUCompiler.jl\GPUCompiler.jl\src\runtime.jl:88
>>>               jl_B252.tmp(text#0.o):(julia_box_uint8_16861)
>>> referenced by D:\a\GPUCompiler.jl\GPUCompiler.jl\src\runtime.jl:88
>>>               jl_B252.tmp(text#0.o):(julia_box_int8_16844)
>>> referenced by D:\a\GPUCompiler.jl\GPUCompiler.jl\src\runtime.jl:88
>>>               jl_B252.tmp(text#0.o):(julia_box_float64_16886)
>>> referenced 9 more times

[vchuravy]

Yes but no... This is a tricky design issue/interaction.

I believe this is the same issue as why I want to use the custom method table for all GPU intrinsics and it should be equivalent to adding precompile()

Essentially precompilation tracking discards the CI and then uses the MI alone.

[vchuravy]

So it would be good to confirm if this is the same as:

JuliaLang/julia#54155

[codecov]

Codecov Report

Attention: Patch coverage is 73.91304% with 6 lines in your changes missing coverage. Please review.

Project coverage is 66.27%. Comparing base (28d96c1) to head (75c2112).

Files	Patch %	Lines
src/precompile.jl	70.00%	6 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #601      +/-   ##
==========================================
+ Coverage   65.77%   66.27%   +0.50%     
==========================================
  Files          24       24              
  Lines        3348     3339       -9     
==========================================
+ Hits         2202     2213      +11     
+ Misses       1146     1126      -20     

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

[maleadt]

So it would be good to confirm if this is the same as:

JuliaLang/julia#54155

Yep, just tested this out, and I can confirm it fixes the precompilation issue.

[vchuravy]

Okay... Then we need to decide if this is the right fix...
I wasn't convinced at the time.

Also just reminded me of JuliaLang/julia#55072 (comment)

[maleadt]

It's also surprising that this only triggers on Windows, no?

[vchuravy]

It's also surprising that this only triggers on Windows, no?

A little bit...

[maleadt]

Workaround: don't build the runtime, so that the simple kernel compiled as part of the precompilation workload doesn't generate GPU-specific code. Seems to keep most of the latency advantage:

❯ hyperfine -w 1 " julia --project examples/kernel.jl"
Benchmark 1:  julia --project examples/kernel.jl
  Time (mean ± σ):     977.3 ms ±  12.8 ms    [User: 1042.0 ms, System: 1110.2 ms]
  Range (min … max):   966.3 ms … 1010.9 ms    10 runs