targets: Spack targets can now be fine-grained microarchitectures

Spack can now:

- label ppc64, ppc64le, x86_64, etc. builds with specific
  microarchitecture-specific names, like 'haswell', 'skylake' or
  'icelake'.

- detect the host architecture of a machine from /proc/cpuinfo or similar
  tools.

- Understand which microarchitectures are compatible with which (for
  binary reuse)

- Understand which compiler flags are needed (for GCC, so far) to build
  binaries for particular microarchitectures.

All of this is managed through a JSON file (microarchitectures.json) that
contains detailed auto-detection, compiler flag, and compatibility
information for specific microarchitecture targets.  The `llnl.util.cpu`
module implements a library that allows detection and comparison of
microarchitectures based on the data in this file.

The `target` part of Spack specs is now essentially a Microarchitecture
object, and Specs' targets can be compared for compatibility as well.
This allows us to label optimized binary packages at a granularity that
enables them to be reused on compatible machines.  Previously, we only
knew that a package was built for x86_64, NOT which x86_64 machines it
was usable on.

Currently this feature supports Intel, Power, and AMD chips. Support for
ARM is forthcoming.

Specifics:

- Add microarchitectures.json with descriptions of architectures

- Relaxed semantic of compiler's "target" attribute.  Before this change
  the semantic to check if a compiler could be viable for a given target
  was exact match. This made sense as the finest granularity of targets
  was architecture families.  As now we can target micro-architectures,
  this commit changes the semantic by interpreting as the architecture
  family what is stored in the compiler's "target" attribute. A compiler
  is then a viable choice if the target being concretized belongs to the
  same family. Similarly when a new compiler is detected the architecture
  family is stored in the "target" attribute.

- Make Spack's `cc` compiler wrapper inject target-specific flags on the
  command line

- Architecture concretization updated to use the same algorithm as
  compiler concretization

- Micro-architecture features, vendor, generation etc. are included in
  the package hash.  Generic architectures, such as x86_64 or ppc64, are
  still dumped using the name only.

- If the compiler for a target is not supported exit with an intelligible
  error message. If the compiler support is unknown don't try to use
  optimization flags.

- Support and define feature aliases (e.g., sse3 -> ssse3) in
  microarchitectures.json and on Microarchitecture objects. Feature
  aliases are defined in targets.json and map a name (the "alias") to a
  list of rules that must be met for the test to be successful. The rules
  that are available can be extended later using a decorator.

- Implement subset semantics for comparing microarchitectures (treat
  microarchitectures as a partial order, i.e. (a < b), (a == b) and (b <
  a) can all be false.

- Implement logic to automatically demote the default target if the
  compiler being used is too old to optimize for it. Updated docs to make
  this behavior explicit.  This avoids surprising the user if the default
  compiler is older than the host architecture.

This commit adds unit tests to verify the semantics of target ranges and
target lists in constraints. The implementation to allow target ranges
and lists is minimal and doesn't add any new type.  A more careful
refactor that takes into account the type system might be due later.

Co-authored-by: Gregory Becker <becker33.llnl.gov>
This commit is contained in:
Massimiliano Culpo
2019-06-19 15:47:07 +02:00
committed by Todd Gamblin
parent dfabf5d6b1
commit 3c4322bf1a
50 changed files with 3039 additions and 539 deletions

