Converting interpolation operation in neural network to depthwise convolution
Converting interpolation operations to depthwise convolutions in neural networks addresses inefficiencies by optimizing resource utilization, resulting in substantial performance improvements for bi-linear and bi-cubic interpolations.
Patent Information
- Application Number
- PCT/CN2024/097714
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-06-06
- Publication Date
- 2025-12-11
AI Technical Summary
Interpolation operations in neural networks, such as bi-linear and bi-cubic interpolation, are bottlenecks due to suboptimal utilization of compute resources, leading to inefficiencies and increased computational costs, especially in resource-constrained devices.
Convert interpolation operations to depthwise convolutions, which are performed on input tensors using DNN accelerators' DPUs, optimizing resource utilization and improving efficiency.
This approach achieves a significant performance uplift, with approximately 10X-fold improvement for bi-linear interpolation and 7X-fold improvement for bi-cubic interpolation, enhancing computational efficiency and reducing power consumption.
Smart Images

Figure CN2024097714_11122025_PF_FP_ABST
Abstract
Description
CONVERTING INTERPOLATION OPERATION IN NEURAL NETWORK TO DEPTHWISE CONVOLUTIONTechnical Field
[0001] This disclosure relates generally to neural networks (also referred to as “deep neural networks” or “DNN” ) , and more specifically, performing interpolation operations in DNNs by converting the interpolation operations to depthwise convolutions.Background
[0002] DNNs are used extensively for a variety of artificial intelligence (AI) applications ranging from computer vision to speech recognition and natural language processing due to their ability to achieve high accuracy. However, the high accuracy comes at the expense of significant computation cost. DNNs have extremely high computing demands as there can be a large number of operations as well as a large amount of data to read and write. Therefore, techniques to improve efficiency of DNNs are needed.Brief Description of the Drawings
[0003] Embodiments will be readily understood by the following detailed description in conjunction with the accompanying drawings. To facilitate this description, like reference numerals designate like structural elements. Embodiments are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings.
[0004] FIG. 1 illustrates an example DNN, in accordance with various embodiments.
[0005] FIG. 2A illustrates an example convolution, in accordance with various embodiments.
[0006] FIG. 2B illustrates an example interpolation operation, in accordance with various embodiments.
[0007] FIG. 3 is a block diagram of a DNN system, in accordance with various embodiments.
[0008] FIG. 4 is a block diagram of a DNN module, in accordance with various embodiments.
[0009] FIG. 5 illustrates an example sparse cell, in accordance with various embodiments.
[0010] FIG. 6 illustrates an example sparse cell array, in accordance with various embodiments.
[0011] FIG. 7 illustrates an example processing element (PE) , in accordance with various embodiments.
[0012] FIG. 8 illustrates an interpolation operation converted to an upsampling operation and a depthwise convolution, in accordance with various embodiments.
[0013] FIGS. 9A and 9B illustrate a depthwise convolution performed for vertical scaling, in accordance with various embodiments.
[0014] FIGS. 10A and 10B illustrate a depthwise convolution performed for horizontal scaling, in accordance with various embodiments.
[0015] FIG. 11A is a flowchart of a method of executing an interpolation operation in a DNN, in accordance with various embodiments.
[0016] FIG. 11B is a flowchart of another method of executing an interpolation operation in a DNN, in accordance with various embodiments.
[0017] FIG. 12 is a block diagram of an example computing device, in accordance with various embodiments.Detailed Description
[0018] Overview
[0019] The last decade has witnessed a rapid rise in AI based data processing, particularly based on DNNs. DNNs are widely used in the domains of computer vision, speech recognition, image, and video processing mainly due to their ability to achieve beyond human-level accuracy. A DNN typically includes a sequence of layers. A DNN layer may include one or more deep learning operations (also referred to as “neural network operations” ) , such as convolution, interpolation, layer normalization, batch normalization, SoftMax operation, pooling, elementwise operation, linear operation, nonlinear operation, and so on.
[0020] Input or output data of neural network operations may be arranged in data structures called tensors. Taking a convolutional layer for example, the input tensors include an activation tensor (also referred to as “input feature map (IFM) ” or “input activation tensor” ) including one or more activations (also referred to as “input elements” ) and a weight tensor. The weight tensor may be a kernel (a 2D weight tensor) , a filter (a 3D weight tensor) , or a group of filters (a 4D weight tensor) . A convolution may be performed on the input activation tensor and weight tensor to compute an output activation tensor in the convolutional layer.
[0021] A tensor is a data structure having multiple elements across one or more dimensions. Examples of tensors include vector (which is one-dimensional (1D) tensor) , matrix (which is two-dimensional (2D) tensor) , three-dimensional (3D) tensors, four-dimensional (4D) tensors, and even higher dimensional tensors. A dimension of a tensor may correspond to an axis, e.g., an axis in a coordinate system. A dimension may be measured by the number of data points along the axis. The dimensions of a tensor may define the shape of the tensor. A DNN layer may receive one or more input tensors and compute an output tensor from the one or more input tensors. In some embodiments, a 3D tensor may have an X-dimension, a Y-dimension, and Z-dimension. The X-dimension of a tensor may be the horizontal dimension, the length of which may be the width of the tensor; the Y-dimension may be the vertical dimension, the length of which may be the height of the tensor; and the Z-dimension may be the channel dimension, the length of which may be the number of channels. The coordinates of the elements along a dimension may be integers in an inclusive range from 0 to (L-1) , where L is the length of the tensor in the dimension. For instance, the x coordinate of the first element in a row may be 0, the x coordinate of the second element in a row may be 1, and so on. Similarly, the y coordinate of the first element in a column may be 0, the y coordinate of the second element in a column may be 1, and so on. A 4D tensor may have a fourth dimension, which may indicate the number of batches in the operation.
[0022] Tensors in DNNs can be saved in X-major (e.g., XYZ or XZY format) , Y-major formats (e.g., YXZ or YZX format) , or Z-major formats (e.g., ZXY or ZYX format) . The format of a tensor may define the order in which the data points in the tensor are stored, written, or read. The first character may represent the dimension in which data points are contiguous in memory. The second character may represent the dimension in which data points can be accessed after the contiguous data points are accessed in memory. The third character may represent the dimension in which data points are accessed after the data points in the dimension represented by the second character are exhausted. Taking the ZXY format for example, the access order first starts in the Z-dimension, then moves to the X-dimension, and finally moves to the Y-dimension. Data points in the tensor are contiguous in memory in the Z-dimension, meaning data points having the same (x, y) coordinates are contiguous in memory. Using tensor permutation, the tensor may be read from memory in a different format.
[0023] The significant improvements in DNN model size and accuracy coupled with the rapid increase in computing power of execution platforms have led to the adoption of DNN applications even within resource constrained mobile and edge devices that have limited energy availability. DNN models may be executed, e.g., for training or inference, by DNN accelerators. A DNN accelerator may be or include one or more data processing units (DPUs) . A DPU may also be referred to as a compute block or compute tile. A DPU may include PEs that can carry out neural network operations.
