moyix

#!/usr/bin/env python3
from flask import Flask, request, redirect
import os
import subprocess
import sys
import yaml
import MySQLdb

app = Flask(__name__)

moyix

{
    "fp16": {
        "enabled": true,
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },
    "optimizer": {

moyix

Total instructions:     48093488942
Total SSE instructions:   100105422
Total XMM instructions:   877832653
Totals by sanitizer:
   ASAN: SSE:   39197160, XMM:  308790743
   MSAN: SSE:   29922931, XMM:  342062480
  UBSAN: SSE:   30985331, XMM:  226979430
All projects per sanitizer, sorted by percent of SSE instructions:
ASAN:                               SSE Instr /      Total =   Pct ↓   Wilson
    simd                         :    1122000 /   63479115 =  1.77 % ( 1.76 %)

moyix

The ffast and the Furious

This is a small and admittedly contrived demo showing how some weird but safe code could become vulnerable if run in an environment where some shared library has changed the FPU's FTZ/DAZ bits to force denormals to zero.

To run it:

# Create an empty file
$ touch gofast.c      

moyix

#===============================================================
#
#
#===============================================================
import shutil
from setuptools import find_packages, setup
from os.path import exists,join,relpath
import os
import stat

moyix

#!/usr/bin/env python

import sys, os
import platform
import ctypes as ct
import mmap
from enum import Enum
import importlib
import functools
import errno

moyix

import sys
import os
import re
import json
import zipfile
from collections import defaultdict, namedtuple
from collections.abc import Mapping
from email.parser import HeaderParser
from email.policy import compat32
from base64 import urlsafe_b64decode

moyix

#!/usr/bin/env python

import os
import sys
import subprocess as sp
import tempfile
import hashlib
script_dir = os.path.dirname(os.path.realpath(__file__))
sys.path.append(script_dir)
from fast_check_for_ffast_math import check_file

moyix

import sys
import mmap
from elftools.elf.elffile import ELFFile, ELFError
import struct

set_fast_math_code = bytes.fromhex('0fae5c24fc814c24fc408000000fae5424fcc3')

def load_bytes_from_elf(bindata, elf, vaddr, size):
    try:
        paddr = next(iter(elf.address_offsets(vaddr)))

moyix

$ objdump -s -j .init_array ./jaxlib/xla_extension.so | sed -e '1,/Contents/ d' | cut -c 10-44 | xxd -r -p | od -A none -w8 -t x8 --endian=little | addr2line -a -f -e ./jaxlib/xla_extension.so | paste -sd '  \n' | c++filt
0x000000000084c5e0 __cpu_indicator_init /dt9-src/libgcc/config/i386/cpuinfo.c:434
0x000000000084ca20 frame_dummy crtstuff.c:?
0x000000000079c440 _GLOBAL__sub_I_xla.cc xla.cc:?
0x000000000079c540 _GLOBAL__sub_I_dlpack.cc dlpack.cc:?
0x000000000079c5f0 _GLOBAL__sub_I_mlir.cc mlir.cc:?
0x000000000079c620 _GLOBAL__sub_I_ops.cc ops.cc:?
0x000000000079c650 _GLOBAL__sub_I_approx_topk.cc approx_topk.cc:?
0x000000000079c680 _GLOBAL__sub_I_approx_topk_shape.cc approx_topk_shape.cc:?
0x000000000079c6b0 _GLOBAL__sub_I_lu_decomposition.cc lu_decomposition.cc:?