bench/benchBlasGemm.cpp

// g++ -O3 -DNDEBUG -I.. -L /usr/lib64/atlas/ benchBlasGemm.cpp -o benchBlasGemm -lrt -lcblas
// possible options:
//    -DEIGEN_DONT_VECTORIZE
//    -msse2

// #define EIGEN_DEFAULT_TO_ROW_MAJOR
#define _FLOAT

#include <iostream>

#include <Eigen/Core>
#include "BenchTimer.h"

// include the BLAS headers
extern "C" {
#include <cblas.h>
}
#include <string>

#ifdef _FLOAT
typedef float Scalar;
#define CBLAS_GEMM cblas_sgemm
#else
typedef double Scalar;
#define CBLAS_GEMM cblas_dgemm
#endif

typedef Eigen::Matrix<Scalar, Eigen::Dynamic, Eigen::Dynamic> MyMatrix;
void bench_eigengemm(MyMatrix& mc, const MyMatrix& ma, const MyMatrix& mb, int nbloops);
void check_product(int M, int N, int K);
void check_product(void);

int main(int argc, char* argv[]) {
// disable SSE exceptions
#ifdef __GNUC__
  {
    int aux;
    asm("stmxcsr   %[aux]           \n\t"
        "orl       $32832, %[aux]   \n\t"
        "ldmxcsr   %[aux]           \n\t"
        :
        : [aux] "m"(aux));
  }
#endif

  int nbtries = 1, nbloops = 1, M, N, K;

  if (argc == 2) {
    if (std::string(argv[1]) == "check")
      check_product();
    else
      M = N = K = atoi(argv[1]);
  } else if ((argc == 3) && (std::string(argv[1]) == "auto")) {
    M = N = K = atoi(argv[2]);
    nbloops = 1000000000 / (M * M * M);
    if (nbloops < 1) nbloops = 1;
    nbtries = 6;
  } else if (argc == 4) {
    M = N = K = atoi(argv[1]);
    nbloops = atoi(argv[2]);
    nbtries = atoi(argv[3]);
  } else if (argc == 6) {
    M = atoi(argv[1]);
    N = atoi(argv[2]);
    K = atoi(argv[3]);
    nbloops = atoi(argv[4]);
    nbtries = atoi(argv[5]);
  } else {
    std::cout << "Usage: " << argv[0] << " size  \n";
    std::cout << "Usage: " << argv[0] << " auto size\n";
    std::cout << "Usage: " << argv[0] << " size nbloops nbtries\n";
    std::cout << "Usage: " << argv[0] << " M N K nbloops nbtries\n";
    std::cout << "Usage: " << argv[0] << " check\n";
    std::cout << "Options:\n";
    std::cout << "    size       unique size of the 2 matrices (integer)\n";
    std::cout << "    auto       automatically set the number of repetitions and tries\n";
    std::cout << "    nbloops    number of times the GEMM routines is executed\n";
    std::cout << "    nbtries    number of times the loop is benched (return the best try)\n";
    std::cout << "    M N K      sizes of the matrices: MxN  =  MxK * KxN (integers)\n";
    std::cout << "    check      check eigen product using cblas as a reference\n";
    exit(1);
  }

  double nbmad = double(M) * double(N) * double(K) * double(nbloops);

  if (!(std::string(argv[1]) == "auto")) std::cout << M << " x " << N << " x " << K << "\n";

  Scalar alpha, beta;
  MyMatrix ma(M, K), mb(K, N), mc(M, N);
  ma = MyMatrix::Random(M, K);
  mb = MyMatrix::Random(K, N);
  mc = MyMatrix::Random(M, N);

  Eigen::BenchTimer timer;

  // we simply compute c += a*b, so:
  alpha = 1;
  beta = 1;

