Composable format would create many blocks that might read the same buffer, and currently the cache_read/cache_write schedule primitives will rewrite buffers in all blocks.
However, the schedules applied to blocks generated by different formats should be treated independently, to enable this behavior, we create a block_filter attribute in ScheduleState class.
Whenever we only want to schedule a subset of blocks, we can annotate them with keywords, and set block filters in schedule:
sch.annotate(block, "group_1", 1)
sch.set_block_filter("group_1")
# START OF SUBSET SCHEDULES
# ...
# END OF SUBSET SCHEDULES
sch.unset_block_filter()
Composable format would create many blocks that might read the same buffer, and currently the
cache_read/cache_writeschedule primitives will rewrite buffers in all blocks.However, the schedules applied to blocks generated by different formats should be treated independently, to enable this behavior, we create a
block_filterattribute inScheduleStateclass.Whenever we only want to schedule a subset of blocks, we can annotate them with keywords, and set block filters in schedule: