- Registered on: 12/21/2018
- Last connection: 11/20/2019
- 06:59 PM GROMACS Bug #3211 (New): Lincs GPU kernel with DD fails for some cases with out-of-bounds shared-memory access
In the lincs GPU kernel we have code like this (L201 in lincs_cuda.cu)
int c1 = gm_coupledConstraints...
- 12:41 PM GROMACS Bug #3164 (In Progress): mdrun-mpi-test with separate PME ranks and PP-PME CPU comm crashes
- This is fixed by https://gerrit.gromacs.org/c/gromacs/+/14223
- 04:43 PM GROMACS Feature #3160: implement direct comm for different src/target memory spaces
Sorry, what I said above wan't quite right. One PP task wasn't taking the path not because of the pasted condition ...
- 01:03 PM GROMACS Feature #3022 (Resolved): modernized naming for new GPU developments
- 05:26 PM GROMACS Task #3158 (In Progress): use MPI_Isend() in sendFToPpCudaDirect and receiveCoordinatesFromPpCudaDirect
- 05:25 PM GROMACS Feature #3022 (In Progress): modernized naming for new GPU developments
- Addressed in https://gerrit.gromacs.org/c/gromacs/+/14162
- 03:34 PM GROMACS Task #3158: use MPI_Isend() in sendFToPpCudaDirect and receiveCoordinatesFromPpCudaDirect
- Implemented for coordinates in https://gerrit.gromacs.org/c/gromacs/+/14159
- 03:15 PM GROMACS Feature #3160: implement direct comm for different src/target memory spaces
>> extend to support case where PME is on CPU and PP is on GPU.
> Have code offline that almost works, just faili...
- 11:26 AM GROMACS Task #3159 (In Progress): eliminate regression due to moving gmx_pme_send_coordinates()
- Direct CPU->GPU copy already works in master branch by setting sendCoordinatesFromGpu=true.
With NVLINK on DGX,...
- 11:12 AM GROMACS Bug #3100 (Closed): crash with GPU comm DD
Also available in: Atom