Abstract: A method of responding to questions about documents by a computing system includes: (a) initially sending queries to a large language model (LLM), each query including a request for any snippets within a corresponding document that provide evidence in response to a question, the LLM returning a respective response to each query; (b) in response to a heuristic based on the responses returned by the LLM indicating a readiness for optimization, training a new model on the queries and respective responses, the new model being more computationally efficient than the LLM; (c) performing a validation operation on additional responses returned by the new model; and (d) in response to the validation operation indicating that the new model has satisfied a condition, sending future queries to the new model instead of the LLM.

Abstract: A method includes: (a) for each of a set of queries, receiving a specification of that query; (b) receiving an indication of a set of clinical notes; (c) for each query, prompting an LLM system based on the respective specification of that query and receiving a response from the LLM system to each query for each clinical note, each response including a label and evidence from that clinical note supporting the label; (d) for at least one query, displaying the label and evidence for each clinical note generated in response to the at least one query; (e) in response to displaying, receiving a revised specification of the at least one query; and (f) prompting the LLM system based on the revised specification of the at least one query and receiving an updated response from the LLM system to the at least one query for each clinical note, each updated response including an updated label and updated evidence from that clinical note supporting the updated label.

Abstract: A data processing system includes a master, a central request agent, and a plurality of snoopers communicatively coupled to a system fabric for communicating requests subject to retry. The master issues on the system fabric a multicast request intended for the plurality of snoopers. The central request agent receives the multicast request on the system fabric, assigns the multicast request to a particular state machine among a plurality of state machines in the central request agent, and provides the master a coherence response indicating successful completion of the multicast request. The central request agent repetitively issues on the system fabric a multicast request in association with a machine identifier identifying the particular state machine until a coherence response indicates the multicast request is successfully received by all of the plurality of snoopers.

Abstract: A data processing system includes a master, a central request agent, and a plurality of snoopers communicatively coupled to a system fabric for communicating requests subject to retry. The master issues on the system fabric a multicast request intended for the plurality of snoopers. The central request agent receives the multicast request on the system fabric, assigns the multicast request to a particular state machine among a plurality of state machines in the central request agent, and provides the master a coherence response indicating successful completion of the multicast request. The central request agent repetitively issues on the system fabric a multicast request in association with a machine identifier identifying the particular state machine until a coherence response indicates the multicast request is successfully received by all of the plurality of snoopers.

Abstract: A first instruction for processing by a processor core is received. Whether the instruction is a larx is determined. Responsive to determining the instruction is a larx, whether a cacheline associated with the larx is locked is determined. Responsive to determining the cacheline associated with the larx is not locked, the cacheline associated with the larx is locked and a counter associated with a first thread of the processor core is started. The first thread is processing the first instruction.

Abstract: A plurality of entries including address translation information are buffered in a data structure in a processor core. At least first and second translation entry invalidation requests specifying different first and second addresses are checked against all of the entries in the data structure. The checking includes accessing and checking at least a first entry in the data structure for an address match with the first address but not the second address, thereafter concurrently checking at least a second entry for an address match with both the first and second addresses, and thereafter completing checking for the first address and accessing and checking the first entry for an address match with the second address but not the first address. The processor core invalidates any entry in the data structure for which the checking detects an address match.

Abstract: An arbiter gathers translation invalidation requests assigned to state machines of a lower-level cache into a set for joint handling in a processor core. The gathering includes selection of one of the set of gathered translation invalidation requests as an end-of-sequence (EOS) request. The arbiter issues to the processor core a sequence of the gathered translation invalidation requests terminating with the EOS request. Based on receipt of each of the gathered requests, the processor core invalidates any translation entries providing translation for the addresses specified by the translation invalidation requests and marks memory-referent requests dependent on the invalidated translation entries. Based on receipt of the EOS request and in response to all of the marked memory-referent requests draining from the processor core, the processor core issues a completion request to the lower-level cache indicating completion of servicing by the processor core of the set of gathered translation invalidation requests.

Abstract: A first instruction for processing by a processor core is received. Whether the instruction is a larx is determined. Responsive to determining the instruction is a larx, whether a cacheline associated with the larx is locked is determined. Responsive to determining the cacheline associated with the larx is not locked, the cacheline associated with the larx is locked and a counter associated with a first thread of the processor core is started. The first thread is processing the first instruction.

Abstract: A coherent data processing system includes a system fabric communicatively coupling a plurality of coherence participants and fabric control logic. The fabric control logic quantifies congestion on the system fabric based on coherence messages associated with commands issued on the system fabric. Based on the congestion on the system fabric, the fabric control logic determines a rate of request issuance applicable to a set of coherence participants among the plurality of coherence participants. The fabric control logic issues at least one rate command to set a rate of request issuance to the system fabric of the set of coherence participants.

Abstract: A data processing system includes a plurality of processor cores each supported by a respective one of a plurality of vertical cache hierarchies. Based on receiving on a system fabric a cache injection request requesting injection of a data into a cache line identified by a target real address, the data is written into a cache in a first vertical cache hierarchy among the plurality of vertical cache hierarchies. Based on a value in a field of the cache injection request, a distribute field is set in a directory entry of the first vertical cache hierarchy. Upon eviction of the cache line the first vertical cache hierarchy, a determination is made whether the distribute field is set. Based on determining the distribute field is set, a lateral castout of the cache line from the first vertical cache hierarchy to a second vertical cache hierarchy is performed.

