Mixed prec CG algorithm has been modified to precompute precision change workspaces
As the original Test_dwf_mixedcg_prec has been coopted to do a performance stability and reproducibility test, requiring the single-prec CG to be run 200 times, I have created a new version of Test_dwf_mixedcg_prec in the solver subdirectory that just does the mixed vs double CG test
Added a new, much faster implementation of precision change that uses (optionally) a precomputed workspace containing pointer offsets that is device resident, such that all lattice copying occurs only on the device and no host<->device transfer is required, other than the pointer table. It also avoids the need to unpack and repack the fields using explicit lane copying. When this new precisionChange is called without a workspace, one will be computed on-the-fly; however it is still considerably faster than the original implementation.
In the special case of using double2 and when the Grids are the same, calls to the new precisionChange will automatically use precisionChangeFast, such that there is a single API call for all precision changes.
Reliable update and mixed-prec multishift have been modified to precompute precision change workspaces
Renamed the original precisionChange as precisionChangeOrig
Fixed incorrect pointer offset bug in copyLane
Added a test and a benchmark for precisionChange
Added a test for reliable update CG
Modified the Wilson flow adaptive smearing step size update to implement the original Ramos definition of the distance, where previously it used the norm of a difference which scales with the volume and so would choose too coarse or too fine steps depending on the volume. This is based on Chulwoo's code.
Added a test comparing adaptive (with tuneable tolerance) to iterative Wilson flow smearing on a random gauge configuration.
(The above required exposing the gauge implementation to the HMC class through the Integrator class)
Made the random shift optional (default on) through a parameter in HMCparameters
Modified ConjugateBC::CshiftLink such that it supports any shift in -L < shift < L rather than just +-1
Added a tester for the BC-respecting Cshift
Fixed a missing system header include in SSE4 intrinsics wrapper
Fixed sumD_cpu for single-prec types performing an incorrect conversion to a single-prec data type at the end, that fails to compile on some systems
Rework of WilsonFlow class
Fixed logic error in smear method where the step index was initialized to 1 rather than 0, resulting in the logged output value of tau being too large by epsilon
Previously smear_adaptive would maintain the current value of tau as a class member variable whereas smear would compute it separately; now both methods maintain the current value internally and it is updated by the evolve_step routines. Both evolve methods are now const.
smear_adaptive now also maintains the current value of epsilon internally, allowing it to be a const method and also allowing the same class instance to be reused without needing to be reset
Replaced the fixed evaluation of the plaquette energy density and plaquette topological charge during the smearing with a highly flexible general strategy where the user can add arbitrary measurements as functional objects that are evaluated at an arbitrary frequency
By default the same plaquette-based measurements are performed, but additional example functions are provided where the smearing is performed with different choices of measurement that are returned as an array for further processing
Added a method to compute the energy density using the Cloverleaf approach which has smaller discretization errors
Added a new tensor utility operation, copyLane, which allows for the copying of a single SIMD lane between two instances of the same tensor type but potentially different precisions
To LocalCoherenceLanczos, added the option to compute the high/low eval of the fine operator on every restart to aid in tuning the Chebyshev
Added Test_field_array_io which demonstrates and tests a single-file write of an arbitrary array of fields
Added Test_evec_compression which generates evecs using Lanczos and attempts to compress them using the local coherence technique
Added Test_compressed_lanczos_gparity which demonstrates the local coherence Lanczos for G-parity BCs
Added HMC main programs for the 40ID and 48ID G-parity lattices
Added a bounds-check function for the RHMC with arbitrary power
Added a pseudofermion action for the rational ratio with an arbitrary power and a mixed-precision variant of the same. The existing one-flavor rational ratio class now uses the general class under the hood
To support testing of the two-flavor even-odd ratio pseudofermion, separated the functionality of generating the random field and performing the heatbath step, and added a method to obtain the pseudofermion field
Added a new HMC runner start type: CheckpointStartReseed, which reseeds the RNG from scratch, allowing for the creation of new evolution streams from an existing checkpoint. Added log output of seeds used when the RNG is seeded.
