Open ax3l opened 10 years ago
The problem is caused by the available shared memory. each block is called for a super cell with max(TILE_X, TILE_Y, TILE_Z) times 1024 elements in phase space.
with the default setting (max=8) and float that uses 32KB shared (<48KB). Using a super cell with a larger edge or using double usually blows the maximum available shared mem.
One could add a compile-time check now, if the setting will fit the shared memory available.
An extension proposed by @psychocoderHPC and me would be to walk over the particle list N-times and to split the local phase space section into N slices in momentum direction.
Oh I already transformed the "bug report" in an to-be-implemented feature. So let me open this again :)
oh sry for closing I pressed the wrong bottom on my mobile phone
classic first world hipster problem :smile:
Implement a more variable phase space binning.
First step:
float_PS
and for changed super cells ("long edge problem").num_pbins
inphaseSpace.param
(feature)The phase space plugin does not run it's kernels for a 2D sim with 8x16 or 16x16 tile sizes.(but works fine for 3D 8x8x4; 2D 8x8)Fixed with #480