xref: /petsc/src/benchmarks/streams/SSEVersion.c (revision abdd934af67c1cc360c4d0cfcf2de16e188d627e)

static const char help[] = "STREAM benchmark specialized for SSE2\n\\n";

/* Note: this file has been modified significantly from its original version */
#include <emmintrin.h>
#include <petsctime.h>
#include <petscsys.h>
#if defined(HAVE_NUMA)
  #include <numa.h>
#endif
#include <float.h>

#if !defined(SSE2)
  #define SSE2 1
#endif
#if !defined(__SSE2__)
  #error SSE2 instruction set is not enabled, try adding -march=native to CFLAGS or disable by adding -DSSE2=0
#endif
#if !defined(PREFETCH_NTA) /* Use software prefetch and set non-temporal policy so that lines evicted from L1D will not subsequently reside in L2 or L3. */
  #define PREFETCH_NTA 1
#endif
#if !defined(STATIC_ALLOC) /* Statically allocate the vectors. Most platforms do not find physical pages when memory is allocated, therefore the faulting strategy still affects performance. */
  #define STATIC_ALLOC 0
#endif
#if !defined(FAULT_TOGETHER) /* Faults all three vectors together which usually interleaves DRAM pages in physical memory. */
  #define FAULT_TOGETHER 0
#endif
#if !defined(USE_MEMCPY) /* Literally call memcpy(3) for the COPY benchmark. Some compilers detect the unoptimized loop as memcpy and call this anyway. */
  #define USE_MEMCPY 0
#endif

/*
 * Program: Stream
 * Programmer: Joe R. Zagar
 * Revision: 4.0-BETA, October 24, 1995
 * Original code developed by John D. McCalpin
 *
 * This program measures memory transfer rates in MB/s for simple
 * computational kernels coded in C.  These numbers reveal the quality
 * of code generation for simple uncacheable kernels as well as showing
 * the cost of floating-point operations relative to memory accesses.
 *
 * INSTRUCTIONS:
 *
 *       1) Stream requires a good bit of memory to run.  Adjust the
 *          value of 'N' (below) to give a 'timing calibration' of
 *          at least 20 clock-ticks.  This will provide rate estimates
 *          that should be good to about 5% precision.
 */

#define N      4000000
#define NTIMES 100
#define OFFSET 0

#define HLINE "-------------------------------------------------------------\n"

#if !defined(MIN)
  #define MIN(x, y) ((x) < (y) ? (x) : (y))
#endif
#if !defined(MAX)
  #define MAX(x, y) ((x) > (y) ? (x) : (y))
#endif

#if STATIC_ALLOC
double a[N + OFFSET], b[N + OFFSET], c[N + OFFSET];
#endif

static int    checktick(void);
static double Second(void);

int main(int argc, char *argv[])
{
  const char  *label[4]   = {"Copy", "Scale", "Add", "Triad"};
  const double bytes[4]   = {2 * sizeof(double) * N, 2 * sizeof(double) * N, 3 * sizeof(double) * N, 3 * sizeof(double) * N};
  double       rmstime[4] = {0}, maxtime[4] = {0}, mintime[4] = {FLT_MAX, FLT_MAX, FLT_MAX, FLT_MAX};
  int          quantum;
  int          BytesPerWord, j, k, size;
  PetscInt     node = -1;
  double       scalar, t, times[4][NTIMES];
#if !STATIC_ALLOC
  double *PETSC_RESTRICT a, *PETSC_RESTRICT b, *PETSC_RESTRICT c;
#endif

  PetscCall(PetscInitialize(&argc, &argv, 0, help));
  PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size));
  PetscCall(PetscOptionsGetInt(NULL, NULL, "-node", &node, NULL));
  /* --- SETUP --- determine precision and check timing --- */

  PetscPrintf(PETSC_COMM_WORLD, HLINE);
  BytesPerWord = sizeof(double);
  PetscPrintf(PETSC_COMM_WORLD, "This system uses %d bytes per DOUBLE PRECISION word.\n", BytesPerWord);

  PetscPrintf(PETSC_COMM_WORLD, HLINE);
  PetscPrintf(PETSC_COMM_WORLD, "Array size = %d, Offset = %d\n", N, OFFSET);
  PetscPrintf(PETSC_COMM_WORLD, "Total memory required = %.1f MB per process.\n", (3 * N * BytesPerWord) / 1048576.0);
  PetscPrintf(PETSC_COMM_WORLD, "Each test is run %d times, but only\n", NTIMES);
  PetscPrintf(PETSC_COMM_WORLD, "the *best* time for each is used.\n");