[0024] DNN accelerators are typically designed to speed up the process of carrying out DNNs. The process of carrying out DNNs is a process of executing DNNs that includes executing the layers or neural network operations in the DNNs. Some DNN accelerators are also referred to as neural processing units (NPUs) , AI processors, or AI processing units. An example DNN accelerator includes a DPU (DPU) providing computation capacity, a memory providing data storage capacity, and a direct memory access (DMA) module allowing pipelined movement of parameters and data between the memory and a memory local to the DPU ( “local memory” ) . The DPU is a compute unit, which may also be referred to as “compute block” or “compute tile. ” Some DNN accelerators may have multiple DPUs, i.e., multiple tiles. A DPU may include an array of PEs, a Data Signal Processor (DSP) , a local memory, other components, or some combination thereof.
[0025] However, many neural network operations include or are associated with pre-processing or post-processing steps that are typically run on CPU (Central Processing Unit) . This can cause additional cost of device switching (e.g., NPU to CPU and vice-versa) as well as memory sharing, which significantly impacts pipeline latency. Interpolation operations, such as bi-cubic or bi-linear interpolation, are extensively utilized in the pre-processing and post-processing stages of deep learning pipelines. These operations may include resize (upsampling or downsampling) along with different interpolation modes and different coordinate axis transformations. In pre-processing, interpolation operations can adjust sizes of IFMs, which include input images, to meet the fixed-size requirements of DNNs, ensuring consistency across datasets and inferences. Bi-cubic interpolation, known for using a higher-order polynomial to achieve smoother transitions between pixels, is usually favored for high-quality image resizing. It can be especially beneficial when images need to retain detail after scaling, making it a preferred choice for tasks that demand high-resolution inputs or outputs, such as in super-resolution tasks.
[0026] Inside the architecture of many DNNs, interpolation operations can play a crucial role in manipulating the spatial dimensions of feature maps. Bi-linear interpolation is commonly employed within network layers for tasks that involve resizing feature maps. This is evident in models dealing with image data, where adjusting the spatial dimensions of feature maps is often necessary for processing through subsequent layers or for integrating feature maps of different scales. Models like DeepLab-V3, which can perform segmentation tasks, leverage bi-linear interpolation to refine spatial details and merge feature maps effectively. Unet and its variants, Super-resolution networks, Generative adversarial networks (GAN) , FPN Feature pyramid networks (FPN) , and vision transformers are some other examples of networks. Furthermore, bi-linear interpolation can be utilized for coordinate transformations in operations such as grid sample and warp. Bi-cubic interpolation is used in sophisticated models requiring high-fidelity image processing. This includes advanced super-resolution DNNs where the quality of the upscaled images is paramount. Bi-cubic interpolation can produce smoother and more visually appealing images. It can be suitable choice for enhancing the output quality in these high-resolution tasks.
[0027] The resize scale factor of interpolation in DNNs, which determines how much a feature map is scaled, is not confined to integer values or powers of 2. This flexibility can be crucial for customizing the network to handle a wide range of input sizes combined with different and for fine-tuning the model's performance. Further, the ability to use non-integer scale factors can be important in DNNs where feature maps may need to be resized at various stages to maintain or alter the spatial resolution, depending on the network's design and objectives.
[0028] However, interpolate operations can be the bottlenecks in end-to-end use-cases because these operations usually cannot fully utilize the compute resources. Currently available methods to implement interpolation operations are using fixed function accelerators or to implement scaling algorithms using CPUs. Fixed function accelerators can result in silicon area cost, and CPUs can provide additional memory copies between devices. In certain cases, bi-linear and bi-cubic interpolations are mapped to the general-purpose DSP on the NPU. Though mapping to the SHAVE (Streaming Hybrid Architecture vector engine) DSP is better than doing the operation on the Host CPU. This mapping is not optimal, in terms of both performance and power.
[0029] Embodiments of the present disclosure may improve on at least some of the challenges and issues described above by mapping interpolation operations, including bi-linear interpolation and bi-cubic interpolation, in DNNs to DPUs as depthwise convolutions.
[0030] In various embodiments of the present disclosure, an interpolation operation may be converted to one or more depthwise convolutions. A depthwise convolution is performed on an input tensor and a kernel to compute an output tensor. Different from standard convolutions, depthwise convolutions are channel-separable operations and do not collapse the channel dimension. The input tensor of the depthwise convolution may be the IFM of the interpolation operation or generated by upsampling the IFM of the interpolation operation. The output tensor of the depthwise convolution may be the output feature map (OFM) of the interpolation operation or an intermediate feature map that is used as the input of a subsequent depthwise convolution to compute the OFM of the interpolation operation. In embodiments where the IFM has multiple channels, the depthwise convolution may be performed on each of the channels so that the channels will be present in the OFM.
[0031] The kernel may be generated from one or more parameters of the interpolation operation. For instance, one or more scale factors of an interpolation operation may be determined. A scale factor may be determined using a dimension of the IFM of the interpolation operation and a dimension of an OFM of the interpolation operation. In an example, a vertical scale factor may be determined using the height of the IFM and the height of the OFM. A horizontal scale factor may be determined using the width of the IFM and the width of the OFM. Scaling coefficients or intermediate parameters may be computed from the scale factor (s) . Further, the kernel may be generated from the scaling coefficients or intermediate parameters. The computation of the scale factors may be conducted offline, e,g., in the compilation stage before the DNN execution is started. The computation of the scaling coefficients, intermediate parameters, or kernel may be performed either in the compilation stage or at runtime.
[0032] During DNN execution, depthwise convolutions converted from interpolation operations may be performed by DPUs in DNN accelerators. A DPU can efficiently perform convolutions. For example, the DPU can accelerate multiply-accumulate (MAC) operations in convolutions based on sparsity in data. As another example, the DPU can load or drain data into the compute elements (e.g., PEs) in a format that leads to optimal utilization of the compute elements. Mapping interpolation operations to DPUs as depthwise convolutions can improve the efficiency in execution of interpolation operations and solve the problem of interpolation operations becoming the bottleneck of efficiency in DNN execution. The DPU block is usually the most powerful and power efficient block in the NPU. Mapping these functions to the DPU allows to get better performance / watt. The performance uplift from the proposed NPU mapping (as compared to CPU at operator level) can be approximately 10X-fold improvement for bi-linear interpolation and a 7X-fold improvement for bi-cubic interpolation in some embodiments.
[0033] For purposes of explanation, specific numbers, materials and configurations are set forth in order to provide a thorough understanding of the illustrative implementations. However, it will be apparent to one skilled in the art that the present disclosure may be practiced without the specific details or / and that the present disclosure may be practiced with only some of the described aspects. In other instances, well known features are omitted or simplified in order not to obscure the illustrative implementations.
[0034] Further, references are made to the accompanying drawings that form a part hereof, and in which is shown, by way of illustration, embodiments that may be practiced. It is to be understood that other embodiments may be utilized, and structural or logical changes may be made without departing from the scope of the present disclosure. Therefore, the following detailed description is not to be taken in a limiting sense.
[0035] Various operations may be described as multiple discrete actions or operations in turn, in a manner that is most helpful in understanding the claimed subject matter. However, the order of description should not be construed as to imply that these operations are necessarily order dependent. In particular, these operations may not be performed in the order of presentation. Operations described may be performed in a different order from the described embodiment. Various additional operations may be performed or described operations may be omitted in additional embodiments.
[0036] For the purposes of the present disclosure, the phrase “A or B” or the phrase "A and / or B" means (A) , (B) , or (A and B) . For the purposes of the present disclosure, the phrase “A, B, or C” or the phrase "A, B, and / or C" means (A) , (B) , (C) , (A and B) , (A and C) , (B and C) , or (A, B, and C) . The term "between, " when used with reference to measurement ranges, is inclusive of the ends of the measurement ranges.