  // bench cblas
  // ROWS_A, COLS_B, COLS_A, 1.0,  A, COLS_A, B, COLS_B, 0.0, C, COLS_B);
  if (!(std::string(argv[1]) == "auto")) {
    timer.reset();
    for (uint k = 0; k < nbtries; ++k) {
      timer.start();
      for (uint j = 0; j < nbloops; ++j)
#ifdef EIGEN_DEFAULT_TO_ROW_MAJOR
        CBLAS_GEMM(CblasRowMajor, CblasNoTrans, CblasNoTrans, M, N, K, alpha, ma.data(), K, mb.data(), N, beta,
                   mc.data(), N);
#else
        CBLAS_GEMM(CblasColMajor, CblasNoTrans, CblasNoTrans, M, N, K, alpha, ma.data(), M, mb.data(), K, beta,
                   mc.data(), M);
#endif
      timer.stop();
    }
    if (!(std::string(argv[1]) == "auto"))
      std::cout << "cblas: " << timer.value() << " (" << 1e-3 * floor(1e-6 * nbmad / timer.value()) << " GFlops/s)\n";
    else
      std::cout << M << " : " << timer.value() << " ; " << 1e-3 * floor(1e-6 * nbmad / timer.value()) << "\n";
  }

  // clear
  ma = MyMatrix::Random(M, K);
  mb = MyMatrix::Random(K, N);
  mc = MyMatrix::Random(M, N);

  // eigen
  //   if (!(std::string(argv[1])=="auto"))
  {
    timer.reset();
    for (uint k = 0; k < nbtries; ++k) {
      timer.start();
      bench_eigengemm(mc, ma, mb, nbloops);
      timer.stop();
    }
    if (!(std::string(argv[1]) == "auto"))
      std::cout << "eigen : " << timer.value() << " (" << 1e-3 * floor(1e-6 * nbmad / timer.value()) << " GFlops/s)\n";
    else
      std::cout << M << " : " << timer.value() << " ; " << 1e-3 * floor(1e-6 * nbmad / timer.value()) << "\n";
  }

  std::cout << "l1: " << Eigen::l1CacheSize() << std::endl;
  std::cout << "l2: " << Eigen::l2CacheSize() << std::endl;

  return 0;
}

using namespace Eigen;

void bench_eigengemm(MyMatrix& mc, const MyMatrix& ma, const MyMatrix& mb, int nbloops) {
  for (uint j = 0; j < nbloops; ++j) mc.noalias() += ma * mb;
}

#define MYVERIFY(A, M)                  \
  if (!(A)) {                           \
    std::cout << "FAIL: " << M << "\n"; \
  }
void check_product(int M, int N, int K) {
  MyMatrix ma(M, K), mb(K, N), mc(M, N), maT(K, M), mbT(N, K), meigen(M, N), mref(M, N);
  ma = MyMatrix::Random(M, K);
  mb = MyMatrix::Random(K, N);
  maT = ma.transpose();
  mbT = mb.transpose();
  mc = MyMatrix::Random(M, N);

  MyMatrix::Scalar eps = 1e-4;

  meigen = mref = mc;
  CBLAS_GEMM(CblasColMajor, CblasNoTrans, CblasNoTrans, M, N, K, 1, ma.data(), M, mb.data(), K, 1, mref.data(), M);
  meigen += ma * mb;
  MYVERIFY(meigen.isApprox(mref, eps), ". * .");

  meigen = mref = mc;
  CBLAS_GEMM(CblasColMajor, CblasTrans, CblasNoTrans, M, N, K, 1, maT.data(), K, mb.data(), K, 1, mref.data(), M);
  meigen += maT.transpose() * mb;
  MYVERIFY(meigen.isApprox(mref, eps), "T * .");

  meigen = mref = mc;
  CBLAS_GEMM(CblasColMajor, CblasTrans, CblasTrans, M, N, K, 1, maT.data(), K, mbT.data(), N, 1, mref.data(), M);
  meigen += (maT.transpose()) * (mbT.transpose());
  MYVERIFY(meigen.isApprox(mref, eps), "T * T");

  meigen = mref = mc;
  CBLAS_GEMM(CblasColMajor, CblasNoTrans, CblasTrans, M, N, K, 1, ma.data(), M, mbT.data(), N, 1, mref.data(), M);
  meigen += ma * mbT.transpose();
  MYVERIFY(meigen.isApprox(mref, eps), ". * T");
}

void check_product(void) {
  int M, N, K;
  for (uint i = 0; i < 1000; ++i) {
    M = internal::random<int>(1, 64);
    N = internal::random<int>(1, 768);
    K = internal::random<int>(1, 768);
    M = (0 + M) * 1;
    std::cout << M << " x " << N << " x " << K << "\n";
    check_product(M, N, K);
  }
}