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
(1) The original conception of Multicore.build would operate on individual index range elements unless an index was SPLIT. In the conception of Par, the index is always split. Is this reasonable, or should we incorporate the non-tiled API as well?
Since the initial Par API always tiles the index dimensions (SPLIT is the default), a non-tiled invocation might be requested with eg MasterPar.POINTWISE.
(2) The original conception of Multicore.build allowed the index space to contain hints to aid load balancing. It would be useful to import that idea, probably, or at least experiment with it to see if it really affects performance.
The text was updated successfully, but these errors were encountered:
One interesting notion of a hint is that if the kernel is using SIMD, it will be useful to specify boundaries for the index space in the sense that we would like boundaries to be divisible by 4 (say). If the index space is itself 4-aligned and the kernel code knows this then only a pure SIMD kernel is required. Otherwise, the kernel would have to handle non-aligned subspaces but we would only generate those at the edges.
(1) The original conception of Multicore.build would operate on individual index range elements unless an index was SPLIT. In the conception of Par, the index is always split. Is this reasonable, or should we incorporate the non-tiled API as well?
Since the initial Par API always tiles the index dimensions (SPLIT is the default), a non-tiled invocation might be requested with eg MasterPar.POINTWISE.
(2) The original conception of Multicore.build allowed the index space to contain hints to aid load balancing. It would be useful to import that idea, probably, or at least experiment with it to see if it really affects performance.
The text was updated successfully, but these errors were encountered: