Ability to freeze compiled kernels

[maleadt]

I'm told Clima would appreciate this:

$ time jl examples/vadd.jl
julia --color=yes --quiet examples/vadd.jl  17,83s user 0,87s system 100% cpu 18,603 total

$ time JULIA_DEBUG=GPUCompiler JULIA_GPUCOMPILER_DISKCACHE=true jl examples/vadd.jl
JULIA_DEBUG=GPUCompiler JULIA_GPUCOMPILER_FROZEN=true julia --color=yes    19,01s user 0,88s system 98% cpu 20,185 total

$ time JULIA_DEBUG=GPUCompiler JULIA_GPUCOMPILER_DISKCACHE=true jl examples/vadd.jl
┌ Debug: Loading compiled kernel for kernel vadd(CuDeviceArray{Float32,2,1}, CuDeviceArray{Float32,2,1}, CuDeviceArray{Float32,2,1}) from /home/tim/Julia/depot/scratchspaces/61eb1bfa-7361-4325-ad38-22787b887f55/kernels/8468340805782850700.jls
└ @ GPUCompiler ~/Julia/pkg/GPUCompiler/src/cache.jl:51
JULIA_DEBUG=GPUCompiler JULIA_GPUCOMPILER_DISKCACHE=true julia --color=yes    14,57s user 0,98s system 99% cpu 15,630 total

[codecov]

Codecov Report

Merging #89 into master will decrease coverage by 2.72%.
The diff coverage is 2.70%.

@@            Coverage Diff             @@
##           master      #89      +/-   ##
==========================================
- Coverage   66.89%   64.16%   -2.73%     
==========================================
  Files          31       20      -11     
  Lines        1637     1387     -250     
==========================================
- Hits         1095      890     -205     
+ Misses        542      497      -45     

Impacted Files	Coverage Δ
src/cache.jl	0.00% <0.00%> (ø)
src/reflection.jl	6.38% <ø> (+0.06%)	⬆️
src/GPUCompiler.jl	100.00% <100.00%> (ø)
src/irgen.jl	60.83% <0.00%> (-25.88%)	⬇️
src/validation.jl	84.84% <0.00%> (-4.55%)	⬇️
src/runtime.jl	65.45% <0.00%> (-3.64%)	⬇️
src/precompile.jl	77.14% <0.00%> (-2.86%)	⬇️
src/gcn.jl	68.88% <0.00%> (-2.54%)	⬇️
test/spirv.jl
examples/kernel.jl
... and 9 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update a0f686f...a90f7bb. Read the comment docs.

[maleadt]

cc @jpsamaroo this would change cached_compilation and split the driver argument in two separate phases: compile, returning (serializable) assembly, and linked to perform the run-time link step (if objects are precompilable too, we should maybe rename this, but the fact remains that you only want to load those objects once too).

[jpsamaroo]

I've implemented support in JuliaGPU/AMDGPU.jl#53, just waiting to see if it passes CI. Otherwise LGTM!

[jpsamaroo]

This also needs a minor version bump due to API changes.

[jpsamaroo]

With or without kwargs, I'm cool with this.

[maleadt]

I moved them over as you asked, so it should be fine now?