Kamlet Core
Kamlet Router
This router connects the kamlet to the kamlet mesh network. It connects the kamlets, memlets, nemlets and lamlet. The main purpose of this network is to allow the lamlet to broadcast kinstructions to the kamlet and to allow the kamlet to send addresses to the memlets.
Instruction Buffer
This is a FIFO buffer to even out latency differences from the lamlet to the different kamlets. This is important for instructions that require synchronization between the kamlets.
Renamer
The renamer maps the 48 logical registers (32 architectural and 16 temporary) to the 48 physical registers. It also tracks which of the 16 temporary registers are free, by processing the FreeRegister kinstructions.
The Renamer and Reservation Station live at the kamlet level rather than in the lamlet because freeing a physical register requires knowing when every jamlet slice of that register has completed its reads and writes. The non-determinism of non-local operations makes this impractical above the kamlet level.
Reservation Station
The Reservation Station holds 16 kinstructions that are waiting for their operands, cache lines and resources (e.g. division unit) to become available. If there are any in-flight operations that access the same cache line (i.e. older kinstructions in the Reservation Station or kinstructions in the Shared Waiting Table) then that is considered as WR or WW conflicts and the kinstruction is not released from the RS. The exception to this is if both kinstructions share a writeset identifier (see Custom Instructions), then it is assumed that the two accesses do not conflict. The setting of writeset identifiers for instructions is supported by custom instructions.
Kinstructions are divided into those that can be locally executed (i.e. they just move data between the local cache, local register slice and execution units), and those that involve data movement beyond their jamlet such as register-register permutations, non-aligned loads and stores, or loads and stores that require access to cache-lines that are not already present in the cache.
When a kinstruction is released from the reservation station it is either directly executed by the jamlets or goes into the Kamlet Transfer Engine depending on whether it is a local or non-local operation.
Kamlet Transfer Engine
Non-local kinstructions are sent to the Kamlet Transfer Engine, which can hold the state for 16 non-local kinstructions. This module drives the Jamlet Transfer Engines located in each jamlet and also uses the Synchronizer to make sure that kamlets are synchronized for non-local operations when necessary.
The entry for each kinstruction in the Kamlet Transfer Engine consists of:
| Field | Width (bits) |
|---|---|
| Instr Ident | 7 |
| Writeset | 5 |
| Cache Slot | 10 |
| KInstr Details | 24 |
See Transfer System.
Synchronizer
The synchronizer connects the kamlet to the synchronization network. The Kamlet Transfer Engine uses it so that kinstructions in the table can wait until synchronization points with the other kamlets and the lamlet. See Synchronization.
Kamlet Cache Engine
Each kamlet is connected to a memory via a memlet. All accesses to that memory go through the kamlet and the Kamlet Cache Engine keeps track of the cache meta data, and processes these requests. It interacts closely with the Jamlet Cache Engines which are located in each jamlet.
See Cache System.
Kamlet TLB
Jamlets will query the kamlet TLB to map logical addresses to physical addresses and to get the lane order and EW of a stripe when that information did not arrive with the instruction. The kamlet TLB will query the Lamlet TLB when it does not have the information in its cache.