alifarazz

When you're converting from microseconds to milliseconds, or go from higher precision to a lower one, or downscaling in any form; you have to be careful not to just do a divide. You also need to adjust the y-intercept. There are better methods for higher-accuracy downscaling methods, but they're more complex than just an add and a divide (also, the divide is a division by a constant, so it gets implemented as a multiplication and a bunch of logical shifts).

Blue line is the high-precision that we're converting from, orange line is what we get if we just int divide by scale, green line is what we get if we first add half of the divisor first and then divide. As you see, the orange line is closer to the ground truth that we're converting from. However, it overshoots the ground truth at of the time, which is something to be aware of.

alifarazz

#!/usr/bin/env bash

# Get path to current script, 
if [ -n "${SLURM_JOB_ID:-}" ] ; then
    SCRIPT_PATH="$(scontrol show job "$SLURM_JOB_ID" | awk -F= '/Command=/{print $2}')"
else
    SCRIPT_PATH="$(readlink -f "${BASH_SOURCE[0]}")"
fi

SCRIPT_DIR="$(dirname "${SCRIPT_PATH}")"

alifarazz

On an Intel i7-13620H, running Linux 6.12.52-lts and gcc 15.2.1 without sanitizers:

$ ./umonitor-umwait-semaphore 16 3
Semaphore initialized with count = 3.
Spawning 16 worker threads...

Main thread: Releasing an additional permit after some time.

All workers finished.

alifarazz

• How to compile:

  • $ g++ -DBIG_INPUT -Ofast -mtune=native -march=native -pipe -funroll-loops -Wall -Wextra -Wcast-align -Wchar-subscripts -Wshadow -Wunused dijkstra.cc -o dijkstra -flto
    or
  • $ clang++ -stdlib=libc++ -DBIG_INPUT -Ofast -mtune=native -march=native -funroll-loops -Wall -Wextra -Wcast-align -Wchar-subscripts -Wshadow -Wunused dijkstra.cc -o dijkstra -flto

• Results on a Intel Core i7-13620H, compiled with gcc, with 1024x1024 input:

  $ numactl --physcpubind=2 hyperfine -N -i -w 30 -r 100 './dijkstra'
  Benchmark 1: ./dijkstra

Time (mean ± σ): 3.2 ms ± 0.0 ms [User: 2.9 ms, System: 0.3 ms]

alifarazz

# in $HOME/.config/tmux/tmux.conf
# set-option -g default-shell "path to fish"
# I copied most of this off someone's blog.

# # remap prefix from 'C-b' to 'C-a'
# unbind C-b
# set-option -g prefix C-a
# bind-key C-a send-prefix

# split panes using | and -

alifarazz

#!/usr/bin/env bash
# SPDX-License-Identifier: AGPL-3.0-or-later

######### NOTE: HOW TO USE #########################################
# $ bash sync-with-remote.bash <path to .(git)ignore file> <local path to watch for>


set -euEo pipefail
# set -x  # show every command being executed, used for debug

alifarazz

// g++ -std=c++20 cpu_burn.cc -o cpu_burn

#include <bits/stdc++.h>

constexpr size_t maxn = 192;
constexpr int iter = 1024 * 1024;

std::array<std::jthread, maxn> threads;
std::array<unsigned int, (1024ull * 1024 * 512)> data;

alifarazz

/// SPDX-License-Identifier: GPL-2.0-or-later
#define _GNU_SOURCE

#include <linux/prctl.h> /* Definition of PR_* constants */
#include <sys/prctl.h>

#include <sys/syscall.h> /* for __NR_write */
#include <ucontext.h>    /* REG_RBP, REG_RSP */
#include <unistd.h>      /* for write(2) */

alifarazz

// g++ -O3 -Wall -Wextra -Wconversion -Wshadow -Wpedantic -std=c++20 -march=alderlake umwait.cc -o umwait.elf
// ./umwait.elf or ./umwait.elf w
#include <atomic>
#include <iostream>
#include <latch>
#include <thread>

#include <immintrin.h>
#include <x86intrin.h>

alifarazz

#include <cstdio>

int f(int *ap, float *bp) {
  *ap = 1;
  *bp = 2.0f;
  return *ap;
}

int main() {
  int a = 0; float b = 0.0f;