symmetry-aware GW accumulation with X2C support by gauravharsha · Pull Request #15 · Green-Phys/green-gpu

gauravharsha · 2026-06-09T22:45:12Z

Summary

Adds full Brillouin-zone symmetry reduction to the GW GPU kernel, including the X2C (two-component spinor / SOC) path. Self-energy is now accumulated over IBZ k-stars and q-stars with on-the-fly rotation of G(k) (orbital point-group rep) and P(q) (auxiliary-basis rep), reproducing the no-symmetry result to ~1e-5 on a cubic Ar def2-svp 3×3×1 reference (9 ink no-symm → 5 ink TRS-only → 3 ink full symmetry).

Highlights

IBZ/star-based GW accumulation with cu_symmetry orchestrating both the AO-basis k-point rotation and the auxiliary-basis q-point rotation.
X2C symmetry support: q-space transform P(q_full) = U_q · P(q_ibz) · U_q† applied via three GEMMs in compute_second_tau_contraction. U_q_conj is precomputed once on device. TR branch reuses the same kernel with operand reassignment — no runtime branching in the contraction itself.
Unified scalar/X2C contraction kernel — the two paths were byte-identical after the convention fixes, so the _2C variant was merged into the scalar function.
All four spinor blocks computed (aa, bb, ab, ba) — the Hermitian relation Σ_ba = (Σ_ab)† does not hold at a single (k, τ) snapshot under SOC, so the prior shortcut was incorrect.
New regression test: GW_X2C_Ar_Symmetry runs HF+GW on cubic Ar def2-svp with no-symm / TRS-only / full-symm symmetry settings and asserts agreement per (t, k_ibz).

Notes for reviewers

Commit 7/9 (cugw: fix X2C cuBLAS row/col-major conventions; …) is the largest squashed unit — it contains the chronological progression of bug fixes for the row-major↔col-major confusion in the q-space transform GEMMs. Worth reviewing the final state of compute_second_tau_contraction (the in-source convention block explains the OP_T choices) rather than the diff itself.
Memory accounting in gw_gpu_kernel.cpp::GW_check_devices_free_space was verified against actual cudaMallocs in cugw_qpt/cugw_qkpt/cu_symmetry; the estimate is conservative with a 20% reserve.
Commit 8/9 leaves a TODO for missing Hermitization after GETRS in compute_Pq_lu — flagging for follow-up but not blocking this PR.

Use value_AO in copy_Gk_2c and get_dm_fbz; restore off-diagonal Madelung in add_Ewald. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…blocks; precompute U_q_conj

…q_lu CPU reference Hermitizes (I-P0)⁻¹ post-solve (gw_cpu_kernel.cpp:249); GPU does not. Documenting for future investigation if tol=1e-8 fails after the X2C symmetry fixes land at tol=1e-5. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Copilot

Pull request overview

Adds full Brillouin-zone symmetry reduction to the GPU GW path (including X2C/SOC), accumulating over IBZ k/q stars with on-the-fly rotations, and introduces an Ar X2C regression test dataset + test coverage to validate symmetry vs no-symmetry agreement.

Changes:

Implement/standardize k- and q-space symmetry transforms for GW (including X2C), including precomputed conj(U_q) buffers for correct complex-U handling.
Merge scalar/X2C second-τ contraction into a unified kernel path and accumulate all four X2C spinor blocks.
Add GW_X2C_Ar_Symmetry / HF_X2C_Ar_Symmetry regression tests and associated input/reference data.

Reviewed changes

Copilot reviewed 14 out of 22 changed files in this pull request and generated 11 comments.

