Commit 75a4ea7a authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Merge branch 'traversers-optimizations' into 'develop' 

Traversers optimizations

See merge request !20
parents 3b73e43e 5f4d085f

CMakeLists.txt

+7 −0
Original line number Diff line number Diff line
@@ -21,6 +21,7 @@ set(WITH_CUDA_ARCH "auto" CACHE STRING "Build for these CUDA architectures") 
option(WITH_OPENMP "Build with OpenMP support" ON)

option(WITH_GMP "Build with GMP support" OFF)

option(WITH_TESTS "Build tests" ON)

option(WITH_PROFILING "Enable code profiling compiler flags" OFF )

option(WITH_COVERAGE "Enable code coverage reports from unit tests" OFF)

option(WITH_EXAMPLES "Compile the 'examples' directory" ON)

option(WITH_TOOLS "Compile the 'src/Tools' directory" ON)
@@ -243,6 +244,11 @@ if( OPENMP_FOUND AND ${WITH_OPENMP} ) 
   set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DHAVE_OPENMP ${OpenMP_CXX_FLAGS}" )

endif()



if( ${WITH_PROFILING} )

    set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g" )

    set( CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} --generate-line-info")

endif()



find_package( DCMTK )

if( DCMTK_FOUND )

   set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DHAVE_DCMTK_H" )
@@ -414,6 +420,7 @@ message( " WITH_CUDA_ARCH = ${WITH_CUDA_ARCH}" ) 
message( "   WITH_OPENMP = ${WITH_OPENMP}" )

message( "   WITH_GMP = ${WITH_GMP}" )

message( "   WITH_TESTS = ${WITH_TESTS}" )

message( "   WITH_PROFILING = ${WITH_PROFILING}" )

message( "   WITH_COVERAGE = ${WITH_COVERAGE}" )

message( "   WITH_EXAMPLES = ${WITH_EXAMPLES}" )

message( "   WITH_TOOLS = ${WITH_TOOLS}" )

build

+4 −0
Original line number Diff line number Diff line
@@ -22,6 +22,7 @@ WITH_CUDA_ARCH="auto" 
WITH_OPENMP="yes"

WITH_GMP="no"

WITH_TESTS="yes"

WITH_PROFILING="no"

WITH_COVERAGE="no"

WITH_EXAMPLES="yes"

WITH_PYTHON="yes"
@@ -57,6 +58,7 @@ do 
        --with-openmp=*                  ) WITH_OPENMP="${option#*=}" ;;

        --with-gmp=*                     ) WITH_GMP="${option#*=}" ;;

        --with-tests=*                   ) WITH_TESTS="${option#*=}" ;;

        --with-profiling=*               ) WITH_PROFILING="${option#*=}" ;;

        --with-coverage=*                ) WITH_COVERAGE="${option#*=}" ;;

        --with-examples=*                ) WITH_EXAMPLES="${option#*=}" ;;

        --with-tools=*                   ) WITH_TOOLS="${option#*=}" ;;
@@ -95,6 +97,7 @@ if [[ ${HELP} == "yes" ]]; then 
    echo "   --with-openmp=yes/no                  Enables OpenMP. 'yes' by default."

    echo "   --with-gmp=yes/no                     Enables the wrapper for GNU Multiple Precision Arithmetic Library. 'no' by default."

    echo "   --with-tests=yes/no                   Enables unit tests. 'yes' by default."

    echo "   --with-profiling=yes/no               Enables code profiling compiler falgs. 'no' by default."

    echo "   --with-coverage=yes/no                Enables code coverage reports for unit tests. 'no' by default (lcov is required)."

    echo "   --with-examples=yes/no                Compile the 'examples' directory. 'yes' by default."

    echo "   --with-tools=yes/no                   Compile the 'src/Tools' directory. 'yes' by default."
@@ -165,6 +168,7 @@ cmake_command=( 
         -DWITH_OPENMP=${WITH_OPENMP}

         -DWITH_GMP=${WITH_GMP}

         -DWITH_TESTS=${WITH_TESTS}

         -DWITH_PROFILING=${WITH_PROFILING}

         -DWITH_COVERAGE=${WITH_COVERAGE}

         -DWITH_EXAMPLES=${WITH_EXAMPLES}

         -DWITH_TOOLS=${WITH_TOOLS}

src/Benchmarks/BLAS/array-operations.h

+12 −12
Original line number Diff line number Diff line
@@ -72,9 +72,9 @@ benchmarkArrayOperations( Benchmark & benchmark, 
      resultDevice = (int) deviceArray == deviceArray2;

   };

   benchmark.setOperation( "comparison (operator==)", 2 * datasetSize );

   benchmark.time( reset1, "CPU", compareHost );

   benchmark.time< Devices::Host >( reset1, "CPU", compareHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", compareCuda );

   benchmark.time< Devices::Cuda >( reset1, "GPU", compareCuda );

#endif
@@ -87,9 +87,9 @@ benchmarkArrayOperations( Benchmark & benchmark, 
   benchmark.setOperation( "copy (operator=)", 2 * datasetSize );

