Abstract: A processing system includes a first processor configured to issue a first request in a first format, an adapter configured to receive the first request in the first format and send the first request in a second format, and a memory coherency interconnect configured to receive the first request in the second format and determine whether the first request in the second format is for a translation lookaside buffer (TLB) operation or a non-TLB operation based on information in the first request in the second format. When the first request in the second format is for a TLB operation, the interconnect routes the first request in the second format to a TLB global ordering point (GOP). When the first request in the second format is not for a TLB operation, the interconnect routes the first request in the second format to a non-TLB GOP.

Abstract: A data processing system includes a plurality of switch points interconnected by a write data network and a write request network. Each switch point includes write request switch circuitry having write request ingress ports and write request egress ports coupled to the write request network and arbitration circuitry configured to grant a write request received at one of the write request ingress ports access to one of the write request egress ports. Each switch point also includes write data switch circuitry having write data ingress ports and write data egress ports coupled to the write data network. In response to the write request arbitration circuitry granting the write request, allowing write data from the write data ingress port corresponding to the one of the write request ingress ports to be provided at the write data egress port which corresponds to the one of the write request egress ports.

Abstract: A method of operating a cache-coherent computing system includes storing first state information corresponding to a first reservation for a first exclusive access to a first memory address requested by a first thread executing on a first processor of a first plurality of processors. The method includes transmitting an output atomic response transaction indicating a status of the first reservation to a coherency interconnection in response to issuance of the first exclusive access to the coherency interconnection. The output atomic response transaction is based on first state information.

Abstract: A data processing system includes a network of interconnected switch points having a plurality of edge switch points located at an edge of the network; a plurality of network interface controllers, wherein each edge switch point of the plurality of edge points is coupled to a corresponding network interface controller of the plurality of network interface controllers; a plurality of target controllers; and a crossbar switch coupled between the plurality of network interface controllers and the plurality of target controllers. The crossbar switch is configured to communicate read/write signals between any one of the plurality of network interface controllers and any one of the plurality of target controllers.

Abstract: A method of operating a cache-coherent computing system includes storing first state information corresponding to a first reservation for a first exclusive access to a first memory address requested by a first thread executing on a first processor of a first plurality of processors. The method includes transmitting an output atomic response transaction indicating a status of the first reservation to a coherency interconnection in response to issuance of the first exclusive access to the coherency interconnection. The output atomic response transaction is based on first state information.

Abstract: A processing system includes a first processor configured to issue a first request in a first format, an adapter configured to receive the first request in the first format and send the first request in a second format, and a memory coherency interconnect configured to receive the first request in the second format and determine whether the first request in the second format is for a translation lookaside buffer (TLB) operation or a non-TLB operation based on information in the first request in the second format. When the first request in the second format is for a TLB operation, the interconnect routes the first request in the second format to a TLB global ordering point (GOP). When the first request in the second format is not for a TLB operation, the interconnect routes the first request in the second format to a non-TLB GOP.

Abstract: A data processing system includes a plurality of switch points interconnected by a write data network and a write request network. Each switch point includes write request switch circuitry having write request ingress ports and write request egress ports coupled to the write request network and arbitration circuitry configured to grant a write request received at one of the write request ingress ports access to one of the write request egress ports. Each switch point also includes write data switch circuitry having write data ingress ports and write data egress ports coupled to the write data network. In response to the write request arbitration circuitry granting the write request, allowing write data from the write data ingress port corresponding to the one of the write request ingress ports to be provided at the write data egress port which corresponds to the one of the write request egress ports.

Abstract: A method and apparatus for calculating a victim way that is always the least recently used way. More specifically, in an m-set, n-way set associative cache, each way in a cache set comprises a valid bit that indicates that the way contains valid data. The valid bit is set when a way is written and cleared upon being invalidated, e.g., via a snoop address, The cache system comprises a cache LRU circuit which comprises an LRU logic unit associated with each cache set. The LRU logic unit comprises a FIFO of n-depth (in certain embodiments, the depth corresponds to the number of ways in the cache) and m-width. The FIFO performs push, pop and collapse functions. Each entry in the FIFO contains the encoded way number that was last accessed.

Abstract: Ordered write transactions from requester devices to multiple target devices are controlled using switch point networks. The requester device and the multiple target devices for the write transactions are coupled to a network of interconnected switch points. Write requests are generated for a plurality of parcels associated with a block of data to be written. The write requests have a particular order associated with an order in which the parcels are to be written, and these write requests are provided to the switch point interconnection network in the particular order. At least one of the switch points is then used to control the flow of write requests to the multiple target devices such that the particular order is maintained. In one embodiment, the target devices are memory devices, and the particular order is based upon the AXI (Advanced eXtensible Interface) protocol.

Abstract: A plurality of processing units are interconnected by a coherency network in accordance with a directed spanning tree. Each processing unit that is a leaf of the directed spanning tree includes processing circuitry to provide a coherency response in response to a snoop request. Each processing unit which is not a root or leaf of the directed spanning tree includes switch point circuitry having one or more ingress ports coupled to neighboring processing units in accordance with the directed spanning tree. The switch point circuitry includes a coherency tracking table configured to store a combined coherency response in response to a particular snoop request based on one or more coherency responses received at the one or more ingress ports from the neighboring processing units.

Abstract: A network having a plurality of switch points, each switch point having both a main multi-stage pipeline and a look-ahead pipeline between input ports and output ports of the plurality of switch points is described. The look-ahead pipeline has fewer pipeline stages than the main multi-stage pipeline. Look-ahead information and corresponding packet are received at an input port. A first stage look-ahead request is generated from the look-ahead information. A second stage look-ahead request is generated in response to the first stage look-ahead request being not granted. And in response to the second stage look-ahead request being granted, transmitting a packet through the switch point using the look-ahead pipeline.