EOFA changes:
To support mixed-precision inversion, generalized the class to maintain a separate solver for the L and R operators in the heatbath (separate solvers are already implemented for the other stages)
To support mixed-precision, the action of setting the operator shift coefficients is now maintained in a virtual function. A derived class for mixed-precision solvers ensures the coefficients are applied to both the double and single-prec operators
The ||^2 of the random source is now stored by the heatbath and compared to the initial action when it is computed. These should be equal but may differ if the rational bounds are not chosen correctly, hence serving as a useful and free test
Fixed calculation of M_eofa (previously incomplete and #if'd out)
Added functionality to compute M_eofa^-1 to complement the calculation of M_eofa (both are equally expensive!)
To support testing, separated the functionality of generating the random field and performing the heatbath step, and added a method to obtain the pseudofermion field
Added a test program which computes the G-parity force using the 1 and 2 flavor implementations and compares the result. Test supports DWF, EOFA and DSDR actions, chosen by a command line option.
The Mobius EOFA force test now also checks the rational approximation used for the heatbath
Added a test program for the mixed precision EOFA compared to the double-prec implementation,
G-parity HMC test now applied GPBC in the y direction and not the t direction (GPBC in t are no longer supported) and checkpoints after every configuration
Added a test program which computes the two-flavor G-parity action (via RHMC) with both the 1 and 2 flavor implementations and checks they agree
Added a test program to check the implementation of M_eofa^{-1}
To the gauge implementations, added CshiftLink functions calling into the appropriate operation for the BC in a given direction.
GaugeTransform, FourierAcceleratedGaugeFixer and WilsonLoops::FieldStrength no longer implicitly assume periodic boundary conditions; instead the shifted link is obtained using CshiftLink and is aware of the gauge implementation.
Added an assert-check to ensure that the gauge fixing converges within the specified number of steps.
Added functionality to compute the timeslice averaged plaquette
Added functionality to compute the 5LI topological charge and timeslice topological charge
Added a check of the properties of the charge conjugation matrix C=-gamma_2 gamma_4 to Test_gamma
Fixed const correctness for Replicate
Modified Test_fft_gfix to support either conjugate or periodic BCs, optionally disabling Fourier-accelerated gauge fixing, and tuning of alpha using cmdline options
Added storage of final true residual in mixed-prec CG and enhanced log output
Fixed const correctness of multi-shift constructor
Added a mixed precision variant of the multi-shift algorithm that uses a single precision operator and applies periodic reliable update to the residual
Added tests/solver/Test_dwf_multishift_mixedprec to test the above
Fixed local coherence lanczos using the (large!) max approx to the chebyshev eval as the scale from which to judge the quality of convergence, resulting a test that always passes
Added a method to local coherence lanczos class that returns the fine eval/evec pair
Added iterative log output to power method
Added optional disabling of the plaquette check in Nerscio to support loading old G-parity configs which have a factor of 2 error in the plaquette
G-parity Dirac op no longer allows GPBC in the time direction; instead we toggle between periodic and antiperiodic
Replaced thread_for G-parity 5D force insertion implementation with accelerator_for version capable of running on GPUs
Generalized tests/lanczos/Test_dwf_lanczos to support regular DWF as well as Gparity, with the action chosen by a command line option
Modified tests/forces/Test_dwf_gpforce,Test_gpdwf_force,Test_gpwilson_force to use GPBC a spatial direction rather than the t-direction, and antiperiodic BCs for time direction
tests/core/Test_gparity now supports using APBC in time direction using command line toggle
Audited the code conventions (again) with the CPS momentum denominator
and added anti periodic in time to the Test_mobius_force.cc and
tested the Test_dwf_gpforce.
Promoted thesee to test full HMC hamiltonian, tr P^2/2 + phidag MdagM phi
with the same pdot and Udot as audited in the Integrator.h etc...
With full comments and sources for factors.
