Repository navigation

Fix GPU single-precision DAV convergence, stress accuracy and meta-GGA XC contributions - #8204

Open

dyzheng wants to merge 27 commits into

deepmodeling:developfrom

dyzheng:develop-single-precision-fixes

Open

Fix GPU single-precision DAV convergence, stress accuracy and meta-GGA XC contributions#8204
dyzheng wants to merge 27 commits into
deepmodeling:developfrom
dyzheng:develop-single-precision-fixes

Commits on Oct 23, 2026
- Fix(xc): use precision-matched gcar/kvec_c in grad_wfc
  dyzheng
  committed
  0c04322
- Fix(pw): accumulate NLCC isotropic stress in double precision
  
  yangchaoss
  authored and
  dyzheng
  committed
  3e298de
- Fix(pw): accumulate kinetic stress in double precision
  
  yangchaoss
  authored and
  dyzheng
  committed
  1d53cc1
- Fix(pw): accumulate mGGA stress in double precision
  
  yangchaoss
  authored and
  dyzheng
  committed
  9297a5d
- Fix(pw): reduce single-precision stress accumulation error
  
  yangchaoss
  authored and
  dyzheng
  committed
  a239501
- Fix(elecstate): accumulate GPU k-point densities in double precision
  
  yangchaoss
  authored and
  dyzheng
  committed
  d5e25be
- Fix(cuda): bound long single-precision conjugate reductions
  
  yangchaoss
  authored and
  dyzheng
  committed
  1364cd5
- Fix(hsolver): use double precision for GPU-single DAV reduced solves
  
  yangchaoss
  authored and
  dyzheng
  committed
  f84cd12
- Fix(hsolver): form CUDA float Davidson bases with single-pass QR
  
  yangchaoss
  authored and
  dyzheng
  committed
  92b0d04
- Fix(pw): add missing single-precision GPU instantiations
  
  dyzheng
  authored and
  dyzheng
  committed
  f027775
- Fix(hsolver): skip near-zero residual directions in single precision
  
  dyzheng
  authored and
  dyzheng
  committed
  39cc018
- Fix(hsolver): use single-precision MPI datatype in Davidson broadcast
  
  dyzheng
  authored and
  dyzheng
  committed
  c5a497c
- Fix(pw): keep double deeq_nc/qq_so buffers for single-precision force
  
  dyzheng
  authored and
  dyzheng
  committed
  680e7a4
- Fix(hsolver): adapt CUDA float QR to HSOperator API
  dyzheng
  committed
  2f79b53
- Fix(pw): adapt single-precision DeltaSpin kernels to develop layout
  dyzheng
  committed
  396cecb
- Fix(pw): keep double deeq_nc/qq_so buffers valid on GPU single
  dyzheng
  committed
  1578569
- Refactor(hsolver): pass use_uspp explicitly to DiagoDavid
  dyzheng
  committed
  add1a06
- Fix(hsolver): warn when skipping degenerate Davidson directions
  dyzheng
  committed
  663b14c
- Fix(pyabacus): pass use_uspp to the DiagoDavid adapter
  dyzheng
  committed
  4905849
- Style(elecstate): remove trailing whitespace
  dyzheng
  committed
  f76b7e0
- Refactor(hsolver): use guarded bcast_data in DiagoDavid
  dyzheng
  committed
  d6746f0
- Fix(hsolver): compact rejected Davidson directions
  dyzheng
  committed
  794e7c4
- Fix(hsolver): compact degenerate directions in DavSubspace
  dyzheng
  committed
  0732671
- Test(hsolver): add exact/approximate Davidson warm-start regression
  dyzheng
  committed
  5c1f755
- Test(cuda): cover blocked Cgemm reductions
  dyzheng
  committed
  39cf212
- Perf(xc): reuse gcar/kvec buffers across bands in stress_mgga
  dyzheng
  committed
  9d41dad
- Fix(hsolver): keep empty blocks in the Davidson collective path
  dyzheng
  committed
  812c285

Commits

Commits on Oct 23, 2026