Abstract: Piggy-back snoops are used for non-coherent memory transactions in distributed processing systems. Coherent and non-coherent memory transactions are received from a plurality of processing cores within a distributed processing system. Non-coherent snoop information for the non-coherent memory transactions is combined with coherent snoop information for the coherent memory transactions to form expanded snoop messages. The expanded snoop messages are then output to a snoop bus interconnect during snoop cycles for the distributed processing system. As such, when the processing cores monitor the snoop bus interconnect, the processing cores receive the non-coherent snoop information along with coherent snoop information within the same snoop cycle. While this piggy-backing of non-coherent snoop information with coherent snoop information uses an expanded snoop bus interconnect, usage of the coherent snoop bandwidth is significantly reduced thereby improving overall performance of the distributed processing system.

Abstract: A network having a plurality of switch points, each switch point having both a main multi-stage pipeline and a look-ahead pipeline between input ports and output ports of the plurality of switch points is described. The look-ahead pipeline has fewer pipeline stages than the main multi-stage pipeline. Look-ahead information and corresponding packet are received at an input port. A first stage look-ahead request is generated from the look-ahead information. A second stage look-ahead request is generated in response to the first stage look-ahead request being not granted. And in response to the second stage look-ahead request being granted, transmitting a packet through the switch point using the look-ahead pipeline.

Abstract: A plurality of processing units are interconnected by a coherency network in accordance with a directed spanning tree. Each processing unit that is a leaf of the directed spanning tree includes processing circuitry to provide a coherency response in response to a snoop request. Each processing unit which is not a root or leaf of the directed spanning tree includes switch point circuitry having one or more ingress ports coupled to neighboring processing units in accordance with the directed spanning tree. The switch point circuitry includes a coherency tracking table configured to store a combined coherency response in response to a particular snoop request based on one or more coherency responses received at the one or more ingress ports from the neighboring processing units.

Abstract: A data processing system includes a network of interconnected switch points having a plurality of edge switch points located at an edge of the network; a plurality of network interface controllers, wherein each edge switch point of the plurality of edge points is coupled to a corresponding network interface controller of the plurality of network interface controllers; a plurality of target controllers; and a crossbar switch coupled between the plurality of network interface controllers and the plurality of target controllers. The crossbar switch is configured to communicate read/write signals between any one of the plurality of network interface controllers and any one of the plurality of target controllers.

Abstract: Ordered write transactions from requester devices to multiple target devices are controlled using switch point networks. The requester device and the multiple target devices for the write transactions are coupled to a network of interconnected switch points. Write requests are generated for a plurality of parcels associated with a block of data to be written. The write requests have a particular order associated with an order in which the parcels are to be written, and these write requests are provided to the switch point interconnection network in the particular order. At least one of the switch points is then used to control the flow of write requests to the multiple target devices such that the particular order is maintained. In one embodiment, the target devices are memory devices, and the particular order is based upon the AXI (Advanced eXtensible Interface) protocol.

Abstract: Piggy-back snoops are used for non-coherent memory transactions in distributed processing systems. Coherent and non-coherent memory transactions are received from a plurality of processing cores within a distributed processing system. Non-coherent snoop information for the non-coherent memory transactions is combined with coherent snoop information for the coherent memory transactions to form expanded snoop messages. The expanded snoop messages are then output to a snoop bus interconnect during snoop cycles for the distributed processing system. As such, when the processing cores monitor the snoop bus interconnect, the processing cores receive the non-coherent snoop information along with coherent snoop information within the same snoop cycle. While this piggy-backing of non-coherent snoop information with coherent snoop information uses an expanded snoop bus interconnect, usage of the coherent snoop bandwidth is significantly reduced thereby improving overall performance of the distributed processing system.

Abstract: A processor provides memory request and a coherency state value for a coherency granule associated with a memory request. The processor further provides either a first indicator or a second indicator depending on whether the coherency state value represents a cumulative coherency state for a plurality of caches of the processor. The first indicator and the second indicator identify the coherency state value as representing a cumulative coherency state or a potentially non-cumulative coherency state, respectively. If the second indicator is provided, a transaction management module determines whether to request the cumulative coherency state for the coherency granule in response to receiving the second indicator. The transaction management module then provides an indicator of the request for the cumulative coherency state to the processor in response to determining to request the cumulative coherency state.

Abstract: A method and apparatus for calculating a victim way that is always the least recently used way. More specifically, in an m-set, n-way set associative cache, each way a cache set comprises a valid bit that indicates that the way contains valid data. The valid bit is set when a way is written and cleared upon being invalidated, e.g., via a snoop address, The cache system comprises a cache LRU circuit which comprises an LRU logic unit associated with each cache set. The LRU logic unit comprises a FIFO of n-depth (in certain embodiments, the depth corresponds to the number of ways in the cache) and m-width. The FIFO performs push, pop and collapse functions. Each entry in the FIFO contains the encoded way number that was last accessed.

Abstract: A method includes communicating a first message between processors of a multiprocessor system via a coherency interconnect, whereby the first message includes coherency information. The method further includes communicating a second message between processors of the multiprocessor system via the coherency interconnect, whereby the second message includes interprocessor message information. A system includes a coherency interconnect and a processor. The processor includes an interface configured to receive messages from the coherency interconnect, each message including one of coherency information or interprocessor message information. The processor further includes a coherency management module configured to process coherency information obtained from at least one of the messages and an interrupt controller configured to generate an interrupt based on interprocessor message information obtained from at least one of the messages.