   // copyBasetime is used later inside HAVE_CUDA guard, so the compiler will

   // complain when compiling without CUDA

   const double copyBasetime = benchmark.time( reset1, "CPU", copyAssignHostHost );

   const double copyBasetime = benchmark.time< Devices::Host >( reset1, "CPU", copyAssignHostHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", copyAssignCudaCuda );

   benchmark.time< Devices::Cuda >( reset1, "GPU", copyAssignCudaCuda );

#endif
@@ -101,8 +101,8 @@ benchmarkArrayOperations( Benchmark & benchmark, 
   };

#ifdef HAVE_CUDA

   benchmark.setOperation( "copy (operator=)", datasetSize, copyBasetime );

   benchmark.time( reset1, "CPU->GPU", copyAssignHostCuda );

   benchmark.time( reset1, "GPU->CPU", copyAssignCudaHost );

   benchmark.time< Devices::Cuda >( reset1, "CPU->GPU", copyAssignHostCuda );

   benchmark.time< Devices::Cuda >( reset1, "GPU->CPU", copyAssignCudaHost );

#endif
@@ -113,9 +113,9 @@ benchmarkArrayOperations( Benchmark & benchmark, 
      deviceArray.setValue( 3.0 );

   };

   benchmark.setOperation( "setValue", datasetSize );

   benchmark.time( reset1, "CPU", setValueHost );

   benchmark.time< Devices::Host >( reset1, "CPU", setValueHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", setValueCuda );

   benchmark.time< Devices::Cuda >( reset1, "GPU", setValueCuda );

#endif
@@ -132,9 +132,9 @@ benchmarkArrayOperations( Benchmark & benchmark, 
#endif

   };

   benchmark.setOperation( "allocation (setSize)", datasetSize );

   benchmark.time( resetSize1, "CPU", setSizeHost );

   benchmark.time< Devices::Host >( resetSize1, "CPU", setSizeHost );

#ifdef HAVE_CUDA

   benchmark.time( resetSize1, "GPU", setSizeCuda );

   benchmark.time< Devices::Cuda >( resetSize1, "GPU", setSizeCuda );

#endif
@@ -151,9 +151,9 @@ benchmarkArrayOperations( Benchmark & benchmark, 
#endif

   };

   benchmark.setOperation( "deallocation (reset)", datasetSize );

   benchmark.time( setSize1, "CPU", resetSizeHost );

   benchmark.time< Devices::Host >( setSize1, "CPU", resetSizeHost );

#ifdef HAVE_CUDA

   benchmark.time( setSize1, "GPU", resetSizeCuda );

   benchmark.time< Devices::Cuda >( setSize1, "GPU", resetSizeCuda );

#endif



   return true;

src/Benchmarks/BLAS/spmv.h

+2 −2
Original line number Diff line number Diff line
@@ -161,9 +161,9 @@ benchmarkSpMV( Benchmark & benchmark, 
   };



   benchmark.setOperation( datasetSize );

   benchmark.time( reset, "CPU", spmvHost );

   benchmark.time< Devices::Host >( reset, "CPU", spmvHost );

#ifdef HAVE_CUDA

   benchmark.time( reset, "GPU", spmvCuda );

   benchmark.time< Devices::Cuda >( reset, "GPU", spmvCuda );

#endif



   return true;

src/Benchmarks/BLAS/vector-operations.h

+30 −30
Original line number Diff line number Diff line
@@ -64,7 +64,7 @@ benchmarkVectorOperations( Benchmark & benchmark, 
      deviceVector.setValue( 1.0 );

#endif

      // A relatively harmless call to keep the compiler from realizing we

      // don't actually do any useful work with the result of the reduciton.

      // don't actually do any useful work with the result of the reduction.

      srand48(resultHost);

      resultHost = resultDevice = 0.0;

   };
@@ -90,9 +90,9 @@ benchmarkVectorOperations( Benchmark & benchmark, 
      resultDevice = deviceVector.max();

   };

   benchmark.setOperation( "max", datasetSize );

   benchmark.time( reset1, "CPU", maxHost );

   benchmark.time< Devices::Host >( reset1, "CPU", maxHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", maxCuda );

   benchmark.time< Devices::Cuda >( reset1, "GPU", maxCuda );

#endif
@@ -103,9 +103,9 @@ benchmarkVectorOperations( Benchmark & benchmark, 
      resultDevice = deviceVector.min();

   };

   benchmark.setOperation( "min", datasetSize );

   benchmark.time( reset1, "CPU", minHost );

   benchmark.time< Devices::Host >( reset1, "CPU", minHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", minCuda );

   benchmark.time< Devices::Cuda >( reset1, "GPU", minCuda );

#endif
@@ -125,10 +125,10 @@ benchmarkVectorOperations( Benchmark & benchmark, 
   };

