louchenyao

// nvcc bench_seq.cu -O3 -gencode arch=compute_70,code=sm_70
// Result on V100:
// bench_seq        max throughput: 769.856 GiB/s
// bench_seq_unroll max throughput: 833.067 GiB/s

#include <iostream>

__device__
int64_t reduce(int64_t *buf, uint64_t s) {
    // The more effecient way is to reduce within the wrap firstly, but it's not the bottleneck

louchenyao

CREATE TABLE date_ (
	d_datekey          INTEGER,
	d_date             TEXT ENCODING DICT(32),
	d_dayofweek        TEXT ENCODING DICT(8),
	d_month            TEXT ENCODING DICT(8),
	d_year             SMALLINT,
	d_yearmonthnum     INTEGER,
	d_yearmonth        TEXT ENCODING DICT(16),
	d_daynuminweek     SMALLINT,
	d_daynuminmonth    SMALLINT,

louchenyao

#include <algorithm>
#include <cstdio>
#include <cstdlib>

template <typename T>
void cmp(T &a, T &b) {
    if (a > b) {
        std::swap(a, b);
    }
}

louchenyao

#include <cassert>
#include <cstdio>
#include <cmath>
#include <random>

int zipf(double alpha, int n, std::mt19937_64 &rng)
{
  static bool first = true;      // Static first time flag
  static double c = 0;          // Normalization constant
  static double *sum_probs;     // Pre-calculated sum of probabilities

louchenyao

// This is a C++ CUDA port of the Zipf Distribution generator from https://github.com/jonhoo/rust-zipf/blob/master/src/lib.rs, written in Rust.

#include <cassert>
#include <cstdio>
#include <cmath>
#include <curand_kernel.h>

namespace Zipf {
    
struct ZipfDist {

louchenyao

// clang++ two_or_three.cpp -O3 --std=c++17

// results:
// two  : 130.3 ns
// two+  : 131.0 ns
// three: 133.2 ns

#include <stdlib.h>
#include <stdio.h>
#include <string.h>