[0037] The description uses the phrases "in an embodiment" or "in embodiments, " which may each refer to one or more of the same or different embodiments. The terms "comprising, " "including, " "having, " and the like, as used with respect to embodiments of the present disclosure, are synonymous. The disclosure may use perspective-based descriptions such as "above, " "below, " "top, " "bottom, " and "side" to explain various features of the drawings, but these terms are simply for ease of discussion, and do not imply a desired or required orientation. The accompanying drawings are not necessarily drawn to scale. Unless otherwise specified, the use of the ordinal adjectives “first, ” “second, ” and “third, ” etc., to describe a common object, merely indicates that different instances of like objects are being referred to and are not intended to imply that the objects so described must be in a given sequence, either temporally, spatially, in ranking or in any other manner.
[0038] In the following detailed description, various aspects of the illustrative implementations will be described using terms commonly employed by those skilled in the art to convey the substance of their work to others skilled in the art.
[0039] The terms “substantially, ” “close, ” “approximately, ” “near, ” and “about, ” generally refer to being within + / -20%of a target value as described herein or as known in the art. Similarly, terms indicating orientation of various elements, e.g., “coplanar, ” “perpendicular, ” “orthogonal, ” “parallel, ” or any other angle between the elements, generally refer to being within + / -5-20%of a target value as described herein or as known in the art.
[0040] In addition, the terms “comprise, ” “comprising, ” “include, ” “including, ” “have, ” “having” or any other variation thereof, are intended to cover a non-exclusive inclusion. For example, a method, process, device, or DNN accelerator that comprises a list of elements is not necessarily limited to only those elements but may include other elements not expressly listed or inherent to such method, process, device, or DNN accelerators. Also, the term “or” refers to an inclusive “or” and not to an exclusive “or. ”
[0041] The systems, methods and devices of this disclosure each have several innovative aspects, no single one of which is solely responsible for all desirable attributes disclosed herein. Details of one or more implementations of the subject matter described in this specification are set forth in the description below and the accompanying drawings.
[0042] Example DNN
[0043] FIG. 1 illustrates an example DNN 100, in accordance with various embodiments. The DNN 100 may be executed by a DNN accelerator, e.g., the DNN accelerator 302 in FIG. 3. In an example, the DNN 100 may be a convolution-based DNN. In other examples, the DNN 100 may be other types of DNNs. For the purpose of illustration, the DNN 100 includes a sequence of layers comprising a plurality of convolutional layers 110 (individually referred to as “convolutional layer 110” ) , a plurality of pooling layers 120 (individually referred to as “pooling layer 120” ) , and a plurality of fully-connected layers 130 (individually referred to as “fully-connected layer 130” ) . In other embodiments, the DNN 100 may include fewer, more, or different layers. In an execution of the DNN 100, the layers of the DNN 100 execute tensor computation that includes many tensor operations, such as convolutions, interpolations, pooling operations, elementwise operations (e.g., elementwise addition, elementwise multiplication, etc. ) , other types of tensor operations, or some combination thereof.
[0044] The convolutional layers 110 summarize the presence of features in inputs to the DNN 100. The convolutional layers 110 function as feature extractors. The first layer of the DNN 100 is a convolutional layer 110. In an example, a convolutional layer 110 performs a convolution on an input tensor 140 (also referred to as IFM 140) and a filter 150. As shown in FIG. 1, the IFM 140 is represented by a 7×7×3 three-dimensional (3D) matrix. The IFM 140 includes 3 input channels, each of which is represented by a 7×7 two-dimensional (2D) matrix. The 7×7 2D matrix includes 7 input elements (also referred to as input points) in each row and 7 input elements in each column. The filter 150 is represented by a 3×3×3 3D matrix. The filter 150 includes 3 kernels, each of which may correspond to a different input channel of the IFM 140. A kernel is a 2D matrix of weights, where the weights are arranged in columns and rows. A kernel can be smaller than the IFM. In the embodiments of FIG. 1, each kernel is represented by a 3×3 2D matrix. The 3×3 kernel includes 3 weights in each row and 3 weights in each column. Weights can be initialized and updated by backpropagation using gradient descent. The magnitudes of the weights can indicate importance of the filter 150 in extracting features from the IFM 140.
[0045] The convolution includes MAC operations with the input elements in the IFM 140 and the weights in the filter 150. The convolution may be a standard convolution 163 or a depthwise convolution 183. In the standard convolution 163, the whole filter 150 slides across the IFM 140. All the input channels are combined to produce an output tensor 160 (also referred to as OFM 160) . The OFM 160 is represented by a 5×5 2D matrix. The 5×5 2D matrix includes 5 output elements (also referred to as output points) in each row and 5 output elements in each column. For the purpose of illustration, the standard convolution includes one filter in the embodiments of FIG. 1. In embodiments where there are multiple filters, the standard convolution may produce multiple output channels in the OFM 160.
[0046] The multiplication applied between a kernel-sized patch of the IFM 140 and a kernel may be a dot product. A dot product is the elementwise multiplication between the kernel-sized patch of the IFM 140 and the corresponding kernel, which is then summed, always resulting in a single value. Because it results in a single value, the operation is often referred to as the “scalar product. ” Using a kernel smaller than the IFM 140 is intentional as it allows the same kernel (set of weights) to be multiplied by the IFM 140 multiple times at different points on the IFM 140. Specifically, the kernel is applied systematically to each overlapping part or kernel-sized patch of the IFM 140, left to right, top to bottom. The result from multiplying the kernel with the IFM 140 one time is a single value. As the kernel is applied multiple times to the IFM 140, the multiplication result is a 2D matrix of output elements. As such, the 2D output matrix (i.e., the OFM 160) from the standard convolution 163 is referred to as an OFM.
[0047] In the depthwise convolution 183, the input channels are not combined. Rather, MAC operations are performed on an individual input channel and an individual kernel and produce an output channel. As shown in FIG. 1, the depthwise convolution 183 produces a depthwise output tensor 180. The depthwise output tensor 180 is represented by a 5×5×3 3D matrix. The depthwise output tensor 180 includes 3 output channels, each of which is represented by a 5×5 2D matrix. The 5×5 2D matrix includes 5 output elements in each row and 5 output elements in each column. Each output channel is a result of MAC operations of an input channel of the IFM 140 and a kernel of the filter 150. For instance, the first output channel (patterned with dots) is a result of MAC operations of the first input channel (patterned with dots) and the first kernel (patterned with dots) , the second output channel (patterned with horizontal strips) is a result of MAC operations of the second input channel (patterned with horizontal strips) and the second kernel (patterned with horizontal strips) , and the third output channel (patterned with diagonal stripes) is a result of MAC operations of the third input channel (patterned with diagonal stripes) and the third kernel (patterned with diagonal stripes) . In such a depthwise convolution, the number of input channels equals the number of output channels, and each output channel corresponds to a different input channel. The input channels and output channels are referred to collectively as depthwise channels. After the depthwise convolution, a pointwise convolution 193 is then performed on the depthwise output tensor 180 and a 1×1×3 tensor 190 to produce the OFM 160.
