An arithmetic processing device includes arithmetic processing units, each having a calculator unit; a scheduler that controls a push instruction to write data to a register file in one of the arithmetic processing units and a pull instruction to read data from the register file; a pull request bus to which the scheduler outputs a pull request and which is connected to the arithmetic processing units; a push request bus to which the scheduler outputs a push request and which is connected to the arithmetic processing units; and a pull data bus that inputs, into the scheduler, pull data read from the register file in response to the pull request. Each of the arithmetic processing units includes a pull data turn-back bus that propagates pull data read from its register file to the pull data bus.

An addition mask value generator is provided. A first operation circuit is configured to obtain first intermediate data according to first output mask value and fourth output mask value. A second operation circuit is configured to obtain the addition output mask value of a first mask group according to first intermediate data and fourth input mask value. A third operation circuit is configured to obtain second intermediate data according to second output mask value and third output mask value. A fourth operation circuit is configured to obtain the addition output mask value of a second mask group according to second intermediate data and second input mask value. The first and second addition input mask values of first mask group are first and second input mask values. The first and second addition input mask values of second mask group are third input mask value and first intermediate data.

There is provided execution circuitry. Storage circuitry retains a stored state of the execution circuitry. Operation receiving circuitry receives, from issue circuitry, an operation signal corresponding to an operation to be performed that accesses the stored state of the execution circuitry from the storage circuitry. Functional circuitry seeks to perform the operation in response to the operation signal by accessing the stored state of the execution circuitry from the storage circuitry. Delete request receiving circuitry receives a deletion signal and in response to the deletion signal, deletes the stored state of the execution circuitry from the storage circuitry. State loss indicating circuitry responds to the operation signal when the stored state of the execution circuitry is not present and is required for the operation by indicating an error. In addition, there is provided a data processing apparatus comprising issue circuitry to issue an operation to execution circuitry. The execution circuitry stores a stored state that is accessed during performance of the operation and error detecting circuitry detects an indication of an error from the execution circuitry that the stored state is required for performance of the operation and that the stored state has been deleted.

A novel and useful system and method of improved power performance and lowered memory requirements for an artificial neural network based on packing memory utilizing several structured sparsity mechanisms. The invention applies to neural network (NN) processing engines adapted to implement mechanisms to search for structured sparsity in weights and activations, resulting in a considerably reduced memory usage. The sparsity guided training mechanism synthesizes and generates structured sparsity weights. A compiler mechanism within a software development kit (SDK), manipulates structured weight domain sparsity to generate a sparse set of static weights for the NN. The structured sparsity static weights are loaded into the NN after compilation and utilized by both the structured weight domain sparsity mechanism and the structured activation domain sparsity mechanism. The application of structured sparsity lowers the span of search options and creates a relatively loose coupling between the data and control planes.

An exemplary SIMD computing system comprises a SIMD processing element (SPE) configured to perform a selected operation on a portion of a processor input data word, with the operation selected by control signals read from a control memory location addressed by a decoded instruction. The SPE may comprise one or more adder, multiplier, or multiplexer coupled to the control signals. The control signals may comprise one or more bit read from the control memory. The control memory may be an MxN (M rows by N columns) memory having M possible SIMD operations and N control signals. Each instruction decoded may select an SPE operation from among N rows. A plurality of SPEs may receive the same control signals. The control memory may be rewritable, advantageously permitting customizable SIMD operations that are reconfigurable by storing in the control memory locations control signals designed to cause the SPE to perform selected operations.

One embodiment provides for a compute apparatus comprising a decode unit to decode a single instruction into a decoded instruction that specifies multiple operands including a multi-bit input value and a ternary weight associated with a neural network and an arithmetic logic unit including a multiplier, an adder, and an accumulator register. To execute the decoded instruction, the multiplier is to perform a multiplication operation on the multi-bit input based on the ternary weight to generate an intermediate product and the adder is to add the intermediate product to a value stored in the accumulator register and update the value stored in the accumulator register.

Embodiments of systems, methods, and apparatuses for heterogeneous computing are described. In some embodiments, a hardware heterogeneous scheduler dispatches instructions for execution on one or more plurality of heterogeneous processing elements, the instructions corresponding to a code fragment to be processed by the one or more of the plurality of heterogeneous processing elements, wherein the instructions are native instructions to at least one of the one or more of the plurality of heterogeneous processing elements.

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 a set of cache lines, line type bits configured to store an indication that a corresponding cache line of the set of cache lines is configured to store write-miss data, and an eviction controller configured to flush stored write-miss data based on the line type bits.

Methods, apparatuses, and systems for in- or near-memory processing are described. Strings of bits (e.g., vectors) may be fetched and processed in logic of a memory device without involving a separate processing unit. Operations (e.g., arithmetic operations) may be performed on numbers stored in a bit-serial way during a single sequence of clock cycles. Arithmetic may thus be performed in a single pass as numbers are bits of two or more strings of bits are fetched and without intermediate storage of the numbers. Vectors may be fetched (e.g., identified, transmitted, received) from one or more bit lines. Registers of the memory array may be used to write (e.g., store or temporarily store) results or ancillary bits (e.g., carry bits or carry flags) that facilitate arithmetic operations. Circuitry near, adjacent, or under the memory array may employ XOR or AND (or other) logic to fetch, organize, or operate on the data.

A discrete three-dimensional (3-D) processor comprises first and second dice. The first die comprises 3-D memory (3D-M) arrays, whereas the second die comprises logic circuits and at least an off-die peripheral-circuit component of the 3D-M array(s). Typical off-die peripheral-circuit component could be an address decoder, a sense amplifier, a programming circuit, a read-voltage generator, a write-voltage generator, a data buffer, or a portion thereof.