Commit 049ac6fc authored by Martin Perdacher's avatar Martin Perdacher
Browse files

added non-blocking l-list

parent 5333e66b

cmake/CMakeLists.txt

+23 −7
Original line number Diff line number Diff line 
cmake_minimum_required(VERSION 3.6)

project(blasJoin)



set(UTIL_SOURCES util/timer.cpp util/dataIo.cpp util/arguments.cpp util/allocation.cpp)

set(UTIL_SOURCES util/dataIo.cpp util/arguments.cpp util/allocation.cpp measure/energy.cpp measure/papicalls.cpp measure/timer.cpp)

set(SOURCE_FILES main.cpp blasJoin/blasJoin.cpp ${UTIL_SOURCES})



#####################

# build type: Release

#####################

# NDDEBUG turns off asserts

set(CMAKE_CXX_FLAGS "-std=c++11 ")



if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")

  set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} -march=knl -mtune=knl -fpic -ffast-math -DNDEBUG -O3 -DNDDEBUG -fopenmp -lmkl_core -lmkl_intel_lp64 -lmkl_intel_thread")

  set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} -march=knl -mtune=knl -fpic -ffast-math -DNDEBUG -O3 -DNDDEBUG -fopenmp -lmkl_core -lmkl_intel_lp64 -lmkl_intel_thread -liomp5 -lboost_system")

elseif ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Intel")

  set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} -xmic-avx512 -qopenmp -qopt-report=2 -DNDEBUG -O3 -liomp5 -lpthread -lmkl_core -lmkl_intel_lp64 -lmkl_intel_thread")

  set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} -xmic-avx512 -qopenmp -DNDEBUG -O3 -lmkl_intel_thread  -liomp5 -lpthread -lmkl_core -lmkl_intel_lp64 -lboost_system")

  # set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} -qopt-report=2")



  # xeon phi (knl) specific:

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

  # set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} -lmemkind")

  add_definitions(-DDEBUG)

endif()
@@ -24,11 +28,12 @@ add_definitions(-DNUM_THREADS=64) 
####################

# cmake -D CMAKE_BUILD_TYPE:STRING=Debug ..

# and ignore the warning: "Manually-specified variables were not used by the project: CMAKE_BUID_TYPE"

set(CMAKE_CXX_FLAGS_DEBUG "-std=c++11")



if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")

  set(CMAKE_CXX_FLAGS_DEBUG "-std=c++11 -march=knl -mtune=knl -fpic -ffast-math -O0 -fopenmp")

  set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -march=knl -mtune=knl -fpic -ffast-math -O0 -fopenmp -lboost_system")

elseif ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Intel")

  set(CMAKE_CXX_FLAGS_DEBUG  "-std=c++11 -xmic-avx512 -fpic -qopenmp -axCOMMON-AVX512 -lmemkind -lmkl_core -lmkl_intel_lp64 -lmkl_intel_thread -liomp5 -lpthread -g -debug all -save-temps -Wl, -O0 -fstack-security-check")

  set(CMAKE_CXX_FLAGS_DEBUG  "${CMAKE_CXX_FLAGS_DEBUG} -xmic-avx512 -fpic -qopenmp -axCOMMON-AVX512 -lmemkind -lmkl_core -lmkl_intel_lp64 -lmkl_intel_thread -liomp5 -lpthread -g -debug all -save-temps -Wl, -O0 -fstack-security-check -lboost_system")

endif()



# adding MKL include directory
@@ -37,8 +42,10 @@ include_directories($ENV{MKLROOT}/include) 
# xeon-phi

link_directories($ENV{MKLROOT}/lib/intel64)



include_directories($ENV{BOOST_HOME}/include/boost)

link_directories($ENV{BOOST_HOME}/lib)



# libraries

# mkl libraries

find_library ( mkl_lp64_LIB NAMES libmkl_intel_lp64.a

        PATHS $ENV{MKLROOT} PATH_SUFFIXES lib)

find_library ( mkl_core_LIB NAMES libmkl_core.a
@@ -48,5 +55,14 @@ find_library ( mkl_thread_LIB NAMES libmkl_intel_thread.a 
find_library ( mkl_omp_LIB NAMES libiomp5.a