[0048] The OFM 160 is then passed to the next layer in the sequence. In some embodiments, the OFM 160 is passed through an activation function. An example activation function is rectified linear unit (ReLU) . ReLU is a calculation that returns the value provided as input directly, or the value zero if the input is zero or less. The convolutional layer 110 may receive several images as input and calculate the convolution of each of them with each of the kernels. This process can be repeated several times. For instance, the OFM 160 is passed to the subsequent convolutional layer 110 (i.e., the convolutional layer 110 following the convolutional layer 110 generating the OFM 160 in the sequence) . The subsequent convolutional layers 110 perform a convolution on the OFM 160 with new kernels and generate a new feature map. The new feature map may also be normalized and resized. The new feature map can be kernelled again by a further subsequent convolutional layer 110, and so on.
[0049] In some embodiments, a convolutional layer 110 has four hyperparameters: the number of kernels, the size F kernels (e.g., a kernel is of dimensions F×F×D pixels) , the S step with which the window corresponding to the kernel is dragged on the image (e.g., a step of one means moving the window one pixel at a time) , and the zero-padding P (e.g., adding a black contour of P pixels thickness to the input image of the convolutional layer 110) . The convolutional layers 110 may perform various types of convolutions, such as 2-dimensional convolution, dilated or atrous convolution, spatial separable convolution, depthwise separable convolution, transposed convolution, and so on. The DNN 100 includes 16 convolutional layers 110. In other embodiments, the DNN 100 may include a different number of convolutional layers.
[0050] The pooling layers 120 down-sample feature maps generated by the convolutional layers, e.g., by summarizing the presence of features in the patches of the feature maps. A pooling layer 120 is placed between two convolution layers 110: a preceding convolutional layer 110 (the convolution layer 110 preceding the pooling layer 120 in the sequence of layers) and a subsequent convolutional layer 110 (the convolution layer 110 subsequent to the pooling layer 120 in the sequence of layers) . In some embodiments, a pooling layer 120 is added after a convolutional layer 110, e.g., after an activation function (e.g., ReLU, etc. ) has been applied to the OFM 160.
[0051] A pooling layer 120 receives feature maps generated by the preceding convolution layer 110 and applies a pooling operation to the feature maps. The pooling operation reduces the size of the feature maps while preserving their important characteristics. Accordingly, the pooling operation improves the efficiency of the DNN and avoids over-learning. The pooling layers 120 may perform the pooling operation through average pooling (calculating the average value for each patch on the feature map) , max pooling (calculating the maximum value for each patch of the feature map) , or a combination of both. The size of the pooling operation is smaller than the size of the feature maps. In various embodiments, the pooling operation is 2×2 pixels applied with a stride of two pixels, so that the pooling operation reduces the size of a feature map by a factor of 2, e.g., the number of pixels or values in the feature map is reduced to one quarter the size. In an example, a pooling layer 120 applied to a feature map of 6×6 results in an output pooled feature map of 3×3. The output of the pooling layer 120 is inputted into the subsequent convolution layer 110 for further feature extraction. In some embodiments, the pooling layer 120 operates upon each feature map separately to create a new set of the same number of pooled feature maps.
[0052] The fully-connected layers 130 are the last layers of the DNN. The fully-connected layers 130 may be convolutional or not. The fully-connected layers 130 receive an input operand. The input operand defines the output of the convolutional layers 110 and pooling layers 120 and includes the values of the last feature map generated by the last pooling layer 120 in the sequence. The fully-connected layers 130 apply a linear combination and an activation function to the input operand and generate a vector. The vector may contain as many elements as there are classes: element i represents the probability that the image belongs to class i. Each element is therefore between 0 and 1, and the sum of all is worth one. These probabilities are calculated by the last fully-connected layer 130 by using a logistic function (binary classification) or a SoftMax function (multi-class classification) as an activation function. In some embodiments, the fully-connected layers 130 multiply each input element by weight, make the sum, and then apply an activation function (e.g., logistic if N=2, SoftMax if N>2) . This is equivalent to multiplying the input operand by the matrix containing the weights.
[0053] FIG. 2A illustrates an example convolution, in accordance with various embodiments. The convolution may be a deep learning operation in a convolutional layer of a DNN, e.g., a convolutional layer 110 in FIG. 1. The convolution can be executed on an activation tensor 210 and filters 220 (individually referred to as “filter 220” ) . The filters may constitute a weight tensor of the convolution. The result of the convolution is an output tensor 230. In some embodiments, the convolution is performed by a DNN accelerator. An example of the DNN accelerator may be the DNN accelerator 302 in FIG. 3. For instance, the convolution may be performed by one or more DPUs 330 in the DNN accelerator 302.
[0054] The activation tensor 210 may be computed in a previous layer of the DNN. In some embodiments (e.g., embodiments where the convolutional layer is the first layer of the DNN) , the activation tensor 210 may be an image. In the embodiments of FIG. 2A, the activation tensor 210 includes activations (also referred to as “input activations, ” “elements, ” or “input elements” ) arranged in a 3D matrix. The activation tensor 210 may also be referred to as an input tensor of the convolution. An input element is a data point in the activation tensor 210. The activation tensor 210 has a spatial size Hin×Win×Cin, where Hin is the height of the 3D matrix (i.e., the length along the Y axis, which indicates the number of activations in a column in the 3D matrix of each input channel) , Win is the width of the 3D matrix (i.e., the length along the X axis, which indicates the number of activations in a row in the 2D matrix of each input channel) , and Cin is the depth of the 3D matrix (i.e., the length along the Z axis, which indicates the number of input channels) . For the purpose of simplicity and illustration, the activation tensor 210 has a spatial size of 7×7×3, i.e., the activation tensor 210 includes three input channels and each input channel has a 7×7 2D matrix. Each input element in the activation tensor 210 may be represented by a (X, Y, Z) coordinate. In other embodiments, the height, width, or depth of the activation tensor 210 may be different.
[0055] Each filter 220 includes weights arranged in a 3D matrix. The values of the weights may be determined through training the DNN. A filter 220 has a spatial size Hf×Wf×Cf, where Hf is the height of the filter (i.e., the length along the Y axis, which indicates the number of weights in a column in each kernel) , Wf is the width of the filter (i.e., the length along the X axis, which indicates the number of weights in a row in each kernel) , and Cf is the depth of the filter (i.e., the length along the Z axis, which indicates the number of channels) . In some embodiments, Cf equals Cin. For purpose of simplicity and illustration, each filter 220 in FIG. 2A has a spatial size of 2×3×3, i.e., the filter 220 includes 2 convolutional kernels with a spatial size of 2×3. In other embodiments, the height, width, or depth of the filter 220 may be different. The spatial size of the convolutional kernels is smaller than the spatial size of the 2D matrix of each input channel in the activation tensor 210.
[0056] An activation or weight may take one or more bytes in a memory. The number of bytes for an activation or weight may depend on the data format. For example, when the activation or weight has an INT8 format, the activation takes one byte. When the activation or weight has a FP16 format, the activation or weight takes two bytes. Other data formats may be used for activations or weights.
