Merge pull request #19639 from pabloferz/pz/inline-broadcast

Reduce allocations in broadcast

Author: stevengj

base/broadcast.jl

@@ -126,7 +126,7 @@ Base.@propagate_inbounds _broadcast_getindex(::Any, A, I) = A[I]
 # of keeps). The first two type parameters are to ensure specialization.
 @generated function _broadcast!{K,ID,AT,nargs}(f, B::AbstractArray, keeps::K, Idefaults::ID, As::AT, ::Type{Val{nargs}}, iter)
     quote
-        $(Expr(:meta, :noinline))
+        $(Expr(:meta, :inline))
         # destructure the keeps and As tuples
         @nexprs $nargs i->(A_i = As[i])
         @nexprs $nargs i->(keep_i = keeps[i])
@@ -147,7 +147,7 @@ end
 # and then copy in chunks into the output
 @generated function _broadcast!{K,ID,AT,nargs}(f, B::BitArray, keeps::K, Idefaults::ID, As::AT, ::Type{Val{nargs}}, iter)
     quote
-        $(Expr(:meta, :noinline))
+        $(Expr(:meta, :inline))
         # destructure the keeps and As tuples
         @nexprs $nargs i->(A_i = As[i])
         @nexprs $nargs i->(keep_i = keeps[i])
@@ -241,9 +241,8 @@ function broadcast_t(f, ::Type{Any}, shape, iter, As...)
     B[I] = val
     return _broadcast!(f, B, keeps, Idefaults, As, Val{nargs}, iter, st, 1)
 end
-@inline function broadcast_t(f, T, shape, iter, As...)
+@inline function broadcast_t{nargs}(f, T, shape, iter, As::Vararg{Any,nargs})
     B = similar(Array{T}, shape)
-    nargs = length(As)
     keeps, Idefaults = map_newindexer(shape, As)
     _broadcast!(f, B, keeps, Idefaults, As, Val{nargs}, iter)
     return B

base/linalg/bidiag.jl

@@ -559,23 +559,30 @@ function eigvecs{T}(M::Bidiagonal{T})
     n = length(M.dv)
     Q = Array{T}(n, n)
     blks = [0; find(x -> x == 0, M.ev); n]
+    v = zeros(T, n)
     if M.isupper
         for idx_block = 1:length(blks) - 1, i = blks[idx_block] + 1:blks[idx_block + 1] #index of eigenvector
-            v=zeros(T, n)
+            fill!(v, zero(T))
             v[blks[idx_block] + 1] = one(T)
             for j = blks[idx_block] + 1:i - 1 #Starting from j=i, eigenvector elements will be 0
                 v[j+1] = (M.dv[i] - M.dv[j])/M.ev[j] * v[j]
             end
-            Q[:, i] = v/norm(v)
+            c = norm(v)
+            for j = 1:n
+                Q[j, i] = v[j] / c
+            end
         end
     else
         for idx_block = 1:length(blks) - 1, i = blks[idx_block + 1]:-1:blks[idx_block] + 1 #index of eigenvector
-            v = zeros(T, n)
+            fill!(v, zero(T))
             v[blks[idx_block+1]] = one(T)
             for j = (blks[idx_block+1] - 1):-1:max(1, (i - 1)) #Starting from j=i, eigenvector elements will be 0
                 v[j] = (M.dv[i] - M.dv[j+1])/M.ev[j] * v[j+1]
             end
-            Q[:,i] = v/norm(v)
+            c = norm(v)
+            for j = 1:n
+                Q[j, i] = v[j] / c
+            end
         end
     end
     Q #Actually Triangular

base/linalg/diagonal.jl

@@ -338,9 +338,9 @@ svdvals(D::Diagonal) = [svdvals(v) for v in D.diag]
 function svd{T<:Number}(D::Diagonal{T})
     S   = abs.(D.diag)
     piv = sortperm(S, rev = true)
-    U   = Array(Diagonal(D.diag ./ S))
+    U   = Diagonal(D.diag ./ S)
     Up  = hcat([U[:,i] for i = 1:length(D.diag)][piv]...)
-    V   = eye(D)
+    V   = Diagonal(ones(D.diag))
     Vp  = hcat([V[:,i] for i = 1:length(D.diag)][piv]...)
     return (Up, S[piv], Vp)
 end