Unify gemm and lazy_gemm benchmarks

bench/perf_monitoring/gemm/changesets.txt

@@ -42,4 +42,6 @@ before-evaluators
 6984:45f26866c091   # rm dynamic loop swapping, adjust lhs's micro panel height to fully exploit L1 cache
 6986:a675d05b6f8f   # blocking heuristic: block on the rhs in L1 if the lhs fit in L1.
 7013:f875e75f07e5   # organize a little our default cache sizes, and use a saner default L1 outside of x86 (10% faster on Nexus 5)
+7591:09a8e2186610   # 3.3-alpha1
+7650:b0f3c8f43025   # help clang inlining
 

bench/perf_monitoring/gemm/gemm.cpp

@@ -53,7 +53,7 @@ int main(int argc, char **argv)
 {
   std::vector<double> results;
   
-  std::ifstream settings("settings.txt");
+  std::ifstream settings("gemm_settings.txt");
   long m, n, k;
   while(settings >> m >> n >> k)
   {

bench/perf_monitoring/gemm/lazy_gemm.cpp

@@ -0,0 +1,97 @@
+#include <iostream>
+#include <fstream>
+#include <vector>
+#include <Eigen/Core>
+#include "../../BenchTimer.h"
+using namespace Eigen;
+
+#ifndef SCALAR
+#error SCALAR must be defined
+#endif
+
+typedef SCALAR Scalar;
+
+template<typename MatA, typename MatB, typename MatC>
+inline void lazy_gemm(const MatA &A, const MatB &B, MatC &C)
+{
+  escape((void*)A.data());
+  escape((void*)B.data());
+  C.noalias() += A.lazyProduct(B);
+  escape((void*)C.data());
+}
+
+template<int m, int n, int k, int TA>
+EIGEN_DONT_INLINE
+double bench()
+{
+  typedef Matrix<Scalar,m,k,TA> MatA;
+  typedef Matrix<Scalar,k,n> MatB;
+  typedef Matrix<Scalar,m,n> MatC;
+
+  MatA A(m,k);
+  MatB B(k,n);
+  MatC C(m,n);
+  A.setRandom();
+  B.setRandom();
+  C.setZero();
+
+  BenchTimer t;
+
+  double up = 1e7*4/sizeof(Scalar);
+  double tm0 = 10, tm1 = 20;
+
+  double flops = 2. * m * n * k;
+  long rep = std::max(10., std::min(10000., up/flops) );
+  long tries = std::max(tm0, std::min(tm1, up/flops) );
+
+  BENCH(t, tries, rep, lazy_gemm(A,B,C));
+
+  return 1e-9 * rep * flops / t.best();
+}
+
+template<int m, int n, int k>
+double bench_t(int t)
+{
+  if(t)
+    return bench<m,n,k,RowMajor>();
+  else
+    return bench<m,n,k,0>();
+}
+
+EIGEN_DONT_INLINE
+double bench_mnk(int m, int n, int k, int t)
+{
+  int id = m*10000 + n*100 + k;
+  switch(id) {
+    case  10101 : return bench_t< 1, 1, 1>(t); break;
+    case  20202 : return bench_t< 2, 2, 2>(t); break;
+    case  30303 : return bench_t< 3, 3, 3>(t); break;
+    case  40404 : return bench_t< 4, 4, 4>(t); break;
+    case  50505 : return bench_t< 5, 5, 5>(t); break;
+    case  60606 : return bench_t< 6, 6, 6>(t); break;
+    case  70707 : return bench_t< 7, 7, 7>(t); break;
+    case  80808 : return bench_t< 8, 8, 8>(t); break;
+    case  90909 : return bench_t< 9, 9, 9>(t); break;
+    case 101010 : return bench_t<10,10,10>(t); break;
+    case 111111 : return bench_t<11,11,11>(t); break;
+    case 121212 : return bench_t<12,12,12>(t); break;
+  }
+  return 0;
+}
+
+int main(int argc, char **argv)
+{
+  std::vector<double> results;
+  
+  std::ifstream settings("lazy_gemm_settings.txt");
+  long m, n, k, t;
+  while(settings >> m >> n >> k >> t)
+  {
+    //std::cerr << "  Testing " << m << " " << n << " " << k << std::endl;
+    results.push_back( bench_mnk(m, n, k, t) );
+  }
+  
+  std::cout << RowVectorXd::Map(results.data(), results.size());
+  
+  return 0;
+}