[0057] In the convolution, each filter 220 slides across the activation tensor 210 and generates a 2D matrix for an output channel in the output tensor 230. In the embodiments of FIG. 2A, the 2D matrix has a spatial size of 5×5. The output tensor 230 includes activations (also referred to as “output activations, ” “elements, ” or “output element” ) arranged in a 3D matrix. An output activation is a data point in the output tensor 230. The output tensor 230 has a spatial size Hout×Wout×Cout, where Hout is the height of the 3D matrix (i.e., the length along the Y axis, which indicates the number of output activations in a column in the 2D matrix of each output channel) , Wout is the width of the 3D matrix (i.e., the length along the X axis, which indicates the number of output activations in a row in the 2D matrix of each output channel) , and Cout is the depth of the 3D matrix (i.e., the length along the Z axis, which indicates the number of output channels) . Cout may equal the number of filters 220 in the convolution. Hout and Wout may depend on the heights and weights of the activation tensor 210 and each filter 220. In an example where the kernel size is 1×1, Hout and Woutmay equal to Hin and Win, respectively.
[0058] As a part of the convolution, MAC operations can be performed on a 2×3×3 subtensor 215 (which is highlighted with a dotted pattern in FIG. 2A) in the activation tensor 210 and each filter 220. The result of the MAC operations on the subtensor 215 and one filter 220 is an output activation. In some embodiments (e.g., embodiments where the convolution is an integral convolution) , an output activation may include 8 bits, e.g., one byte. In other embodiments (e.g., embodiments where the convolution is a floating-point convolution) , an output activation may include more than one byte. For instance, an output element may include two bytes.
[0059] After the MAC operations on the subtensor 215 and all the filters 220 are finished, a vector 235 is produced. The vector 235 is highlighted with a dotted pattern in FIG. 2A. The vector 235 includes a sequence of output activations, which are arranged along the Z axis. The output activations in the vector 235 have the same (x, y) coordinate, but the output activations correspond to different output channels and have different Z coordinates. The dimension of the vector 235 along the Z axis may equal the total number of output channels in the output tensor 230. After the vector 235 is produced, further MAC operations are performed to produce additional vectors till the output tensor 230 is produced. In the embodiments of FIG. 2A, the output tensor 230 is computed in a Z-major format. When the output tensor 230 is computed in the ZXY format, the vector that is adjacent to the vector 235 along the X axis may be computed right after the vector 235. When the output tensor 230 is computed in the ZYX format, the vector that is adjacent to the vector 235 along the Y axis may be computed right after the vector 235. The output tensor 230 may be permuted, e.g., by the drain module 390, and stored in a memory (e.g., the local memory 340) in an X-major format or Y-major format.
[0060] In some embodiments, the MAC operations on a 3×3×3 subtensor (e.g., the subtensor 215) and a filter 220 may be performed by a plurality of MAC units. One or more MAC units may receive an input operand (e.g., an activation operand 217 shown in FIG. 2A) and a weight operand (e.g., the weight operand 227 shown in FIG. 2A) . The activation operand 217 includes a sequence of activations having the same (x, y) coordinate but different z coordinates. The activation operand 217 includes an activation from each of the input channels in the activation tensor 210. The weight operand 227 includes a sequence of weights having the same (x, y) coordinate but different z coordinates. The weight operand 227 includes a weight from each of the channels in the filter 220. Activations in the activation operand 217 and weights in the weight operand 227 may be sequentially fed into a MAC unit. The MAC unit may receive an activation and a weight ( “an activation-weight pair” ) at a time and multiple the activation and the weight. The position of the activation in the activation operand 217 may match the position of the weight in the weight operand 227. The activation and weight may correspond to the same channel.
[0061] Activations or weights may be floating-point numbers. Floating-point numbers may have various data formats, such as FP32, FP16, BF16, and so on. A floating-point number may be a positive or negative number with a decimal point. A floating-point number may be represented by a sequence of bits that includes one or more bits representing the sign of the floating-point number (e.g., positive or negative) , bits representing an exponent of the floating-point number, and bits representing a mantissa of the floating-point number. The mantissa is the part of a floating-point number that represents the significant digits of that number. The mantissa is multiplied by the base raised to the exponent to give the actual value of the floating-point number.
[0062] In some embodiments, the output activations in the output tensor 230 may be further processed based on one or more activation functions before they are written into the memory or inputted into the next layer of the DNN. The processing based on the one or more activation functions may be at least part of the post processing of the convolution. In some embodiments, the post processing may include one or more other computations, such as offset computation, bias computation, and so on. The results of the post processing may be stored in a local memory of the compute block and be used as input to the next DNN layer. In some embodiments, the input activations in the activation tensor 210 may be results of post processing of the previous DNN layer.
[0063] FIG. 2B illustrates an example interpolation operation, in accordance with various embodiments. The interpolation operation uses an IFM 250 to generate an OFM 260. As shown in FIG. 3, the OFM 260 has a larger size than the IFM 250. Even though the IFM 250 and the OFM 260 are 2D tensors, the feature maps may be 3D tensors in other embodiments. For instance, the IFM 250 or the OFM 260 may include multiple channels in a third dimension. The interpolation operation may include estimation of new values within the range of values in the IFM 250. The new values are pixels in the OFM 260. The interpolation operation may be a bi-linear interpolation or a bi-cubic interpolation. In some embodiments, bi-linear interpolation uses four neighbors (e.g., a 2×2 neighborhood) to determine an output pixel. Bi-cubic interpolation uses 16 neighbors (e.g., a 4×4 neighborhood) to determine an output pixel. Weight distribution in bi-cubic interpolation may be different from that in bi-linear interpolation.
[0064] In the interpolation operation, a horizontal scale factor and a vertical scale factor may be determined based on dimensions of the IFM and the OFM. The horizontal scale factor may be denoted as xscale=IFMwidt / OFMwidt, where IFMwidt is the width of the IFM, and OFMwidth is the width of the OFM. The vertical scale factor may be denoted as yscale=IFMheigh / OFMheight, where IFMheigh is the height of the IFM, and OFMheigh is the height of the OFM. For the purpose of simplicity and illustration, the IFM 250 has a spatial size of 2×2, and the OFM 260 has a spatial size of 4×4. The interpolation operation has a scale factor of 2 in both the horizontal direction and the vertical direction. In other embodiments, the IFM 250 or the OFM 260 may have a different size. The scale factor may be different in the horizontal direction or the vertical direction.
[0065] The interpolation operation may determine values of each output pixel in the OFM. The coordinate of an output pixel in the OFM may be denoted as (xout, yout) . The corresponding position of the output pixel in the IFM may be denoted as (xin, yin) , where xin=xout×xscale, and yin=yout×yscale. A group of input pixels in the IFM may be identified based on the corresponding coordinate of the output pixel in the IFM. The value of the output pixel may be determined based on the identified input pixels.
[0066] In some embodiments (e.g., embodiments in which the interpolation operation is a bi-linear interpolation) , four neighboring input pixels may be identified for a single output pixel. The four input pixels may be in a 2×2 grid. The coordinates of the four input pixels may be denoted as (x1, y1) , (x2, y1) , (x1, y2) , and (x2, y2) , respectively. The coordinates of the four input pixels in the IFM may be determined based on the coordinate of the output pixel in the IFM and the dimensions of the IFM. In some embodiments, x1=int (xin) , y1=int (yin) , x2=min (int (xin) +1, IFMwidth-1) , and y2=min (int (yin) +1, IFMheight-1) ,
[0067] where int () denotes a function that converts an input value into an integer number, and min () denotes a function that returns the lowest value of the input values. The four input pixels (e.g., their values) may be retrieved from the IFM based on their positions in the IFM.
