Skip to content

Test GPUArrays reverse #2832

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 23 commits into
base: master
Choose a base branch
from
+858 −799
Open

Test GPUArrays reverse #2832

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
23 commits
Select commit Hold shift + click to select a range
3652aeb
KernelIntrinsics
christiangnrd Oct 22, 2025
05bdd5c
Tweak
christiangnrd Oct 22, 2025
b86ce95
dogfood
christiangnrd Oct 22, 2025
70944da
CI
christiangnrd Oct 22, 2025
0d49f17
Apply suggestion from @christiangnrd
christiangnrd Oct 22, 2025
762e591
Apply suggestion from @christiangnrd
christiangnrd Oct 22, 2025
e133fbd
Apply suggestion from @christiangnrd
christiangnrd Oct 22, 2025
f3413ba
Fix
christiangnrd Oct 22, 2025
0474a67
shtgbr
christiangnrd Oct 22, 2025
0c1ea74
Fix
christiangnrd Oct 22, 2025
14f1c27
another fix
christiangnrd Oct 22, 2025
0a3e029
fix
christiangnrd Oct 22, 2025
d652d0b
dfgbsg
christiangnrd Oct 22, 2025
53dfbf0
dgbsg
christiangnrd Oct 22, 2025
08baa0f
Finally?
christiangnrd Oct 22, 2025
fed145c
Revert "shtgbr"
christiangnrd Oct 22, 2025
46cbc05
Reapply "shtgbr"
christiangnrd Oct 22, 2025
d341f2d
Revert KIKernel
christiangnrd Oct 22, 2025
6a2c51b
fix
christiangnrd Nov 3, 2025
b63f875
New
christiangnrd Nov 4, 2025
f61fd57
Test GPUArrays `reverse`
christiangnrd Aug 4, 2025
e2f192a
temp
christiangnrd Aug 5, 2025
69871e5
trhws
christiangnrd Nov 3, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
883 changes: 449 additions & 434 deletions .buildkite/pipeline.yml
View file
Open in desktop

Large diffs are not rendered by default.

6 changes: 5 additions & 1 deletion Project.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,6 +44,10 @@ EnzymeCore = "f151be2c-9106-41f4-ab19-57ee4f262869"
SparseMatricesCSR = "a0a7dd2c-ebf4-11e9-1f05-cf50bc540ca1"
SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"

[sources]
GPUArrays = {rev = "reverse", url = "https://github.com/christiangnrd/GPUArrays.jl"}
KernelAbstractions = {rev = "intrinsics", url = "https://github.com/christiangnrd/KernelAbstractions.jl"}

[extensions]
ChainRulesCoreExt = "ChainRulesCore"
EnzymeCoreExt = "EnzymeCore"
Expand All @@ -67,7 +71,7 @@ ExprTools = "0.1"
GPUArrays = "11.2.4"
GPUCompiler = "1.4"
GPUToolbox = "0.3, 1"
KernelAbstractions = "0.9.38"
KernelAbstractions = "0.10"
LLVM = "9.3.1"
LLVMLoopInfo = "1"
LazyArtifacts = "1"
Expand Down
1 change: 1 addition & 0 deletions perf/Project.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,5 +2,6 @@
BenchmarkTools = "6e4b80f9-dd63-53aa-95a3-0cdb28fa8baf"
HTTP = "cd3eb016-35fb-5094-929b-558a96fad6f3"
JSON = "682c06a0-de6a-54ab-a142-c8b1cf79cde6"
Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
StableRNGs = "860ef19b-820b-49d6-a774-d7a799459cd3"
StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
230 changes: 115 additions & 115 deletions perf/array.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,37 +19,37 @@ gpu_vec_ints = reshape(gpu_mat_ints, length(gpu_mat_ints))
gpu_mat_bools = CuArray(rand(rng, Bool, m, n))
gpu_vec_bools = reshape(gpu_mat_bools, length(gpu_mat_bools))

group["construct"] = @benchmarkable CuArray{Int}(undef, 1)
# group["construct"] = @benchmarkable CuArray{Int}(undef, 1)

group["copy"] = @async_benchmarkable copy($gpu_mat)
# group["copy"] = @async_benchmarkable copy($gpu_mat)

