mlxd / gpue

GPU Gross-Pitaevskii Equation numerical solver for rapidly rotating Bose-Einstein condensates. Further development and maintenance from https://github.com/GPUE-group/GPUE

Hi there 👋, I'm Lee

Senior Quantum Software Developer @ Xanadu.ai

Languages and Tools:

gpue's Issues

Allow read-in of parameters from XML input

Input of necessary parameters from an XML file would allow the generation of arbitrary potentials, wavefunctions, etc, and allow a more robust and easier to use code-base.

Remove unecessary files/functions/etc

Lots of old and unnecessary source files, functions and variables exist across the code. These should be removed for cleanliness.

BETA: Vortex lattice centre failing when 1 more than 1 vortex now present

The issue shows up with the move to std::vector for the vortex tracking routines. Any more than a single vortex causes the system to fail. @ # #

BETA: Vortex tracking fails at strange times during the evolution

Following the move to std::vector the tracking seems perfectly sane, but tends to locate hundreds of vortices at unexpected times during the evolution. This may likely be a sizing issue with the radial masking, but previous attempts to examine did not rectify. Need to investigate further.

Use half-step in, full step evolution

Use the symmetry of the split-operator method to remove unnecessary transforms between evolution steps. Will require a re-working of the angular momentum section.

Perform half step into the algorithm, then perform the remaining evolution in full steps, until time to print out a wavefunction, when the half-step out is required.

Will remove some unneeded FFT's, multiplications and scalings per evolution step.

Implement scaling function calls as cuFFT callbacks

This will remove unnecessary on-device memory transfers and host-kernel calls. Not sure of the expected performance increase, but it should be beneficial.

Graph does not preserve order between steps

The graph generation routine does not maintain the same UID order between simulations. Not a problem for initial selections and trajectories using python, but will be a problem if performing analysis on graphs expecting initial order.

Use pinned memory for wavefunction and evolution operators

The transfers should be faster for the printout, and when replacing the evolution operators. Use of pinned memory should help with this.

line 647 of splitop, change grid to block

this fixes issues with memory, I promise.

Add dissipation switch

Add a flag that can be used to set a dissipative term for the GPE.