View File

@@ -848,18 +848,111 @@ that executables will run without the need to set ``LD_LIBRARY_PATH``.
Architecture specifiers
^^^^^^^^^^^^^^^^^^^^^^^
The architecture can be specified by using the reserved
words ``target`` and/or ``os`` (``target=x86-64 os=debian7``). You can also
use the triplet form of platform, operating system and processor.
Each node in the dependency graph of a spec has an architecture attribute.
This attribute is a triplet of platform, operating system and processor.
You can specify the elements either separately, by using
the reserved keywords ``platform``, ``os`` and ``target``:
.. code-block:: console
$ spack install libelf platform=linux
$ spack install libelf os=ubuntu18.04
$ spack install libelf target=broadwell
or together by using the reserved keyword ``arch``:
.. code-block:: console
$ spack install libelf arch=cray-CNL10-haswell
Users on non-Cray systems won't have to worry about specifying the architecture.
Spack will autodetect what kind of operating system is on your machine as well
as the processor. For more information on how the architecture can be
used on Cray machines, see :ref:`cray-support`
Normally users don't have to bother specifying the architecture
if they are installing software for their current host as in that case the
values will be detected automatically.
.. admonition:: Cray machines
The situation is a little bit different for Cray machines and a detailed
explanation on how the architecture can be set on them can be found at :ref:`cray-support`
.. _support-for-microarchitectures:
"""""""""""""""""""""""""""""""""""""""
Support for specific microarchitectures
"""""""""""""""""""""""""""""""""""""""
Spack knows how to detect and optimize for many specific microarchitectures
(including recent Intel, AMD and IBM chips) and encodes this information in
the ``target`` portion of the architecture specification. A complete list of
the microarchitectures known to Spack can be obtained in the following way:
.. command-output:: spack arch --known-targets
When a spec is installed Spack matches the compiler being used with the
microarchitecture being targeted to inject appropriate optimization flags
at compile time. Giving a command such as the following:
.. code-block:: console
$ spack install zlib%gcc@9.0.1 target=icelake
will produce compilation lines similar to:
.. code-block:: console
$ /usr/bin/gcc-9 -march=icelake-client -mtune=icelake-client -c ztest10532.c
$ /usr/bin/gcc-9 -march=icelake-client -mtune=icelake-client -c -fPIC -O2 ztest10532.
...
where the flags ``-march=icelake-client -mtune=icelake-client`` are injected
by Spack based on the requested target and compiler.
If Spack knows that the requested compiler can't optimize for the current target
or can't build binaries for that target at all, it will exit with a meaningful error message:
.. code-block:: console
$ spack install zlib%gcc@5.5.0 target=icelake
==> Error: cannot produce optimized binary for micro-architecture "icelake" with gcc@5.5.0 [supported compiler versions are 8:]
When instead an old compiler is selected on a recent enough microarchitecture but there is
no explicit ``target`` specification, Spack will optimize for the best match it can find instead
of failing:
.. code-block:: console
$ spack arch
linux-ubuntu18.04-broadwell
$ spack spec zlib%gcc@4.8
Input spec
--------------------------------
zlib%gcc@4.8
Concretized
--------------------------------
zlib@1.2.11%gcc@4.8+optimize+pic+shared arch=linux-ubuntu18.04-haswell
$ spack spec zlib%gcc@9.0.1
Input spec
--------------------------------
zlib%gcc@9.0.1
Concretized
--------------------------------
zlib@1.2.11%gcc@9.0.1+optimize+pic+shared arch=linux-ubuntu18.04-broadwell
In the snippet above, for instance, the microarchitecture was demoted to ``haswell`` when
compiling with ``gcc@4.8`` since support to optimize for ``broadwell`` starts from ``gcc@4.9:``.
Finally if Spack has no information to match compiler and target, it will
proceed with the installation but avoid injecting any microarchitecture
specific flags.
.. warning::
Currently Spack doesn't print any warning to the user if it has no information
on which optimization flags should be used for a given compiler. This behavior
might change in the future.
.. _sec-virtual-dependencies:

