Build for vanilla armadillo

Note, this may lead to reduced performance as I've had to remove calls
to custom cholbacksub and solve_tri functions. Will look into whether
Armadillo correctly implements these performance gains yet...

Author: jonfink

CMakeLists.txt

@@ -1,61 +1,50 @@
-cmake_minimum_required(VERSION 2.4.6)
-include($ENV{ROS_ROOT}/core/rosbuild/rosbuild.cmake)
+cmake_minimum_required(VERSION 2.8.3)
+project(gp)
+set(CMAKE_BUILD_TYPE Debug)
 
-# Set the build type.  Options are:
-#  Coverage       : w/ debug symbols, w/o optimization, w/ code-coverage
-#  Debug          : w/ debug symbols, w/o optimization
-#  Release        : w/o debug symbols, w/ optimization
-#  RelWithDebInfo : w/ debug symbols, w/ optimization
-#  MinSizeRel     : w/o debug symbols, w/ optimization, stripped binaries
-#set(ROS_BUILD_TYPE RelWithDebInfo)
-set(ROS_BUILD_TYPE Debug)
-
-rosbuild_init()
-
-#set the default path for built executables to the "bin" directory
-set(EXECUTABLE_OUTPUT_PATH ${PROJECT_SOURCE_DIR}/bin)
-#set the default path for built libraries to the "lib" directory
-set(LIBRARY_OUTPUT_PATH ${PROJECT_SOURCE_DIR}/lib)
+find_package(ARMADILLO REQUIRED)
 
+include_directories(
+ ${ARMADILLO_INCLUDE_DIRS})
 
 #common commands for building c++ executables and libraries
-rosbuild_add_library(${PROJECT_NAME} 
-  src/GP.cc 
+add_library(${PROJECT_NAME}
+  src/GP.cc
   src/KernelFunction.cc
   src/MeanFunction.cc
   src/CGOptimizer.cc)
-target_link_libraries(${PROJECT_NAME} 
-  armadillo
+target_link_libraries(${PROJECT_NAME}
+  ${ARMADILLO_LIBRARIES}
   gsl)
 if(APPLE)
   set_target_properties(${PROJECT_NAME}
     PROPERTIES LINK_FLAGS "${LINK_FLAGS} -framework Accelerate -framework vecLib")
 endif()
 
-rosbuild_add_library(SPGP
-  src/SPGP.cc 
-  src/GP.cc 
+add_library(SPGP
+  src/SPGP.cc
+  src/GP.cc
   src/KernelFunction.cc
   src/MeanFunction.cc
   src/CGOptimizer.cc)
 target_link_libraries(SPGP
-  armadillo
+  ${ARMADILLO_LIBRARIES}
   gsl)
 if(APPLE)
   set_target_properties(SPGP
     PROPERTIES LINK_FLAGS "${LINK_FLAGS} -framework Accelerate -framework vecLib")
 endif()
 
-rosbuild_add_executable(test src/test.cc)
-target_link_libraries(test armadillo ${PROJECT_NAME})
+add_executable(test_gp src/test.cc)
+target_link_libraries(test_gp ${ARMADILLO_LIBRARIES} ${PROJECT_NAME})
 if(APPLE)
-  set_target_properties(test
+  set_target_properties(test_gp
     PROPERTIES LINK_FLAGS "${LINK_FLAGS} -framework Accelerate")
 endif()
 
-rosbuild_add_executable(test_spgp src/test_spgp.cc)
-target_link_libraries(test_spgp armadillo SPGP)
+add_executable(test_spgp src/test_spgp.cc)
+target_link_libraries(test_spgp ${ARMADILLO_LIBRARIES} SPGP)
 if(APPLE)
   set_target_properties(test_spgp
     PROPERTIES LINK_FLAGS "${LINK_FLAGS} -framework Accelerate")
-endif()
+endif()

Makefile

@@ -125,7 +125,8 @@ void GP::ComputeAlpha()
     timer.tic();
 #endif
     this->alpha.set_size(y.n_cols);
-    cholbacksub(alpha, this->L, trans(this->y - this->meanvals));
+    // cholbacksub(alpha, this->L, trans(this->y - this->meanvals));
+    solve(alpha, trimatu(this->L), trans(this->y - this->meanvals));
     this->need_to_compute_alpha = false;
 #ifdef BENCHMARK
     printf("%f seconds to compute alpha\n", timer.toc());
@@ -138,7 +139,8 @@ void GP::ComputeW()
   if(this->need_to_compute_w) {
     ComputeChol();
     this->W.set_size(this->L.n_rows, this->L.n_cols);
-    cholbacksub(W, this->L, eye<Mat<REAL> >(K.n_rows, K.n_cols));
+    // cholbacksub(W, this->L, eye<Mat<REAL> >(K.n_rows, K.n_cols));
+    solve(W, trimatu(this->L), eye<Mat<REAL> >(K.n_rows, K.n_cols));
     this->need_to_compute_w = false;
   }
 }
@@ -321,7 +323,8 @@ void GP::Predict(const Col<REAL> &Xs, REAL &mu, REAL &var)
 
