CpuId is a package for the Julia programming language that enables you to
query the availability of specific CPU features with minimal run-time cost
using the assembly instruction
Status: considered a pre-beta version, ready for you to try out.
Works on Julia 0.5 and 0.6, on Linux, Mac and Windows with Intel compatible CPUs.
Besides the obvious reason to gather information for diagnostics, the CPU provides valuable information when aiming at increasing the efficiency of code. Such use-cases could be to tailor the size of working sets of data according to the available cache sizes, to detect when the code is executed in a virtual machine (hypervisor), or to determine the size of the largest SIMD registers available to choose the best algorithm for the current hardware.
This information is obtained by directly querying the CPU through the
assembly instruction which operates only using CPU register. In fact,
determining simple boolean feature flags through
cpuid can be even faster
than accessing a global variable in Julia, in particular if caches are cold.
Also, this provides a portable way to adapt code to unknown hardware if Julia
code is compiled into a static system image (sysimg), where constant globals are
not an option.
Same information may of course be collected from various sources, from Julia
itself or from the operating system, e.g. on Linux from
below for a few alternatives. However, the
is perfectly portable and highly efficient.
CpuId is a registered Julia package; use the package manager to install:
Or, if you're keen to get some intermediate updates, clone from GitHub master branch:
For practical examples on how CpuId could be used in production, see the follow-up package CpuHints.jl which aims at improving benchmarking accuracy by giving the CPU hints on when to load and store from and to caches. Further, see the package Perf.jl which is in an early development stage but aims at bringing Linux performance monitoring capabilities aka 'perf' to Julia.
See the diagnostic summary on your CPU by typing
julia> using CpuId julia> cpuinfo() Cpu Property Value ╾───────────────────╌───────────────────────────────────────────────────────────╼ Brand Intel(R) Xeon(R) CPU E3-1225 v5 @ 3.30GHz Vendor :Intel Architecture :Skylake Model Family: 6, Model: 94, Stepping: 3, Type: 0 Cores 4 physical cores, 4 logical cores (on executing CPU) No Hyperthreading detected Clock Frequencies 3300 / 3700 MHz (base/max), 100 MHz bus Data Cache Level 1:3 : (32, 256, 8192) kbytes 64 byte cache line size Address Size 48 bits virtual, 39 bits physical SIMD 256 bit = 32 byte max. SIMD vector size Time Stamp Counter TSC is accessible via `rdtsc` TSC runs at constant rate (invariant from clock frequency) Perf. Monitoring Performance Monitoring Counters (PMC) revision 4 Available hardware counters per logical core: 3 fixed-function counters of 48 bit width 8 general-purpose counters of 48 bit width Hypervisor No
This release covers a selection of fundamental and higher level functionality:
cpuinfo()generates the summary shown above (markdown string).
cpuvendor()allow the identification of the CPU.
cpuarchitecture()tries to infer the microarchitecture, currently only of Intel CPUs.
cpucores_total()to determine the number of physical and logical cores on the currently executing CPU, which typically share L3 caches and main memory bandwidth. If the result of both functions is equal, then the CPU does not use of hyperthreading.
physical_address_size()return the number of bits used in pointers. Useful when stealing a few bits from a pointer.
cachelinesize()gives the size in bytes of one cache line, which is typically 64 bytes.
cachesize()returns a tuple with the sizes of the data caches in bytes.
cpu_bus_frequency()give - if supported by the CPU, the base, maximum and bus clock frequencies. Use
has_cpu_frequencies()to check whether this property is supported.
hypervised()returns true when the CPU indicates that a hypervisor is running the operating system, aka a virtual machine. In that case,
hvvendor()may be invoked to get the, well, hypervisor vendor, and
hvversion()returns a dictionary of additional version tags.
hvinfo()generates a markdown summary of same dictionary.
simdbytes()return the size of the largest SIMD register available on the executing CPU.
perf_revision()to query the revision number of hardware performance monitoring counters, along with
perf_gen_bits()to determine the number and bit width of available fixed-function and general purpose counters per logical core.
cpucycle_id()let you directly get the CPU's time stamp counter, which is increased for every CPU clock cycle. This is a method to perform low overhead micro-benchmarking; though, technically, this uses the
rdtscpinstructions rather than
cpufeature(::Symbol)permits asking for the availability of a specific feature, and
cpufeaturetable()gives a complete overview of all detected features with a brief explanation, as shown below.
html julia> cpufeaturetable() Cpu Feature Description ╾────────────╌───────────────────────────────────────────────────────────────╼ 3DNowP 3D Now PREFETCH and PREFETCHW instructions ACPI Thermal monitor and software controlled clock facilities (MSR) ADX Intel ADX (Multi-Precision Add-Carry Instruction Extensions) AES AES encryption instruction set AHF64 LAHF and SAHF in PM64 APIC APIC on-chip (Advanced Programmable Interrupt Controller) AVX 256bit Advanced Vector Extensions, AVX AVX2 SIMD 256bit Advanced Vector Extensions 2 BMI1 Bit Manipulation Instruction Set 1 BMI2 Bit Manipulation Instruction Set 2 CLFLUSH CLFLUSHOPT Instructions CLFSH CLFLUSH instruction (SSE2) CMOV Conditional move CMOV and FCMOV instructions CX16 CMPXCHG16B instruction CX8 CMPXCHG8 instruction (64bit compare and exchange) ...
cpuid instruction is a generic way provided by the CPU vendor to obtain
basic hardware information. It provides data in form of boolean bit fields,
integer fields and strings, all packed in the returned CPU registers EAX, EBX,
ECX and EDX. Which information is returned is determined by the so called leaf,
which is defined by setting the input register EAX to a specific 32 bit integer
value before executing the instruction. The extent and kind of information
obtainable via this facility has changed quite a lot over the past decade and
still evolves with every CPU generation. Thus, not all information is available
on every CPU model, and certainly everything is vendor dependent.
