- Generated by
1.17.0
|
Orrery
A GPU-accelerated N-body gravitational simulator
|
How a range of indices is divided between workers. More...
Go to the source code of this file.
Classes | |
| struct | orrery::backend::IndexRange |
| A half-open range of indices. More... | |
Functions | |
| constexpr IndexRange | orrery::backend::equal_share (core::Index count, unsigned workers, unsigned worker) noexcept |
| The share of [0, count) belonging to worker out of workers. | |
Variables | |
| constexpr core::Index | orrery::backend::kPartitionGrain = core::kCacheLineBytes / sizeof(core::Real) |
| How many scalars fit in a cache line, and so the granularity every partition boundary is a multiple of. | |
How a range of indices is divided between workers.
Both partitioning schemes start from the same equal division and differ in what happens afterwards, so the division itself is written once here. It is also the place where a decision deferred in Phase 2 is finally spent.
core/aligned_allocator.hpp explains that aligning a component array to a cache line is only half of avoiding false sharing between threads, and that the other half, choosing partition boundaries on line multiples, belongs to the scheduler. This is that scheduler. If two workers were given ranges that met in the middle of a cache line, both would write to that line, and each write would take exclusive ownership of it away from the other core. The two threads would ping the line between their caches for the whole region while touching none of the same elements. Rounding every boundary out to a line multiple costs at most 63 bytes of imbalance per worker and removes the effect entirely.
|
nodiscardconstexprnoexcept |
The share of [0, count) belonging to worker out of workers.
The division is over whole cache lines rather than over indices, and the remainder is spread one line at a time across the first few workers rather than dropped on the last one. Both matter for the same reason: this is the partition the static scheme lives with for the whole region, so any imbalance built into it is imbalance the measurement will attribute to the hardware.
An out-of-range worker gets an empty range, which is what a pool with more workers than there is work should hand the surplus.
|
inlineconstexpr |
How many scalars fit in a cache line, and so the granularity every partition boundary is a multiple of.
Eight in the default double-precision build and sixteen under the single-precision one, which is correct in both cases: what matters is the line, not the number of elements that happen to fill it.
1.17.0