SparseMatricesCSR Dispatch

[Abdelrahman912]

This PR adds a new extension module SparseMatricesCSRExt that enables dispatching SparseMatrixCSR from SparseMatricesCSR.jl to CuSparseMatrixCSR.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.62%. Comparing base (3d44e32) to head (0e80594).
Report is 4 commits behind head on master.

Additional details and impacted files

@@             Coverage Diff             @@
##           master    #2720       +/-   ##
===========================================
+ Coverage   77.34%   89.62%   +12.28%     
===========================================
  Files         153      153               
  Lines       13108    13195       +87     
===========================================
+ Hits        10138    11826     +1688     
+ Misses       2970     1369     -1601     

☔ 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.

[github-actions]

Your PR no longer requires formatting changes. Thank you for your contribution!

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: 0e80594	Previous: 7a83380	Ratio
latency/precompile	42831148863 ns	42926176632.5 ns	1.00
latency/ttfp	7055807029 ns	7106699468 ns	0.99
latency/import	3375773699 ns	3389948934 ns	1.00
integration/volumerhs	9597397 ns	9608052 ns	1.00
integration/byval/slices=1	147336 ns	147162 ns	1.00
integration/byval/slices=3	426120 ns	425761 ns	1.00
integration/byval/reference	145426 ns	145313 ns	1.00
integration/byval/slices=2	286891 ns	286425 ns	1.00
integration/cudadevrt	103861 ns	103604 ns	1.00
kernel/indexing	14551 ns	14311 ns	1.02
kernel/indexing_checked	15115.5 ns	15150.5 ns	1.00
kernel/occupancy	719.3142857142857 ns	707.8943661971831 ns	1.02
kernel/launch	2538 ns	2395.4444444444443 ns	1.06
kernel/rand	18953 ns	18678 ns	1.01
array/reverse/1d	20039 ns	19617 ns	1.02
array/reverse/2d	24584 ns	24183.5 ns	1.02
array/reverse/1d_inplace	11270 ns	10894 ns	1.03
array/reverse/2d_inplace	12872 ns	12860 ns	1.00
array/copy	21059 ns	20975 ns	1.00
array/iteration/findall/int	159756.5 ns	157315 ns	1.02
array/iteration/findall/bool	140032 ns	138316 ns	1.01
array/iteration/findfirst/int	154689 ns	154273.5 ns	1.00
array/iteration/findfirst/bool	156124 ns	154976.5 ns	1.01
array/iteration/scalar	72823 ns	73010 ns	1.00
array/iteration/logical	220835.5 ns	215446 ns	1.03
array/iteration/findmin/1d	42153 ns	41421 ns	1.02
array/iteration/findmin/2d	94512 ns	94396 ns	1.00
array/reductions/reduce/1d	45220 ns	43605.5 ns	1.04
array/reductions/reduce/2d	51768.5 ns	45135 ns	1.15
array/reductions/mapreduce/1d	40197.5 ns	40394 ns	1.00
array/reductions/mapreduce/2d	52388.5 ns	51842 ns	1.01
array/broadcast	21246 ns	21077 ns	1.01
array/copyto!/gpu_to_gpu	12645 ns	11034 ns	1.15
array/copyto!/cpu_to_gpu	218110 ns	216527 ns	1.01
array/copyto!/gpu_to_cpu	284544 ns	283217 ns	1.00
array/accumulate/1d	110400 ns	108949 ns	1.01
array/accumulate/2d	81360 ns	80328 ns	1.01
array/construct	1237.2 ns	1243.7 ns	0.99
array/random/randn/Float32	48013 ns	47595 ns	1.01
array/random/randn!/Float32	25166 ns	25281 ns	1.00
array/random/rand!/Int64	27536 ns	27337 ns	1.01
array/random/rand!/Float32	8832 ns	8944.333333333334 ns	0.99
array/random/rand/Int64	34423 ns	33758 ns	1.02
array/random/rand/Float32	13124 ns	13235 ns	0.99
array/permutedims/4d	61784.5 ns	61033.5 ns	1.01
array/permutedims/2d	55629 ns	54927 ns	1.01
array/permutedims/3d	56340 ns	55899 ns	1.01
array/sorting/1d	2778494 ns	2777029.5 ns	1.00
array/sorting/by	3369759 ns	3367967 ns	1.00
array/sorting/2d	1087242 ns	1085746 ns	1.00
cuda/synchronization/stream/auto	1005 ns	1030 ns	0.98
cuda/synchronization/stream/nonblocking	7507.5 ns	8027.8 ns	0.94
cuda/synchronization/stream/blocking	803.6382978723404 ns	803.8265306122449 ns	1.00
cuda/synchronization/context/auto	1161.3 ns	1173.1 ns	0.99
cuda/synchronization/context/nonblocking	8219.7 ns	7706.2 ns	1.07
cuda/synchronization/context/blocking	899.6078431372549 ns	906.5853658536586 ns	0.99

This comment was automatically generated by workflow using github-action-benchmark.

