test tweaks, use single precision in choose_num_threads

Author: chriselrod

src/codegen/lower_threads.jl

@@ -165,11 +165,17 @@ end
 #     block_per_m, blocks_per_n
 # end
 if Sys.ARCH === :x86_64
-    @inline choose_num_threads(C::Float64, NT::UInt, x::Base.BitInteger) = _choose_num_threads(Base.mul_float_fast(C, 0.05460264079015985), NT, x)
+  @inline function choose_num_threads(C::T, NT::UInt, x::Base.BitInteger) where {T<:Union{Float32,Float64}}
+    _choose_num_threads(Base.mul_float_fast(T(C), T(0.05460264079015985)), NT, x)
+  end
 else
-    @inline choose_num_threads(C::Float64, NT::UInt, x::Base.BitInteger) = _choose_num_threads(Base.mul_float_fast(C, 0.05460264079015985 * 0.25), NT, x)
+  @inline function choose_num_threads(C::T, NT::UInt, x::Base.BitInteger) where {T<:Union{Float32,Float64}}
+    _choose_num_threads(Base.mul_float_fast(C, T(0.05460264079015985) * T(0.25)), NT, x)
+  end
+end
+@inline function _choose_num_threads(C::T, NT::UInt, x::Base.BitInteger) where {T<:Union{Float32,Float64}}
+  min(Base.fptoui(UInt, Base.ceil_llvm(Base.mul_float_fast(C, Base.sqrt_llvm_fast(Base.uitofp(T, x))))), NT)
 end
-@inline _choose_num_threads(C::Float64, NT::UInt, x::Base.BitInteger) = min(Base.fptoui(UInt, Base.ceil_llvm(Base.mul_float_fast(C, Base.sqrt_llvm(Base.uitofp(Float64, x))))), NT)
 function push_loop_length_expr!(q::Expr, ls::LoopSet)
     l = 1
     ndynamic = 0
@@ -342,7 +348,7 @@ function thread_one_loops_expr(
     _num_threads > 1 || return avx_body(ls, UNROLL)
     choose_nthread = Expr(:(=), Symbol("#nthreads#"), _num_threads)
   else
-    choose_nthread = :(_choose_num_threads($c, $ntmax))
+    choose_nthread = :(_choose_num_threads($(Float32(c)), $ntmax))
     push_loop_length_expr!(choose_nthread, ls)
     choose_nthread = Expr(:(=), Symbol("#nthreads#"), choose_nthread)
   end
@@ -474,7 +480,7 @@ function thread_two_loops_expr(
     _num_threads > 1 || return avx_body(ls, UNROLL)
     choose_nthread = Expr(:(=), Symbol("#nthreads#"), _num_threads)
   else
-    choose_nthread = :(_choose_num_threads($c, $ntmax))
+    choose_nthread = :(_choose_num_threads($(Float32(c)), $ntmax))
     push_loop_length_expr!(choose_nthread, ls)
     choose_nthread = Expr(:(=), Symbol("#nthreads#"), choose_nthread)
   end

test/outer_reductions.jl

@@ -1,5 +1,5 @@
 
-function awmean_lv(x::Array{T1}, σ::Array{T2}) where {T1<:Number,T2<:Number}
+function awmean_lv(x::AbstractArray{T1}, σ::AbstractArray{T2}) where {T1<:Number,T2<:Number}
   n = length(x)
   T3 = promote_type(T1,T2)
   T = sizeof(T3) ≤ 4 ? Float32 : Float64
@@ -17,7 +17,7 @@ function awmean_lv(x::Array{T1}, σ::Array{T2}) where {T1<:Number,T2<:Number}
   wσ = sqrt(one(T) / sum_of_weights)
   return wx, wσ, mswd
 end
-function awmean_simd(x::Array{T1}, σ::Array{T2}) where {T1<:Number,T2<:Number}
+function awmean_simd(x::AbstractArray{T1}, σ::AbstractArray{T2}) where {T1<:Number,T2<:Number}
   n = length(x)
   T3 = promote_type(T1,T2)
   T = sizeof(T3) ≤ 4 ? Float32 : Float64
@@ -39,18 +39,18 @@ end
 function test_awmean(::Type{T}) where {T}
     for n ∈ 2:100
       if T <: Integer
-        x = rand(T(-100):T(100), n)
-        σ = rand(T(1):T(10), n)
+        x = view(rand(T(-100):T(100), n + 32), 17:n+16)
+        σ = view(rand(T(1):T(10), n + 32), 17:n+16)
       else
-        x = randn(T, n)
-        σ = rand(T, n)
+        x = view(randn(T, n + 32), 17:n+16)
+        σ = view(rand(T, n + 32), 17:n+16)
       end
       wx, wσ, mswd = awmean_simd(x, σ)
       @test iszero(@allocated((wxlv, wσlv, mswdlv) = awmean_lv(x, σ)))
       wxlv, wσlv, mswdlv = awmean_lv(x, σ)
-      @test wx ≈ wxlv
-      @test wσ ≈ wσlv
-      @test mswd ≈ mswdlv
+      isfinite(wx)   && @test wx ≈ wxlv
+      isfinite(wσ)   && @test wσ ≈ wσlv
+      isfinite(mswd) && @test mswd ≈ mswdlv
     end
 end