  /* Get initial value for system clock. */

#if !STATIC_ALLOC
  if (node == -1) {
    posix_memalign((void **)&a, 64, N * sizeof(double));
    posix_memalign((void **)&b, 64, N * sizeof(double));
    posix_memalign((void **)&c, 64, N * sizeof(double));
  } else if (node == -2) {
    a = malloc(N * sizeof(double));
    b = malloc(N * sizeof(double));
    c = malloc(N * sizeof(double));
  #if defined(HAVE_NUMA)
  } else {
    a = numa_alloc_onnode(N * sizeof(double), node);
    b = numa_alloc_onnode(N * sizeof(double), node);
    c = numa_alloc_onnode(N * sizeof(double), node);
  #endif
  }
#endif
#if FAULT_TOGETHER
  for (j = 0; j < N; j++) {
    a[j] = 1.0;
    b[j] = 2.0;
    c[j] = 0.0;
  }
#else
  for (j = 0; j < N; j++) a[j] = 1.0;
  for (j = 0; j < N; j++) b[j] = 2.0;
  for (j = 0; j < N; j++) c[j] = 0.0;
#endif

  PetscPrintf(PETSC_COMM_WORLD, HLINE);

  if ((quantum = checktick()) >= 1) PetscPrintf(PETSC_COMM_WORLD, "Your clock granularity/precision appears to be %d microseconds.\n", quantum);
  else PetscPrintf(PETSC_COMM_WORLD, "Your clock granularity appears to be less than one microsecond.\n");

  t = Second();
  for (j = 0; j < N; j++) a[j] = 2.0E0 * a[j];
  t = 1.0E6 * (Second() - t);

  PetscPrintf(PETSC_COMM_WORLD, "Each test below will take on the order of %d microseconds.\n", (int)t);
  PetscPrintf(PETSC_COMM_WORLD, "   (= %d clock ticks)\n", (int)(t / quantum));
  PetscPrintf(PETSC_COMM_WORLD, "Increase the size of the arrays if this shows that\n");
  PetscPrintf(PETSC_COMM_WORLD, "you are not getting at least 20 clock ticks per test.\n");

  PetscPrintf(PETSC_COMM_WORLD, HLINE);

  PetscPrintf(PETSC_COMM_WORLD, "WARNING -- The above is only a rough guideline.\n");
  PetscPrintf(PETSC_COMM_WORLD, "For best results, please be sure you know the\n");
  PetscPrintf(PETSC_COMM_WORLD, "precision of your system timer.\n");
  PetscPrintf(PETSC_COMM_WORLD, HLINE);

  /* --- MAIN LOOP --- repeat test cases NTIMES times --- */

  scalar = 3.0;
  for (k = 0; k < NTIMES; k++) {
    MPI_Barrier(PETSC_COMM_WORLD);
    /* ### COPY: c <- a ### */
    times[0][k] = Second();
    MPI_Barrier(PETSC_COMM_WORLD);
#if USE_MEMCPY
    memcpy(c, a, N * sizeof(double));
#elif SSE2
    for (j = 0; j < N; j += 8) {
      _mm_stream_pd(c + j + 0, _mm_load_pd(a + j + 0));
      _mm_stream_pd(c + j + 2, _mm_load_pd(a + j + 2));
      _mm_stream_pd(c + j + 4, _mm_load_pd(a + j + 4));
      _mm_stream_pd(c + j + 6, _mm_load_pd(a + j + 6));
  #if PREFETCH_NTA
      _mm_prefetch(a + j + 64, _MM_HINT_NTA);
  #endif
    }
#else
    for (j = 0; j < N; j++) c[j] = a[j];
#endif
    MPI_Barrier(PETSC_COMM_WORLD);
    times[0][k] = Second() - times[0][k];

    /* ### SCALE: b <- scalar * c ### */
    times[1][k] = Second();
    MPI_Barrier(PETSC_COMM_WORLD);
#if SSE2
    {
      __m128d scalar2 = _mm_set1_pd(scalar);
      for (j = 0; j < N; j += 8) {
        _mm_stream_pd(b + j + 0, _mm_mul_pd(scalar2, _mm_load_pd(c + j + 0)));
        _mm_stream_pd(b + j + 2, _mm_mul_pd(scalar2, _mm_load_pd(c + j + 2)));
        _mm_stream_pd(b + j + 4, _mm_mul_pd(scalar2, _mm_load_pd(c + j + 4)));
        _mm_stream_pd(b + j + 6, _mm_mul_pd(scalar2, _mm_load_pd(c + j + 6)));
  #if PREFETCH_NTA
        _mm_prefetch(c + j + 64, _MM_HINT_NTA);
  #endif
      }
    }
#else
    for (j = 0; j < N; j++) b[j] = scalar * c[j];
#endif
    MPI_Barrier(PETSC_COMM_WORLD);
    times[1][k] = Second() - times[1][k];