Show a summary per file

File	Description
test/data/GW_X2C_Ar/make_g_input.sh	Helper script to run the Ar symmetry cases (needs portability tweaks).
test/data/GW_X2C_Ar/init.sh	Helper script to regenerate Ar inputs (needs portability/safety tweaks).
test/data/GW_X2C_Ar/atom.dat	Test structure input for Ar.
test/data/GW_X2C_Ar/a.dat	Lattice vectors for cubic Ar test cell.
test/cu_solver_test.cpp	Adds HF/GW X2C Ar symmetry regression checks and adjusts solver args.
src/hf_gpu_kernel.cpp	Uses k-symmetry `value_AO` path for X2C DM expansion and fixes Ewald block handling.
src/gw_gpu_kernel.cpp	Updates device memory estimate for storing both `U_q` and `conj(U_q)`.
src/green/gpu/cugw_qkpt.h	Updates second-τ contraction API/docs (U/U_conj operands).
src/green/gpu/cu_symmetry.h	Adds accessors for device `conj(U_q)` buffers.
src/cugw_utils.cu	Wires U/U_conj into second-τ contraction and selects q-conjugated P buffers (one bug found).
src/cugw_qpt.cu	Documents known missing Hermitization in GPU LU path (TODO).
src/cugw_qkpt.cu	Reworks second-τ contraction GEMM chain + removes the `_2C` variant.
src/cu_symmetry.cu	Allocates and initializes device `conj(U_q)` buffers.
.github/workflows/test.yaml	Reruns failed tests with output-on-failure for easier CI debugging.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

codecov-commenter · 2026-06-09T23:23:45Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.70%. Comparing base (de553fd) to head (ffc1ba0).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #15      +/-   ##
==========================================
+ Coverage   95.32%   96.70%   +1.37%     
==========================================
  Files          13       13              
  Lines         877      849      -28     
==========================================
- Hits          836      821      -15     
+ Misses         41       28      -13

Flag	Coverage Δ
unittests	`96.70% <100.00%> (+1.37%)`	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

After the unified second-tau kernel, the two helpers differed only by an extra G(k1_ibz) → G(k1_full) AO rotation (scalar only). Also fixes a bug where the scalar variant hardcoded need_minus_q=false, so a symmetric scalar GW with TR-active q-points would have used Pqk_tQP_ instead of Pqk_tQP_conj_. The AO transform moves to the call site under !_X2C, with a TODO to implement the X2C nso×nso G rotation on the GPU. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

- cugw_qkpt: throw if exactly one of {U, U_conj} is null in compute_second_tau_contraction (operand pair must be all-or-nothing); tighten U parameter docstring. - cu_symmetry: rewrite the q_p0_transforms_conj comment so it points readers to the kernel's convention block rather than asserting a specific OP ↔ math correspondence at the storage site — the kernel actually uses OP_T on this buffer at GEMM 2c, which doesn't match the "OP_N → U_q†" reading the prior comment implied. - cu_solver_test: cast result.size() to int for MPI_Bcast count (silent narrowing; safe in test, but flagged by -Wconversion). - GW_X2C_Ar test data scripts: replace hard-coded \$HOME paths with GREEN_MBPT_PYTHON / GREEN_MBPT_BIN env vars (or local ScriptDir / GWBin overrides), with :? failure messages if neither is set. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

egull

Overall looks good. Have a look at the optional changes (especially the doubled memory allocations?) and then commit.

If you want to keep the code paths for double and float, we should do an if/else before the malloc/free to save memory.

Template cu_symmetry on class prec (mirroring cugw_utils<prec> and gw_qkpt<prec>) using cu_type_map<std::complex<prec>>::cuda_type for the device buffers. The five paired _d_/_f_ pointer members collapse to single cuda_complex* members, the four paired accessors collapse to two (q_p0_transform, q_p0_transform_conj), and the two transform_k_ao_device / transform_k_ao_device_2c overloads collapse to single methods. Explicit instantiation for <float> and <double> in cu_symmetry.cu. Consequences: - cugw_utils.cu drops the _sp ? _d : _f ternary at the only consumer. - cugw_utils<prec>::_cu_symmetry now holds cu_symmetry<prec>. - gw_gpu_kernel.cpp's cugw_extra budget reflects single-precision storage instead of summing both precisions. - test/cu_symmetry_test.cu uses cu_symmetry<double> for the existing double-precision roundtrip. Wrong-precision calls are now type errors at the boundary rather than runtime guards. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Turn test_symmetry_transform_roundtrip into a function template on class prec so both cu_symmetry<double> and cu_symmetry<float> are exercised end-to-end. The CPU reference (cpu_transform) stays in std::complex<double>; the helper casts to/from the device cuda_complex only at upload/download boundaries. Two REQUIREs in the Symmetry_Transform Catch2 section, both at 1e-4 (measured ~5e-5 on GPU host for both precisions on the DFT Fock test data). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Defensively symmetrize the LU solver's output to match the CPU reference's 0.5*(T + T.adjoint()) step. With P0 Hermitian (enforced above) and I-P0 Hermitian, the product is mathematically Hermitian; this suppresses float roundoff that would otherwise be amplified by the U_q rotation in the second-tau kernel. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

