Skip to content

Remove Metal warning, align default tol, and improve Metal gather #91

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

Merged
eloualiche merged 7 commits into from
Nov 3, 2026
Merged

Remove Metal warning, align default tol, and improve Metal gather #91

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
473e4ff
Remove Metal.allowscalar call that warns at precompile
matthieugomez Nov 2, 2026
01094a2
Update Metal timings in benchmark comments
matthieugomez Nov 2, 2026
d0eca8f
Use the documented default tol in prebuilt solver methods
matthieugomez Nov 2, 2026
ae2f491
Run benchmarks at tol = 1e-6
matthieugomez Nov 2, 2026
9787160
Choose the Metal gather from group sizes and row order
matthieugomez Nov 2, 2026
56bd172
Use the fused scatter kernel for the forward Metal mul!
matthieugomez Nov 3, 2026
1c3ffc8
Remove CUDA.allowscalar call at module top level
matthieugomez Nov 3, 2026
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
2 changes: 1 addition & 1 deletion Project.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
name = "FixedEffects"
uuid = "c8885935-8500-56a7-9867-7708b20db0eb"
version = "3.5.0"
version = "3.5.1"

[deps]
GroupedArrays = "6407cd72-fade-4a84-8a1e-56e431fc1533"
Expand Down
64 changes: 32 additions & 32 deletions benchmarks/benchmark.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,8 +2,8 @@
# Benchmarks for solve_residuals! across backends (:cpu, :CUDA, :Metal).
# Problems are defined once; every available backend runs the same
# specifications. Each specification residualizes 6 columns through one
# solver, which is how FixedEffectModels calls this package (first run
# includes compilation).
# solver at tol = 1e-6, which is how FixedEffectModels calls this package by
# default (first run includes compilation).
# Usage: julia --project -t auto benchmarks/benchmark.jl
##############################################################################

Expand Down Expand Up @@ -64,33 +64,33 @@ println("="^60)

feM = AbstractFixedEffectSolver{Float64}(fes_simple, uweights(length(cols_simple[1])), Val{:cpu})
println("Simple (N=10M, 100k×100), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_simple], feM; progress_bar = false) # ~1.1 s
@time solve_residuals!([copy(c) for c in cols_simple], feM; progress_bar = false, tol = 1e-6) # ~0.73 s
println("Simple (N=10M, 100k×100), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_simple], feM; progress_bar = false) # ~0.8 s
@time solve_residuals!([copy(c) for c in cols_simple], feM; progress_bar = false, tol = 1e-6) # ~0.60 s

feM = AbstractFixedEffectSolver{Float64}(fes_hard, uweights(N), Val{:cpu})
println("Hard (N=800k, 40k×5k), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false) # ~13.6 s
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6) # ~3.0 s
println("Hard (N=800k, 40k×5k), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false) # ~13.8 s
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6) # ~3.1 s

feM = AbstractFixedEffectSolver{Float64}(fes_hard_interact, uweights(N), Val{:cpu})
println("Hard (N=800k, interacted), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false) # ~22 s
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6) # ~6.1 s
println("Hard (N=800k, interacted), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false) # ~26 s
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6) # ~5.4 s

feM = AbstractFixedEffectSolver{Float64}(fes_hard_3way, uweights(N), Val{:cpu})
println("Hard 3-way (N=800k, 40k×5k×36), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false) # ~20.6 s
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6) # ~4.1 s
println("Hard 3-way (N=800k, 40k×5k×36), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false) # ~19.9 s
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6) # ~4.1 s

feM = AbstractFixedEffectSolver{Float64}(fes_large, uweights(N_large), Val{:cpu})
println("Large (N=10M, 500k×50k, sorted layout), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_large], feM; progress_bar = false) # ~4.4 s
@time solve_residuals!([copy(c) for c in cols_large], feM; progress_bar = false, tol = 1e-6) # ~2.7 s
println("Large (N=10M, 500k×50k, sorted layout), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_large], feM; progress_bar = false) # ~4.1 s
@time solve_residuals!([copy(c) for c in cols_large], feM; progress_bar = false, tol = 1e-6) # ~2.7 s

