Test the non-AVX, AVX2 and AVX-512 speeds for various types of CPU intensive loops, across various active core counts. Currently it is Linux only, but the basic testing mechanism could be ported to OSX and Windows as well (help welcome).
make
You should load the msr
kernel module if it is not already loaded. This is as simple as:
modprobe msr
Or as complex as (if you want nice messages about what happened):
lsmod | grep -q msr && echo "MSR already loaded" || { echo "Loading MSR module"; sudo modprobe msr ; }
You get the most info running as root (since we can read various MSRs to calculate the frequency directly):
sudo ./avx-turbo
You can also run it without root, but you only get the "Mops" reading (but this can be read directly as frequency for the 1-latency tests).
The default behavior for ./avx-turbo is to run tests with various thread counts, but with the same test on each thread. For example, the avx256_fma
test means that the same FMA-using test code will be run on each test thread.
An alternate approach is availe with so-called spec-based tests. This lets you specificy exactly what each thread in a test will run. The general form of a specification is: test1/thead-count1[,test2/thread-count2[,...]]
. For example,
if you run sudo ./avx-turbo --spec avx256_fma/1,scalar_iadd/3
you'll get one copy of avx256_fma
and three copies of scalar_iadd
running in parallel.
This mode is useful to testing that happens when not all cores are doing the same thing.
Try:
./avx-turbo --help
for a summary of some options something like this:
./avx-turbo {OPTIONS}
avx-turbo: Determine AVX2 and AVX-512 downclocking behavior
OPTIONS:
-h, --help Display this help menu
--force-tsc-calibrate Force manual TSC calibration loop, even
if cpuid TSC Hz is available
--no-pin Don't try to pin threads to CPU - gives
worse results but works around affinity
issues on TravisCI
--verbose Output more info
--no-barrier Don't sync up threads before each test
(no real purpose)
--list List the available tests and their
descriptions
--allow-hyperthreads By default we try to filter down the
available cpus to include only physical
cores, but with this option we'll use
all logical cores meaning you'll run two
tests on cores with hyperthreading
--test=[TEST-ID] Run only the specified test (by ID)
--spec=[SPEC] Run a specific type of test specified by
a specification string
--iters=[ITERS] Run the test loop ITERS times (default
100000)
--min-threads=[MIN] The minimum number of threads to use
--max-threads=[MAX] The maximum number of threads to use
--warmup-ms=[MILLISECONDS] Warmup milliseconds for each thread
after pinning (default 100)
The output looks like this:
Running as root : [YES]
CPU supports AVX2 : [YES]
CPU supports AVX-512: [NO ]
cpuid = eax = 2, ebx = 216, ecx = 0, edx = 0
cpu: family = 6, model = 94, stepping = 3
tsc_freq = 2592.0 MHz (from cpuid leaf 0x15)
Will test up to 4 CPUs
============================== Threads: 1 ==============================
ID | Description | Mops | A/M-ratio | A/M-MHz | M/tsc-ratio
scalar_iadd | Scalar integer adds | 2594 | 1.00 | 2592 | 1.00
avx128_iadd | 128-bit integer adds | 2594 | 1.00 | 2592 | 1.00
avx128_imul | 128-bit integer muls | 519 | 1.00 | 2592 | 1.00
avx128_fma | 128-bit 64-bit FMAs | 649 | 1.00 | 2592 | 1.00
avx256_iadd | 256-bit integer adds | 2594 | 1.00 | 2592 | 1.00
avx256_imul | 256-bit integer muls | 519 | 1.00 | 2592 | 1.00
avx256_fma | 256-bit serial DP FMAs | 648 | 1.00 | 2592 | 1.00
avx256_fma_t | 256-bit parallel DP FMAs | 5189 | 1.00 | 2592 | 1.00
=========================================================================
============================== Threads: 2 ==============================
ID | Description | Mops | A/M-ratio | A/M-MHz | M/tsc-ratio
scalar_iadd | Scalar integer adds | 2593, 2593 | 1.00, 1.00 | 2592, 2592 | 1.00, 1.00
avx128_iadd | 128-bit integer adds | 2594, 2594 | 1.00, 1.00 | 2592, 2592 | 1.00, 1.00
avx128_imul | 128-bit integer muls | 519, 519 | 1.00, 1.00 | 2592, 2592 | 1.00, 1.00
avx128_fma | 128-bit 64-bit FMAs | 648, 649 | 1.00, 1.00 | 2592, 2592 | 1.00, 1.00
avx256_iadd | 256-bit integer adds | 2594, 2594 | 1.00, 1.00 | 2592, 2592 | 1.00, 1.00
avx256_imul | 256-bit integer muls | 519, 519 | 1.00, 1.00 | 2592, 2592 | 1.00, 1.00
avx256_fma | 256-bit serial DP FMAs | 648, 648 | 1.00, 1.00 | 2592, 2592 | 1.00, 1.00
avx256_fma_t | 256-bit parallel DP FMAs | 5188, 5189 | 1.00, 1.00 | 2592, 2592 | 1.00, 1.00
=========================================================================
============================== Threads: 3 ==============================