        PATHS $ENV{MKLROOT} PATH_SUFFIXES lib)



# papi library

set(PAPI_PREFIX "/usr/local")

find_library(PAPI_LIBRARIES

    # Pick the static library first for easier run-time linking.

    NAMES libpapi.a papi PATHS ${PAPI_PREFIX}/lib

)



add_executable(blasJoin ${SOURCE_FILES})

add_definitions(-DCOUNT_ONLY)

add_executable(blasJoinCountOnly ${SOURCE_FILES})

target_link_libraries(blasJoin ${PAPI_LIBRARIES})

cmake/blasJoin/blasJoin.cpp

+17 −6
Original line number Diff line number Diff line
@@ -2,13 +2,12 @@ 
#include "blasJoin.h"



/* returns energy consumption [watthours] */

double blasJoin(const double *x, const size_t N, const size_t D, const double EPS, const unsigned int THREADS, const size_t BLOCKSIZE, size_t *joinCounts, CounterBin * hwcounters){

void blasJoin(const double *x, const size_t N, const size_t D, const double EPS, const unsigned int THREADS, const size_t BLOCKSIZE, size_t *joinCounts, CounterBin * hwcounters, boost::lockfree::queue<join_pair> &jpartners){



    assert( BLOCKSIZE < N && BLOCKSIZE < 21000 && BLOCKSIZE > 1);



    double elapsed=0.0;

    // CUtilTimer timer;

    Hioki energy_consume;

    PapiBin papi_bin;

    double *iresult = NULL;
@@ -23,7 +22,6 @@ double blasJoin(const double *x, const size_t N, const size_t D, const double EP 
    iresult = (double*) ddr_alloc(BLOCKSIZE * BLOCKSIZE * sizeof(double));



    // timer.start();

    energy_consume.reset(); energy_consume.start();

    papi_bin.start();



    const double EPS_SQUARED = (EPS * EPS) / 4.0 ;
@@ -64,6 +62,14 @@ double blasJoin(const double *x, const size_t N, const size_t D, const double EP 
                     * count only join-partners (positive ones)

                    */

                    accum +=  - ( (long long) floor( - ( iresult[k * BLOCKSIZE + l] + p[i*BLOCKSIZE+k] + p[j*BLOCKSIZE+l] ) ) >> 63);

#ifndef COUNT_ONLY

                    if ( iresult[k * BLOCKSIZE + l] + p[i*BLOCKSIZE+k] + p[j*BLOCKSIZE+l] < 0 ){

                        join_pair p;

                        p.p1 = i * BLOCKSIZE + k;

                        p.p2 = j * BLOCKSIZE + l;

                        while ( ! jpartners.push(p) );

                    }

#endif

                }

            }
@@ -89,6 +95,14 @@ double blasJoin(const double *x, const size_t N, const size_t D, const double EP 
        for ( int k = 1 ; k < blockRow ; k++ ){

            for ( int l = 0 ; l < k ; l++ ){

                accum += - ( (long long) floor( - ( iresult[k * BLOCKSIZE + l] + p[i*BLOCKSIZE+k] + p[i*BLOCKSIZE+l] ) ) >> 63);

#ifndef COUNT_ONLY

                if ( iresult[k * BLOCKSIZE + l] + p[i*BLOCKSIZE+k] + p[i*BLOCKSIZE+l] < 0 ){

                    join_pair p;

                    p.p1 = i * BLOCKSIZE + k;

                    p.p2 = i * BLOCKSIZE + l;

                    while (!jpartners.push(p));

                }

#endif

            }

        }

        *joinCounts += accum;
@@ -97,11 +111,8 @@ double blasJoin(const double *x, const size_t N, const size_t D, const double EP 


    // timer.stop();

    papi_bin.stop();

    energy_consume.stop();

