Accera v1.2 Reference
accera.CacheStrategy
| type | description |
|---|---|
accera.CacheStrategy.BLOCKED |
Every thread copies a contiguous block of memory based on their thread index. e.g. If 100 elements are cached by 10 threads, thread 0 copies elements [0, 10), thread 1 copies elements [10, 20) and so on. |
accera.CacheStrategy.STRIPED |
Every thread copies a part of their contribution in a round-robin fashion. e.g. In the previous example, thread 0 will now copy elements [0, 2), [20, 22), [40, 42), [60, 62) and [80, 82), thread 1 will copy [2, 4), [22, 24), [42, 44), [62, 64) and [82, 84) and so on. The minimum number of contiguous elements that each thread copies is governed by the vectorization parameter, which in this example is 2. |
The effects of different caching strategies can be noticed as performance characteristics arising out of overhead caused by bank conflicts, memory coalescing etc.
Last update:
2023-04-17