   // trans(L) is LOWER TRIANGULAR
   Col<REAL> v(kstar.n_elem);
-  solve_tri(v, trans(L), kstar, false);
+  // solve_tri(v, trans(L), kstar, false);
+  solve(v, trimatu(trans(L)), kstar);
   REAL kxsxs = this->kernel->Eval(Xs, Xs);
   var = kxsxs - dot(v, v) + this->s2_n;
 }
@@ -375,12 +378,15 @@ void GP::PredictGradient(const Col<REAL> &Xs, Col<REAL> &grad, Col<REAL> &vargra
   Col<REAL> v(kstar.n_elem);
   Col<REAL> tmp(kstar.n_elem);
   Mat<REAL> dv(kstar.n_elem, grad.n_elem);
-  solve_tri(v, trans(L), kstar, false);
+  // solve_tri(v, trans(L), kstar, false);
+  solve(v, trimatu(trans(L)), kstar);
 
   //printf("About to solve for dv\n");
-  solve_tri(tmp, trans(L), trans(dkstar.row(0)), false);
+  // solve_tri(tmp, trans(L), trans(dkstar.row(0)), false);
+  solve(tmp, trimatu(trans(L)), trans(dkstar.row(0)));
   dv.col(0) = tmp;
-  solve_tri(tmp, trans(L), trans(dkstar.row(1)), false);
+  // solve_tri(tmp, trans(L), trans(dkstar.row(1)), false);
+  solve(tmp, trimatu(trans(L)), trans(dkstar.row(1)));
   dv.col(1) = tmp;
   //dv.print("dv: ");
 
@@ -408,7 +414,7 @@ void GP::OptimizeNoiseParam(REAL &noise_param, int max_iterations)
 
   Col<REAL> noise_params(1);
   noise_params(0) = noise_param;
-  
+
   opt.Initialize(noise_params, &f_eval_noise, &df_eval_noise, &fdf_eval_noise, 0.5, 0.1);
   opt.Optimize(noise_params, max_iterations);
 }
@@ -432,7 +438,7 @@ void GP::OptimizeKernelParam(Col<REAL> &kernel_param, int max_iterations)
 double f_eval_mean(const gsl_vector *x, void *param)
 {
   GP *gp_obj = reinterpret_cast<GP*>(param);
-  
+
   Col<REAL> mean_param(gp_obj->GetMeanFunction()->GetParamDim());
   for(unsigned int i=0; i < mean_param.n_elem; ++i) {
     mean_param(i) = gsl_vector_get(x, i);
@@ -441,14 +447,14 @@ double f_eval_mean(const gsl_vector *x, void *param)
   gp_obj->SetMeanFuncParams(mean_param);
 
   double ret = -gp_obj->ComputeLikelihood();
-  
+
   return ret;
 }
 
 void df_eval_mean(const gsl_vector *x, void *param, gsl_vector *g)
 {
   GP *gp_obj = reinterpret_cast<GP*>(param);
-  
+
   Col<REAL> mean_param(gp_obj->GetMeanFunction()->GetParamDim());
   for(unsigned int i=0; i < mean_param.n_elem; ++i) {
     mean_param(i) = gsl_vector_get(x, i);
@@ -472,26 +478,26 @@ void fdf_eval_mean(const gsl_vector *x, void *param, double *f, gsl_vector *g)
 double f_eval_kernel(const gsl_vector *x, void *param)
 {
   GP *gp_obj = reinterpret_cast<GP*>(param);
-  
+
   Col<REAL> kernel_param(gp_obj->GetKernelFunction()->GetParamDim());
   for(unsigned int i=0; i < kernel_param.n_elem; ++i) {
     kernel_param(i) = gsl_vector_get(x, i);
     if(kernel_param(i) < 1e-6) {
       return 1e6;
     }
   }
-  
+
   gp_obj->SetKernelFuncParams(kernel_param);
 
   double ret = -gp_obj->ComputeLikelihood();
-  
+
   return ret;
 }
 
 void df_eval_kernel(const gsl_vector *x, void *param, gsl_vector *g)
 {
   GP *gp_obj = reinterpret_cast<GP*>(param);
-  
+
   Col<REAL> kernel_param(gp_obj->GetKernelFunction()->GetParamDim());
   for(unsigned int i=0; i < kernel_param.n_elem; ++i) {
     kernel_param(i) = gsl_vector_get(x, i);
@@ -515,17 +521,17 @@ void fdf_eval_kernel(const gsl_vector *x, void *param, double *f, gsl_vector *g)
 double f_eval_noise(const gsl_vector *x, void *param)
 {
   GP *gp_obj = reinterpret_cast<GP*>(param);
-  
+
   REAL noise_param;
   noise_param = gsl_vector_get(x, 0);
   if(noise_param < 1e-6) {
     return 1e6;
   }
-  
+
   gp_obj->SetNoise(noise_param);
 
   double ret = -gp_obj->ComputeLikelihood();
-  
+
   return ret;
 }