gpu_mat2 = copy(gpu_mat)
let group = addgroup!(group, "copyto!")
group["cpu_to_gpu"] = @async_benchmarkable copyto!($gpu_mat, $cpu_mat)
group["gpu_to_cpu"] = @async_benchmarkable copyto!($cpu_mat, $gpu_mat)
group["gpu_to_gpu"] = @async_benchmarkable copyto!($gpu_mat2, $gpu_mat)
end
# gpu_mat2 = copy(gpu_mat)
# let group = addgroup!(group, "copyto!")
# group["cpu_to_gpu"] = @async_benchmarkable copyto!($gpu_mat, $cpu_mat)
# group["gpu_to_cpu"] = @async_benchmarkable copyto!($cpu_mat, $gpu_mat)
# group["gpu_to_gpu"] = @async_benchmarkable copyto!($gpu_mat2, $gpu_mat)
# end

let group = addgroup!(group, "iteration")
group["scalar"] = @benchmarkable CUDA.@allowscalar [$gpu_vec[i] for i in 1:10]
# let group = addgroup!(group, "iteration")
# group["scalar"] = @benchmarkable CUDA.@allowscalar [$gpu_vec[i] for i in 1:10]

group["logical"] = @benchmarkable $gpu_vec[$gpu_vec_bools]
# group["logical"] = @benchmarkable $gpu_vec[$gpu_vec_bools]

let group = addgroup!(group, "findall")
group["bool"] = @benchmarkable findall($gpu_vec_bools)
group["int"] = @benchmarkable findall(isodd, $gpu_vec_ints)
end
# let group = addgroup!(group, "findall")
# group["bool"] = @benchmarkable findall($gpu_vec_bools)
# group["int"] = @benchmarkable findall(isodd, $gpu_vec_ints)
# end

let group = addgroup!(group, "findfirst")
group["bool"] = @benchmarkable findfirst($gpu_vec_bools)
group["int"] = @benchmarkable findfirst(isodd, $gpu_vec_ints)
end
# let group = addgroup!(group, "findfirst")
# group["bool"] = @benchmarkable findfirst($gpu_vec_bools)
# group["int"] = @benchmarkable findfirst(isodd, $gpu_vec_ints)
# end

let group = addgroup!(group, "findmin") # findmax
group["1d"] = @async_benchmarkable findmin($gpu_vec)
group["2d"] = @async_benchmarkable findmin($gpu_mat; dims=1)
end
end
# let group = addgroup!(group, "findmin") # findmax
# group["1d"] = @async_benchmarkable findmin($gpu_vec)
# group["2d"] = @async_benchmarkable findmin($gpu_mat; dims=1)
# end
# end

let group = addgroup!(group, "reverse")
group["1d"] = @async_benchmarkable reverse($gpu_vec)
Expand All @@ -62,94 +62,94 @@ let group = addgroup!(group, "reverse")
group["2dL_inplace"] = @async_benchmarkable reverse!($gpu_mat_long; dims=2)
end

group["broadcast"] = @async_benchmarkable $gpu_mat .= 0f0

# no need to test inplace version, which performs the same operation (but with an alloc)
let group = addgroup!(group, "accumulate")
let group = addgroup!(group, "Float32")
group["1d"] = @async_benchmarkable accumulate(+, $gpu_vec)
group["dims=1"] = @async_benchmarkable accumulate(+, $gpu_mat; dims=1)
group["dims=2"] = @async_benchmarkable accumulate(+, $gpu_mat; dims=2)

group["dims=1L"] = @async_benchmarkable accumulate(+, $gpu_mat_long; dims=1)
group["dims=2L"] = @async_benchmarkable accumulate(+, $gpu_mat_long; dims=2)
end
let group = addgroup!(group, "Int64")
group["1d"] = @async_benchmarkable accumulate(+, $gpu_vec_ints)
group["dims=1"] = @async_benchmarkable accumulate(+, $gpu_mat_ints; dims=1)
group["dims=2"] = @async_benchmarkable accumulate(+, $gpu_mat_ints; dims=2)

group["dims=1L"] = @async_benchmarkable accumulate(+, $gpu_mat_long_ints; dims=1)
group["dims=2L"] = @async_benchmarkable accumulate(+, $gpu_mat_long_ints; dims=2)
end
end

let group = addgroup!(group, "reductions")
let group = addgroup!(group, "reduce")
let group = addgroup!(group, "Float32")
group["1d"] = @async_benchmarkable reduce(+, $gpu_vec)
group["dims=1"] = @async_benchmarkable reduce(+, $gpu_mat; dims=1)
group["dims=2"] = @async_benchmarkable reduce(+, $gpu_mat; dims=2)
group["dims=1L"] = @async_benchmarkable reduce(+, $gpu_mat_long; dims=1)
group["dims=2L"] = @async_benchmarkable reduce(+, $gpu_mat_long; dims=2)
end
let group = addgroup!(group, "Int64")
group["1d"] = @async_benchmarkable reduce(+, $gpu_vec_ints)
group["dims=1"] = @async_benchmarkable reduce(+, $gpu_mat_ints; dims=1)
group["dims=2"] = @async_benchmarkable reduce(+, $gpu_mat_ints; dims=2)
group["dims=1L"] = @async_benchmarkable reduce(+, $gpu_mat_long_ints; dims=1)
group["dims=2L"] = @async_benchmarkable reduce(+, $gpu_mat_long_ints; dims=2)
end
end

