fix[next-dace]: Fix Memory Layout for CPU#2459
fix[next-dace]: Fix Memory Layout for CPU#2459philip-paul-mueller merged 11 commits intoGridTools:mainfrom
Conversation
| unit_strides_kind = ( | ||
| gtx_common.DimensionKind.HORIZONTAL if gpu else gtx_common.DimensionKind.VERTICAL | ||
| ) | ||
| unit_strides_kind = gtx_common.DimensionKind.HORIZONTAL |
There was a problem hiding this comment.
Why does that make sense? You cannot assume anything...
There was a problem hiding this comment.
Or is that just for transients? Then I would change the comment assume -> set or something.
There was a problem hiding this comment.
There are two things here, first the name is bad and should be probably something else.
However the value selection is correct, one could even argue that it is probably the only one that make sense.
The reason for this is that the maximal numbers of blocks is different for each direction, because (for ICON) size(horizontal) >>> size(vertical) one would get launch errors otherwise.
edopao
changed the title
…escription. If the leading kind is not known then it will not reorder strides nor the iteration order. However, for cetain reasons (launch errors) we have to set one for GPU in that case.
edopao
left a comment
edopao
left a comment
There was a problem hiding this comment.
LGTM, only one refactoring suggestion.
src/gt4py/next/program_processors/runners/dace/transformations/auto_optimize.py
Show resolved
Hide resolved
|
There is some variability but I think there is nothing pathological going on. 
|
3 participants
Before the optimizer was assuming that the memory allocation for GPU and CPU was different, i.e. that in CPU the stride 1 dimension is associated with the vertical dimension while for GPU it is associated with the horizontal dimension. However, this is wrong and in both cases stride 1 is associated with the horizontal dimension.
This PR fixes this and now the loop order and the memory layout for transients assumes that stride 1 is associated to the horizontal dimension.
Note that the current implementation assumes that there is only one horizontal dimension.