View File

@@ -3213,6 +3213,127 @@ the two functions is that ``satisfies()`` tests whether spec
constraints overlap at all, while ``in`` tests whether a spec or any
of its dependencies satisfy the provided spec.
^^^^^^^^^^^^^^^^^^^^^^^
Architecture specifiers
^^^^^^^^^^^^^^^^^^^^^^^
As mentioned in :ref:`support-for-microarchitectures` each node in a concretized spec
object has an architecture attribute which is a triplet of ``platform``, ``os`` and ``target``.
Each of these three items can be queried to take decisions when configuring, building or
installing a package.
""""""""""""""""""""""""""""""""""""""""""""""
Querying the platform and the operating system
""""""""""""""""""""""""""""""""""""""""""""""
Sometimes the actions to be taken to install a package might differ depending on the
platform we are installing for. If that is the case we can use conditionals:
.. code-block:: python
if spec.platform == 'darwin':
# Actions that are specific to Darwin
args.append('--darwin-specific-flag')
and branch based on the current spec platform. If we need to make a package directive
conditional on the platform we can instead employ the usual spec syntax and pass the
corresponding constraint to the appropriate argument of that directive:
.. code-block:: python
class Libnl(AutotoolsPackage):
conflicts('platform=darwin', msg='libnl requires FreeBSD or Linux')
Similar considerations are also valid for the ``os`` part of a spec's architecture.
For instance:
.. code-block:: python
class Glib(AutotoolsPackage)
patch('old-kernels.patch', when='os=centos6')
will apply the patch only when the operating system is Centos 6.
.. note::
Even though experienced Python programmers might recognize that there are other ways
to retrieve information on the platform:
.. code-block:: python
if sys.platform == 'darwin':
# Actions that are specific to Darwin
args.append('--darwin-specific-flag')
querying the spec architecture's platform should be considered the preferred. The key difference
is that a query on ``sys.platform``, or anything similar, is always bound to the host on which the
interpreter running Spack is located and as such it won't work correctly in environments where
cross-compilation is required.
"""""""""""""""""""""""""""""""""""""
Querying the target microarchitecture
"""""""""""""""""""""""""""""""""""""
The third item of the architecture tuple is the ``target`` which abstracts the information on the
CPU microarchitecture. A list of all the targets known to Spack can be obtained via the
command line:
.. command-output:: spack arch --known-targets
Within directives each of the names above can be used to match a particular target:
.. code-block:: python
class Julia(Package):
# This patch is only applied on icelake microarchitectures
patch("icelake.patch", when="target=icelake")
in a similar way to what we have seen before for ``platform`` and ``os``.
Where ``target`` objects really shine though is when they are used in methods
called at configure, build or install time. In that case we can test targets
for supported features, for instance:
.. code-block:: python
if 'avx512' in spec.target:
args.append('--with-avx512')
The snippet above will append the ``--with-avx512`` item to a list of arguments only if the corresponding
feature is supported by the current target. Sometimes we need to take different actions based
on the architecture family and not on the specific microarchitecture. In those cases
we can check the ``family`` attribute:
.. code-block:: python
if spec.target.family == 'ppc64le':
args.append('--enable-power')
Possible values for the ``family`` attribute are displayed by ``spack arch --known-targets``
under the "Generic architectures (families)" header.
Finally it's possible to perform actions based on whether the current microarchitecture
is compatible with a known one:
.. code-block:: python
if spec.target > 'haswell':
args.append('--needs-at-least-haswell')
The snippet above will add an item to a list of configure options only if the current
architecture is a superset of ``haswell`` or, said otherwise, only if the current
architecture is a later microarchitecture still compatible with ``haswell``.
.. admonition:: Using Spack on unknown microarchitectures
If Spack is used on an unknown microarchitecture it will try to perform a best match
of the features it detects and will select the closest microarchitecture it has
information for. In case nothing matches, it will create on the fly a new generic
architecture. This is done to allow users to still be able to use Spack
for their work. The software built won't be probably as optimized as it could but just
as you need a newer compiler to build for newer architectures, you may need newer
versions of Spack for new architectures to be correctly labeled.
^^^^^^^^^^^^^^^^^^^^^^
Accessing Dependencies
^^^^^^^^^^^^^^^^^^^^^^