let group = addgroup!(group, "mapreduce")
let group = addgroup!(group, "Float32")
group["1d"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_vec)
group["dims=1"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat; dims=1)
group["dims=2"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat; dims=2)
group["dims=1L"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat_long; dims=1)
group["dims=2L"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat_long; dims=2)
end
let group = addgroup!(group, "Int64")
group["1d"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_vec_ints)
group["dims=1"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat_ints; dims=1)
group["dims=2"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat_ints; dims=2)
group["dims=1L"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat_long_ints; dims=1)
group["dims=2L"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat_long_ints; dims=2)
end
end

# used by sum, prod, minimum, maximum, all, any, count
end

let group = addgroup!(group, "random")
let group = addgroup!(group, "rand")
group["Float32"] = @async_benchmarkable CUDA.rand(Float32, m*n)
group["Int64"] = @async_benchmarkable CUDA.rand(Int64, m*n)
end

let group = addgroup!(group, "rand!")
group["Float32"] = @async_benchmarkable CUDA.rand!($gpu_vec)
group["Int64"] = @async_benchmarkable CUDA.rand!($gpu_vec_ints)
end

let group = addgroup!(group, "randn")
group["Float32"] = @async_benchmarkable CUDA.randn(Float32, m*n)
end

let group = addgroup!(group, "randn!")
group["Float32"] = @async_benchmarkable CUDA.randn!($gpu_vec)
end
end

let group = addgroup!(group, "sorting")
group["1d"] = @async_benchmarkable sort($gpu_vec)
group["2d"] = @async_benchmarkable sort($gpu_mat; dims=1)
group["by"] = @async_benchmarkable sort($gpu_vec; by=sin)
end

let group = addgroup!(group, "permutedims")
group["2d"] = @async_benchmarkable permutedims($gpu_mat, (2,1))
group["3d"] = @async_benchmarkable permutedims($gpu_arr_3d, (3,1,2))
group["4d"] = @async_benchmarkable permutedims($gpu_arr_4d, (2,1,4,3))
end
# group["broadcast"] = @async_benchmarkable $gpu_mat .= 0f0

# # no need to test inplace version, which performs the same operation (but with an alloc)
# let group = addgroup!(group, "accumulate")
# let group = addgroup!(group, "Float32")
# group["1d"] = @async_benchmarkable accumulate(+, $gpu_vec)
# group["dims=1"] = @async_benchmarkable accumulate(+, $gpu_mat; dims=1)
# group["dims=2"] = @async_benchmarkable accumulate(+, $gpu_mat; dims=2)

# group["dims=1L"] = @async_benchmarkable accumulate(+, $gpu_mat_long; dims=1)
# group["dims=2L"] = @async_benchmarkable accumulate(+, $gpu_mat_long; dims=2)
# end
# let group = addgroup!(group, "Int64")
# group["1d"] = @async_benchmarkable accumulate(+, $gpu_vec_ints)
# group["dims=1"] = @async_benchmarkable accumulate(+, $gpu_mat_ints; dims=1)
# group["dims=2"] = @async_benchmarkable accumulate(+, $gpu_mat_ints; dims=2)

# group["dims=1L"] = @async_benchmarkable accumulate(+, $gpu_mat_long_ints; dims=1)
# group["dims=2L"] = @async_benchmarkable accumulate(+, $gpu_mat_long_ints; dims=2)
# end
# end

# let group = addgroup!(group, "reductions")
# let group = addgroup!(group, "reduce")
# let group = addgroup!(group, "Float32")
# group["1d"] = @async_benchmarkable reduce(+, $gpu_vec)
# group["dims=1"] = @async_benchmarkable reduce(+, $gpu_mat; dims=1)
# group["dims=2"] = @async_benchmarkable reduce(+, $gpu_mat; dims=2)
# group["dims=1L"] = @async_benchmarkable reduce(+, $gpu_mat_long; dims=1)
# group["dims=2L"] = @async_benchmarkable reduce(+, $gpu_mat_long; dims=2)
# end
# let group = addgroup!(group, "Int64")
# group["1d"] = @async_benchmarkable reduce(+, $gpu_vec_ints)
# group["dims=1"] = @async_benchmarkable reduce(+, $gpu_mat_ints; dims=1)
# group["dims=2"] = @async_benchmarkable reduce(+, $gpu_mat_ints; dims=2)
# group["dims=1L"] = @async_benchmarkable reduce(+, $gpu_mat_long_ints; dims=1)
# group["dims=2L"] = @async_benchmarkable reduce(+, $gpu_mat_long_ints; dims=2)
# end
# end

