In various examples, a VPU and associated components may be optimized to improve VPU performance and throughput. For example, the VPU may include a min/max collector, automatic store predication functionality, a SIMD data path organization that allows for inter-lane sharing, a transposed load/store with stride parameter functionality, a load with permute and zero insertion functionality, hardware, logic, and memory layout functionality to allow for two point and two by two point lookups, and per memory bank load caching capabilities. In addition, decoupled accelerators may be used to offload VPU processing tasks to increase throughput and performance, and a hardware sequencer may be included in a DMA system to reduce programming complexity of the VPU and the DMA system. The DMA and VPU may execute a VPU configuration mode that allows the VPU and DMA to operate without a processing controller for performing dynamic region based data movement operations.

Technologies for addressing individual bits in memory include a device having a memory that includes partitions that each have tiles, in which each tile stores an individual bit. The device also includes circuitry to receive a request to access (e.g., read or write) a sequence of bits in a partition. The request specifies a logical row or column address. A corresponding tile is determined from the logical row or column address and for each bit in the sequence. The corresponding tile is accessed to read or write the bit therein.

The application provides a processing method and device. Weights and input neurons are quantized respectively, and a weight dictionary, a weight codebook, a neuron dictionary, and a neuron codebook are determined. A computational codebook is determined according to the weight codebook and the neuron codebook. Meanwhile, according to the application, the computational codebook is determined according to two types of quantized data, and the two types of quantized data are combined, which facilitates data processing.

A streaming engine employed in a digital data processor may specify a fixed read-only data stream defined by plural nested loops. An address generator produces address of data elements for the nested loops. A steam head register stores data elements next to be supplied to functional units for use as operands. A stream template register independently specifies a linear address or a circular address mode for each of the nested loops.

Embodiments detailed herein relate to matrix operations. In particular, the loading of a matrix (tile) from memory. For example, support for a loading instruction is described in the form of decode circuitry to decode an instruction having fields for an opcode, a destination matrix operand identifier, and source memory information, and execution circuitry to execute the decoded instruction to load groups of strided data elements from memory into configured rows of the identified destination matrix operand to memory.

Methods, apparatus, systems and articles of manufacture are disclosed facilitate read-modify-write support in a coherent victim cache with parallel data paths. An example apparatus includes a random-access memory configured to be coupled to a central processing unit via a first interface and a second interface, the random-access memory configured to obtain a read request indicating a first address to read via a snoop interface, an address encoder coupled to the random-access memory, the address encoder to, when the random-access memory indicates a hit of the read request, generate a second address corresponding to a victim cache based on the first address, and a multiplexer coupled to the victim cache to transmit a response including data obtained from the second address of the victim cache.

A caching system including a first sub-cache and a second sub-cache in parallel with the first sub-cache, wherein the second sub-cache includes: line type bits configured to store an indication that a corresponding cache line of the second sub-cache is configured to store write-miss data, and an eviction controller configured to evict a cache line of the second sub-cache storing write-miss data based on an indication that the cache line has been fully written.

Apparatuses, systems, and techniques to select cache policies. In at least one embodiment, a system causes one or more cache policies of one or more caches to be selected based, at least in part, on one or more neural networks to use data stored in the one or more caches.

Embodiments detailed herein relate to matrix operations. In particular, the loading of a matrix (tile) from memory. For example, support for a loading instruction is described in at least a form of decode circuitry to decode an instruction having fields for an opcode, a source matrix operand identifier, and destination memory information, and execution circuitry to execute the decoded instruction to store each data element of configured rows of the identified source matrix operand to memory based on the destination memory information.

A streaming engine employed in a digital data processor specifies a fixed read only data stream defined by plural nested loops. An address generator produces address of data elements. A steam head register stores data elements next to be supplied to functional units for use as operands. The streaming engine stores an early address of next to be fetched data elements and a late address of a data element in the stream head register for each of the nested loops. The streaming engine stores an early loop counts of next to be fetched data elements and a late loop counts of a data element in the stream head register for each of the nested loops.