Code

Feature/reproducible#1446

Code

Open

maddyscientist wants to merge 40 commits intodeveloplattice/quda:developfrom feature/reproduciblelattice/quda:feature/reproducibleCopy head branch name to clipboard

+4,576-3,204

Commits

Commits on Jul 18, 2023

• Fix compiler warning with dbldble

  maddyscientist

  committed
  ccb1c73

Commits on Jul 19, 2023

• Add array copy assignment from one type of array to another

  maddyscientist

  committed
  2049be6

• Remove use of zero function and fix caxpyxmazMR functor for when the reduction type is changed

  maddyscientist

  committed
  81566c8

• Make math_helper.cuh safe to include in non CUDA-aware compiler

  maddyscientist

  committed
  ce5d396

• Add doubledouble support for host, add complex-number support, remove legacy zero helper function

  maddyscientist

  committed
  7a4e04f

Commits on Jul 20, 2023

• Modify reduction kernels to use device_reduce_t and not double for internal computation

  maddyscientist

  committed
  2d67d97

• Use same underlying reduction type on host as device

  maddyscientist

  committed
  feccf89

• Move get_scalar<deviation_t> overload to float_Vector.h

  maddyscientist

  committed
  d70303a

• Add *= and /= overloads for doubledouble

  maddyscientist

  committed
  4a7061a

• Fix heavy quark residual norm for non-double reduction type

  maddyscientist

  committed
  7e40280

• Add various functions to doubledouble needed for generic deployment

  maddyscientist

  committed
  2a80b2f

• Commence the slog that generizes the host-side scalar precision: introduce new CMake set types: real_t (QUDA_SCALAR_TYPE) - the host side scalar precision, complex_t the complex version of this (re…

  Show description for e9089e1
  maddyscientist

  committed
  e9089e1

Commits on Jul 21, 2023

• Add isfinite method for doubledouble

  maddyscientist

  committed
  a4e8f76

Commits on Jul 22, 2023

• 99% of double -> real_t replacement now done (MPI reductions not yet done)

  maddyscientist

  committed
  a7cc5f7

• Updated ReduceArg::complete function to work when real_t and device_reduce_t are different types, e.g., double vs doubledouble

  maddyscientist

  committed
  008c632

• Remove some legacy code

  maddyscientist

  committed
  dc62b01

Commits on Jul 23, 2023

• Fix some issues

  maddyscientist

  committed
  3324b05

Commits on Jul 25, 2023

• Add missing cast operator to deviation_t::operator= when copying from a different type (needed when copying from deviation_t<doubledouble> to deviation_t<double> for example

  maddyscientist

  committed
  a16ff6c

• Add ostream << overlead for doubledouble type

  maddyscientist

  committed
  2b5bac8

• Update CUDA block_reduce_helper.h atomic types to work with doubledouble (need to split into 64-bit words) and small generic cleanup

  maddyscientist

  committed
  9d69abd

• Add support for doubledouble heterogeneous-atomic reductions. This also updates the coalesced writing to sysmem to work with large reduce_t types, such that sizeof(device_reduce_t) / sizeof(atomic_…

  Show description for 7e21a5b
  maddyscientist

  committed
  7e21a5b

• transform_reduce now respects device_reduce_t and real_t

  maddyscientist

  committed
  d5f914d

• Add initial support for multi-process doubledouble reductions: only QMP at present and just a simple gather method for now

  maddyscientist

  committed
  1a73132

• Multi-process reduction now uses device_reduce_t with the conversion to real_t done after the multi-process reduction

  maddyscientist

  committed
  d76e57c

• Updates for blas_test: use same basis for host and device to allow for direct comparisons, use max error not error sum when multiple norms are used to check correctness, print out the deviation whe…

  Show description for 27ba8de
  maddyscientist

  committed
  27ba8de

Commits on Jul 26, 2023

• Minor comment clean up

  maddyscientist

  committed
  4b5aa52

Commits on Jul 28, 2023

• Add single gpu support for doubledouble

  maddyscientist

  committed
  bcde6ad

• Small fix for doubledouble::operator>

  maddyscientist

  committed
  2ee73a6

Commits on Aug 15, 2023

• Initial version of reproduction reductions, fully works but a few limitations representing this being WIP (bin bounds LUT repeatadly recomputed on the host, bin bounds LUT presently in explicit con…

  Show description for 9789820
  maddyscientist

  committed
  9789820

• Merge branch 'feature/gaugefield_unity' of github.com:lattice/quda into feature/reproducible

  maddyscientist

  committed
  67514d0

• Fix io_test when not all precision compiled

  maddyscientist

  committed
  d455000

• Fix compiler warning

  maddyscientist

  committed
  030836d

• Reenable explicit zero support with rfa_t (fixes dilution_test)

  maddyscientist

  committed
  08b9776

• Fix gauge loop trace when using doubledouble precision reductions

  maddyscientist

  committed
  64ed607

• Fix doubledouble multi-GPU compilation (missing comm_allreduce_max function)

  maddyscientist

  committed
  ba96720

• Fix gauge_path_test loop trace test when using doubledouble reduction variables (remove bad cast)

  maddyscientist

  committed
  b7687b4

Commits on Aug 23, 2023

• Rework of reproducible reductions to pre-compute the bins when initializing the reduction. All tests passing, but seemingly break when a SANITIZE build is used

  maddyscientist

  committed
  bc74e7b

• Minor optimization of det_trace kernel

  maddyscientist

  committed
  6a60bc3

• Fix compiler warning

  maddyscientist

  committed
  a8085dc

• Merge branch 'feature/gaugefield_unity' of github.com:lattice/quda into feature/reproducible

  maddyscientist

  committed
  a413153