* develop: (26 commits)
Added the ability to apply a custom "filter" to the conjugate momentum in the Integrator classes, applied both after refresh and after applying the forces Added a conjugate momentum "filter" that applies a phase to each site. With sites set to 1.0 or 0.0 this acts as a mask and enables, for example, the freezing of inactive gauge links in DDHMC Added tests/forces/Test_momentum_filter demonstrating the use of the filter to freeze boundary links
Correct misleading ac help string
Enable performance counting in WilsonFermion like in others
changed back A2AUtils warning
changed if and accelerator_for - no runtime errors any more
Mac OS (Darwin) sed -i flag for in-place editing differs from posix / gnu
Seems the intention with AutoConf produced Grid/Config.h was to use sed to translate standard PACKAGE_ #defines into GRID_ however due to missing '' after -i this hasn't been working. Perhaps it is too late to fix this, since we don't know who/what is relying on this downstream? ... but if they are, and AutoConf is being used, then likely these #defines have just been redefined anyway. Seems reasonable to redefine PACKAGE and VERSION as well, as none of these macros are used throughout Grid or Hadrons.
Fixed compile issues with maxLocalNorm2 for non-scalar lattices maxLocalNorm2 test now reuses the random field
MADWF 5d source option for hadrons - look at Grid of source Abort on GPU error
maxLocalNorm2()
change back benchmark_ITT
prettify
Flop cout matches DiRAC-ITT-2020
revert changes
merge develop
fixes
weird bug in 2pt function...
revert changes
final version, tested on CPU and GPU
bugfix
...
Added a conjugate momentum "filter" that applies a phase to each site. With sites set to 1.0 or 0.0 this acts as a mask and enables, for example, the freezing of inactive gauge links in DDHMC
Added tests/forces/Test_momentum_filter demonstrating the use of the filter to freeze boundary links
Added a general implementation of the Remez algorithm for producing arbitrary rational polynomial approximation with optional restriction to even/odd polynomials
Added implementation of computation of ZMobius parameters
Added Test_zMADWF_prec to test ZMobius in MADWF
* develop: (27 commits)
Update README.md
result layout standardised, iterator size more elegant
updated syntac in Test_hadrons_spectrum
chroma-regression test now prints difference correctly
baryon input strings are now pairs of pairs of gammas - still ugly!!
second update to pull request
Changing back interface for Gamma3pt
Removing old debug code
Changes to A2Autils
suggested changes for 1st pull request implemented
changed input parameters for easier use
Should compile everywhere now
changed baryon interface
added author information
ready for pull request
code compiling now - still need to test
Baryons module works in 1 of 3 cases - still need SlicedProp and Msource part!!
thread_for caused the problems - slow for loop for now
still bugfix
weird bug...
...
# Conflicts:
# Hadrons/Modules.hpp
# Hadrons/modules.inc
This compiles and looks right ... but may need some testing
* develop: (762 commits)
Tensor ambiguous fix
Fix for GCC preprocessor/pragma handling bug
Trips up NVCC for reasons I dont understand on summit
Fix GCC complaint
Zero() change
Force a couple of things to compile on NVCC
Remove debug code
nvcc error suppress
Merge develop
Reduction finished and hopefully fixes CI regression fail on single precisoin and force
Double precision variants for summation accuracy
Update todo list
Freeze the seed
Fix compiling of MSource::Gauss for single precision
Think the reduction is now sorted and cleaned up
Fix force term
Printing improvement
GPU reduction fix and also exit backtrace option
GPU friendly
Simplify the comms benchmark
...
# Conflicts:
# Grid/communicator/SharedMemoryMPI.cc
# Grid/qcd/action/fermion/WilsonKernelsAsm.cc
# Grid/qcd/action/fermion/implementation/StaggeredKernelsAsm.h
# Grid/qcd/smearing/StoutSmearing.h
# Hadrons/Modules.hpp
# Hadrons/Utilities/Contractor.cc
# Hadrons/modules.inc
# tests/forces/Test_dwf_force_eofa.cc
# tests/forces/Test_dwf_gpforce_eofa.cc
