mtorture -- Machine Torture

Determining clock speed

Linux

lscpu
# CPU flags, look for sse4.2, avx, avx2, avx512
cat /proc/cpuinfo  | grep flags | tail -n1
su -
echo performance > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
cat /sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_cur_freq
cat /sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_max_freq
# Enter frequency into Timing.cxx (note, this is in kHz, multiply by 1000)
# Run the tests ...
echo ondemand > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor

OSX

sysctl machdep.cpu.features
sysctl hw.cpufrequency_min hw.cpufrequency hw.cpufrequency_max
# Enter the max value into Timing.cxx (note, this is in Hz, copy as is)

External documentation of interest

• Lists of instruction latencies and throughput for Intel, AMD and VIA CPUs: http://www.agner.org/optimize/instruction_tables.pdf

• Intel intrinsics guide: https://software.intel.com/sites/landingpage/IntrinsicsGuide/

Running tests

Makefile

Variables:

• OPT -- set optimiztion level, default is -O3
• USER_CPPFLAGS -- set extra cpp flags, especially -D for compilation
• USER_CXXFLAGS -- whatever you want to pass to compiler
• USER_LDFLAGS -- whatever you want to pass to linker

There is a special target echo that will print out the relevant varables, e.g:

  make CXX:=my-c++ OPT:=-O2 USER_CPPFLAGS:="-DTEST_FUNC=sum3_cube -DFOO=bar" echo

Defines used in tests

• TEST_FUNC -- the default function that gets called on the test object

To use this from a script, you'd have to do something like this:

  make USER_CPPFLAGS:="-DTEST_FUNC=sum3_cube" -W t1.cxx t1 t1-mic

Global variables and environment variables

Format: global type and name | environment name | default value, see also common.h.

• double g_test_duration | TEST_DURATION | 1.0 -- duration of each test in seconds

• double g_pre_test_frac | PRE_TEST_FRAC | 0.01 -- fraction of the above to use for loop calibration

• int g_n_vec_min | N_VEC_MIN | 8 -- starting vector size for the test

• int g_n_vec_max | N_VEC_MAX | 64*1024*1024 -- ending vector size