extend monitoring benchmarks with transpose matrix-vector and triangular matrix-vectors.

bench/perf_monitoring/gemm/changesets.txt

@@ -25,13 +25,13 @@ before-evaluators
 #6742:0cbd6195e829   # merge default to tensors
 #6747:853d2bafeb8f   # Generalized the gebp apis
 6765:71584fd55762   # Made the blocking computation aware of the l3 cache; Also optimized the blocking parameters to take into account the number of threads used for a computation
-#6781:9cc5a931b2c6   # generalized gemv
+6781:9cc5a931b2c6   # generalized gemv
-#6792:f6e1daab600a   # ensured that contractions that can be reduced to a matrix vector product
+6792:f6e1daab600a   # ensured that contractions that can be reduced to a matrix vector product
 #6844:039efd86b75c   # merge tensor
 6845:7333ed40c6ef   # change prefetching in gebp
 #6856:b5be5e10eb7f   # merge index conversion
-#6893:c3a64aba7c70   # clean blocking size computation
+6893:c3a64aba7c70   # clean blocking size computation
-#6898:6fb31ebe6492   # rotating kernel for ARM
+6898:6fb31ebe6492   # rotating kernel for ARM
 6899:877facace746   # rotating kernel for ARM only
 #6904:c250623ae9fa   # result_of
 6921:915f1b1fc158   # fix prefetching change for ARM
@@ -50,7 +50,7 @@ before-evaluators
 7098:b6f1db9cf9ec   # Bug 992: don't select a 3p GEMM path with non-vectorizable scalar types, this hits unsupported paths in symm/triangular products code
 7591:09a8e2186610   # 3.3-alpha1
 7650:b0f3c8f43025   # help clang inlining
-#8744:74b789ada92a   # Improved the matrix multiplication blocking in the case where mr is not a power of 2 (e.g on Haswell CPUs)
+8744:74b789ada92a   # Improved the matrix multiplication blocking in the case where mr is not a power of 2 (e.g on Haswell CPUs)
 8789:efcb912e4356   # Made the index type a template parameter to evaluateProductBlockingSizes. Use numext::mini and numext::maxi instead of std::min/std::max to compute blocking sizes
 8972:81d53c711775   # Don't optimize the processing of the last rows of a matrix matrix product in cases that violate the assumptions made by the optimized code path
 8985:d935df21a082   # Remove the rotating kernel.
@@ -59,6 +59,11 @@ before-evaluators
 9174:d228bc282ac9   # merge
 9212:c90098affa7b   # Fix performance regression introduced in changeset 8aad8f35c955
 9213:9f1c14e4694b   # Fix performance regression in dgemm introduced by changeset 81d53c711775
-3.3-beta2
+9361:69d418c06999   # 3.3-beta2
-3.3-rc1
+9583:bef509908b9d   # 3.3-rc1
-3.3.0
+9792:26667be4f70b   # 3.3.0
+9942:b1d3eba60130   # Operators += and -= do not resize!
+9943:79bb9887afd4   # Ease compiler job to generate clean and efficient code in mat*vec
+9946:2213991340ea   # Complete rewrite of column-major-matrix * vector product to deliver higher performance of modern CPU.
+9953:21acc0e8d782   # Improve performance of row-major-dense-matrix * vector products for recent CPUs.
+

bench/perf_monitoring/gemm/gemm.cpp

@@ -53,7 +53,10 @@ int main(int argc, char **argv)
 {
   std::vector<double> results;
   
-  std::ifstream settings("gemm_settings.txt");
+  std::string filename = std::string("gemm_settings.txt");
+  if(argc>1)
+    filename = std::string(argv[1]);
+  std::ifstream settings(filename);
   long m, n, k;
   while(settings >> m >> n >> k)
   {

bench/perf_monitoring/gemm/gemv.cpp

@@ -1,18 +1,4 @@
-#include <iostream>
+#include "gemv_common.h"
-#include <fstream>
-#include <vector>
-#include <Eigen/Core>
-#include "../../BenchTimer.h"
-using namespace Eigen;
-
-#ifndef SCALAR
-#error SCALAR must be defined
-#endif
-
-typedef SCALAR Scalar;
-
-typedef Matrix<Scalar,Dynamic,Dynamic> Mat;
-typedef Matrix<Scalar,Dynamic,1>       Vec;
 
 EIGEN_DONT_INLINE
 void gemv(const Mat &A, const Vec &B, Vec &C)
@@ -20,49 +6,7 @@ void gemv(const Mat &A, const Vec &B, Vec &C)
   C.noalias() += A * B;
 }
 
