Advanced Computing Platform for Theoretical Physics

Skip to content

Tags

Tags give the ability to mark specific points in history as being important
  • stab55_20
    f419c77b · docs(mg): prove parity cache asset equivalence · 
    follow stab55_19, 1. prove 22-color and 16-color c128 caches have identical identity/manifest/metadata; 2. record that only construction stats differ; 3. keep parity optimization evidence auditable. [opencode].
  • stab55_19
    8f07dfe0 · docs(mg): record parity cold-start comparison · 
    follow stab55_18, 1. record parity-optimized cold-start estimate about 1.10x; 2. keep cold smoke separate from formal steady-solve 8.3507x. [opencode].
  • stab55_18
    85df3a11 · docs(test): rerun regression after parity probes · 
    follow stab55_17, 1. rerun full dev87 regression after parity-bucket probes; 2. record PASS 5/5 in 756.3 s with QUDA scaled-agreement. [opencode].
  • stab55_17
    d1e3b612 · perf(mg): order colored probes by parity buckets · 
    follow stab55_16, 1. order strict colored probes by 16 parity buckets; 2. reduce large-c128 source colors 22 -> 16 and C=4 calls 132 -> 96 per level; 3. record cold c128 C=4 double setup 974 -> 585.28 s with unchanged residual; 4. update report and solver/benchmark skills. [opencode].
  • stab55_16
    75c31696 · docs(test): record full dev87 regression · 
    follow stab55_15, 1. record full dev87 regression PASS 5/5 in 757.4 s; 2. preserve the distinction between regression coverage and performance claims. [opencode].
  • stab55_15
    ea823a1e · docs(mg): add large-c128 layered trace evidence · 
    follow stab55_14, 1. add large-c128 cache-hit trace evidence; 2. quantify 20 outer iterations and about 1.6x trace overhead; 3. record coarse V-cycle dominance and keep trace diagnostic only. [opencode].
  • stab55_14
    73c08ca5 · test(mg): make cache-hit validation asset based · 
    follow stab55_13, 1. make strict runtime cache-hit validation asset/geometry based; 2. allow an extended C=4-built double cache to serve formal C=1 solve with full SHA verification; 3. record formal c128 16x32x32x48 fair speedup 8.3507x with 20/60 iterations and passing residuals; 4. document cold C=1 setup and BiCGStab caveats. [opencode].
  • stab55_13
    88f8c4e1 · perf(bench): release setup cache before c128 solve · 
    follow stab55_12, 1. release sealed setup allocator cache before steady solve; 2. enable c128 16x32x32x48 full-double dual-side smoke; 3. record PyQCU 2.06395 s vs QUDA 17.35793 s with passing true residuals and explicit smoke-only status; 4. preserve formal C=1 limitation and BiCGStab distortion context. [opencode].
  • stab55_12
    efabde8f · docs(mg): record revalidated MG and BiCGStab ratios · 
    follow stab55_11, 1. revalidate formal c64/c128 MG-vs-MG ratios after builder optimization; 2. record c64-l3 1.9898x, c128 small 12.3768x, c128 medium 3.2332x with passing residuals; 3. add per-side BiCGStab distortion checks and isolate single-coarse-block CPU staging as a capacity experiment. [opencode].
  • stab55_11
    70001933 · perf(mg): vectorize colored Galerkin scatter · 
    follow stab55_10, 1. vectorize strict colored Galerkin support gather/scatter and cache target groups; 2. add explicit c128 CPU block staging and null-basis offload options; 3. verify c128 setup 44.86 s -> 16.67 s with true residual 1.84e-8 unchanged; 4. add CUDA regression, protocol tests, report and skill notes. [opencode].
  • stab55_10
    937849ea · perf(mg): add flat-site inverse chunks · 
    follow stab55_9, 1. add flat-site inverse chunking and compact inverse cache; 2. preserve strict setup correctness under tier0/tier1; 3. record the remaining c128 large fine-dslash OOM. [opencode].
  • stab55_9
    24dc9bbf · perf(mg): chunk fine diagonal inversion · 
    follow stab55_8, 1. chunk fine diagonal inverse to reduce setup peak; 2. record that c128 strict large still exceeds 32 GiB; 3. preserve the QUDA mixed-coarse capacity result as the validated large-coarse path. [opencode].
  • stab55_8
    17c995da · feat(mg): add mixed-coarse QUDA capacity path · 
    follow stab55_7, 1. add mixed-coarse QUDA c128 capacity path; 2. validate large three-level QUDA mixed run; 3. add PyQCU block-precision setup experiment and document its remaining OOM. [opencode].
  • stab55_7
    742125a5 · docs(mg): isolate c128 precision memory limit · 
    follow stab55_6, 1. isolate the c128 large-lattice limit as precision-driven; 2. record the two-level OOM evidence; 3. rebuild the corrected report. [opencode].
  • stab55_6
    78a99e4e · docs(bench): add c64 large-lattice level ablation · 
    follow stab55_5, 1. add same-lattice c64 two-level/three-level ablation; 2. record 1.1226x to 1.9863x level-scaling evidence; 3. rebuild the corrected report. [opencode].
  • stab55_5
    7b72fb32 · perf(mg): make blocked basis export zero-copy · 
    follow stab55_4, 1. make blocked transfer-basis export zero-copy; 2. retain CGS two-buffer setup optimization; 3. document the remaining joint-live-set memory limit for full-double c128 large lattices. [opencode].
  • stab55_4
    b945ebb0 · docs(test): record strict MPI stage-1 preflight · 
    follow stab55_3, 1. record strict MPI stage-1 preflight evidence; 2. clarify that global reduction and fail-closed preflight do not imply distributed MG support; 3. rebuild the corrected report. [opencode].
  • stab55_3
    301352c5 · perf(mg): reduce CGS setup peak allocation · 
    follow stab55_2, 1. reduce CGS setup peak allocation by reusing the source matrix; 2. preserve transfer correctness under tier0/tier1 tests; 3. document that c128 large-lattice setup remains limited by the full double hierarchy. [opencode].
  • stab55_2
    d3a3016f · test(dev87): fix QUDA legacy runner bootstrap · 
    follow stab55_1, 1. restore run_all.py --with-quda by initializing QMP before PyQUDA; 2. keep legacy c64 solve/MG checks on single precision in the precision-12 build; 3. record the 5/5 regression gate in testing skill. [opencode].
  • stab55_1
    489941ac · fix(mg): enable and validate QUDA double multigrid · 
    follow stab55, 1. enable and validate QUDA double MultiGrid; 2. retract unmatched 4.3817x and publish corrected aligned/library speedups; 3. add BiCGStab references, trace-off formal gate, c64/c128 matrices, skills and 2026-09-17 report. [opencode].

Provided by Institute of Theoretical Physics, Chinese Academy of Sciences. Please contact support@itp.ac.cn for more information.