[0068] Further, a horizontal coefficient α and a vertical coefficient β may be determined based on the position of the output pixel in the IFM and the position of one of the input pixels in the IFM. In an example, α=xin-x1, and β=yin-y1. The value of the output pixel may be determined based on the horizontal coefficient, the vertical coefficient, and the values of the identified input pixels. In some embodiments, for a single (x, y) coordinate in the OFM, a vector is generated in the interpolation operation. The vector may be a 1D tensor along the channel dimension of the OFM. Multiple output pixels having the same (x, y) coordinate are computed for multiple channels of the OFM. In an example, the OFM may have three channels corresponding to different colors, such as red, green, and blue. The output pixel in each channel may be computed based on the four input pixels in the same channel. In an example, the value of the output pixel in a channel may be denoted as p= (1-α) × (1-β) ×a+α× (1-β) ×b+ (1-α) ×β×c+α×β×d, where a, b, c, and d may be the values of the four input pixels in the channel, respectively.
[0069] In other embodiments (e.g., embodiments in which the interpolation operation is a bi-cubic interpolation) , 16 neighboring input pixels may be identified for a single output pixel. The 16 input pixels may be in a 4×4 grid. The coordinates of the four input pixels may be denoted as (x1, y1) , (x2, y1) , (x1, y2) , and (x2, y2) , respectively. The coordinates of the four input pixels in the IFM may be determined based on the coordinate of the output pixel in the IFM and the dimensions of the IFM. In some embodiments, x1=max (int (xin) -1, 0) , y1=max (int (yin) -1, 0) , x2=min (int (xin) +2, IFMwidth-1) , and y2=min (int (yin) +2, IFMheight-1)
[0070] Further, fractional parts for interpolation may be computed. The fractional parts may be denoted as dx=xin-int (xin) , and dy=yin-int (yin) .
[0071] An interpolated pixel values for each channel may be initiated. In some embodiments, the initiated interpolated pixel value is zero for each channel. A 4×4 intermediate grid may be determined. Each value in the intermediate grid ( “intermediate value” ) may equal the value of the input pixel with a coordinate of (x1+i, y1+j) , where i indicates a horizontal positional index of the intermediate value in the intermediate grid, and j indicates a vertical positional index of the intermediate value in the intermediate grid. i may be an integer in a range from 0 to 3, which includes 0 and 3. j may also be an integer in a range from 0 to 3, which includes 0 and 3. In some embodiments, the intermediate values may be denoted as value [0] -value
[0015] .
[0072] Cubic interpolation along rows may be performed. For the first row of the intermediate grid, the cubic interpolation result may be denoted as row [0] =value [0] ×dx3+value [1] ×dx2+value [2] ×dx+value [3] . For the second row of the intermediate grid, the cubic interpolation result may be denoted as row [1] =value [4] ×dx3+value [5] ×dx2+value [6] ×dx+value [7] . For the third row of the intermediate grid, the cubic interpolation result may be denoted as row [2] =value [8] ×dx3+value [9] ×dx2+value
[0010] ×dx+value
[0011] . For the fourth row of the intermediate grid, the cubic interpolation result may be denoted as row [3] =value
[0012] ×dx3+value
[0013] ×dx2+value
[0014] ×dx+value
[0015] .
[0073] Also, cubic interpolation along columns may be performed to compute an interpolated pixel value, which may be denoted as pinterpolated=row [0] ×dy3+row [1] ×dy2+row [2] ×dy+row [3] . An integer function may be applied on the interpolated pixel value to compute the value of the output pixel, which may be denoted as p=int (pinterpolated) . In some embodiments, for a single (x, y) coordinate in the OFM, a vector is generated in the bi-cubic interpolation operation. The vector may be a 1D tensor along the channel dimension of the OFM. Multiple output pixels having the same (x, y) coordinate are computed for multiple channels of the OFM. The output pixel in each channel may be computed based on 16 neighboring input pixels in the same channel.
[0074] Example DNN System
[0075] FIG. 3 is a block diagram of a DNN system 300, in accordance with various embodiments. The whole DNN system 300 or a part of the DNN system 300 may be implemented in one or more computing devices, such as the computing device 2000 in FIG. 12. The DNN system 300 can generate and execute DNNs, such as DNNs including interpolation operations. As shown in FIG. 3, the DNN system 300 includes a DNN module 301 and a DNN accelerator 302. In other embodiments, alternative configurations, different or additional components may be included in the DNN system 300. For instance, the DNN system 300 may include multiple DNN modules or multiple DNN accelerators. Further, functionality attributed to a component of the DNN system 300 may be accomplished by a different component included in the DNN system 300 or a different system. In some embodiments, the DNN module 301 and DNN accelerator 302 may include different types of processing units. In an example, the DNN module 301 may be implemented by one or more CPUs. The DNN accelerator 302 may also be referred to as an AI accelerator or an AI processor. The DNN module 301 and DNN accelerator 302 may be implemented in the same chip or separate chips.
[0076] The DNN module 301 facilitates generation and deployment of DNNs. In some embodiments, the DNN module 301 may generate and train DNNs. For instance, the DNN module 301 can define the layered architecture of a DNN. The DNN module 301 can also determine the internal parameters of the DNN through a DNN training process. The DNN module 301 may also determine one or more hyperparameters that define how the DNN is trained. An example hyperparameter is a sparsity ratio that defines the sparsity level of one or more deep learning tensors for the DNN. The DNN module 301 may also compress DNNs, e.g., during or after training. In some embodiments, the DNN module 301 may prune weights in one or more layers of a DNN by changing nonzero valued weight to zeros. The DNN module 301 may prune weights based on a target weight sparsity ratio. A weight sparsity ratio may be the ratio of the number of zero-valued weights to the total number of weights. In an example where the DNN module 301 prunes weight during DNN training, the DNN module 301 may prune weight of a layer to achieve a target sparsity ratio after one or more epochs. The DNN module 301 may prevent the pruned weights from changing values during the rest of the training process. Alternatively, the DNN module 301 may allow the pruned weights to change values so that a pruned, zero-valued weight may have a nonzero value after further training. The DNN module 301 may prune weights of the layer again after one or more additional epochs.
[0077] The DNN module 301 may deploy trained, compressed, or validated DNNs for use in neural network applications. In some embodiments, the DNN module 301 may distribute trained, compressed, or validated DNNs to devices or systems which may use the DNNs to perform tasks (e.g., image classification, motion planning, etc. ) for which the DNNs were trained. In other embodiments, the DNN module 301 may facilitate deployment of the DNNs using the DNN accelerator 302. For instance, the DNN module 301 may receive data from a device or system coupled with the DNN system 300 and input the received data (or data generated by the DNN module 301, e.g., based on the received data) into a DNN. The DNN module 301 may generate instructions (e.g., computer program instructions) that can be executed by the DNN accelerator 302 for DNN execution. The DNN module 301 may receive an output of the DNN from the DNN accelerator 302. The DNN module 301 may transmit the output of the DNN (or a result of processing the output of the DNN by the DNN module 301) to the device or system. In some embodiments, the DNN module 301 may control execution processes of trained, compressed, or validated DNNs. The DNN module 301 may function as a complier for DNNs executed by the DNN accelerator 302. The DNN module 301 may perform compilation of DNNs and generate compilation descriptors, based on which the DNNs may be executed.