#endif

   benchmark.setOperation( "absMax", datasetSize );

   benchmark.time( reset1, "CPU", absMaxHost );

   benchmark.time< Devices::Host >( reset1, "CPU", absMaxHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", absMaxCuda );

   benchmark.time( reset1, "cuBLAS", absMaxCublas );

   benchmark.time< Devices::Cuda >( reset1, "GPU", absMaxCuda );

   benchmark.time< Devices::Cuda >( reset1, "cuBLAS", absMaxCublas );

#endif
@@ -148,10 +148,10 @@ benchmarkVectorOperations( Benchmark & benchmark, 
   };

#endif

   benchmark.setOperation( "absMin", datasetSize );

   benchmark.time( reset1, "CPU", absMinHost );

   benchmark.time< Devices::Host >( reset1, "CPU", absMinHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", absMinCuda );

   benchmark.time( reset1, "cuBLAS", absMinCublas );

   benchmark.time< Devices::Cuda >( reset1, "GPU", absMinCuda );

   benchmark.time< Devices::Cuda >( reset1, "cuBLAS", absMinCublas );

#endif
@@ -162,9 +162,9 @@ benchmarkVectorOperations( Benchmark & benchmark, 
      resultDevice = deviceVector.sum();

   };

   benchmark.setOperation( "sum", datasetSize );

   benchmark.time( reset1, "CPU", sumHost );

   benchmark.time< Devices::Host >( reset1, "CPU", sumHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", sumCuda );

   benchmark.time< Devices::Cuda >( reset1, "GPU", sumCuda );

#endif
@@ -182,10 +182,10 @@ benchmarkVectorOperations( Benchmark & benchmark, 
   };

#endif

   benchmark.setOperation( "l1 norm", datasetSize );

   benchmark.time( reset1, "CPU", l1normHost );

   benchmark.time< Devices::Host >( reset1, "CPU", l1normHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", l1normCuda );

   benchmark.time( reset1, "cuBLAS", l1normCublas );

   benchmark.time< Devices::Cuda >( reset1, "GPU", l1normCuda );

   benchmark.time< Devices::Cuda >( reset1, "cuBLAS", l1normCublas );

#endif
@@ -203,10 +203,10 @@ benchmarkVectorOperations( Benchmark & benchmark, 
   };

#endif

   benchmark.setOperation( "l2 norm", datasetSize );

   benchmark.time( reset1, "CPU", l2normHost );

   benchmark.time< Devices::Host >( reset1, "CPU", l2normHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", l2normCuda );

   benchmark.time( reset1, "cuBLAS", l2normCublas );

   benchmark.time< Devices::Cuda >( reset1, "GPU", l2normCuda );

   benchmark.time< Devices::Cuda >( reset1, "cuBLAS", l2normCublas );

#endif
@@ -217,9 +217,9 @@ benchmarkVectorOperations( Benchmark & benchmark, 
      resultDevice = deviceVector.lpNorm( 3.0 );

   };

   benchmark.setOperation( "l3 norm", datasetSize );

   benchmark.time( reset1, "CPU", l3normHost );

   benchmark.time< Devices::Host >( reset1, "CPU", l3normHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", l3normCuda );

   benchmark.time< Devices::Cuda >( reset1, "GPU", l3normCuda );

#endif
@@ -238,10 +238,10 @@ benchmarkVectorOperations( Benchmark & benchmark, 
   };

#endif

   benchmark.setOperation( "scalar product", 2 * datasetSize );

   benchmark.time( reset1, "CPU", scalarProductHost );

   benchmark.time< Devices::Host >( reset1, "CPU", scalarProductHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", scalarProductCuda );

   benchmark.time( reset1, "cuBLAS", scalarProductCublas );

   benchmark.time< Devices::Cuda >( reset1, "GPU", scalarProductCuda );

   benchmark.time< Devices::Cuda >( reset1, "cuBLAS", scalarProductCublas );

#endif



   /*
@@ -289,10 +289,10 @@ benchmarkVectorOperations( Benchmark & benchmark, 
   };

#endif

   benchmark.setOperation( "scalar multiplication", 2 * datasetSize );

   benchmark.time( reset1, "CPU", multiplyHost );

   benchmark.time< Devices::Host >( reset1, "CPU", multiplyHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", multiplyCuda );

   benchmark.time( reset1, "cuBLAS", multiplyCublas );

   benchmark.time< Devices::Cuda >( reset1, "GPU", multiplyCuda );

   benchmark.time< Devices::Cuda >( reset1, "cuBLAS", multiplyCublas );

#endif
@@ -312,10 +312,10 @@ benchmarkVectorOperations( Benchmark & benchmark, 
   };

#endif

   benchmark.setOperation( "vector addition", 3 * datasetSize );

   benchmark.time( reset1, "CPU", addVectorHost );

   benchmark.time< Devices::Host >( reset1, "CPU", addVectorHost );

#ifdef HAVE_CUDA

   benchmark.time( reset1, "GPU", addVectorCuda );

   benchmark.time( reset1, "cuBLAS", addVectorCublas );

   benchmark.time< Devices::Cuda >( reset1, "GPU", addVectorCuda );

   benchmark.time< Devices::Cuda >( reset1, "cuBLAS", addVectorCublas );

#endif