The flags were always (void)-cast in every implementation because transform_k_ao_device handles all TR rotations on the GPU. Remove them from gw_reader1_callback and gw_reader2_callback typedefs, the four scalar/X2C lambda implementations, and both call sites. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

gauravharsha · 2026-06-11T04:27:37Z

All the changes have been completed. @egull could you take a look at the relevant sections again -- I'm keeping your comments unresolved so you can see and compare. If all looks good, we can merge.

egull · 2026-06-11T13:01:46Z

The template really makes things much cleaner in this case.

gauravharsha and others added 7 commits June 9, 2026 18:47

apply space group symmetry transform in X2C GPU kernels

bc1806e

Use value_AO in copy_Gk_2c and get_dm_fbz; restore off-diagonal Madelung in add_Ewald. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

test: add X2C GPU symmetry consistency test (cubic Ar def2-svp)

001ea26

ci: rerun failed ctest with --output-on-failure to surface logs

0f90939

add params for P_sp and Sigma_sp

a15acad

cugw: fix X2C cuBLAS row/col-major conventions; compute all 4 spinor …

db50d44

…blocks; precompute U_q_conj

cugw: unify scalar and X2C second-tau kernel; drop dev docs/plans

b859c20

gauravharsha force-pushed the symm-x2c-v2 branch from 419e1f9 to b859c20 Compare June 9, 2026 22:47

gauravharsha requested review from Copilot and egull June 9, 2026 22:50

Copilot started reviewing on behalf of gauravharsha June 9, 2026 22:50 View session

Copilot AI reviewed Jun 9, 2026

View reviewed changes

gauravharsha and others added 2 commits June 9, 2026 21:49

gauravharsha changed the title ~~Symm x2c v2~~ symmetry-aware GW accumulation with X2C support Jun 10, 2026

egull requested changes Jun 10, 2026

View reviewed changes

Comment thread src/cu_symmetry.cu

Comment thread src/cugw_qpt.cu

Comment thread src/cugw_utils.cu Outdated

gauravharsha and others added 4 commits June 11, 2026 00:19

gauravharsha force-pushed the symm-x2c-v2 branch from 7fe7e30 to ffc1ba0 Compare June 11, 2026 04:20

egull approved these changes Jun 11, 2026

View reviewed changes

gauravharsha merged commit 1ccc729 into main Jun 11, 2026
1 check passed

Navigation Menu

Uh oh!

symmetry-aware GW accumulation with X2C support - #15

symmetry-aware GW accumulation with X2C support#15
gauravharsha merged 13 commits into
mainfrom
symm-x2c-v2

gauravharsha commented Jun 9, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

codecov-commenter commented Jun 9, 2026 •

edited

Loading

Uh oh!

egull left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

gauravharsha commented Jun 11, 2026

Uh oh!

egull commented Jun 11, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

gauravharsha commented Jun 9, 2026

Summary

Highlights

Notes for reviewers

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

codecov-commenter commented Jun 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

egull left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

gauravharsha commented Jun 11, 2026

Uh oh!

egull commented Jun 11, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

codecov-commenter commented Jun 9, 2026 •

edited

Loading