Abstract: A coherent data processing system includes a system fabric communicatively coupling a plurality of coherence participants and fabric control logic. The fabric control logic quantifies congestion on the system fabric based on coherence messages associated with commands issued on the system fabric. Based on the congestion on the system fabric, the fabric control logic determines a rate of request issuance applicable to a set of coherence participants among the plurality of coherence participants. The fabric control logic issues at least one rate command to set a rate of request issuance to the system fabric of the set of coherence participants.

Abstract: A data processing system includes a plurality of processor cores each supported by a respective one of a plurality of vertical cache hierarchies. Based on receiving on a system fabric a cache injection request requesting injection of a data into a cache line identified by a target real address, the data is written into a cache in a first vertical cache hierarchy among the plurality of vertical cache hierarchies. Based on a value in a field of the cache injection request, a distribute field is set in a directory entry of the first vertical cache hierarchy. Upon eviction of the cache line the first vertical cache hierarchy, a determination is made whether the distribute field is set. Based on determining the distribute field is set, a lateral castout of the cache line from the first vertical cache hierarchy to a second vertical cache hierarchy is performed.

Abstract: A processing unit for a data processing system includes a processor core that issues memory access requests and a cache memory coupled to the processor core. The cache memory includes a reservation circuit that tracks reservations established by the processor core via load-reserve requests and a plurality of read-claim (RC) state machines for servicing memory access requests of the processor core. The cache memory, responsive to receipt from the processor core of a store-conditional request specifying a store target address, allocates an RC state machine among the plurality of RC state machines to process the store-conditional request and transfers responsibility for tracking a reservation for the store target address from the reservation circuit to the RC state machine.

Abstract: A processing unit for a data processing system includes a processor core that issues memory access requests and a cache memory coupled to the processor core. The cache memory includes a reservation circuit that tracks reservations established by the processor core via load-reserve requests and a plurality of read-claim (RC) state machines for servicing memory access requests of the processor core. The cache memory, responsive to receipt from the processor core of a store-conditional request specifying a store target address, allocates an RC state machine among the plurality of RC state machines to process the store-conditional request and transfers responsibility for tracking a reservation for the store target address from the reservation circuit to the RC state machine.

Abstract: An upper level cache receives from an associated processor core a plurality of memory access requests including at least first and second memory access requests of differing first and second classes. Based on class histories associated with the first and second classes of memory access requests, the upper level cache initiates, on the system interconnect fabric, a first interconnect transaction corresponding to the first memory access request without first issuing the first memory access request to the lower level cache via a private communication channel between the upper level cache and the lower level cache.

Abstract: An upper level cache receives from an associated processor core a plurality of memory access requests including at least first and second memory access requests of differing first and second classes. Based on class histories associated with the first and second classes of memory access requests, the upper level cache initiates, on the system interconnect fabric, a first interconnect transaction corresponding to the first memory access request without first issuing the first memory access request to the lower level cache via a private communication channel between the upper level cache and the lower level cache.

Abstract: A cache memory includes a data array, a directory of contents of the data array that specifies coherence state information, and snoop logic that processes operations snooped from a system fabric by reference to the data array and the directory. The snoop logic, responsive to snooping on the system fabric a request of a flush or clean memory access operation of an initiating coherence participant, determines whether the directory indicates the cache memory has coherence ownership of a target address of the request. Based on determining the directory indicates the cache memory has coherence ownership of the target address, the snoop logic provides a coherence response to the request that causes coherence ownership of the target address to be transferred to the initiating coherence participant, such that the initiating coherence participant can protect the target address against conflicting requests.

Abstract: A cache memory includes a data array, a directory of contents of the data array that specifies coherence state information, and snoop logic that processes operations snooped from a system fabric by reference to the data array and the directory. The snoop logic, responsive to snooping on the system fabric a request of a first flush/clean memory access operation that specifies a target address, determines whether or not the cache memory has coherence ownership of the target address. Based on determining the cache memory has coherence ownership of the target address, the snoop logic services the request and thereafter enters a referee mode. While in the referee mode, the snoop logic protects a memory block identified by the target address against conflicting memory access requests by the plurality of processor cores until conclusion of a second flush/clean memory access operation that specifies the target address.

Abstract: A cache memory includes a data array, a directory of contents of the data array that specifies coherence state information, and snoop logic that processes operations snooped from a system fabric by reference to the data array and the directory. The snoop logic, responsive to snooping on the system fabric a request of a first flush/clean memory access operation that specifies a target address, determines whether or not the cache memory has coherence ownership of the target address. Based on determining the cache memory has coherence ownership of the target address, the snoop logic services the request and thereafter enters a referee mode. While in the referee mode, the snoop logic protects a memory block identified by the target address against conflicting memory access requests by the plurality of processor cores until conclusion of a second flush/clean memory access operation that specifies the target address.

Abstract: A cache memory includes a data array, a directory of contents of the data array that specifies coherence state information, and snoop logic that processes operations snooped from a system fabric by reference to the data array and the directory. The snoop logic, responsive to snooping on the system fabric a request of a flush or clean memory access operation of an initiating coherence participant, determines whether the directory indicates the cache memory has coherence ownership of a target address of the request. Based on determining the directory indicates the cache memory has coherence ownership of the target address, the snoop logic provides a coherence response to the request that causes coherence ownership of the target address to be transferred to the initiating coherence participant, such that the initiating coherence participant can protect the target address against conflicting requests.