[0078] The DNN module 301 facilitates converting interpolation operations to depthwise convolutions. The DNN module 301 may determine scale factors of interpolation operations. In some embodiments, the DNN module 301 may determine scale factors offline as part of the compilation stage of DNNs. The DNN module 301 may further generate kernels based on scale factors or generate input tensor of depthwise convolutions. The generation of kernel or input tensor may be wholly or partially conducted offline or during runtime. Also, the DNN module 301, instead of generating kernel or input tensor by itself, may generate instructions that can be executed by one or more components of the DNN accelerator 302 (e.g., the DMA engine 320) for generating kernel or input tensor. Certain aspects of the DNN module 301 are provided below in conjunction with FIG. 4.
[0079] The DNN accelerator 302 executes DNNs provided by the DNN module 301. For instance, the DNN accelerator 302 can execute a DNN by running neural network operations in the DNN. The process of carrying out a neural network operation is also referred to as a process of executing the neural network operation or performing the neural network operation. The execution of the DNN may be for training the DNN or for using the DNN to perform AI tasks. As shown in FIG. 3, the DNN accelerator 302 includes a memory 310, a DMA engine 320, and DPUs 330 (individually referred to as “DPU 330” ) . In other embodiments, alternative configurations, different or additional components may be included in the DNN accelerator 302. For example, the DNN accelerator 302 may include more than one memory 310 or DMA engine 320. As another example, the DNN accelerator 302 may include a single DPU 330. Further, functionality attributed to a component of the DNN accelerator 302 may be accomplished by a different component included in the DNN accelerator 302 or by a different system. A component of the DNN accelerator 302 may be implemented in hardware, software, firmware, or some combination thereof.
[0080] The memory 310 stores data associated with neural network operations performed by the DNN accelerator 302. In some embodiments, the memory 310 may store data to be used by the DPUs 330 for executing neural network operations. The memory 310 may store IFMs. The memory 310 may also store weights, such as weights in kernels of convolutions, which are determined by training DNNs or determined based on scale factors of interpolation operations. The memory 310 may further store outputs of neural network operations, such as OFMs. In some embodiments, the memory 310 includes one or more dynamic random-access memories (DRAMs) .
[0081] The DMA engine 320 facilitates data transfer between the memory 310 and local memories of the DPUs 330. For example, the DMA engine 320 can read data from the memory 310 and write data into a local memory of a DPU 330. As another example, the DMA engine 320 can read data from a local memory of a DPU 330 and write data into the memory 310. For instance, the DMA engine 320 may read IFMs of interpolation operations from the memory 310 and load the IFMs to one or more DPUs 330. The DMA engine 320 may also write OFMs of interpolation operations computed by one or more DPUs 330 to the memory 310. The DMA engine 320 provides a DMA feature that allows the DPU 330 to initiate data transfer between the memory 310 and the local memories of the DPUs 330 and to perform other operations while the data transfer is being conducted. In some embodiments, the DMA engine 320 may read tensors from the memory 310, modify the tensors in a way that is optimized for the DPU 330 before it writes the tensors into the local memories of the DPUs 330. For instance, the DMA engine 320 may reshape permute tensors for generating kernels of depthwise convolution converted from interpolation operations.
[0082] The DPUs 330 perform neural network operations in DNNs. For instance, a DPU 330 may execute a DNN layer by running one or more deep learning operations in the DNN layer. A DPU 330 may execute a layer, or a portion of a layer, at a time. In some embodiments, the operations of the DNN layers may be run by multiple DPUs 330 in parallel. For instance, multiple DPUs 330 may each perform a portion of a workload for a neural network operation. Data may be shared between the DPUs 330. A DPU 330 may also be referred to as a neural processing unit, a compute block, or a compute tile.
[0083] The DPUs 330 may be capable of running various types of neural network operations, such as convolution (including depthwise convolutions) , layer normalization, SoftMax operation, pooling, elementwise operation, linear operation, nonlinear operation, and so on. N=Neural network operations performed by the DPUs 330 include tensor operations, i.e., operations whose inputs are tensors or operations whose outputs are tensors. In an example, the DPU 330 receives an input tensor and one or more convolutional kernels and performs a convolution with the input tensor and convolutional kernels. The result of the convolution may be an output tensor, which can be further computed, e.g., by the DPU 330 or another DPU 330.
[0084] In the embodiments of FIG. 3, each DPU 330 includes a local memory 340, a sparsity mode module 350, a load module 360, a processing engine 370, a post-processing engine 380, and a drain module 390. Some or all the components of the DPU 330 can be implemented on the same chip. In other embodiments, alternative configurations, different or additional components may be included in the DPU 330. Further, functionality attributed to a component of the DPU 330 may be accomplished by a different component included in the DPU 330, a different DPU 330, another component of the DNN accelerator 302, or a different system. A component of the DPU 330 may be implemented in hardware, software, firmware, or some combination thereof.
[0085] The local memory 340 is local to the corresponding DPU 330. In the embodiments of FIG. 3, the local memory 340 is inside the DPU 330. In other embodiments, the local memory 340 may be outside the DPU 330. Data in the local memory 340 may be transferred to or from the memory 310, e.g., through the DMA engine 320. In some embodiments, data in the local memory 340 may be transferred to or from the local memory of another DPU 330. The local memory 340 may store data received, used, or generated by the sparsity mode module 350, the load module 360, the processing engine 370, the post-processing engine 380, or the drain module 390. Examples of the data may include input activations, weights, output activations, sparsity bitmaps, and so on.
[0086] In some embodiments, the local memory 340 may store tensors to be processed by the processing engine 370 or the post-processing engine 380. The tensors may be input tensors of deep learning operations. The local memory 340 may also store tensors generated by the processing engine 370 or the post-processing engine 380. The tensors may be output tensors of deep learning operations. The layout of data points of a tensor in the local memory 340 may depend on the format in which the tensor is stored. In some embodiments, the local memory 340 may store tensors in various formats, including Z-major (e.g., ZXY or ZYX) format, X-major (e.g., XYZ or XZY) format, and Y-major (e.g., YXZ or YZX) format. For a tensor with Z-major format, the local memory 340 may store data points having the same (x, y) coordinate contiguously. For instance, the data points having the same (x, y) coordinate may be stored at a sequence of memory addresses in the local memory 340. For a tensor with the ZXY format or ZYX format, the local memory 340 may store data points having the same (x, y) coordinate contiguously. For instance, the data points having the same (x, y) coordinate may be stored at a sequence of memory addresses in the local memory 340. For a tensor with X-major format, the local memory 340 may store data points having the same (y, z) coordinate contiguously. For a tensor with Y-major format, the local memory 340 may store data points having the same (x, z) coordinate contiguously.
[0087] In some embodiments, the local memory 340 may store dense tensors (e.g., dense activation tensors, dense weight tensors, etc. ) , sparse tensors (e.g., sparse activation tensors, sparse weight tensors, etc. ) , and so on. A dense tensor may be a tensor from which zero-valued elements (if any) are not removed. A dense tensor may be converted to a sparse tensor by removing one or more zero-valued elements in the dense tensor. A sparse tensor may also be referred to as a compressed tensor or packed tensor. The process of converting a dense tensor to a sparse tensor may be referred to as sparsity encoding. Sparsity encoding may also generate a sparsity tensor. Each element in the sparsity tensor may correspond to a different element in the dense tensor and indicate whether the element in the dense tensor is zero or not. The sparsity tensor may indicate positions of elements of the sparse tensor in the dense tensor. The sparsity tensor may be a sparsity bitmap, each element of which is a bit. A sparse tensor may be converted to a dense tensor through a densifying process, in which one or more zeros may be added to the sparse tensor based on the sparsity tensor.
