Baidu USA LLC

Juniper Networks, Inc.

Min Guo

Manjiang Zhang

Shengjin Zhou

Tiffany Zhang

Ji Li

Davy Huang

Xu Zhou

Kwan Oh

Bernard Deadman

Chang Shu

Michael E Marcellin

Nancy G Weintraub

Brian M Cooper

Spandan Shivannaiah

Krishnakumar Angia Ramamoorthy

Thomas P McCarthy

Srinivas Attili

Shalini Ruppa Subramanian

A data format converter rearranges data of an input image for input to a systolic array of multiply and accumulate processing elements. The image has a pixel height and a pixel width in a number of channels equal to a number of colors per pixel. The data format converter rearranges the data to a second, greater number of channels and inputs the second number of channels to one side of the systolic array. The second number of channels is less than or equal to the number of MAC PEs on the one side of the systolic array, and results in greater MAC PE utilization in the systolic array.

Data format transform method to improve AI engine MAC utilization

In one embodiment, an exemplary computer-implemented method of storing point cloud data in an autonomous driving vehicle can include the operations of receiving raw point cloud data from a LiDAR sensor mounted on the autonomous driving vehicle, the raw point cloud data representing cloud data points acquired in response to laser beams emitted at a given angle; retrieving configuration information of the LiDAR sensor, the configuration information including at least a number of laser lines of the LiDAR sensor. The method further includes the operations of constructing, based on the configuration information, a data structure that includes a data entry for each of the cloud data points, the data entry including multiple fields for storing attributes of the cloud data point, each field having a bit width determined based on the configuration information using a predetermined algorithm; and writing the cloud data points to a storage medium using the data structure.

Point cloud format optimized for LiDAR data storage based on device property

Batch normalization (BN) layer fusion and quantization method for model inference in artificial intelligence (AI) network engine are disclosed. A method for a neural network (NN) includes merging batch normalization (BN) layer parameters with NN layer parameters and computing merged BN layer and NN layer functions using the merged BN and NN layer parameters. A rectified linear unit (RELU) function can be merged with the BN and NN layer functions.

Batch normalization layer fusion and quantization method for model inference in AI neural network engine

Embodiments relate to systems and methods to optimize quantization of tensors of an AI model. According to one embodiment, a system receives an AI model having one or more layers. The system receives a number of input data for offline inferencing and applies offline inferencing to the AI model based on the input data to generate offline data distributions for the AI model. The system quantizes one or more tensors of the AI model based on the offline data distributions to generate a low-bit representation AI model, where each layer of the AI model includes the one or more tensors, where the one or more tensors include the one or more tensors. In one embodiment, the system applies online inferencing using the low-bit representation AI model to generate online data distributions for a feature map, and quantizes a feature map tensor based on the online data distributions.

Optimization methods for quantization of neural network models

Memory layout and conversion are disclosed to improve neural network (NN) inference performance. For one example, a NN selects a memory layout for a neural network (NN) among a plurality of different memory layouts based on thresholds derived from performance simulations of the NN. The NN stores multi-dimensional NN kernel computation data using the selected memory layout during NN inference. The memory layouts to be selected can be a channel, height, width, and batches (CHWN) layout, a batches, height, width and channel (NHWC) layout, and a batches, channel, height and width (NCHW) layout. If the multi-dimensional NN kernel computation data is not in the selected memory layout, the NN transforms the multi-dimensional NN kernel computation data for the selected memory layout.

Memory layouts and conversion to improve neural network inference performance

A method and system is disclosed for protecting neural network models by segmenting partitions of the models into segments of pre-configured memory size, hashing the segmented models, and concatenating the hash segments. The concatenated hash segment may be further hashed, encrypted, and stored with the neural network models as an executable loadable file (ELF) in memories external to the neural network prior to the use of the models by the neural network. The models may include model weights of the inference layers and metadata. The model weights and the metadata may be hashed as separate hash segments and concatenated. Segmenting the models into segments of pre-configured memory size and hashing the segmented models offline prior to the operation of the neural network enables rapid validation of the models when the models are used in the inference layers during online operation of the neural network.

Methods to protect neural network models

Embodiments of the invention disclose an integrated circuit and a method for improving utilization of multiply and accumulate (MAC) units on the integrated circuit in an artificial intelligence (AI) engine. In one embodiment, the integrated circuit can include a scheduler for allocating the MAC units to execute a neural network model deployed on the AI engine to process input data. The scheduler includes status information for the MAC units, and can select one or more idle MAC units based on the status information for use to process the feature map slice. The integrated circuit can dynamically map idle MAC units to an input feature map, thereby improving utilization of the MAC units. A pair of linked list, each with a reference head, can be provided in a static random access memory (SRAM) to store only feature map slices and weights for a layer that is currently being processed. When processing a next layer, the two reference heads can be swapped so that output feature map slices for the current layer can be used as input feature maps for the next layer.

Methods for improving AI engine MAC utilization

The disclosed method may include (1) deploying a virtual-tracking technology across a plurality of customer-facing platforms, (2) detecting, via one customer-facing platform, a first touchpoint between a browser of a contact related to a buyer-side business and a seller-side business, (3) generating, by way of the virtual-tracking technology, a cookie that is linked to the buyer-side business, (4) dropping the cookie into the browser of the contact during the first touchpoint, (5) detecting, via another customer-facing platform, a second touchpoint between the browser of the contact and the seller-side business based at least in part on the cookie, and then (6) tracing a customer journey that (A) represents the interactions between the buyer-side business and the seller-side business via the contact and (B) accounts for the first touchpoint and the second touchpoint. Various other systems, methods, and computer-readable media are also disclosed.

Systems and methods for matching and delivering account-level data from multiple customer-facing platforms based on contact-level touchpoints

A method, apparatus, and system for timing synchronization between multiple computing nodes in an autonomous vehicle host system is disclosed. Timing of a first computing node of an autonomous vehicle host system is calibrated based on an external time source. A first timing message is transmitted from the first computing node to a second computing node of the autonomous vehicle host system via a first communication channel between the first computing node and the second computing node. Timing of the second computing node is calibrated based on the first timing message, wherein immediately subsequent to the calibration of timing of the second computing node, timing of the first computing node and of the second computing node is synchronized.

Time synchronization scheme between different computation nodes in autonomous driving system

In one embodiment, a computer-implemented method of performing a secure boot operation in an autonomous driving vehicle includes reading a first marker from a storage device in which the storage device includes a plurality of partitions and at least the first marker. The plurality of partitions includes a first partition including stored software, the first marker associated with the first partition, and wherein the first marker includes a unique identifier and an authentication code. The method further includes determining if the read first marker associated with the first partition is valid during a boot-up operation and executing the stored software in the first partition if the read first marker is determined valid.

Boot failure recovery scheme for hardware-based system of autonomous driving vehicles

A sensor unit includes a sensor interface, host interface, and pre-processing hardware. The sensor interface is coupled to a plurality of cameras configured to capture images around an autonomous driving vehicle (ADV). The host interface is coupled to a perception and planning system. The pre-processing hardware is coupled to the sensor interface to receive images from the plurality of cameras and to perform one or more pre-processing functions on the images and to transmit pre-processed images to the perception and planning system via the host interface. The perception and planning system is configured to perceive a driving environment surrounding the ADV based on the pre-processed images and to plan a path to control the ADV to navigate through the driving environment. The pre-processing functions can adjust for different calibrations and formats across the plurality of cameras.

Growing community of inventors

Sunnyvale, CA, United States of America

Min Guo