You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Reimplemented loops over DataLayout arguments using two communication primitives: foreach_slice and reduce_points, which work the same way on CPUs, multi-threaded CPUs, and GPUs. All DataLayout broadcast expressions and reductions can now be fused through foreach_point/foreach_level/foreach_slab/foreach_column when using statically-sized slices.
Replaced all Field values with a unified VIJHWithF layout type. Data is now indexed in the order [v, i, j, h], and every parent array has a 5-dimensional shape. For backward compatibility, [i, j, f, v, h]DataLayout indexing is still allowed, but downstream code that indexes directly into parent(field) arrays needs to be updated.
Removed the CUDAWithShmemColumnStencilStyle finite difference operator code path. Downstream code that opts in by defining Operators.use_fd_shmem() = true will now be a no-op.
Removed the columnwise! operator, which is not used in any downstream code.