-EIGEN_DONT_INLINE
-double bench(long m, long n)
-{
-  Mat A(m,n);
-  Vec B(n);
-  Vec C(m);
-  A.setRandom();
-  B.setRandom();
-  C.setZero();
-  
-  BenchTimer t;
-  
-  double up = 1e9*4/sizeof(Scalar);
-  double tm0 = 4, tm1 = 10;
-  if(NumTraits<Scalar>::IsComplex)
-  {
-    up /= 4;
-    tm0 = 2;
-    tm1 = 4;
-  }
-  
-  double flops = 2. * m * n;
-  long rep = std::max(1., std::min(100., up/flops) );
-  long tries = std::max(tm0, std::min(tm1, up/flops) );
-  
-  BENCH(t, tries, rep, gemv(A,B,C));
-  
-  return 1e-9 * rep * flops / t.best();
-}
-
 int main(int argc, char **argv)
 {
-  std::vector<double> results;
+  return main_gemv(argc, argv, gemv);
-  
-  std::ifstream settings("gemv_settings.txt");
-  long m, n;
-  while(settings >> m >> n)
-  {
-    //std::cerr << "  Testing " << m << " " << n << " " << k << std::endl;
-    results.push_back( bench(m, n) );
-  }
-  
-  std::cout << RowVectorXd::Map(results.data(), results.size());
-  
-  return 0;
 }

bench/perf_monitoring/gemm/gemv_common.h

@@ -0,0 +1,68 @@
+#include <iostream>
+#include <fstream>
+#include <vector>
+#include <functional>
+#include <Eigen/Core>
+#include "../../BenchTimer.h"
+using namespace Eigen;
+
+#ifndef SCALAR
+#error SCALAR must be defined
+#endif
+
+typedef SCALAR Scalar;
+
+typedef Matrix<Scalar,Dynamic,Dynamic> Mat;
+typedef Matrix<Scalar,Dynamic,1>       Vec;
+
+template<typename Func>
+EIGEN_DONT_INLINE
+double bench(long m, long n, Func &f)
+{
+  Mat A(m,n);
+  Vec B(n);
+  Vec C(m);
+  A.setRandom();
+  B.setRandom();
+  C.setRandom();
+
+  BenchTimer t;
+
+  double up = 1e8/sizeof(Scalar);
+  double tm0 = 4, tm1 = 10;
+  if(NumTraits<Scalar>::IsComplex)
+  {
+    up /= 4;
+    tm0 = 2;
+    tm1 = 4;
+  }
+
+  double flops = 2. * m * n;
+  long rep = std::max(1., std::min(100., up/flops) );
+  long tries = std::max(tm0, std::min(tm1, up/flops) );
+
+  BENCH(t, tries, rep, f(A,B,C));
+
+  return 1e-9 * rep * flops / t.best();
+}
+
+template<typename Func>
+int main_gemv(int argc, char **argv, Func& f, const std::string &setting_filename)
+{
+  std::vector<double> results;
+
+  std::string filename = std::string("gemv_settings.txt");
+  if(argc>1)
+    filename = std::string(argv[1]);
+  std::ifstream settings(setting_filename);
+  long m, n;
+  while(settings >> m >> n)
+  {
+    //std::cerr << "  Testing " << m << " " << n << std::endl;
+    results.push_back( bench(m, n, f) );
+  }
+
+  std::cout << RowVectorXd::Map(results.data(), results.size());
+
+  return 0;
+}

bench/perf_monitoring/gemm/gemvt.cpp

@@ -0,0 +1,12 @@
+#include "gemv_common.h"
+
+EIGEN_DONT_INLINE
+void gemv(const Mat &A, Vec &B, const Vec &C)
+{
+  B.noalias() += A.transpose() * C;
+}
+
+int main(int argc, char **argv)
+{
+  return main_gemv(argc, argv, gemv);
+}

bench/perf_monitoring/gemm/lazy_gemm.cpp

@@ -84,7 +84,10 @@ int main(int argc, char **argv)
 {
   std::vector<double> results;
   
-  std::ifstream settings("lazy_gemm_settings.txt");
+  std::string filename = std::string("lazy_gemm_settings.txt");
+  if(argc>1)
+    filename = std::string(argv[1]);
+  std::ifstream settings(filename);
   long m, n, k, t;
   while(settings >> m >> n >> k >> t)
   {

bench/perf_monitoring/gemm/make_plot.sh

@@ -5,6 +5,7 @@
 # and generates $1.pdf
 WHAT=$1
 bench=$2
+settings_file=$3
 
 header="rev "
 while read line
@@ -12,7 +13,7 @@ do
   if [ ! -z '$line' ]; then
     header="$header  \"$line\""
   fi
-done < $bench"_settings.txt"
+done < $settings_file
 
 echo $header > $WHAT.out.header
 cat $WHAT.out >> $WHAT.out.header
@@ -25,7 +26,7 @@ echo "set xtics rotate 1" >> $WHAT.gnuplot
 echo "set term pdf color rounded enhanced fontscale 0.35 size 7in,5in" >> $WHAT.gnuplot
 echo set output "'"$WHAT.pdf"'" >> $WHAT.gnuplot
 
-col=`cat $bench"_settings.txt" | wc -l`
+col=`cat $settings_file | wc -l`
 echo "plot for [col=2:$col+1] '$WHAT.out.header' using 0:col:xticlabels(1) with lines" >> $WHAT.gnuplot
 echo " " >>  $WHAT.gnuplot
 

bench/perf_monitoring/gemm/run.sh

@@ -1,17 +1,22 @@
 #!/bin/bash
 
-# ./run.sh gemm
+# ./run.sh gemm gemm_settings.txt
-# ./run.sh lazy_gemm
+# ./run.sh lazy_gemm lazy_gemm_settings.txt
+# ./run.sh gemv gemv_settings.txt
+# ./run.sh trmv_up gemv_square_settings.txt
+# ...
 
