Dev m4 optimization - #29
Open
DiamonDinoia wants to merge 198 commits into
Open
Conversation
Collaborator
|
How's the FINUFFT speed vs NFFT3 working out for you? |
Wentzell
force-pushed
the
DEV_NEW_ALPHA_REB
branch
from
June 18, 2024 21:07
d7823eb to
1efbbd1
Compare
Wentzell
force-pushed
the
DEV_M4_OPTIMIZATION
branch
from
June 18, 2024 21:07
d450d90 to
791f048
Compare
Member
Hey @ahbarnett, switching to the FINUFFT backend has been a major performance improvement! |
Wentzell
force-pushed
the
DEV_M4_OPTIMIZATION
branch
from
June 25, 2024 16:40
8565f44 to
607b619
Compare
Wentzell
force-pushed
the
DEV_M4_OPTIMIZATION
branch
from
August 9, 2024 19:51
19c0e6a to
e57982b
Compare
* [ghactions] Update runner images and compiler version
Wentzell
force-pushed
the
DEV_M4_OPTIMIZATION
branch
7 times, most recently
from
May 2, 2025 15:02
03852e0 to
405df56
Compare
Wentzell
force-pushed
the
DEV_NEW_ALPHA_REB
branch
from
May 2, 2025 15:12
2e7d310 to
7ffa19f
Compare
Wentzell
force-pushed
the
DEV_M4_OPTIMIZATION
branch
from
May 2, 2025 15:13
405df56 to
d7e36bd
Compare
…and release branches - for version switcher on the sphinx doc page
Set html_baseurl so every version build emits a rel="canonical" link to the latest/ copy of each page. This tells search engines to prefer the latest docs over older/unstable versions (and de-duplicates the identical latest/ and newest-release folders). Assisted-by: Claude <noreply@anthropic.com>
sphinx_rtd_theme 3.x ships an empty versions.html (the flyout is RTD-addons only). Override it with _templates/versions.html to revive the bottom-left 'v: <version>' widget, populated from a static html_context['versions'] list. Styled by the theme's existing .rst-versions CSS, toggled by theme.js. Assisted-by: Claude <noreply@anthropic.com>
- remove latest - use a dedicated cmake var VERSION_SUFFIX to distinguish unstable vs. release branches - display unstable in case of non-release branches
# Conflicts: # doc/conf.py.in
# Conflicts: # CMakeLists.txt # README.md # doc/conf.py.in
The documentation is published at triqs.github.io/<repo> while PROJECT_NAME is triqs_<repo>, so the canonical URL and all version-switcher links pointed to non-existent pages. Use the actual website name instead. Assisted-by: Claude <noreply@anthropic.com>
triqs_hartree_fock is resolved through external_dependency, which tries find_package first unless Build_Deps=Always. When it instead builds from source it does so as a subproject without EXCLUDE_FROM_ALL, so its own install rules already ship the python module -- along with version.py and its cmake config, which the rule here missed. The module was therefore installed twice. The rule also pointed unconditionally at deps/triqs_hartree_fock, a path that only exists when the sources happen to be checked out there: not when external_dependency git-clones them into the build tree, and not when find_package resolved the package from the system. Dropping the rule makes every mode correct by construction. Verified that a from-source install ships an unchanged set of triqs_hartree_fock files, now once instead of twice, and that a Build_Deps=IfNotFound build against an installed triqs_hartree_fock 4.0.0 resolves it from the system and installs none of it. Assisted-by: Claude <noreply@anthropic.com> squash
The <triqs/arrays/block_matrix.hpp> include in types.hpp was dead: block_matrix_t is std::vector<matrix<...>> and the triqs::arrays::block_matrix<T> template it pulled in is used nowhere in ctint. It only reached the bundled shim under c++/triqs/, which is not installed, so a downstream C++ consumer of the installed ctint headers failed to resolve the header. Drop the include and delete the now-unreferenced shim (and the empty c++/triqs/ tree), closing the install gap without shipping the shim into the triqs/ include namespace. Assisted-by: Claude <noreply@anthropic.com>
poet was pulled in for a single utility: a compile-time unrolled loop over [0, N). Reimplementing it as detail::static_for in nfft_buf.hpp is a handful of lines and removes a dependency, its cmake resolution and its git clone. Assisted-by: Claude <noreply@anthropic.com>
M3ph_tau, M3pp_tau, M3xph_tau, chi2_tau and chiAB_tau include nfft_buf.hpp but use nothing from it. Only the six *_iw measures hold nfft_buf_t members, which is also exactly the set of objects that reference FINUFFT symbols. Assisted-by: Claude <noreply@anthropic.com>
nfft_buf.hpp is an installed public header, so including finufft.h and <xsimd/xsimd.hpp> forced both dependencies' headers into our own prefix. Everything that needs them now lives in a new nfft_buf.cpp, which explicitly instantiates nfft_buf_t for the three Ranks its static_assert allows. The finufft_opts struct becomes a local of each constructor rather than a member: finufft_makeplan copies what it needs out of it and nothing reads it afterwards, so it never had to be one. That is a prerequisite for dropping the finufft.h include, and it also removes a ~15-field struct from every nfft_buf_t, of which the M3/M4 measures hold n_blocks * n_orb^2. The FINUFFT plan is reached through an opaque handle whose deleter is defined out of line, so the header needs no FINUFFT declaration and the unique_ptr member stays pointer-sized. push_back stays inline: it runs per Monte Carlo accumulation and only reaches the new translation-unit boundary through do_nfft(), once per nfft_buf_size pushes. do_direct_bitwise() is now constrained with requires(Rank == 1) instead of asserting it. The explicit instantiations instantiate every unconstrained member, and that body is valid for Rank 1 only. rebind() also needed triqs/utility/exceptions.hpp to be included explicitly: TRIQS_ASSERT used to arrive transitively in the translation units that happened to instantiate it. Assisted-by: Claude <noreply@anthropic.com>
The prime-sum and odd-exponent machinery used to set up the direct kernels' per-
target exponent lists was a block of static members of nfft_buf_t, but none of it
depends on Rank -- it was templated by accident, and instantiated three times for
no reason. Nothing outside nfft_buf.{hpp,cpp} refers to any of it, and the only
uses are in the non-uniform-target constructor and the two kernel dispatch sites.