    *hwcounters = papi_bin.getBin();



    ddr_free(iresult);

    ddr_free(p);



    return energy_consume.getWH();

}

cmake/blasJoin/blasJoin.h

+8 −1
Original line number Diff line number Diff line
@@ -16,7 +16,14 @@ 
#include "../measure/papicalls.h"

#include "../util/dataIo.h"



#include <boost/lockfree/queue.hpp>

#include <boost/atomic.hpp>



double blasJoin(const double *x, const size_t N, const size_t D, const double EPS, const unsigned int THREADS, const size_t BLOCKSIZE, size_t *joinCounts, CounterBin * hwcounters);

struct join_pair {

    size_t p1;

    size_t p2;

};



void blasJoin(const double *x, const size_t N, const size_t D, const double EPS, const unsigned int THREADS, const size_t BLOCKSIZE, size_t *joinCounts, CounterBin * hwcounters, boost::lockfree::queue<join_pair> &jpartners);



#endif

cmake/main.cpp

+14 −34
Original line number Diff line number Diff line 


// https://software.intel.com/en-us/forums/intel-c-compiler/topic/776876



#include <stdio.h>

#include <omp.h>



#include <boost/lockfree/queue.hpp>

#include <boost/atomic.hpp>





#include "util/allocation.h"

#include "util/arguments.h"

#include "util/dataIo.h"

#include "blasJoin/blasJoin.h"

#include "measure/energy.h"





int main(int argc, char** argv) {

    CounterBin hwcounters;
@@ -21,55 +27,29 @@ int main(int argc, char** argv) { 


    parsing_args(argc, argv, &N, &EPS, &D, &threads, &blocksize, filename, &isBinary);



    N = N * 1000;



    if ( threads != 0 ){

        omp_set_num_threads(threads);

    }



    x = (double*) ddr_alloc(sizeof (double)*N * D);

    boost::lockfree::queue<join_pair> queue(10000);



    // size_t threads = 1;

    // size_t N=10, D=3;

    // size_t blocksize=2, joinCounts=0;

    // double EPS=0.3;

    // double x[30] = { 1.0000, 2.0000, 3.0000,

    //                  1.1000, 2.2000, 3.0000,

    //                  5.0000, 6.0000, 7.0000,

    //                  5.1000, 6.2000, 7.1000,

    //                  1.0000, 4.8000, 6.3000,

    //                  1.0000, 14.8000, 1.3000,

    //                  3.0000, 17.0000, 2.0000,

    //                  14.000, 13.2000, 1.9000,

    //                  14.000, 1.00000, 15.000,

    //                  1.000, 1.000, 1.10000

    //              };

    x = (double*) ddr_alloc(sizeof (double)* N * D);



    if ( strcmp(filename,"" ) == 0) {

        // printf("random_init\n");

        random_init(x,N,D);

    }else{

        // printf("read_file\n");

        read_file(x, N, D, filename, isBinary);

    }



    // printf("# name: A\n");

    // printf("# type: matrix\n");

    // printf("# rows: 8000\n");

    // printf("# columns: 64\n");

    //

    // for ( int i = N - 8000 ; i < N ; i++ ){

    //     for ( int j = 0 ; j < D ; j++ ){

    //         printf("%f ", x[i*D + j]);

    //     }

    //     printf("\n");

    // }



    blasJoin( x, N, D, EPS, threads, blocksize, &joinCounts, &hwcounters, queue);



    watthours = blasJoin( x, N, D, EPS, threads, blocksize, &joinCounts, &hwcounters);



#pragma omp parallel

{

    if (  omp_get_thread_num() == 0 ){

        printf("%ld; %ld; %f; %ld; %d; %f; %lu; %lld; %f\n", N, D, EPS, blocksize, omp_get_num_threads(), hwcounters.rtime, joinCounts, hwcounters.l1, watthours);

        printf("%ld;%ld;%f;%ld;%d;%f;%lu;%f\n", N, D, EPS, blocksize, omp_get_num_threads(), hwcounters.rtime, joinCounts, hwcounters.whours);

    }

}