The present disclosure provides a computation device. The computation device is configured to perform a machine learning computation, and includes an operation unit, a controller unit, and a conversion unit. The storage unit is configured to obtain input data and a computation instruction. The controller unit is configured to extract and parse the computation instruction from the storage unit to obtain one or more operation instructions, and to send the one or more operation instructions and the input data to the operation unit. The operation unit is configured to perform operations on the input data according to one or more operation instructions to obtain a computation result of the computation instruction. In the examples of the present disclosure, the input data involved in machine learning computations is represented by fixed-point data, thereby improving the processing speed and efficiency of training operations.

Embodiments of apparatuses and methods for variable-length instruction steering to instruction decode clusters are disclosed. In an embodiment, an apparatus includes a decode cluster and chunk steering circuitry. The decode cluster includes multiple instruction decoders. The chunk steering circuitry is to break a sequence of instruction bytes into a plurality of chunks, create a slice from a one or more of the plurality of chunks based on one or more indications of a number of instructions in each of the one or more of the plurality of chunks, wherein the slice has a variable size and includes a plurality of instructions, and steer the slice to the decode cluster.

The present disclosure provides a computation device. The computation device is configured to perform a machine learning computation, and includes an operation unit, a controller unit, and a conversion unit. The storage unit is configured to obtain input data and a computation instruction. The controller unit is configured to extract and parse the computation instruction from the storage unit to obtain one or more operation instructions, and to send the one or more operation instructions and the input data to the operation unit. The operation unit is configured to perform operations on the input data according to one or more operation instructions to obtain a computation result of the computation instruction. In the examples of the present disclosure, the input data involved in machine learning computations is represented by fixed-point data, thereby improving the processing speed and efficiency of training operations.

A system and method for reducing pipeline latency. In one embodiment, a processing system includes a processing pipeline. The processing pipeline includes a plurality of processing stages. Each stage is configured to further processing provided by a previous stage. A first of the stages is configured to perform a first function in a pipeline cycle. A second of the stages is disposed downstream of the first of the stages, and is configured to perform, in a pipeline cycle, a second function that is different from the first function. The first of the stages is further configured to selectably perform the first function and the second function in a pipeline cycle, and bypass the second of the stages.

A processor architecture arrangement for emulated shared memory (ESM) architectures is disclosed. The arrangement has a number of multi-threaded processors, each provided with an interleaved inter-thread pipeline and a plurality of functional units for carrying out arithmetic and logical operations on data. The pipeline has at least two operatively parallel pipeline branches. The first pipeline branch includes a first sub-group of the plurality of functional units, such as ALUs (arithmetic logic unit) for carrying out integer operations. The second pipeline branch includes non-overlapping subgroup of the plurality of functional units, such as FPUs (floating point unit) for carrying out floating point operations. One or more of the functional units of at least the second sub-group are located operatively in parallel with the memory access segment of the pipeline.

A system and method for avoiding write back collisions. The system receives a plurality of instructions at a pipeline queue. Next an issue queue determines a number of cycles for each instruction of the plurality of instructions. The issue queue further determines if a collision will occur between at least two of the instructions. Additionally, the system determines in response to a collision between at least two of the instructions, a number of cycles to delay at least one of the at least two instructions. The instructions are then executed. The system then places the results of the instruction for instructions that had a calculated delay in a result buffer for the determined number of cycles of delay. After the determined number of cycles of delay, the system sends the results to a results mux. Once received at the results mux the results are written back to the register file.

Devices and techniques for variable pipeline length in a barrel-multithreaded processor are described herein. A completion time for an instruction can be determined prior to insertion into a pipeline of a processor. A conflict between the instruction and a different instruction based on the completion time can be detected. Here, the different instruction is already in the pipeline and the conflict detected when the completion time equals the previously determined completion time for the different instruction. A difference between the completion time and an unconflicted completion time can then be calculated and completion of the instruction delayed by the difference.

In an embodiment, a coprocessor may include a bypass indication which identifies execution circuitry that is not used by a given processor instruction, and thus may be bypassed. The corresponding circuitry may be disabled during execution, preventing evaluation when the output of the circuitry will not be used for the instruction. In another embodiment, the coprocessor may implement a grid of processing elements in rows and columns, where a given coprocessor instruction may specify an operation that causes up to all of the processing elements to operate on vectors of input operands to produce results. Implementations of the coprocessor may implement a portion of the processing elements. The coprocessor control circuitry may be designed to operate with the full grid or partial grid, reissuing instructions in the partial grid case to perform the requested operation. In still another embodiment, the coprocessor may be able to fuse vector mode operations.

Apparatus and methods are disclosed for implementing block-based processors having custom function blocks, including field-programmable gate array (FPGA) implementations. In some examples of the disclosed technology, a dynamically configurable scheduler is configured to issue at least one block-based processor instruction. A custom function block is configured to receive input operands for the instruction and generate ready state data indicating completion of a computation performed for the instruction by the respective custom function block.

An integrated circuit may include elastic datapaths or pipelines, through which software threads or iterations of loops, may be executed. Throttling circuitry may be coupled along an elastic pipeline in the integrated circuit. The throttling circuitry may include dependency detection circuitry that dynamically detect memory dependency issues that may arise during runtime. To mitigate these dependency issues, the throttling circuitry may assert stall signals to upstream stages in the pipeline. Additionally, the throttling circuitry may control the pipeline to resolve a store operation prior to a corresponding load operation in order to avoid store/load conflicts. In an embodiment, the throttling circuitry may include a validator circuit, a rewind block, a revert block, and a flush block. The throttling circuitry may pass speculative iterations through the rewind block, and later validate the speculative iterations using the validator block.