From d7cd0235259128a6eef8aa8087ce43a0aa2dc86e Mon Sep 17 00:00:00 2001
From: Carlo Baldassi <carlobaldassi@gmail.com>
Date: Thu, 12 Nov 2015 00:09:30 -0500
Subject: [PATCH] Speed up BitArray packing

Changes in convert(::BitArray, ::AbstractArray) and
dumpbitcache():
* avoid branches in for loops
* pack 8 Bools at a time
* use inbounds
---
 base/bitarray.jl  |  8 ++------
 base/broadcast.jl | 28 +++++++++++++++++++++++-----
 2 files changed, 25 insertions(+), 11 deletions(-)

diff --git a/base/bitarray.jl b/base/bitarray.jl
index 845b040ce7498..83194d470dcef 100644
--- a/base/bitarray.jl
+++ b/base/bitarray.jl
@@ -310,21 +310,17 @@ function convert{T,N}(::Type{BitArray{N}}, A::AbstractArray{T,N})
     ind = 1
     @inbounds begin
         for i = 1:length(Bc)-1
-            u = UInt64(1)
             c = UInt64(0)
             for j = 0:63
-                A[ind]!=0 && (c |= u)
+                c |= (UInt64(A[ind] != 0) << j)
                 ind += 1
-                u <<= 1
             end
             Bc[i] = c
         end
-        u = UInt64(1)
         c = UInt64(0)
         for j = 0:_mod64(l-1)
-            A[ind]!=0 && (c |= u)
+            c |= (UInt64(A[ind] != 0) << j)
             ind += 1
-            u <<= 1
         end
         Bc[end] = c
     end
diff --git a/base/broadcast.jl b/base/broadcast.jl
index 32666ed6a32c8..efb0eb1173c5b 100644
--- a/base/broadcast.jl
+++ b/base/broadcast.jl
@@ -103,16 +103,34 @@ end
 const bitcache_chunks = 64 # this can be changed
 const bitcache_size = 64 * bitcache_chunks # do not change this
 
+function bpack(z::UInt64)
+    z |= z >>> 7
+    z |= z >>> 14
+    z |= z >>> 28
+    z &= 0xFF
+    return z
+end
+
 function dumpbitcache(Bc::Vector{UInt64}, bind::Int, C::Vector{Bool})
     ind = 1
     nc = min(bitcache_chunks, length(Bc)-bind+1)
-    for i = 1:nc
-        u = UInt64(1)
+    C8 = reinterpret(UInt64, C)
+    nc8 = (nc >>> 3) << 3
+    @inbounds for i = 1:nc8
+        c = UInt64(0)
+        for j = 0:8:63
+            c |= (bpack(C8[ind]) << j)
+            ind += 1
+        end
+        Bc[bind] = c
+        bind += 1
+    end
+    ind = (ind-1) << 3 + 1
+    @inbounds for i = (nc8+1):nc
         c = UInt64(0)
-        for j = 1:64
-            C[ind] && (c |= u)
+        for j = 0:63
+            c |= (UInt64(C[ind]) << j)
             ind += 1
-            u <<= 1
         end
         Bc[bind] = c
         bind += 1