Split splines across multiple GPUs on a given node#1101

Merged

prckent merged 40 commits intoQMCPACK:developfrom atillack:split_splines_gpu

Dec 7, 2018

+1,817-338

Commits on Mar 21, 2018

Splitting the memory based on relative MPI ranks/GPUs
Andreas Tillack
committed

Commits on May 31, 2018

Almost working implementation.
Andreas Tillack
committed

Commits on Jun 21, 2018

Current version of split splines. Issues are it still does not get quite the right final results.
Andreas Tillack
committed

Commits on Jun 25, 2018

First working version (with correct results). Performance still has to be tweaked.
Andreas Tillack
committed

Commits on Jul 3, 2018

Added NVTX debugging to code.
Andreas Tillack
committed

Commits on Jul 9, 2018

Performance improvements using cudaMemPrefetchAsync for managed memory.
Andreas Tillack
committed

Commits on Aug 8, 2018

Improved performance and fixed some bugs.
Andreas Tillack
committed

Commits on Sep 24, 2018

Minor updates to EinsplineSetCuda.cpp.
Andreas Tillack
committed

Commits on Oct 8, 2018

Merge remote-tracking branch 'upstream/develop' into split_splines_gpu
Andreas Tillack
committed

Commits on Oct 9, 2018

Add switch for splitting splines (gpusharing="yes/no" in einspline section). Also restored non-split splines performance back to reference.
Andreas Tillack
committed
Code cleanup (removal of NVTX)
Andreas Tillack
committed
Added complex double eval function stub for split splines code and variable name typo for QMC_COMPLEX case.
Andreas Tillack
committed
Minor cleanup
Andreas Tillack
committed
More split spline stub functions to (hopefully) fix test compiles.
Andreas Tillack
committed

Commits on Oct 11, 2018

Output error and abort when UVM and Split Splines are used together (will work in the future, but not now).
Andreas Tillack
committed

Commits on Nov 6, 2018

Fixed QMC_COMPLEX code path in PhaseFactors.cu. One kernel was not split spline enabled.
Andreas Tillack
committed
Merge remote-tracking branch 'upstream/develop' into split_splines_gpu
Andreas Tillack
committed
Changed Cuda Pos/Real/Value Type to CTS::Pos/Real/Value Type in split spline code.
Andreas Tillack
committed
Fixed MPI related compilation issue.
Andreas Tillack
committed

Commits on Nov 9, 2018

Fixed bug in number of split splines calculation.
Andreas Tillack
committed

Commits on Nov 14, 2018

Added separate managed memory allocation function for device vector.
Andreas Tillack
committed
More recognizable names for cudapos2 (cudaphasepos) and hostPos2 (hostPhasePos) and small bugfix (I forgot a line of code).
Andreas Tillack
committed

Commits on Nov 15, 2018

In split spline code path spline kernels on originating rank are now also in their own stream.
Andreas Tillack
committed

Commits on Nov 19, 2018

Code cleanup and first bugfixes for non-mixed precision code path.
Andreas Tillack
committed

Commits on Nov 20, 2018

Added double precision complex split spline code path.
Andreas Tillack
committed

Commits on Nov 26, 2018

Simplified device_rank_numbers loop and cleaned up device_group_numbers assignment.
Andreas Tillack
committed

Commits on Nov 28, 2018

Implemented single and double precision real wave function split spline code path.
Andreas Tillack
committed

Commits on Nov 30, 2018

Commits on Dec 2, 2018

Add test for running CUDA MPS daemon (tested on Titan, SummitDev, and Summit). This allows a warning in case CUDA MPS is not available and the user tries to use split splines.
Andreas Tillack
committed
Fix merge conflicts.
Andreas Tillack
committed
This time for real.
Andreas Tillack
committed
Merge remote-tracking branch 'upstream/develop' into split_splines_gpu
Andreas Tillack
committed
Minor cleanup.
Andreas Tillack
committed

Commits on Dec 3, 2018

Simplified check for CUDA MPS running.
Andreas Tillack
committed

Commits on Dec 5, 2018

Commits on Dec 6, 2018

Improve feature description
prckent
authored

Commits on Dec 7, 2018