This Julia package also provides the
cpucycle() function which allows getting
the current time stamp counter (TSC), which is determined by emitting a
instruction. Similarly to
cpuid, it only requires CPU registers and is thus,
if inlined, probably the lowest overhead method to perform micro-benchmarking.
The behaviour on non-Intel CPUs is currently unknown; though technically a crash
of Julia could be expected, theoretically, a rather large list of CPUs support
cpuid instruction. Tip: Just try it and report back.
There are plenty of different CPUs, and in particular the
has numerous corner cases, which this package does not address, yet.
cpuid instruction can only provide information for the executing
physical CPU, called a package. To obtain information on all packages, and all
physical and logical cores, the executing program must be pinned sequentially to
each and every core, and gather its properties. This is how
the operating system obtain that kind information.
In most situations, this is not really required. Even on machines with multiple CPUs, the CPUs are typically of the same model. Furthermore, it is in most cases only relevant for the currently running process whether that it is sharing its cache, rather than knowing all the details about other CPUs in the machine.
Finally, quite a bit of the really interesting information that the CPU collects
is stored in the so called machine specific registers (MSR) or control registers
(CR), which require special 'ring 0' program execution privileges, and which are
not available through the
cpuid instruction. For instance the actual CPU
clock frequencies are stored there.Typically, only the kernel and root have
these privileges, but not normal user processes and threads.
Why aren't all infos available that are seen e.g. in
Many of those features, flags and properties reside in the so called machine
specific registers (MSR), which are only accessible to privileged programs
running in the so called ring 0, such as the Linux kernel itself. Thus,
short answer: You're not allowed to.
The results obtained by
CpuId functions are inconsistent with my hardware!
Try other programs whether they give the same information or differ. If they
differ, then you found a bug. See below for some
alternatives, in particular the Linux command line tool
My hypervisor is not detected! Either you're not really running a hypervisor, e.g. Bash on Windows is not a virtual machine, or there is a feature missing. Raise an issue on GitHub.
When running a hypervisor the presented information is wrong!
Yeah, well, hypervisor vendors are free to provide the
by intercepting calls to that instruction. Not all vendors comply, and some
even permit the user to change what is reported. Thus, expect some
surprises when a hypervisor is detected.
rdtsc; that is not
True, but who cares. Both are valuable when diagnosing performance issues
and trying to perform micro benchmarks based on specific hardware features.
On Linux, most of the information may be obtained by reading from the
tree, in particular
/proc/cpuinfo, which eventually also invokes the
man 4 cpuid to get a brief description of this kernel
On many Linux distributions, there is also the command line tool cpuid, which essentially does exactly the same. On
Ubuntu, you would install it using
sudo apt install cpuid, then use it to show
a summary by simply typing
Then, of course, there are a few functions in Julia Base. These are
Base.Sys.cpu_summary(), as well as the global
Base.Sys.cpu_name. These are mostly
provided by wrapping libuv. In particular
CPU_CORES is the reason for this
module: It's intrinsically unclear whether that number includes hyperthreading
cores, or whether it is referring to real physical cores of the current machine.
The Julia package Hwloc.jl provides similar and more information primarily directed towards the topology of your CPUs, viz. number of CPU packages, physical & logical cores and associated caches, along with a number of features to deal with thread affinity. However, it also pulls in additional external binary dependencies in that it relies on hwloc, which also implies quite some computational overhead. Whether this is an issue in the first place depends much on your use-case.
CpuId takes a different approach in that it talks directly to the CPU. For
instance, asking the CPU for its number of cores or whether it supports AVX2 can
be achieved in probably 250..500 CPU cycles, thanks to Julia's JIT-compilation
approach and inlining. For comparison, 100..200 CPU cycles is roughly loading
one integer from main memory, or one or two integer divisions. Calling any
external library function is at least one order more cycles. This allows moving
such feature checks much closer or even directly in a hot zone (which, however,
might also hint towards a questionable coding pattern). Also, CpuId gives
additional feature checks, such as whether your executing on a virtual machine,
which again may or may not influence how you set up your high performance
computing tasks in a more general way. Finally, the
exposes this low-level interface to the users, enabling them to make equally
fast and reliable run-time feature checks on new or other hardware.
This Julia package CpuId is published as open source and licensed under the MIT "Expat" License.
Show that you like this package by giving it a GitHub star. Thanks! You're also highly welcome to report successful usage or any issues via GitHub, and to open pull requests to extend the current functionality.
about 2 months ago