Fix type instability with static sized arrays (#54)

* Fix type instability with static sized arrays

* Bump patch version

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "Octavian"
 uuid = "6fd5a793-0b7e-452c-907f-f8bfe9c57db4"
 authors = ["Mason Protter", "Chris Elrod", "Dilum Aluthge", "contributors"]
-version = "0.2.2"
+version = "0.2.3"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

src/block_sizes.jl

@@ -223,7 +223,7 @@ Splits both `M` and `N` into blocks when trying to spawn a large number of threa
 
     Miter = clamp(div_fast(M, W*StaticInt{mᵣ}() * MᵣW_mul_factor), 1, _nspawn)
     nspawn = div_fast(_nspawn, Miter)
-    Niter = if (nspawn ≤ 1) & (Miter < _nspawn)
+    if (nspawn ≤ 1) & (Miter < _nspawn)
         # rebalance Miter
         Miter = cld_fast(_nspawn, cld_fast(_nspawn, Miter))
         nspawn = div_fast(_nspawn, Miter)

src/integerdivision.jl

@@ -1,9 +1,10 @@
 @inline cld_fast(x, y) = cld(x, y)
-@inline function cld_fast(x::I, y) where {I <: Integer}
+@inline function cld_fast(x::I, y) where {I <: Base.BitInteger}
     d = div_fast(x, y)
     (d + (d * unsigned(y) != unsigned(x))) % I
 end
-cld_fast(::StaticInt{N}, ::StaticInt{M}) where {N,M}= (StaticInt{N}() + StaticInt{M}() + One()) ÷ StaticInt{M}()
+@inline cld_fast(::StaticInt{N}, y) where {N} = cld_fast(N, y)
+cld_fast(::StaticInt{N}, ::StaticInt{M}) where {N,M}= (StaticInt{N}() + StaticInt{M}() - One()) ÷ StaticInt{M}()
 @inline function divrem_fast(x::I, y) where {I <: Integer}
     ux = unsigned(x); uy = unsigned(y)
     d = Base.udiv_int(ux, uy)

test/integer_division.jl

@@ -0,0 +1,5 @@
+
+@test @inferred(Octavian.cld_fast(Octavian.StaticInt(6), Octavian.StaticInt(2))) == @inferred(Octavian.cld_fast(11, Octavian.StaticInt(4))) == @inferred(Octavian.cld_fast(Octavian.StaticInt(7), 3)) == @inferred(Octavian.cld_fast(9, 4)) == 3
+@test @inferred(Octavian.cld_fast(Octavian.StaticInt(7), Octavian.StaticInt(2))) == @inferred(Octavian.cld_fast(13, Octavian.StaticInt(4))) == @inferred(Octavian.cld_fast(Octavian.StaticInt(8), 2)) == @inferred(Octavian.cld_fast(15, 4)) == 4
+
+

test/runtests.jl

@@ -9,7 +9,7 @@ import Random
 import Test
 import VectorizationBase
 
-using Test: @testset, @test, @test_throws
+using Test: @testset, @test, @test_throws, @inferred
 
 include("test_suite_preamble.jl")
 
@@ -20,6 +20,7 @@ Random.seed!(123)
 
 include("block_sizes.jl")
 include("init.jl")
+include("integer_division.jl")
 include("macrokernels.jl")
 include("matmul_coverage.jl")
 include("utils.jl")