Abstract: An electronic device receives a single instruction to apply a neural network operation to a set of M-bit elements stored in one or more input vector registers to initiate a sequence of computational operations related to a neural network. In response to the single instruction, the electronic device implements the neural network operation on the set of M-bit elements to generate a set of P-bit elements by obtaining the set of M-bit elements from the one or more input vector registers, quantizing each of the set of M-bit elements from M bits to P bits, and packing the set of P-bit elements into an output vector register. P is smaller than M. In some embodiments, the neural network operation is a quantization operation including at least a multiplication with a quantization factor and an addition with a zero point.

Abstract: Certain aspects of the present disclosure provide techniques and apparatus for efficiently accessing memory in a computing system. An example method includes organizing a plurality of physical memory banks having a base size into a plurality of logical memory banks. A request to execute operations on the plurality of physical memory banks is received. The request to execute the operations comprises a request to interact with data having a sample width based on the base size. Responsive to receiving the request to execute the operations, the operations are executed on one or more logical memory banks of the plurality of logical memory banks via a memory crossbar shared across the plurality of logical memory banks. An amount of the data on which the operations are executed is a multiple of the sample width, and each logical memory bank has a size based on the base size and a multiplier value.

Abstract: Certain aspects of the present disclosure provide techniques and apparatus for quantized machine learning. A quantized input matrix is accessed at a layer of a neural network, and a first interim value is generated in an accumulator by performing matrix multiplication, using the accumulator, of the quantized input matrix and a quantized weight matrix associated with the layer of the neural network. The first interim value is normalized based at least in part on one or more leading sign bits of the first interim value, and the normalized first interim value is dequantized. A second interim value is generated by applying a rounded right-shift operation to the dequantized normalized first interim value, and activation data is generated by applying an activation function to the second interim value.

Abstract: A device includes a vector register file, a memory, and a processor. The vector register file includes a plurality of vector registers. The memory is configured to store a permutation instruction. The processor is configured to access a periodicity parameter of the permutation instruction. The periodicity parameter indicates a count of a plurality of data sources that contain source data for the permutation instruction. The processor is also configured to execute the permutation instruction to, for each particular element of multiple elements of a first permutation result register of the plurality of vector registers, select a data source of the plurality of data sources based at least in part on the count of the plurality of data sources and populate the particular element based on a value in a corresponding element of the selected data source.

Abstract: A device includes a processor that includes a rotation vector register file, a second vector register file, and multiply-accumulate circuitry (MAC). The rotation vector register file includes a rotation vector register. The rotation vector register file is configured to rotate data in the rotation vector register. The second vector register file includes a source vector register. The MAC is configured to receive first input data from the rotation vector register file and second input data from the source vector register.

Abstract: This application is directed to using a single instruction to initiate a sequence of computational operations related to a neural network. An electronic device receives a single instruction to apply a neural network operation to a set of M-bit elements stored in one or more input vector registers. In response to the single instruction, the electronic device implements the neural network operation on the set of M-bit elements to generate a set of P-bit elements by obtaining the set of M-bit elements from the one or more input vector registers, quantizing each of the set of M-bit elements from M bits to P bits, and packing the set of P-bit elements into an output vector register. P is smaller than M. In some embodiments, the neural network operation is a quantization operation including at least a multiplication with a quantization factor and an addition with a zero point.

Abstract: This application is directed to using a single instruction to initiate a sequence of computational operations related to a neural network activation function. An electronic device receives a single instruction to apply a linear activation operation to a set of first elements stored in one or more input vector registers. In response to the single instruction, the linear activation operation is implemented on the set of first elements to generate a set of output elements. For each first element, the electronic device detects a sign value of the respective first element, selects a respective scalar from one or more scalars based on the sign value, and applies the linear activation operation on the respective first element based on the selected respective scalar and a bias value to generate a respective element of the set of output elements. The electronic device quantizes the set of output elements.

Abstract: A device includes a vector register file, a memory, and a processor. The vector register file includes a plurality of vector registers. The memory is configured to store a permutation instruction. The processor is configured to access a periodicity parameter of the permutation instruction. The periodicity parameter indicates a count of a plurality of data sources that contain source data for the permutation instruction. The processor is also configured to execute the permutation instruction to, for each particular element of multiple elements of a first permutation result register of the plurality of vector registers, select a data source of the plurality of data sources based at least in part on the count of the plurality of data sources and populate the particular element based on a value in a corresponding element of the selected data source.