Move it verbatim into a file-local anonymous namespace in nfft_buf.cpp, dropping
the now-meaningless `static` and giving the whole block internal linkage. An
anonymous namespace rather than detail:: so that express_as_prime_sum does not
become an exported symbol of libtriqs_ctint_c.a.
This takes another 96 lines out of the installed header, which is now 262 lines
against 908 on unstable. The one behavioural difference is that the static_assert
validating the prime-sum table now fires when ctint is compiled rather than in
every translation unit that includes the header; it checks an internal invariant
of our own algorithm, so consumers gain nothing by re-verifying it.
Assisted-by: Claude <noreply@anthropic.com>
Two things in the installed header served no consumer: - All three nfft_buf_t deduction guides are dead. Every construction site in c++/, test/ and the measures names the Rank explicitly as nfft_buf_t<1>, <2> or <3>, so class template argument deduction is never exercised. Two of the three also carried pre-existing clang-format violations. - to_array_vector exists solely so the Rank==1 convenience constructor can delegate to the general one, so move it into detail rather than leave it as a free function in triqs::utility. Assisted-by: Claude <noreply@anthropic.com>
Now that nfft_buf.cpp is the only translation unit touching either library, they are pure build-time dependencies. Mark them PRIVATE and wrap them in $<BUILD_INTERFACE:>: CMake records the private dependencies of a static library as $<LINK_ONLY:...> in INTERFACE_LINK_LIBRARIES, which would both re-export them and make install(EXPORT) fail on targets that are not in our export set. That alone would leave the finufft symbols undefined in the installed archive, so absorb finufft's objects into it via $<TARGET_OBJECTS:>. libtriqs_ctint_c.a then defines finufft_makeplan and friends itself, and nothing of finufft needs to be installed, exported or found again downstream. This is also why finufft stays BUILD_ALWAYS. fftw is all that is left over; triqs_ctint-config.cmake resolves it with find_library at consumer configure time rather than baking in the absolute paths that finufft's own fftlibs list carries. The install tree goes from 158 files to 51, with no third-party entry left in it, and the exported interface names only triqs. Assisted-by: Claude <noreply@anthropic.com>
conda-forge builds have no network, so every dependency has to come from the host environment or from a source tarball declared in the recipe: - triqs_hartree_fock is now a host and run requirement rather than something we bundle, and -DBuild_Deps=IfNotFound lets external_dependency() find it there instead of cloning it. - fftw is both host and run: it is the one external dependency of the finufft code absorbed into our archive, a DT_NEEDED of the python modules and a link-time dependency of downstream C++ consumers. - finufft has no C++ package on conda-forge, only the python bindings, so fetch its release tarball into deps/finufft where external_dependency() picks it up without a clone. Assisted-by: Claude <noreply@anthropic.com>
The reference was stored by the constructor but never read, so clang warned about it with -Wunused-private-field. The constructor keeps the parameter, as every measure is built through the same interface. Assisted-by: Claude <noreply@anthropic.com>
Match the style used across the other TRIQS repositories. The badge target is unchanged: the concept DOI 10.5281/zenodo.15598402. Assisted-by: Claude <noreply@anthropic.com>
Zenodo sends the badge SVG with cache-control: no-cache behind a 120 req/min per-IP rate limit. GitHub's camo proxy therefore re-fetches it on every render from a shared IP pool, often getting a 429 that it surfaces as a 502 broken image. shields.io sends max-age=432000, so camo caches it instead. The link target stays the concept DOI. Assisted-by: Claude <noreply@anthropic.com>
The embedded plot images made the notebooks the largest source of tracked text in the repository, so GitHub reported Jupyter Notebook as the dominant language at 58%. Clearing the outputs shrinks them from 601K to 21K and leaves all cell sources untouched. Assisted-by: Claude <noreply@anthropic.com> (cherry picked from commit 8dc03fc)
measures/iw_accumulate.hpp adds xsimd kernels for the inner frequency accumulation of M3pp, M3ph, M4, M4pp and M4ph. The five measure translation units call the kernels instead of scalar loops. Clang needs -ffp-contract=fast to emit the fma. xsimd itself needs no new dependency entry: finufft already fetches it via CPM and c++/triqs_ctint links it through $<BUILD_INTERFACE:xsimd>.
…layer The frequency and orbital loops in iw_accumulate.hpp run as one loop nest, so the batch loads happen once per tile instead of once per inner iteration. The forwarding wrapper around the kernels is gone; the measures call the kernel directly.
DiamonDinoia
force-pushed
the
DEV_M4_OPTIMIZATION
branch
from
August 21, 2026 18:47
c1109de to
1046043
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
9 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Used clang builtin to vectorize iw4 accumulation.
Optimizations are disabled if gcc is used.