    /* ### ADD: c <- a + b ### */
    times[2][k] = Second();
    MPI_Barrier(PETSC_COMM_WORLD);
#if SSE2
    {
      for (j = 0; j < N; j += 8) {
        _mm_stream_pd(c + j + 0, _mm_add_pd(_mm_load_pd(a + j + 0), _mm_load_pd(b + j + 0)));
        _mm_stream_pd(c + j + 2, _mm_add_pd(_mm_load_pd(a + j + 2), _mm_load_pd(b + j + 2)));
        _mm_stream_pd(c + j + 4, _mm_add_pd(_mm_load_pd(a + j + 4), _mm_load_pd(b + j + 4)));
        _mm_stream_pd(c + j + 6, _mm_add_pd(_mm_load_pd(a + j + 6), _mm_load_pd(b + j + 6)));
  #if PREFETCH_NTA
        _mm_prefetch(a + j + 64, _MM_HINT_NTA);
        _mm_prefetch(b + j + 64, _MM_HINT_NTA);
  #endif
      }
    }
#else
    for (j = 0; j < N; j++) c[j] = a[j] + b[j];
#endif
    MPI_Barrier(PETSC_COMM_WORLD);
    times[2][k] = Second() - times[2][k];

    /* ### TRIAD: a <- b + scalar * c ### */
    times[3][k] = Second();
    MPI_Barrier(PETSC_COMM_WORLD);
#if SSE2
    {
      __m128d scalar2 = _mm_set1_pd(scalar);
      for (j = 0; j < N; j += 8) {
        _mm_stream_pd(a + j + 0, _mm_add_pd(_mm_load_pd(b + j + 0), _mm_mul_pd(scalar2, _mm_load_pd(c + j + 0))));
        _mm_stream_pd(a + j + 2, _mm_add_pd(_mm_load_pd(b + j + 2), _mm_mul_pd(scalar2, _mm_load_pd(c + j + 2))));
        _mm_stream_pd(a + j + 4, _mm_add_pd(_mm_load_pd(b + j + 4), _mm_mul_pd(scalar2, _mm_load_pd(c + j + 4))));
        _mm_stream_pd(a + j + 6, _mm_add_pd(_mm_load_pd(b + j + 6), _mm_mul_pd(scalar2, _mm_load_pd(c + j + 6))));
  #if PREFETCH_NTA
        _mm_prefetch(b + j + 64, _MM_HINT_NTA);
        _mm_prefetch(c + j + 64, _MM_HINT_NTA);
  #endif
      }
    }
#else
    for (j = 0; j < N; j++) a[j] = b[j] + scalar * c[j];
#endif
    MPI_Barrier(PETSC_COMM_WORLD);
    times[3][k] = Second() - times[3][k];
  }

  /* --- SUMMARY --- */

  for (k = 0; k < NTIMES; k++)
    for (j = 0; j < 4; j++) {
      rmstime[j] = rmstime[j] + (times[j][k] * times[j][k]);
      mintime[j] = MIN(mintime[j], times[j][k]);
      maxtime[j] = MAX(maxtime[j], times[j][k]);
    }

  PetscPrintf(PETSC_COMM_WORLD, "%8s:  %11s  %11s  %11s  %11s  %11s\n", "Function", "Rate (MB/s)", "Total (MB/s)", "RMS time", "Min time", "Max time");
  for (j = 0; j < 4; j++) {
    rmstime[j] = sqrt(rmstime[j] / (double)NTIMES);
    PetscPrintf(PETSC_COMM_WORLD, "%8s: %11.4f  %11.4f  %11.4f  %11.4f  %11.4f\n", label[j], 1.0e-06 * bytes[j] / mintime[j], size * 1.0e-06 * bytes[j] / mintime[j], rmstime[j], mintime[j], maxtime[j]);
  }
  PetscCall(PetscFinalize());
  return 0;
}

static double Second()
{
  double t;
  PetscTime(&t);
  return t;
}

#define M 20
static int checktick(void)
{
  int    i, minDelta, Delta;
  double t1, t2, timesfound[M];

  /*  Collect a sequence of M unique time values from the system. */

  for (i = 0; i < M; i++) {
    t1 = Second();
    while ((t2 = Second()) - t1 < 1.0E-6) { }
    timesfound[i] = t1 = t2;
  }

  /*
   * Determine the minimum difference between these M values.
   * This result will be our estimate (in microseconds) for the
   * clock granularity.
   */

  minDelta = 1000000;
  for (i = 1; i < M; i++) {
    Delta    = (int)(1.0E6 * (timesfound[i] - timesfound[i - 1]));
    minDelta = MIN(minDelta, MAX(Delta, 0));
  }

  return minDelta;
}