SpMV.cpp

/**
Copyright (c) 2015, Intel Corporation. All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
    * Redistributions of source code must retain the above copyright
      notice, this list of conditions and the following disclaimer.
    * Redistributions in binary form must reproduce the above copyright
      notice, this list of conditions and the following disclaimer in the
      documentation and/or other materials provided with the distribution.
    * Neither the name of the Intel Corporation nor the
      names of its contributors may be used to endorse or promote products
      derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL INTEL CORPORATION BE LIABLE FOR ANY
DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/

#include <algorithm>

#include "CSR.hpp"

using namespace std;

namespace SpMP
{

template<class T>
static void SpMV_(
  int m,
  T *w,
  T alpha,
  const int *rowptr, const int *colidx, const T* values,
  const T *x,
  T beta,
  const T *y,
  T gamma)
{
  assert(w != x);

  int base = rowptr[0];

  rowptr -= base;
  colidx -= base;
  if (values) values -= base;

  w -= base;
  x -= base;
  y -= base;

//#define MEASURE_LOAD_BALANCE
#ifdef MEASURE_LOAD_BALANCE
  double barrierTimes[omp_get_max_threads()];
  double tBegin = omp_get_wtime();
#endif

#pragma omp parallel
  {
    int iBegin, iEnd;
    getLoadBalancedPartition(&iBegin, &iEnd, rowptr + base, m);
    iBegin += base;
    iEnd += base;

    if (1 == alpha && 0 == beta && 0 == gamma) {
      if (values) {
        for (int i = iBegin; i < iEnd; ++i) {
          T sum = 0;
          for (int j = rowptr[i]; j < rowptr[i + 1]; ++j) {
            sum += values[j]*x[colidx[j]];
          }
          w[i] = sum;
        }
      }
      else {
        // pattern-only matrix: assume all non-zero values are 1
        for (int i = iBegin; i < iEnd; ++i) {
          T sum = 0;
          for (int j = rowptr[i]; j < rowptr[i + 1]; ++j) {
            sum += x[colidx[j]];
          }
          w[i] = sum;
        }
      }
    }
    else if (-1 == alpha && 1 == beta && 0 == gamma) {
      if (values) {
        for (int i = iBegin; i < iEnd; ++i) {
          T sum = y[i];
          for (int j = rowptr[i]; j < rowptr[i + 1]; ++j) {
            sum -= values[j]*x[colidx[j]];
          }
          w[i] = sum;
        }
      }
      else {
        for (int i = iBegin; i < iEnd; ++i) {
          T sum = y[i];
          for (int j = rowptr[i]; j < rowptr[i + 1]; ++j) {
            sum -= x[colidx[j]];
          }
          w[i] = sum;
        }
      }
    }
    else {
      if (values) {
        for (int i = iBegin; i < iEnd; ++i) {
          T sum = 0;
          for (int j = rowptr[i]; j < rowptr[i + 1]; ++j) {
            sum += values[j]*x[colidx[j]];
          }
          w[i] = alpha*sum + beta*y[i] + gamma;
        }
      }
      else {
        // pattern-only matrix: assume all non-zero values are 1
        for (int i = iBegin; i < iEnd; ++i) {
          T sum = 0;
          for (int j = rowptr[i]; j < rowptr[i + 1]; ++j) {
            sum += x[colidx[j]];
          }
          w[i] = alpha*sum + beta*y[i] + gamma;
        }
      }
    }
#ifdef MEASURE_LOAD_BALANCE
    double t = omp_get_wtime();
    int tid = omp_get_thread_num();
    int nthreads = omp_get_num_threads();
#pragma omp barrier
    barrierTimes[tid] = omp_get_wtime() - t;

#pragma omp barrier
#pragma omp master
    {
      double tEnd = omp_get_wtime();
      double barrierTimeSum = 0;
      for (int i = 0; i < nthreads; ++i) {
        barrierTimeSum += barrierTimes[i];
      }
      printf("%f load imbalance = %f\n", tEnd - tBegin, barrierTimeSum/(tEnd - tBegin)/nthreads);
    }
#undef MEASURE_LOAD_BALANCE
#endif // MEASURE_LOAD_BALANCE
  } // omp parallel
}

void CSR::multiplyWithVector(
  double *w,
  double alpha, const double *x, double beta, const double *y, double gamma)
  const
{
  double *tmp_x = (double *)x;

