From 3c98677376fb33f11fe2f786a49bc97b07743711 Mon Sep 17 00:00:00 2001
From: Benoit Jacob <jacob.benoit.1@gmail.com>
Date: Wed, 10 Oct 2007 06:33:09 +0000
Subject: fix big bug in loop unrolling

---
 src/Core/Dot.h       | 10 +++++-----
 src/Core/MatrixOps.h | 18 +++++-------------
 src/Core/Trace.h     | 13 ++++---------
 3 files changed, 14 insertions(+), 27 deletions(-)

diff --git a/src/Core/Dot.h b/src/Core/Dot.h
index 0d000ea70..ea9653f8a 100644
--- a/src/Core/Dot.h
+++ b/src/Core/Dot.h
@@ -32,12 +32,12 @@ struct EiDotUnroller
   static void run(const Derived1 &v1, const Derived2& v2, typename Derived1::Scalar &dot)
   {
     EiDotUnroller<Index-1, Size, Derived1, Derived2>::run(v1, v2, dot);
-    dot += v1[Index-1] * EiConj(v2[Index-1]);
+    dot += v1[Index] * EiConj(v2[Index]);
   }
 };
 
 template<int Size, typename Derived1, typename Derived2>
-struct EiDotUnroller<1, Size, Derived1, Derived2>
+struct EiDotUnroller<0, Size, Derived1, Derived2>
 {
   static void run(const Derived1 &v1, const Derived2& v2, typename Derived1::Scalar &dot)
   {
@@ -45,8 +45,8 @@ struct EiDotUnroller<1, Size, Derived1, Derived2>
   }
 };
 
-template<int Size, typename Derived1, typename Derived2>
-struct EiDotUnroller<EiDynamic, Size, Derived1, Derived2>
+template<int Index, typename Derived1, typename Derived2>
+struct EiDotUnroller<Index, EiDynamic, Derived1, Derived2>
 {
   static void run(const Derived1 &v1, const Derived2& v2, typename Derived1::Scalar &dot)
   {
@@ -63,7 +63,7 @@ Scalar EiObject<Scalar, Derived>::dot(const OtherDerived& other) const
   assert(IsVector && OtherDerived::IsVector && size() == other.size());
   Scalar res;
   if(SizeAtCompileTime != EiDynamic && SizeAtCompileTime <= 16)
-    EiDotUnroller<SizeAtCompileTime, SizeAtCompileTime, Derived, OtherDerived>
+    EiDotUnroller<SizeAtCompileTime-1, SizeAtCompileTime, Derived, OtherDerived>
       ::run(*static_cast<const Derived*>(this), other, res);
   else
   {
diff --git a/src/Core/MatrixOps.h b/src/Core/MatrixOps.h
index eb1682ab0..1449a3923 100644
--- a/src/Core/MatrixOps.h
+++ b/src/Core/MatrixOps.h
@@ -111,12 +111,8 @@ struct EiMatrixProductUnroller
   static void run(int row, int col, const Lhs& lhs, const Rhs& rhs,
                   typename Lhs::Scalar &res)
   {
-    const int i = Index - 1;
     EiMatrixProductUnroller<Index-1, Size, Lhs, Rhs>::run(row, col, lhs, rhs, res);
-    if(i == Size - 1)
-      res = lhs.read(row, i) * rhs.read(i, col);
-    else
-      res += lhs.read(row, i) * rhs.read(i, col);
+    res += lhs.read(row, Index) * rhs.read(Index, col);
   }
 };
 
@@ -126,16 +122,12 @@ struct EiMatrixProductUnroller<0, Size, Lhs, Rhs>
   static void run(int row, int col, const Lhs& lhs, const Rhs& rhs,
                   typename Lhs::Scalar &res)
   {
-    EI_UNUSED(row);
-    EI_UNUSED(col);
-    EI_UNUSED(lhs);
-    EI_UNUSED(rhs);
-    EI_UNUSED(res);
+    res = lhs.read(row, 0) * rhs.read(0, col);
   }
 };
 
-template<int Size, typename Lhs, typename Rhs>
-struct EiMatrixProductUnroller<EiDynamic, Size, Lhs, Rhs>
+template<int Index, typename Lhs, typename Rhs>
+struct EiMatrixProductUnroller<Index, EiDynamic, Lhs, Rhs>
 {
   static void run(int row, int col, const Lhs& lhs, const Rhs& rhs,
                   typename Lhs::Scalar &res)
@@ -181,7 +173,7 @@ template<typename Lhs, typename Rhs> class EiMatrixProduct
     {
       Scalar res;
       if(Lhs::ColsAtCompileTime != EiDynamic && Lhs::ColsAtCompileTime <= 16)
-        EiMatrixProductUnroller<Lhs::ColsAtCompileTime, Lhs::ColsAtCompileTime, LhsRef, RhsRef>
+        EiMatrixProductUnroller<Lhs::ColsAtCompileTime-1, Lhs::ColsAtCompileTime, LhsRef, RhsRef>
           ::run(row, col, m_lhs, m_rhs, res);
       else
       {
diff --git a/src/Core/Trace.h b/src/Core/Trace.h
index df62118e1..80c5dbed1 100644
--- a/src/Core/Trace.h
+++ b/src/Core/Trace.h
@@ -30,12 +30,8 @@ template<int Index, int Rows, typename Derived> struct EiTraceUnroller
 {
   static void run(const Derived &mat, typename Derived::Scalar &trace)
   {
-    const int i = Index - 1;
     EiTraceUnroller<Index-1, Rows, Derived>::run(mat, trace);
-    if(i == Rows - 1)
-      trace = mat(i, i);
-    else
-      trace += mat(i, i);
+    trace += mat(Index, Index);
   }
 };
 
@@ -43,12 +39,11 @@ template<int Rows, typename Derived> struct EiTraceUnroller<0, Rows, Derived>
 {
   static void run(const Derived &mat, typename Derived::Scalar &trace)
   {
-    EI_UNUSED(mat);
-    EI_UNUSED(trace);
+    trace = mat(0, 0);
   }
 };
 
-template<int Rows, typename Derived> struct EiTraceUnroller<EiDynamic, Rows, Derived>
+template<int Index, typename Derived> struct EiTraceUnroller<Index, EiDynamic, Derived>
 {
   static void run(const Derived &mat, typename Derived::Scalar &trace)
   {
@@ -63,7 +58,7 @@ Scalar EiObject<Scalar, Derived>::trace() const
   assert(rows() == cols());
   Scalar res;
   if(RowsAtCompileTime != EiDynamic && RowsAtCompileTime <= 16)
-    EiTraceUnroller<RowsAtCompileTime, RowsAtCompileTime, Derived>
+    EiTraceUnroller<RowsAtCompileTime-1, RowsAtCompileTime, Derived>
       ::run(*static_cast<const Derived*>(this), res);
   else
   {
-- 
cgit v1.2.3