bench/perf_monitoring/gemm/lazy_gemm_settings.txt

@@ -0,0 +1,15 @@
+1 1 1 0
+2 2 2 0
+3 3 3 0
+4 4 4 0
+4 4 4 1
+5 5 5 0
+6 6 6 0
+7 7 7 0
+7 7 7 1
+8 8 8 0
+9 9 9 0
+10 10 10 0
+11 11 11 0
+12 12 12 0
+12 12 12 1

bench/perf_monitoring/gemm/make_plot.sh

@@ -4,6 +4,7 @@
 # it reads $1.out
 # and generates $1.pdf
 WHAT=$1
+bench=$2
 
 header="rev "
 while read line
@@ -11,7 +12,7 @@ do
   if [ ! -z '$line' ]; then
     header="$header  \"$line\""
   fi
-done < settings.txt
+done < $bench"_settings.txt"
 
 echo $header > $WHAT.out.header
 cat $WHAT.out >> $WHAT.out.header

bench/perf_monitoring/gemm/run.sh

@@ -1,5 +1,8 @@
 #!/bin/bash
 
+# ./run.sh gemm
+# ./run.sh lazy_gemm
+
 # Examples of environment variables to be set:
 #   PREFIX="haswell-fma-"
 #   CXX_FLAGS="-mfma"
@@ -8,6 +11,7 @@
 #   -up : enforce the recomputation of existing data, and keep best results as a merging strategy
 #   -s  : recompute selected changesets only and keep bests
 
+bench=$1
 
 if echo "$*" | grep '\-up' > /dev/null; then
   update=true
@@ -84,7 +88,7 @@ function test_current
   fi
   res=$prev
   count_rev=`echo $prev |  wc -w`
-  count_ref=`cat "settings.txt" |  wc -l`
+  count_ref=`cat $bench"_settings.txt" |  wc -l`
   if echo "$global_args" | grep "$rev" > /dev/null; then
     rev_found=true
   else
@@ -93,7 +97,7 @@ function test_current
 #  echo $update et $selected et $rev_found because $rev et "$global_args"
 #  echo $count_rev et $count_ref
   if [ $update == true ] || [ $count_rev != $count_ref ] || ([ $selected == true ] &&  [ $rev_found == true ]); then
-    if $CXX -O2 -DNDEBUG -march=native $CXX_FLAGS -I eigen_src gemm.cpp -DSCALAR=$scalar -o $name; then
+    if $CXX -O2 -DNDEBUG -march=native $CXX_FLAGS -I eigen_src $bench.cpp -DSCALAR=$scalar -o $name; then
       curr=`./$name`
       if [ $count_rev == $count_ref ]; then
         echo "merge previous $prev"
@@ -113,9 +117,9 @@ function test_current
   fi
 }
 
-make_backup $PREFIX"sgemm"
-make_backup $PREFIX"dgemm"
-make_backup $PREFIX"cgemm"
+make_backup $PREFIX"s"$bench
+make_backup $PREFIX"d"$bench
+make_backup $PREFIX"c"$bench
 
 cut -f1 -d"#" < changesets.txt | grep -E '[[:alnum:]]' | while read rev
 do
@@ -126,27 +130,27 @@ do
     actual_rev=`hg identify | cut -f1 -d' '`
     cd ..
     
-    test_current $actual_rev float                  $PREFIX"sgemm"
-    test_current $actual_rev double                 $PREFIX"dgemm"
-    test_current $actual_rev "std::complex<double>" $PREFIX"cgemm"
+    test_current $actual_rev float                  $PREFIX"s"$bench
+    test_current $actual_rev double                 $PREFIX"d"$bench
+    test_current $actual_rev "std::complex<double>" $PREFIX"c"$bench
   fi
   
 done
 
 echo "Float:"
-cat $PREFIX"sgemm.out"
+cat $PREFIX"s"$bench.out"
 echo ""
 
 echo "Double:"
-cat $PREFIX"dgemm.out"
+cat $PREFIX"d"$bench.out"
 echo ""
 
 echo "Complex:"
-cat $PREFIX"cgemm.out"
+cat $PREFIX"c"$bench.out"
 echo ""
 
-./make_plot.sh $PREFIX"sgemm"
-./make_plot.sh $PREFIX"dgemm"
-./make_plot.sh $PREFIX"cgemm"
+./make_plot.sh $PREFIX"s"$bench $bench
+./make_plot.sh $PREFIX"d"$bench $bench
+./make_plot.sh $PREFIX"c"$bench $bench