Incrimental performance checks

CMake/SuperBuild.cmake

@@ -23,7 +23,7 @@ else()
     message( STATUS "YAML-CPP WAS NOT FOUND, BUILDING" )
     ExternalProject_Add(YAML_CPP
         GIT_REPOSITORY  "https://github.com/jbeder/yaml-cpp.git"
-        GIT_TAG "master" # "yaml-cpp-0.6.2" # "master" 
+        GIT_TAG "master" # "yaml-cpp-0.6.2", # "master" 
         UPDATE_COMMAND ""
         PATCH_COMMAND ""
         PREFIX ${CMAKE_CURRENT_BINARY_DIR}/external/yaml-cpp

CMakeLists.txt

@@ -332,6 +332,10 @@ include_directories("${PROJECT_BINARY_DIR}/include")
 include_directories ("${CMAKE_INSTALL_PREFIX}/include/")
 link_directories ("${CMAKE_INSTALL_PREFIX}/lib/")
 
+##############
+# gprof hack #
+##############
+#set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pg")
 
 ########################################################################
 # Build Modules

Modules/FDEM1D/src/KernelEM1DReflSpec.cpp

@@ -19,6 +19,7 @@ namespace Lemma {
 
         rams = lambda*lambda;
         u = (rams-kk.array()).sqrt(); // CRITICAL
+
         uk = u(0);
         um = u(0);
 
@@ -27,16 +28,14 @@ namespace Lemma {
 
         Zyd.tail<1>() = Zyi.tail<1>();
 
-        // Vectorise, performance is not really any better
+        // Vectorise, performance benchmarks approx. the same as loop w gcc 9 -TI
         cf.segment(1,nlay-2) = (-2.*u.segment(1, nlay-2).array() * LayerThickness.segment(1, nlay-2).array()).exp();
         th.segment(1,nlay-2) = (1.-cf.segment(1, nlay-2).array()) / (1.+cf.segment(1, nlay-2).array());
 
-        /*
-        for (int ilay=1; ilay<nlay-1; ++ilay) {
-            cf(ilay) = std::exp(-(Real)(2.)*u(ilay)*LayerThickness(ilay));
-            th(ilay) = ((Real)(1.)-cf(ilay)) / ((Real)(1.)+cf(ilay));
-        }
-        */
+//         for (int ilay=1; ilay<nlay-1; ++ilay) {
+//             cf(ilay) = std::exp(-(Real)(2.)*u(ilay)*LayerThickness(ilay));
+//             th(ilay) = ((Real)(1.)-cf(ilay)) / ((Real)(1.)+cf(ilay));
+//         }
 
         // recursive, can't vectorize
         for (int N=nlay-2; N >= 1; --N) {
@@ -54,6 +53,7 @@ namespace Lemma {
 
         rams = lambda*lambda;
         u = (rams-kk.array()).sqrt(); // CRITICAL
+
         uk = u(0);
         um = u(0);
 
@@ -62,16 +62,14 @@ namespace Lemma {
 
         Zyd.tail<1>() = Zyi.tail<1>();
 
-        // Vectorise
+        // Vectorise, performance benchmarks approx. the same as loop w/ gcc 9 -TI
         cf.segment(1,nlay-2) = (-2.*u.segment(1, nlay-2).array() * LayerThickness.segment(1, nlay-2).array()).exp();
         th.segment(1,nlay-2) = (1.-cf.segment(1, nlay-2).array()) / (1.+cf.segment(1, nlay-2).array());
 
-        /*
-        for (int ilay=1; ilay<nlay-1; ++ilay) {
-            cf(ilay) = std::exp(-(Real)(2.)*u(ilay)*LayerThickness(ilay));
-            th(ilay) = ((Real)(1.)-cf(ilay)) / ((Real)(1.)+cf(ilay));
-        }
-        */
+//         for (int ilay=1; ilay<nlay-1; ++ilay) {
+//             cf(ilay) = std::exp(-(Real)(2.)*u(ilay)*LayerThickness(ilay));
+//             th(ilay) = ((Real)(1.)-cf(ilay)) / ((Real)(1.)+cf(ilay));
+//         }
 
         // recursive, can't vectorize
         for (int N=nlay-2; N >=1; --N) {

Modules/FDEM1D/testing/BenchKiHa.h

@@ -56,7 +56,7 @@ public:
 		dipole->SetFrequency(0, 4400.1000);
 		//dipole->SetPhase(0);
 		//dipole->SetLocation( (VectorXr(3) << 49, 49, -1e-4).finished() );
-		dipole->SetLocation( 0, 0, -1e-4  );
+		dipole->SetLocation( 0, 0, -1e-1  );
 
 	    // Define model
 	    VectorXcr sigma(8);
@@ -72,17 +72,17 @@ public:
 	    receivers = FieldPoints::NewSP();
 		Vector3r loc;
 
-        Real ox = 2.;
-		Real oy = 2.;
-		Real oz = 2;
+        Real ox = 250.;
+		Real oy = 250.;
+		Real oz = -250.;
 
-		Real dx = 0.62;
-		Real dy = 0.62;
-		Real dz = 0.62;
+		Real dx = 20;
+		Real dy = 20;
+		Real dz = 20;
 
-		int  nx = 20;
-		int  ny = 20;
-		int  nz = 20;
+		int  nx = 13;
+		int  ny = 13;
+		int  nz = 13;
         Delta = nx*ny*nz*1e-10;
 
 		receivers->SetNumberOfPoints(nx*ny*nz);
@@ -118,6 +118,7 @@ public:
 		dipole->SetPolarisation(ZPOLARISATION);
 
         // Put in a unit test that will be slow.
+        std::cout << "MAGNETICDIPOLE Z polarisation" << std::endl;
 	    std::cout << "C++\n";
 
   	    timer.begin();
@@ -153,6 +154,7 @@ public:
 
         // Put in a unit test that will be slow.
 	    std::cout << "C++\n";
+        std::cout << "MAGNETICDIPOLE X polarisation" << std::endl;
 
   	    timer.begin();
 	    EmEarth->MakeCalc3();
@@ -187,6 +189,7 @@ public:
 
         // Put in a unit test that will be slow.
 	    std::cout << "C++\n";
+        std::cout << "MAGNETICDIPOLE Y polarisation" << std::endl;
 
   	    timer.begin();
 	    EmEarth->MakeCalc3();
@@ -221,6 +224,7 @@ public:
 
         // Put in a unit test that will be slow.
 	    std::cout << "C++\n";
+        std::cout << "GROUNDEDELECTRICDIPOLE X polarisation" << std::endl;
 
   	    timer.begin();
 	    EmEarth->MakeCalc3();