Commit 76a342ef authored by Martin Perdacher's avatar Martin Perdacher
Browse files

added vectorization

parent 8dbe9de9

cmake/CMakeLists.txt

+3 −3
Original line number Diff line number Diff line
@@ -9,9 +9,9 @@ set(SOURCE_FILES main.cpp blasJoin/blasJoin.cpp ${UTIL_SOURCES}) 
#####################

# NDDEBUG turns off asserts

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

  set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} -march=knl -mtune=knl -fpic -ffast-math -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")

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

  set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} -qopenmp -axCOMMON-AVX512 -O3 -liomp5 -lpthread -lmkl_core -lmkl_intel_lp64 -lmkl_intel_thread")

  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")

  # xeon phi (knl) specific:

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

endif()
@@ -27,7 +27,7 @@ add_definitions(-DNUM_THREADS=64) 
if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")

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

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

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

endif()



# adding MKL include directory

cmake/blasJoin/blasJoin.cpp

+13 −25
Original line number Diff line number Diff line
@@ -20,8 +20,7 @@ double blasJoin(const double *x, const size_t N, const size_t D, const double EP 
    const double EPS_SQUARED = (EPS * EPS) / 4;



    // pre-compute the values P(i) and P(j)



    __assume_aligned(x, ALIGNMENT);

    #pragma simd

    for ( int i = 0 ; i < N ; i++ ){

        p[i] = EPS_SQUARED - cblas_ddot(D, &x[i*D], 1, &x[i*D], 1) / 2.0;

    }
@@ -36,10 +35,11 @@ double blasJoin(const double *x, const size_t N, const size_t D, const double EP 
    timer.start();

    // lower triangle excluding the main diagonal



    __assume_aligned(x, ALIGNMENT);

    for ( size_t i = 1 ; i < NRC ; i++ ){

    #pragma ivdep

    #pragma vector always

    for ( int i = 1 ; i < NRC ; i++ ){

        blockRow = ( i + 1 == NRC ) ? N - (BLOCKSIZE * i ) : BLOCKSIZE;

        for ( size_t j = 0 ; j < i ; j++ ){

        for ( int j = 0 ; j < i ; j++ ){

            blockCol = ( j + 1 == NRC ) ? N - (BLOCKSIZE * (j + 1) ) : BLOCKSIZE;



            // perform regular matrix multiplication
@@ -52,17 +52,12 @@ double blasJoin(const double *x, const size_t N, const size_t D, const double EP 
            #pragma omp parallel for reduction(+:accum) collapse(2)

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

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

                    // if (i*BLOCKSIZE + l >= N)

                    //     printf("%d * %d + %d\n", i, BLOCKSIZE, l);

                    assert(i*BLOCKSIZE+k < N);

                    assert(j*BLOCKSIZE+l < N);

                    // if (k*blockCol + l >= BLOCKSIZE * BLOCKSIZE )

                    //     printf("%d * %d + %d, %d, %d\n", k, blockCol, l, blockRow, NRC);

                    assert(k*blockCol + l < BLOCKSIZE * BLOCKSIZE);

                    iresult[k*blockCol + l] += ( p[i*BLOCKSIZE+k] + p[j*BLOCKSIZE+l] );

                    // change sign of double to 0 or 1

                    accum +=  ( (long long) iresult[k * blockCol + l] ) >> 63;

                    // accum += copysign(1.0,iresult[k * blockCol + l]);

                    /* with the 63-rshift we find out whether we are +/- and sum it up to efficiently determine count of join-partners

                     * prevent RAW-dependency with one-liner instead of the following:

                        accum +=  ( (long long) iresult[k * NRC + l] ) >> 63;

                        iresult[k*blockCol + l] += ( p[i*BLOCKSIZE + l] + p[i*BLOCKSIZE+k] );

                    */

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

                }

            }
@@ -71,8 +66,7 @@ double blasJoin(const double *x, const size_t N, const size_t D, const double EP 
    }



    // main diagonal

    __assume_aligned(x, ALIGNMENT);

    for ( size_t i = 0 ; i < NRC ; i++ ){

    for ( int i = 0 ; i < NRC ; i++ ){

        blockRow = ( i + 1 == NRC ) ? N - (BLOCKSIZE * i ) : BLOCKSIZE;

        // perform symmetric matrix multiplication on lower triangle only

        // C := alpha*A*A' + beta*C
@@ -85,13 +79,7 @@ double blasJoin(const double *x, const size_t N, const size_t D, const double EP 
        #pragma omp parallel for reduction(+:accum) collapse(2)

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

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

                assert(i*BLOCKSIZE + l < N);

                assert(i*BLOCKSIZE + k < N);

                assert(k*blockCol + l < BLOCKSIZE * BLOCKSIZE);

                iresult[k*blockCol + l] += ( p[i*BLOCKSIZE + l] + p[i*BLOCKSIZE+k] );

                // change sign of double to 0 or 1

                accum +=  ( (long long) iresult[k * NRC + l] ) >> 63;

                // accum += copysign(1.0,iresult[k * blockCol + l]);

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

            }

        }