[Abdelrahman912]

This PR should be ready by now

[maleadt]

Extension packages need to be listed explicitly in the CI pipeline:

CUDA.jl/.buildkite/pipeline.yml

Lines 144 to 251 in a3d6aa4

	- group: ":telescope: Downstream"
	depends_on: "cuda"
	steps:
	#- label: "NNlib.jl"
	# plugins:
	# - JuliaCI/julia#v1:
	# version: "1.11"
	# - JuliaCI/julia-coverage#v1:
	# dirs:
	# - src
	# - lib
	# - examples
	# command: |
	# julia --project -e '
	# using Pkg
	#
	# cuda = pwd()
	# cudnn = joinpath(cuda, "lib", "cudnn")
	# devdir = mktempdir()
	# nnlib = joinpath(devdir, "NNlib")
	#
	# println("--- :julia: Installing TestEnv")
	# Pkg.activate(; temp=true)
	# Pkg.add("TestEnv")
	# using TestEnv
	#
	# println("--- :julia: Installing NNlib")
	# withenv("JULIA_PKG_PRECOMPILE_AUTO" => 0,
	# "JULIA_PKG_DEVDIR" => devdir) do
	# Pkg.develop("NNlib")
	# Pkg.activate(nnlib)
	#
	# try
	# Pkg.develop([PackageSpec(path=cuda), PackageSpec(path=cudnn)])
	# TestEnv.activate()
	# catch err
	# @error "Could not install NNlib" exception=(err,catch_backtrace())
	# exit(3)
	# finally
	# Pkg.activate(nnlib)
	# end
	# end
	#
	# println("+++ :julia: Running tests")
	# Pkg.test(; coverage=true)'
	# env:
	# NNLIB_TEST_CUDA: "true"
	# NNLIB_TEST_CPU: "false"
	# agents:
	# queue: "juliagpu"
	# cuda: "*"
	# if: |
	# build.message =~ /\[only tests\]/ ||
	# build.message =~ /\[only downstream\]/ ||
	# build.message !~ /\[only/ && !build.pull_request.draft &&
	# build.message !~ /\[skip tests\]/ &&
	# build.message !~ /\[skip downstream\]/
	# timeout_in_minutes: 30
	# soft_fail:
	# - exit_status: 3
	- label: "Enzyme.jl"
	plugins:
	- JuliaCI/julia#v1:
	version: "1.10" # XXX: Enzyme.jl is broken on 1.11
	- JuliaCI/julia-coverage#v1:
	dirs:
	- src
	- lib
	- examples
	command: |
	julia -e '
	using Pkg
	println("--- :julia: Instantiating project")
	withenv("JULIA_PKG_PRECOMPILE_AUTO" => 0) do
	# add Enzyme to the test deps
	Pkg.activate("test")
	Pkg.add(["Enzyme", "EnzymeCore"])
	# to check compatibility, also add Enzyme to the main environment
	# (or Pkg.test, which merges both environments, could fail)
	Pkg.activate(".")
	# Try to co-develop Enzyme and KA, if that fails, try just to dev Enzyme
	try
	Pkg.develop([PackageSpec("Enzyme"), PackageSpec("KernelAbstractions")])
	catch err
	try
	Pkg.develop([PackageSpec("Enzyme")])
	catch err
	@error "Could not install Enzyme" exception=(err,catch_backtrace())
	exit(3)
	end
	end
	end
	println("+++ :julia: Running tests")
	Pkg.test(; coverage=true, test_args=`extensions/enzyme`)'
	agents:
	queue: "juliagpu"
	cuda: "*"
	if: |
	build.message =~ /\[only tests\]/ ||
	build.message =~ /\[only downstream\]/ ||
	build.message !~ /\[only/ && !build.pull_request.draft &&
	build.message !~ /\[skip tests\]/ &&
	build.message !~ /\[skip downstream\]/
	timeout_in_minutes: 60
	soft_fail: true

Alternatively, since this is a simple interface package, we could add it to the test project environment and test it unconditionally, much like SpecialFunctions or BFloat16s.

[maleadt]

Tests in test/extensions aren't executed by default:

CUDA.jl/test/runtests.jl

Lines 115 to 117 in c113666

	# package extensions often require additional dependencies,
	# which we don't want to put in our test env by default.
	startswith(test, "extensions") && return false

You'll have to move it to e.g. libraries/cusparse