#pragma once

#include <iostream>
#include <stdio.h>
#include <stdlib.h>
#include <sstream>
#include <fstream>
#include <string>
#include <vector>
#include <unordered_map>
#include <queue>
#include <assert.h>
#include <unistd.h>
#include <chrono>
#include <atomic>
#include <random>

#include <curand.h>
#include <cuda.h>
#include <nccl.h>
#include <cub/util_allocator.cuh>
#include "crystal/crystal.cuh"

#include "tbb/tbb.h"
#include "PerfEvent.hpp"
#include "papi.h"
#include "ssb_utils.h"

using namespace cub;
using namespace std;
using namespace tbb;

// #define SF 42
#define NUM_EVENTS 2
#define NUM_GPU 2
#define ALL_GPU 2
#define NUM_TABLE 5
#define NUM_COLUMN 25
#define NUM_PARTITION NUM_GPU