  // if x and w point to the same array, copy x to a temporary buffer
  // in order to avoid possible race condition in SpMV_
  if (w == x) {
    tmp_x = MALLOC(double, m);
 
    #pragma omp parallel for
    for (int i=0; i<m; i++) {
      tmp_x[i] = x[i];
    }
  }
 
  SpMV_<double>(m, w, alpha, rowptr, colidx, values, tmp_x, beta, y, gamma);
 
  if (w == x) {
    FREE(tmp_x);
  }
}

void CSR::multiplyWithVector(double *w, const double *x) const
{
  return multiplyWithVector(w, 1, x, 0, w, 0);
}

template<class T>
static void SpMDM_(
  int m,
  int k, // width of W, X, and Y
  T *W, int wRowStride, int wColumnStride,
  T alpha,
  const int *rowptr, const int *colidx, const T* values,
  const T *X, int xRowStride, int xColumnStride,
  T beta,
  const T *Y, int yRowStride, int yColumnStride,
  T gamma)
{
  int base = rowptr[0];

  rowptr -= base;
  colidx -= base;
  values -= base;

  W -= base;
  X -= base;
  Y -= base;

//#define MEASURE_LOAD_BALANCE
#ifdef MEASURE_LOAD_BALANCE
  double barrierTimes[omp_get_max_threads()];
  double tBegin = omp_get_wtime();
#endif

#pragma omp parallel
  {
    int iBegin, iEnd;
    getLoadBalancedPartition(&iBegin, &iEnd, rowptr + base, m);
    iBegin += base;
    iEnd += base;

    T sum[k];
    if (1 == alpha && 0 == beta && 0 == gamma && 1 == wRowStride && 1 == xRowStride && 1 == yRowStride) {
      for (int i1 = iBegin; i1 < iEnd; ++i1) {
        for (int i3 = 0; i3 < k; ++i3) {
          sum[i3] = 0;
        }
        for (int i2 = rowptr[i1]; i2 < rowptr[i1 + 1]; ++i2) {
          for (int i3 = 0; i3 < k; ++i3) {
            sum[i3] += values[i2]*X[colidx[i2] + i3*xColumnStride];
          }
        }
        for (int i3 = 0; i3 < k; ++i3) {
          W[i1 + i3*wColumnStride] = sum[i3];
        }
      }
    }
    else {
      for (int i1 = iBegin; i1 < iEnd; ++i1) {
        for (int i3 = 0; i3 < k; ++i3) {
          sum[i3] = 0;
        }
        for (int i2 = rowptr[i1]; i2 < rowptr[i1 + 1]; ++i2) {
          for (int i3 = 0; i3 < k; ++i3) {
            sum[i3] += values[i2]*X[colidx[i2]*xRowStride + i3*xColumnStride];
          }
        }
        for (int i3 = 0; i3 < k; ++i3) {
          W[i1*wRowStride + i3*wColumnStride] =
            alpha*sum[i3] + beta*Y[i1*yRowStride + i3*yColumnStride] + gamma;
        }
      }
    }
#ifdef MEASURE_LOAD_BALANCE
    double t = omp_get_wtime();
#pragma omp barrier
    barrierTimes[tid] = omp_get_wtime() - t;

#pragma omp barrier
#pragma omp master
    {
      double tEnd = omp_get_wtime();
      double barrierTimeSum = 0;
      for (int i = 0; i < nthreads; ++i) {
        barrierTimeSum += barrierTimes[i];
      }
      printf("%f load imbalance = %f\n", tEnd - tBegin, barrierTimeSum/(tEnd - tBegin)/nthreads);
    }
#undef MEASURE_LOAD_BALANCE
#endif // MEASURE_LOAD_BALANCE
  } // omp parallel
}

void CSR::multiplyWithDenseMatrix(
  double *W, int k, int wRowStride, int wColumnStride,
  double alpha,
  const double *X, int xRowStride, int xColumnStride,
  double beta, const double *Y, int yRowStride, int yColumnStride,
  double gamma) const
{
  SpMDM_<double>(
    m, k,
    W, wRowStride, wColumnStride,
    alpha, rowptr, colidx, values,
    X, xRowStride, xColumnStride,
    beta,
    Y, yRowStride, yColumnStride,
    gamma);
}

void CSR::multiplyWithDenseMatrix(double *W, int k, const double *X) const
{
  return multiplyWithDenseMatrix(
    W, k, 1, m,
    1,
    X, 1, n,
    0, W, 1, m,
    0);
}

} // namespace SpMP