 # Examples of environment variables to be set:
 #   PREFIX="haswell-fma-"
 #   CXX_FLAGS="-mfma"
+#   CXX=clang++
 
 # Options:
 #   -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
+settings_file=$2
 
 if echo "$*" | grep '\-up' > /dev/null; then
   update=true
@@ -88,7 +93,7 @@ function test_current
   fi
   res=$prev
   count_rev=`echo $prev |  wc -w`
-  count_ref=`cat $bench"_settings.txt" |  wc -l`
+  count_ref=`cat $settings_file |  wc -l`
   if echo "$global_args" | grep "$rev" > /dev/null; then
     rev_found=true
   else
@@ -98,7 +103,7 @@ function test_current
 #  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 $bench.cpp -DSCALAR=$scalar -o $name; then
-      curr=`./$name`
+      curr=`./$name $settings_file`
       if [ $count_rev == $count_ref ]; then
         echo "merge previous $prev"
         echo "with new       $curr"
@@ -149,8 +154,8 @@ echo "Complex:"
 cat $PREFIX"c""$bench.out"
 echo ""
 
-./make_plot.sh $PREFIX"s"$bench $bench
+./make_plot.sh $PREFIX"s"$bench $bench $settings_file
-./make_plot.sh $PREFIX"d"$bench $bench
+./make_plot.sh $PREFIX"d"$bench $bench $settings_file
-./make_plot.sh $PREFIX"c"$bench $bench
+./make_plot.sh $PREFIX"c"$bench $bench $settings_file
 
 

bench/perf_monitoring/gemm/runall.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+
+# ./runall.sh
+
+# Examples of environment variables to be set:
+#   PREFIX="haswell-fma-"
+#   CXX_FLAGS="-mfma"
+#   CXX=clang++
+
+# Options:
+#   -up : enforce the recomputation of existing data, and keep best results as a merging strategy
+#   -s  : recompute selected changesets only and keep bests
+
+./run.sh gemm gemm_settings.txt $*
+./run.sh lazy_gemm lazy_gemm_settings.txt $*
+./run.sh gemv gemv_settings.txt $*
+./run.sh gemvt gemv_settings.txt $*
+./run.sh trmv_up gemv_square_settings.txt $*
+./run.sh trmv_lo gemv_square_settings.txt $*
+./run.sh trmv_upt gemv_square_settings.txt $*
+./run.sh trmv_lot gemv_square_settings.txt $*
+

bench/perf_monitoring/gemm/trmv_lo.cpp

@@ -0,0 +1,12 @@
+#include "gemv_common.h"
+
+EIGEN_DONT_INLINE
+void trmv(const Mat &A, const Vec &B, Vec &C)
+{
+  C.noalias() += A.triangularView<Lower>() * B;
+}
+
+int main(int argc, char **argv)
+{
+  return main_gemv(argc, argv, trmv);
+}

bench/perf_monitoring/gemm/trmv_lot.cpp

@@ -0,0 +1,12 @@
+#include "gemv_common.h"
+
+EIGEN_DONT_INLINE
+void gemv(const Mat &A, Vec &B, const Vec &C)
+{
+  B.noalias() += A.transpose().triangularView<Lower>() * C;
+}
+
+int main(int argc, char **argv)
+{
+  return main_gemv(argc, argv, trmv);
+}

bench/perf_monitoring/gemm/trmv_up.cpp

@@ -0,0 +1,12 @@
+#include "gemv_common.h"
+
+EIGEN_DONT_INLINE
+void trmv(const Mat &A, const Vec &B, Vec &C)
+{
+  C.noalias() += A.triangularView<Upper>() * B;
+}
+
+int main(int argc, char **argv)
+{
+  return main_gemv(argc, argv, trmv);
+}

bench/perf_monitoring/gemm/trmv_upt.cpp

@@ -0,0 +1,12 @@
+#include "gemv_common.h"
+
+EIGEN_DONT_INLINE
+void trmv(const Mat &A, Vec &B, const Vec &C)
+{
+  B.noalias() += A.transpose().triangularView<Upper>() * C;
+}
+
+int main(int argc, char **argv)
+{
+  return main_gemv(argc, argv, trmv);
+}