# let group = addgroup!(group, "mapreduce")
# let group = addgroup!(group, "Float32")
# group["1d"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_vec)
# group["dims=1"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat; dims=1)
# group["dims=2"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat; dims=2)
# group["dims=1L"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat_long; dims=1)
# group["dims=2L"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat_long; dims=2)
# end
# let group = addgroup!(group, "Int64")
# group["1d"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_vec_ints)
# group["dims=1"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat_ints; dims=1)
# group["dims=2"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat_ints; dims=2)
# group["dims=1L"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat_long_ints; dims=1)
# group["dims=2L"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat_long_ints; dims=2)
# end
# end

# # used by sum, prod, minimum, maximum, all, any, count
# end

# let group = addgroup!(group, "random")
# let group = addgroup!(group, "rand")
# group["Float32"] = @async_benchmarkable CUDA.rand(Float32, m*n)
# group["Int64"] = @async_benchmarkable CUDA.rand(Int64, m*n)
# end

# let group = addgroup!(group, "rand!")
# group["Float32"] = @async_benchmarkable CUDA.rand!($gpu_vec)
# group["Int64"] = @async_benchmarkable CUDA.rand!($gpu_vec_ints)
# end

# let group = addgroup!(group, "randn")
# group["Float32"] = @async_benchmarkable CUDA.randn(Float32, m*n)
# end

# let group = addgroup!(group, "randn!")
# group["Float32"] = @async_benchmarkable CUDA.randn!($gpu_vec)
# end
# end

# let group = addgroup!(group, "sorting")
# group["1d"] = @async_benchmarkable sort($gpu_vec)
# group["2d"] = @async_benchmarkable sort($gpu_mat; dims=1)
# group["by"] = @async_benchmarkable sort($gpu_vec; by=sin)
# end

# let group = addgroup!(group, "permutedims")
# group["2d"] = @async_benchmarkable permutedims($gpu_mat, (2,1))
# group["3d"] = @async_benchmarkable permutedims($gpu_arr_3d, (3,1,2))
# group["4d"] = @async_benchmarkable permutedims($gpu_arr_4d, (2,1,4,3))
# end
26 changes: 13 additions & 13 deletions perf/runbenchmarks.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,13 +16,13 @@ end

# before anything else, run latency benchmarks. these spawn subprocesses, so we don't want
# to do so after regular benchmarks have caused the memory allocator to reserve memory.
@info "Running latency benchmarks"
latency_results = include("latency.jl")
# @info "Running latency benchmarks"
# latency_results = include("latency.jl")

SUITE = BenchmarkGroup()

include("cuda.jl")
include("kernel.jl")
# include("cuda.jl")
# include("kernel.jl")
include("array.jl")

@info "Preparing main benchmarks"
Expand All @@ -34,20 +34,20 @@ GC.gc(true)
CUDA.reclaim()

# benchmark groups that aren't part of the suite
addgroup!(SUITE, "integration")
# addgroup!(SUITE, "integration")

@info "Running main benchmarks"
results = run(SUITE, verbose=true)

# integration tests (that do nasty things, so need to be run last)
@info "Running integration benchmarks"
integration_results = BenchmarkGroup()
integration_results["volumerhs"] = include("volumerhs.jl")
integration_results["byval"] = include("byval.jl")
integration_results["cudadevrt"] = include("cudadevrt.jl")

results["latency"] = latency_results
results["integration"] = integration_results
# @info "Running integration benchmarks"
# integration_results = BenchmarkGroup()
# integration_results["volumerhs"] = include("volumerhs.jl")
# integration_results["byval"] = include("byval.jl")
# integration_results["cudadevrt"] = include("cudadevrt.jl")

# results["latency"] = latency_results
# results["integration"] = integration_results

# write out the results
result_file = length(ARGS) >= 1 ? ARGS[1] : "benchmarkresults.json"
Expand Down
1 change: 1 addition & 0 deletions src/CUDA.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@ module CUDA
using GPUCompiler

using GPUArrays
import KernelAbstractions: KernelIntrinsics as KI

using GPUToolbox

Expand Down
Loading