Apparatuses, systems, and techniques to receive, by a processor of a computer system, one or more operations for a kernel; automatically generate, by the processor, one or more operators that perform the one or more operations on elements of one or more input data structures; and automatically generate, by the processor, the kernel that comprises the one or more operators.

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.

Methods, systems and apparatuses for performing walk operations of single instruction, multiple data (SIMD) instructions are disclosed. One method includes initiating, by a scheduler, a SIMD thread, where the scheduler is operative to schedule the SIMD thread. The method further includes fetching a plurality of instructions for the SIMD thread. The method further includes determining, by a thread arbiter, at least one instruction that is a walk instruction, where the walk instruction iterates a block of instructions for a subset of channels of the SIMD thread, where the walk instruction includes a walk size, and where the walk size is a number of channels in the subset of channels of the SIMD thread that are processed in a walk iteration in association with the walk instruction. The method further includes executing the walk instruction based on the walk size.

Embodiments described herein are generally directed to an improved vector normalization instruction. An embodiment of a method includes responsive to receipt by a GPU of a single instruction specifying a vector normalization operation to be performed on V vectors: (i) generating V squared length values, N at a time, by a first processing unit, by, for each N sets of inputs, each representing multiple component vectors for N of the vectors, performing N parallel dot product operations on the N sets of inputs. Generating V sets of outputs representing multiple normalized component vectors of the V vectors, N at a time, by a second processing unit, by, for each N squared length values of the V squared length values, performing N parallel operations on the N squared length values, wherein each of the N parallel operations implement a combination of a reciprocal square root function and a vector scaling function.

An apparatus to facilitate machine learning matrix processing is disclosed. The apparatus comprises a memory to store matrix data one or more processors to execute an instruction to examine a message descriptor included in the instruction to determine a type of matrix layout manipulation operation that is to be executed, examine a message header included in the instruction having a plurality of parameters that define a two-dimensional (2D) memory surface that is to be retrieved, retrieve one or more blocks of the matrix data from the memory based on the plurality of parameters and a register file including a plurality of registers, wherein the one or more blocks of the matrix data is stored within a first set of the plurality of registers.

The invention relates to a method for processing instructions out-of-order on a processor comprising an arrangement of execution units. The inventive method comprises: 1) looking up operand sources in a Register Positioning Table and setting operand input references of the instruction to be issued accordingly; 2) checking for an Execution Unit (EXU) available for receiving a new instruction; and 3) issuing the instruction to the available Execution Unit and enter a reference of the result register addressed by the instruction to be issued to the Execution Unit into the Register Positioning Table (RPT).

A processor includes: an address generating unit that, when an instruction decoded by a decoding unit is an instruction to execute arithmetic processing on a plurality of operand sets each including a plurality of operands that are objects of the arithmetic processing, in parallel a plurality of times, generates an address set corresponding to each of the operand sets of the arithmetic processing for each time, based on a certain address displacement with respect to the plurality of operands included in each of the operand sets; a plurality of instruction queues that hold the generated address sets corresponding to the respective operand sets, in correspondence to respective processing units; and a plurality of processing units that perform the arithmetic processing in parallel on the operand sets obtained based on the respective address sets outputted by the plurality of instruction queues.

A method is described that includes executing a convolutional neural network layer on an image processor having an array of execution lanes and a two-dimensional shift register. The executing of the convolutional neural network includes loading a plane of image data of a three-dimensional block of image data into the two-dimensional shift register. The executing of the convolutional neural network also includes performing a two-dimensional convolution of the plane of image data with an array of coefficient values by sequentially: concurrently multiplying within the execution lanes respective pixel and coefficient values to produce an array of partial products; concurrently summing within the execution lanes the partial products with respective accumulations of partial products being kept within the two dimensional register for different stencils within the image data; and, effecting alignment of values for the two-dimensional convolution within the execution lanes by shifting content within the two-dimensional shift register array.

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.

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.