Abstract: An Adaptive Memory Mirroring Performance Accelerator (AMMPA) includes a centralized transaction handling block that dynamically maps the most frequently accessed memory regions into faster access memory. The technique creates shadow copies of the most frequently accessed memory regions in memory devices associated with lower latency. The regions for which shadow copies are provided are updated dynamically based on use, and the technique flexible for different memory hierarchies.

Abstract: In an example, a method includes storing code for a first central processing unit (CPU) executing a first application in a first region of a memory, and storing code for a second CPU executing a second application in a second region of the memory. The method includes storing shared code for the first CPU and the second CPU in a third region of the memory. The method includes storing read-write data for the first CPU in a fourth region of the memory and storing read-write data for the second CPU in a fifth region of the memory. The method includes translating a first address from a first unique address space for the first CPU to a shared address space in the third region, and translating a second address from a second unique address space for the second CPU to the shared address space in the third region.

Abstract: An Adaptive Memory Mirroring Performance Accelerator (AMMPA) includes a centralized transaction handling block that dynamically maps the most frequently accessed memory regions into faster access memory. The technique creates shadow copies of the most frequently accessed memory regions in memory devices associated with lower latency. The regions for which shadow copies are provided are updated dynamically based on use, and the technique flexible for different memory hierarchies.

Abstract: An example apparatus includes: a first interface configured to couple to a processor core; a second interface configured to couple to a first memory configured to store an image that includes a set of slices; a third interface coupled to the first interface, the third interface configured to couple to a second memory; a direct memory access circuit coupled to the second interface and the third interface and configured to: receive a transaction from the second interface, wherein the transaction specifies a read of a slice of the set of slices; and based on the transaction: read the slice from the first memory; perform on-the-fly operations to the slice; and store the slice in the second memory.

Abstract: Systems, methods and devices that improve fault detection capability of an imaging/vision hardware accelerator are provided. One such system includes a hardware accelerator, a signature generator, a signature processor, and a controller. These components cooperate to generate first and second output frames based on first and second reference frames, respectively; generate a third output frame based on a use-case frame; generate first and second image signatures based on the first and second output frames, respectively; compare the first image signature to a stored first reference image signature and output a first result; and compare the second image signature to a stored second reference image signature and output a second result. The controller determines, based on the results, whether the hardware accelerator has a fault at either a first time or a second time. When no fault is detected at either time, the controller analyzes the use-case frame for designation as an adaptive reference frame.

Abstract: A system-on-chip (SoC) in which trace data is managed includes a first memory device, a first interface to couple the first memory to a second memory external to the system-on-chip, and a first processing resource coupled to the first interface and the first memory device. The first processing resource includes a data buffer and a first direct access memory (DMA) controller. The first DMA controller transmits data from the data buffer to the first interface over a first channel, and transmits the data from the data buffer with associated trace information for the data to the first memory device over a second channel.

Abstract: An electronic device may be configured to detect a fault in imaging and vision hardware accelerators. The electronic device may include a controller configured to select a first golden input frame of multiple golden input frames to perform a first self-test, and retrieve a first reference image signature corresponding to the first golden input frame. The electronic device may include a hardware accelerator module configured to obtain the first golden input frame, and generate a first output frame based on the first golden input frame. The electronic device may include a signature generator configured to generate a first generated image signature based on the first output frame. The electronic device may include a signature comparison module configured to compare the first generated image signature to the first reference image signature in order to determine whether the hardware accelerator module includes a fault at a first time.

Abstract: A hardware neural network engine which uses checksums of the matrices used to perform the neural network computations. For fault correction, expected checksums are compared with checksums computed from the matrix developed from the matrix operation. The expected checksums are developed from the prior stage of the matrix operations or from the prior stage of the matrix operations combined with the input matrices to a matrix operation. This use of checksums allows reading of the matrices from memory, the dot product of the matrices and the accumulation of the matrices to be fault corrected without triplication of the matrix operation hardware and extensive use of error correcting codes. The nonlinear stage of the neural network computation is done using triplicated nonlinear computational logic. Fault detection is done in a similar manner, with fewer checksums needed and correction logic removed as compared to the fault correction operation.

Abstract: A technique including receiving an image stream for processing; processing the received image stream in a real time mode of operation; outputting an indication that an image processing pipeline has begun processing the received image stream; receiving, in response to the indication, first configuration information associated with test data for testing the image processing pipeline; switching the image processing pipeline to a non-real time mode of operation to process the test data based on the first configuration information during a vertical blanking period of the received image stream; loading the test data from an external memory; switching an input of the image processing pipeline from the image stream to the test data; determining a checksum based on the processed test data; comparing the determined checksum to an expected checksum to determine that the test data was successfully processed; and outputting an indication that the test data was successfully processed.

Abstract: A microcontroller is provided and comprises a central repository, a processing device, and a firewall. Rule repository memory in the central repository stores one or more access rules defining an access permission of a software context to one or more target resources of the microcontroller. The firewall receives a bus transaction initiated based on an instruction and determines whether any access rule stored in memory of the firewall defines the access permission of the software context to a destination resource. If no access rule stored in the firewall memory defines the access permission, the firewall communicates a miss query condition to the central repository. The central repository searches the rule repository memory for an access rule defining the access permission of the software context to the destination resource, and if a related access rule is found, the related access rule is stored in the firewall memory.

Abstract: An electronic device, comprising: a first component configured to transmit a first set of data to a second component by providing a first memory request specifying the first set of data for and an input memory address, and a transaction tracking unit coupled to a first transport interface, the transaction tracking unit configured to: receive the first memory request; transmit a second memory request that specifies at least a first portion of the first set of data, via the first transport interface, to the second component; receive a response to the second memory request from the second component; determine that the response corresponds to the second memory request; and provide, to the first component, an output response based on the received response to the second memory request.

Abstract: Techniques for executing machine learning (ML) models including receiving an indication to run a ML model, receiving synchronization information for organizing the running of the ML model with other ML models, determining, based on the synchronization information, to delay running the ML model, delaying the running of the ML model, determining, based on the synchronization information, a time to run the ML model; and running the ML model at the time.

Abstract: An electronic device, comprising: a first component configured to transmit a first set of data to a second component by providing a first memory request specifying the first set of data for and an input memory address, and a transaction tracking unit coupled to a first transport interface, the transaction tracking unit configured to: receive the first memory request; transmit a second memory request that specifies at least a first portion of the first set of data, via the first transport interface, to the second component; receive a response to the second memory request from the second component; determine that the response corresponds to the second memory request; and provide, to the first component, an output response based on the received response to the second memory request.

Abstract: A hardware neural network engine which uses checksums of the matrices used to perform the neural network computations. For fault correction, expected checksums are compared with checksums computed from the matrix developed from the matrix operation. The expected checksums are developed from the prior stage of the matrix operations or from the prior stage of the matrix operations combined with the input matrices to a matrix operation. This use of checksums allows reading of the matrices from memory, the dot product of the matrices and the accumulation of the matrices to be fault corrected without triplication of the matrix operation hardware and extensive use of error correcting codes. The nonlinear stage of the neural network computation is done using triplicated nonlinear computational logic. Fault detection is done in a similar manner, with fewer checksums needed and correction logic removed as compared to the fault correction operation.