ID | Description | Mops | A/M-ratio | A/M-MHz | M/tsc-ratio
scalar_iadd | Scalar integer adds | 2594, 2594, 2594 | 1.00, 1.00, 1.00 | 2592, 2592, 2592 | 1.00, 1.00, 1.00
avx128_iadd | 128-bit integer adds | 2594, 2594, 2594 | 1.00, 1.00, 1.00 | 2592, 2592, 2592 | 1.00, 1.00, 1.00
avx128_imul | 128-bit integer muls | 519, 519, 519 | 1.00, 1.00, 1.00 | 2592, 2592, 2592 | 1.00, 1.00, 1.00
avx128_fma | 128-bit 64-bit FMAs | 649, 648, 648 | 1.00, 1.00, 1.00 | 2592, 2592, 2592 | 1.00, 1.00, 1.00
avx256_iadd | 256-bit integer adds | 2594, 2594, 2594 | 1.00, 1.00, 1.00 | 2592, 2592, 2592 | 1.00, 1.00, 1.00
avx256_imul | 256-bit integer muls | 519, 519, 519 | 1.00, 1.00, 1.00 | 2592, 2592, 2592 | 1.00, 1.00, 1.00
avx256_fma | 256-bit serial DP FMAs | 649, 648, 649 | 1.00, 1.00, 1.00 | 2592, 2592, 2592 | 1.00, 1.00, 1.00
avx256_fma_t | 256-bit parallel DP FMAs | 5190, 5189, 5190 | 1.00, 1.00, 1.00 | 2592, 2592, 2592 | 1.00, 1.00, 1.00
=========================================================================
============================== Threads: 4 ==============================
ID | Description | Mops | A/M-ratio | A/M-MHz | M/tsc-ratio
scalar_iadd | Scalar integer adds | 2594, 2594, 2594, 2594 | 1.00, 1.00, 1.00, 1.00 | 2592, 2592, 2592, 2592 | 1.00, 1.00, 1.00, 1.00
avx128_iadd | 128-bit integer adds | 2593, 2594, 2594, 2594 | 1.00, 1.00, 1.00, 1.00 | 2592, 2592, 2592, 2592 | 1.00, 1.00, 1.00, 1.00
avx128_imul | 128-bit integer muls | 519, 519, 519, 519 | 1.00, 1.00, 1.00, 1.00 | 2592, 2592, 2592, 2592 | 1.00, 1.00, 1.00, 1.00
avx128_fma | 128-bit 64-bit FMAs | 648, 648, 649, 648 | 1.00, 1.00, 1.00, 1.00 | 2592, 2592, 2592, 2592 | 1.00, 1.00, 1.00, 1.00
avx256_iadd | 256-bit integer adds | 2594, 2594, 2594, 2594 | 1.00, 1.00, 1.00, 1.00 | 2592, 2592, 2592, 2592 | 1.00, 1.00, 1.00, 1.00
avx256_imul | 256-bit integer muls | 519, 519, 519, 519 | 1.00, 1.00, 1.00, 1.00 | 2592, 2592, 2592, 2592 | 1.00, 1.00, 1.00, 1.00
avx256_fma | 256-bit serial DP FMAs | 648, 648, 648, 648 | 1.00, 1.00, 1.00, 1.00 | 2592, 2592, 2592, 2592 | 1.00, 1.00, 1.00, 1.00
avx256_fma_t | 256-bit parallel DP FMAs | 5189, 5189, 5189, 5189 | 1.00, 1.00, 1.00, 1.00 | 2592, 2592, 2592, 2592 | 1.00, 1.00, 1.00, 1.00
=========================================================================
The headings are:
ID
The ID for the test, which you can use with the--test
argument to only run a specific test (handy when you want to focus on one test to read the frequency externally, e.g., viaperf
).Description
Yes, it's a description.Mops
Million operations per second. Every test runs a loop of the same type of instruction and this is how many millions of those instructions were executed per second. This is handy since this value corresponds exactly to frequency in MHz for tests with serially dependent 1-latency instructions, which here are all the "integer adds" tests.A/M
This is the ratio of theAPERF
andMPERF
ratios exposed in an MSR. For details, see the Intel SDM Vol 3, but basically APERF is a free running counter of actual cycles (i.e., varying with the CPU frequency), while MPERF counts at a constant rate, usually the processor's nominal frequency. A ratio of 1.0 therefore means that the CPU was is running, on average, at the nominal frequency during the test (I had turbo off, that's why you see 1.00 everywhere). Lower than 1 means lower than nominal frequencies (e.g., due to running heavy AVX code).A/M-MHz
This is the measured frequency over the duration of the test, based on theAPERF
andMPERF
ratio described above, multiplied by the base (TSC) frequency. Note that this only counts "non-halted" periods, so if the CPU was running at 1000 MHz half the time but halted the other half of the time (due to a frequency transition), you'd see 1000 MHz here, not 500 MHz.M/tsc-ratio
This shows the ration of theMPERF
register to the TSC (time stamp counter) over the duration of the test. These counters count at the same rate, except thatMPERF
only counts "unhalted" cycles, while the TSC counts all cycles, so this ratio gives you an indication of the "lost" cycles due to halt events. A big source of halt events is frequency transitions in the turbo range: on my Skylake client CPU, any time another core starts up, the allowed turbo ratio changes, so the CPU halts for perhaps 20,000 cycles, so with moderate activity I often see ratios of 0.9 which means that 10% of the time my CPU is doing nothing. To get a "true" frequency, you should multiply this ratio by theA/M-MHz
column, which would be the actual average frequency, counting halted periods as zero.