##############################################################################
# Metal
Expand All @@ -103,33 +103,33 @@ if isdefined(Main, :Metal) && Metal.functional()

feM = AbstractFixedEffectSolver{Float32}(fes_simple, uweights(length(cols_simple[1])), Val{:Metal})
println("Simple (N=10M, 100k×100), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_simple], feM; progress_bar = false) # ~9.9 s
@time solve_residuals!([copy(c) for c in cols_simple], feM; progress_bar = false, tol = 1e-6) # ~4.3 s
println("Simple (N=10M, 100k×100), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_simple], feM; progress_bar = false) # ~0.38 s
@time solve_residuals!([copy(c) for c in cols_simple], feM; progress_bar = false, tol = 1e-6) # ~0.34 s

feM = AbstractFixedEffectSolver{Float32}(fes_hard, uweights(N), Val{:Metal})
println("Hard (N=800k, 40k×5k), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false) # ~3.5 s
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6) # ~2.8 s
println("Hard (N=800k, 40k×5k), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false) # ~0.75 s
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6) # ~2.5 s

feM = AbstractFixedEffectSolver{Float32}(fes_hard_interact, uweights(N), Val{:Metal})
println("Hard (N=800k, interacted), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false) # ~3.2 s
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6) # ~4.2 s
println("Hard (N=800k, interacted), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false) # ~1.3 s
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6) # ~3.9 s

feM = AbstractFixedEffectSolver{Float32}(fes_hard_3way, uweights(N), Val{:Metal})
println("Hard 3-way (N=800k, 40k×5k×36), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false) # ~1.6 s
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6) # ~3.6 s
println("Hard 3-way (N=800k, 40k×5k×36), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false) # ~0.95 s
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6) # ~3.5 s

feM = AbstractFixedEffectSolver{Float32}(fes_large, uweights(N_large), Val{:Metal})
println("Large (N=10M, 500k×50k, sorted layout), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_large], feM; progress_bar = false) # ~2.8 s
@time solve_residuals!([copy(c) for c in cols_large], feM; progress_bar = false, tol = 1e-6) # ~1.5 s
println("Large (N=10M, 500k×50k, sorted layout), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_large], feM; progress_bar = false) # ~1.0 s
@time solve_residuals!([copy(c) for c in cols_large], feM; progress_bar = false, tol = 1e-6) # ~1.3 s
end

##############################################################################
Expand All @@ -143,31 +143,31 @@ if isdefined(Main, :CUDA) && CUDA.functional()

feM = AbstractFixedEffectSolver{Float32}(fes_simple, uweights(length(cols_simple[1])), Val{:CUDA})
println("Simple (N=10M, 100k×100), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_simple], feM; progress_bar = false)
@time solve_residuals!([copy(c) for c in cols_simple], feM; progress_bar = false, tol = 1e-6)
println("Simple (N=10M, 100k×100), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_simple], feM; progress_bar = false)
@time solve_residuals!([copy(c) for c in cols_simple], feM; progress_bar = false, tol = 1e-6)

feM = AbstractFixedEffectSolver{Float32}(fes_hard, uweights(N), Val{:CUDA})
println("Hard (N=800k, 40k×5k), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false)
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6)
println("Hard (N=800k, 40k×5k), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false)
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6)

feM = AbstractFixedEffectSolver{Float32}(fes_hard_interact, uweights(N), Val{:CUDA})
println("Hard (N=800k, interacted), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false)
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6)
println("Hard (N=800k, interacted), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false)
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6)

feM = AbstractFixedEffectSolver{Float32}(fes_hard_3way, uweights(N), Val{:CUDA})
println("Hard 3-way (N=800k, 40k×5k×36), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false)
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6)
println("Hard 3-way (N=800k, 40k×5k×36), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false)
@time solve_residuals!([copy(c) for c in cols_hard], feM; progress_bar = false, tol = 1e-6)

feM = AbstractFixedEffectSolver{Float32}(fes_large, uweights(N_large), Val{:CUDA})
println("Large (N=10M, 500k×50k, sorted layout), 6 columns, first run:")
@time solve_residuals!([copy(c) for c in cols_large], feM; progress_bar = false)
@time solve_residuals!([copy(c) for c in cols_large], feM; progress_bar = false, tol = 1e-6)
println("Large (N=10M, 500k×50k, sorted layout), 6 columns, second run:")
@time solve_residuals!([copy(c) for c in cols_large], feM; progress_bar = false)
@time solve_residuals!([copy(c) for c in cols_large], feM; progress_bar = false, tol = 1e-6)
end
1 change: 0 additions & 1 deletion ext/CUDAExt.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,6 @@
module CUDAExt
using FixedEffects, CUDA
using FixedEffects: FixedEffectCoefficients, AbstractWeights, UnitWeights, LinearAlgebra, Adjoint, mul!, rmul!, AbstractFixedEffectLinearMap, copy_internal!, AbsorptionPlan, AbsorbedBlock, _group_permutation, block_width
CUDA.allowscalar(false)

##############################################################################
##
Expand Down
Loading
Loading