IMPORTANT: To view this page as Markdown, append `.md` to the URL (e.g. /get-started.md). For the complete documentation index, see llms.txt.
Skip to main content
For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo struct

HiSparseStateLayout

struct HiSparseStateLayout[page_size: Int, hot_pages: Int, num_groups: Int, max_misses: Int, hash_capacity: Int]

Byte layout of a request's HiSparse state page.

Parameters​

  • ​page_size (Int): Rows per cache page.
  • ​hot_pages (Int): Pages in the hot buffer; it holds hot_pages * page_size slots.
  • ​num_groups (Int): Indexer layers that keep their own LRU state.
  • ​max_misses (Int): Most distinct host-only selections one request may make in one call, summed over its rows.
  • ​hash_capacity (Int): Entries in the per-call hash of those selections; a power of two above max_misses.

Implemented traits​

AnyType, Deinitable, Movable

comptime members​

BYTES​

comptime BYTES = (_align16(Int((add (mul hot_pages, page_size, 4), _align16(Int((add (mul max_misses, 4), _align16(Int((add (mul max_misses, 4), _align16(Int((add (mul max_misses, 8), 64)))))))))))) + Int((mul _align16(Int((add (mul hot_pages, page_size, 8), 16))), num_groups)))

Bytes the state needs; must fit one cache page.

EPOCH​

comptime EPOCH = 0

UInt32 call counter, bumped by every resolve.

GROUP_INIT​

comptime GROUP_INIT = 0

UInt32, nonzero once the group's order is initialized.

GROUP_KEYS​

comptime GROUP_KEYS = 16

Int32[hot_rows], the position each slot holds plus one, or 0.

GROUP_ORDER0​

comptime GROUP_ORDER0 = (Int(16) + (Int(4) * HiSparseStateLayout[page_size, hot_pages, num_groups, max_misses, hash_capacity].hot_rows))

Int16[hot_rows] slots, least recently used first.

GROUP_ORDER1​

comptime GROUP_ORDER1 = (HiSparseStateLayout[page_size, hot_pages, num_groups, max_misses, hash_capacity].GROUP_ORDER0 + (Int(2) * HiSparseStateLayout[page_size, hot_pages, num_groups, max_misses, hash_capacity].hot_rows))

The other order buffer.

GROUP_STRIDE​

comptime GROUP_STRIDE = _align16((HiSparseStateLayout[page_size, hot_pages, num_groups, max_misses, hash_capacity].GROUP_ORDER1 + (Int(2) * HiSparseStateLayout[page_size, hot_pages, num_groups, max_misses, hash_capacity].hot_rows)))

Bytes per group.

GROUP_WHICH​

comptime GROUP_WHICH = 4

UInt32, which order buffer is current.

GROUPS​

comptime GROUPS = _align16(Int((add (mul hot_pages, page_size, 4), _align16(Int((add (mul max_misses, 4), _align16(Int((add (mul max_misses, 4), _align16(Int((add (mul max_misses, 8), 64))))))))))))

Start of the per-group LRU state.

HITS​

comptime HITS = _align16(Int((add (mul max_misses, 4), _align16(Int((add (mul max_misses, 4), _align16(Int((add (mul max_misses, 8), 64)))))))))

UInt32[hot_rows], the epoch of the call that last hit each slot.

hot_rows​

comptime hot_rows = (hot_pages * page_size)

Slots in the hot buffer.

MISS_COUNT​

comptime MISS_COUNT = 4

Int32 misses queued by the last resolve.

MISS_HASH​

comptime MISS_HASH = _align16((Int(64) + (Int(8) * max_misses)))

Int32[max_misses] hash entry of each miss, scratch.

MISS_LIST​

comptime MISS_LIST = 64

Int32[max_misses, 2] of (hot row, host row) pairs.

VICTIMS​

comptime VICTIMS = _align16(Int((add (mul max_misses, 4), _align16(Int((add (mul max_misses, 8), 64))))))

Int32[max_misses] slot taken by each miss, scratch.

Was this page helpful?