From 03d19f3bae65d5dff1417cc4bd4ce32262cdaa3a Mon Sep 17 00:00:00 2001
From: Benoit Jacob <jacob.benoit.1@gmail.com>
Date: Mon, 23 Jun 2008 11:23:05 +0000
Subject: [PATCH] quick temporary fix for a perf issue we just identified with
 vectorization.... now the sum benchmark runs 3x faster with vectorization
 than without.

---
 Eigen/src/Core/Sum.h | 25 ++++++++++++++++++++-----
 1 file changed, 20 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Core/Sum.h b/Eigen/src/Core/Sum.h
index 05a8722dc..c6b7cacce 100644
--- a/Eigen/src/Core/Sum.h
+++ b/Eigen/src/Core/Sum.h
@@ -194,17 +194,32 @@ struct ei_sum_impl<Derived, LinearVectorization, NoUnrolling>
     // do the vectorizable part of the sum
     if(size >= packetSize)
     {
+    asm("#begin");
+
       PacketScalar packet_res;
       packet_res = mat.template packet<Aligned>(0, 0);
       int index;
-      for(index = packetSize; index<alignedSize ; index+=packetSize)
+      if(Derived::IsVectorAtCompileTime)
       {
-        // FIXME the following is not really efficient
-        const int row = rowMajor ? index/innerSize : index%innerSize;
-        const int col = rowMajor ? index%innerSize : index/innerSize;
-        packet_res = ei_padd(packet_res, mat.template packet<Aligned>(row, col));
+        for(index = packetSize; index<alignedSize ; index+=packetSize)
+        {
+          const int row = Derived::RowsAtCompileTime==1 ? 0 : index;
+          const int col = Derived::RowsAtCompileTime==1 ? index : 0;
+          packet_res = ei_padd(packet_res, mat.template packet<Aligned>(row, col));
+        }
+      }
+      else
+      {
+        for(index = packetSize; index<alignedSize ; index+=packetSize)
+        {
+          // FIXME the following is not really efficient
+          const int row = rowMajor ? index/innerSize : index%innerSize;
+          const int col = rowMajor ? index%innerSize : index/innerSize;
+          packet_res = ei_padd(packet_res, mat.template packet<Aligned>(row, col));
+        }
       }
       res = ei_predux(packet_res);
+    asm("#end");
 
       // now we must do the rest without vectorization.
       if(alignedSize == size) return res;