FabioLuporini

#include <stdio.h>
#include "stdlib.h"
#include "math.h"
#include "sys/time.h"
#include "omp.h"

/*
 * clear
 * aompcc -O3 -g -fPIC -Wall -Wno-unused-result -Wno-unused-variable -ffast-math omp_off_leak_mfe.c -o omp_off_leak_mfe
 * ./omp_off_leak_mfe

FabioLuporini

#include <stdio.h>
#include "stdlib.h"
#include "math.h"
#include "sys/time.h"
#include "omp.h"

/*
 * clear
 * aompcc -O3 -g -fPIC -Wall -Wno-unused-result -Wno-unused-variable -ffast-math omp_off_leak_mfe.c -o omp_off_leak_mfe
 * ./omp_off_leak_mfe

FabioLuporini

import numpy as np
from devito import *  # noqa
from examples.seismic import TimeAxis, RickerSource, Receiver

shape = (101, 101)
extent = (1000, 1000)
origin = (0., 0.)

v = np.empty(shape, dtype=np.float32)
v[:, :51] = 1.5

FabioLuporini

/* Compile and run with

clang -O3 -g -fPIC -Wall -Wno-unused-result -Wno-unused-variable -ffast-math -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda mfe.c -o mfe ; ./mfe

*/

#include "omp.h"
#include <stdio.h>

int main(int argc, char* argv[])

FabioLuporini

#include "omp.h"
#include <stdio.h>

int main(int argc, char* argv[])
{
  int t_size = 1;
  int x_size = 4;
  int y_size = 4;
  int z_size = 4;

FabioLuporini

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Time Dependent Born\n",
    "\n",
    "Here we code the propagation time extension Born operators."
   ]

FabioLuporini

#define _POSIX_C_SOURCE 200809L
#include "stdlib.h"
#include "math.h"
#include "sys/time.h"
#include "mpi.h"
#include "openacc.h"

struct dataobj
{
  void *restrict data;

FabioLuporini

# First, make sure icc/mpiicc are in PATH, typically via sourcing :

source /opt/intel/bin/compilervars.sh intel64

# Then :

export CC=icc
export MPICC=mpiicc
pip install mpi4py

FabioLuporini

RUNTIME COMPARISON
------------------

conclusions: master matches domain-allocation switch in both ``core`` and ``yask`` (no slowdowns, no speedups)

reports below for acoustic SO=4 and tti SO=4 with max opt level, openmp, pinning etc

>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
# *domain-allocation switch*: acoustic, core, max opt level

FabioLuporini

  for (int x = x_start, xs0 = x + x_ub0; x < x_finish; x += 1, xs0 = xs0 + 1)
  {
    for (int y = y_start, ys0 = y + y_ub0; y < y_finish; y += 1, ys0 = ys0 + 1)
    {   
      #pragma ivdep
      #pragma omp simd
      for (int z = z_start; z < z_finish; z += 1)
      {   
        float s0 = 1.0e-1F*v[t0][xs0][ys0 - 1][z] - 7.5e-2F*v[t0][xs0][ys0 - 1][z - 1] - 2.5e-2F*v[t0][xs0][ys0 - 1][z + 1]; 
        float s1 = 2.5e-2F*(-v[t0][xs0 - 1][ys0 - 1][z - 1] + v[t0][xs0 + 1][ys0 - 1][z - 1]);