[0088] In some embodiments, the local memory 340 includes one or more SRAMs. The local memory 340 may be byte-addressable, and each memory address identifies a single byte (eight bits) of storage. In some embodiments, the local memory 340 may include memory banks. The number of data banks in the local memory 340 may be 16, 64, 128, 356, 512, 1024, 2048, or other numbers. A memory bank may include a plurality of storage units. In an example, a data bank may include 8, 16, 64, or a different number of storage units. A memory bank or a storage unit in a memory bank may have a memory address. In an example, a storage unit may store a single byte, and data larger than a single byte may be stored in storage units with consecutive memory addresses, i.e., adjacent storage units. For instance, a storage unit can store an integer number in the INT8 format, versus two storage units may be needed to store a number in the FP16 or BF16 format, which has 16 bits. In some embodiments, 16 bits can be transferred from the local memory 340 in a single read cycle. In other embodiments, 16 bits can be transferred from the local memory 340 in multiple read cycles, such as two cycles.
[0089] The sparsity mode module 350 determines sparsity modes in which the DPU 330 operates to execute DNN layers. For instance, the sparsity mode module 350 may determine whether to accelerate a layer based on weight sparsity, activation sparsity, or both. The sparsity mode module 350 select the sparsity mode for a layer from a group of sparsity modes that includes, for example, combined sparsity mode in which the layer is accelerated based on both weight sparsity and activation sparsity, activation sparsity mode in which the layer is accelerated based on activation sparsity but not based on weight sparsity, weight sparsity mode in which the layer is accelerated based on weight sparsity but not based on activation sparsity, and a dense mode in which the layer is not accelerated based on sparsity. In some embodiments (e.g., embodiments where a layer is executed by multiple DPUs 330) , the sparsity mode module 350 may determine the sparsity mode for all the DPUs 330 that executes the layer. In some embodiments, the sparsity mode module 350 may receive configuration parameters from the DNN module 301. A configuration parameter may correspond to a layer and indicate whether to accelerate the layer based on weight sparsity. The sparsity mode module 350 may determine the sparsity mode of the layer based on the configuration parameter.
[0090] The load module 360 loads data from the local memory 340 to the processing engine 370 or to the post-processing engine 380. The load module 360 may read tensors from the local memory 340. The tensors may include sparse activation tensors, sparse weight tensors, activation sparsity tensors, weight sparsity tensors, and so on. In some embodiments, the load module 360 may load data based on the sparsity mode determined by the sparsity mode module 350. The load module 360 may select different data to transmit to the processing engine 370 in different sparsity modes. For instance, the load module 360 may transmit an activation sparsity tensor and a weight sparsity tensor of a layer to the processing engine 370 in the combined sparsity mode, while transmit the activation sparsity tensor but not the weight sparsity tensor to the processing engine 370 in the activation sparsity mode and transmit the weight sparsity tensor but not the activation sparsity tensor to the processing engine 370 in the weight sparsity mode. In the dense mode, the load module 360 does not transmit either the activation sparsity tensor or the weight sparsity tensor to the processing engine 370.
[0091] In some embodiments, the load module 360 may process (e.g., densify) data stored in the local memory 340 before providing the data to the processing engine 370. In an example, the load module 360, while operating in the weight sparsity mode, may densify sparse activation tensors to generate dense activation tensors based on corresponding activation sparsity tensors. For instance, the load module 360 may add one or more zeros into a sparse activation tensor based on an activation sparsity tensor associated with the sparse activation tensor to generate the dense activation tensor. The dense activation tensor includes one or more elements than the sparse activation tensor. The additional element (s) are zero-valued. The load module 360 may identify one or more elements in the activation sparsity tensor that correspond to the zero-valued element (s) , determine the position of each of the zero-valued element (s) in the dense activation tensor, and insert the zero-valued element (s) into the sparse activation tensor based on the determined positions. After the densification, the load module 360 may transmit the dense activation tensors to the processing engine 370. The load module 360 may also transmit corresponding sparse weight tensors and weight sparsity tensors to the processing engine 370. Activation sparsity tensor of the dense activation tensors may not be loaded to the processing engine 370.
[0092] In another example, the load module 360, while operating in the activation sparsity mode, may densify sparse weight tensors to generate dense weight tensors based on corresponding weight sparsity tensors by inserting zeros into sparse weight tensors. The densification of sparse weight tensors may be similar to the densification of sparse activation tensors described above. After the densification, the load module 360 may transmit the dense weight tensors to the processing engine 370. The load module 360 may also transmit corresponding sparse activation tensors and activation sparsity tensors to the processing engine 370. Weight sparsity tensor of the dense weight tensors may not be loaded to the processing engine 370.
[0093] In yet another example, the load module 360, while operating in the dense mode, may densify both sparse weight tensors and sparse activation tensors. The load module 360 may generate the input tensor and weight tensor of the layer and transmit the tensors to the processing engine 370 for executing the layer without sparsity acceleration.
[0094] The processing engine 370 performs operations in DNNs. The processing engine 370 may accelerate neural network operations based on sparsity in data. In some embodiments, the processing engine 370 may operate in a dense mode in which sparsity acceleration is not performed. The processing engine 370 may include one or more processing cells. In some embodiments, the processing cells may be arranged in one or more rows and one or more columns in the processing engine 370. Each processing cell may include PEs that may be arranged in an array that includes rows and columns. All the PEs in the processing engine 370 may constitute a bigger array that includes more rows and columns.
[0095] An example PE may be or may include one or more MAC units that can perform MAC operations. In some embodiments (e.g., embodiments where the DPU 330 executes a convolutional layer) , a computation in an MAC unit may be an MAC operation on an activation operand and a weight operand. The activation operand may be an activation tensor that may include one or more activations in the input tensor of the convolution. Different activations may be in different input channels. The weight operand may be a weight tensor that may include one or more weights in the filter of the convolution. The values of the weights are determined through training the DNN. The weights in the weight operand may be in different input channels.
[0096] In some embodiments, an MAC unit includes one or more multipliers for performing multiplications. An MAC unit may also include one or more accumulators ( “adders” ) for performing accumulations. A column of MAC units is referred to as an MAC column. An MAC column may be associated with one or more MAC lanes. A MAC lane is a path for loading data e.g., by the load module 360, into an MAC column. A MAC lane may be also referred to as a data transmission lane or data loading lane. An MAC column may have multiple MAC lanes. The loading bandwidth of the MAC column is an aggregation of the loading bandwidths of all the MAC lanes associated with the MAC column. With a certain number of MAC lanes, data can be fed into the same number of independent MAC units simultaneously. In some embodiments where an MAC column has four MAC lanes for feeding activations or weights into the MAC column and each MAC lane may have a bandwidth of 16 bytes, the four MAC lanes can have a total loading bandwidth of 64 bytes.
[0097] In some embodiments, the processing engine 370 may be capable of depthwise convolution, standard convolution, or both. In a depthwise convolution, an MAC unit may perform an MAC operation that includes a sequence of multiplications for an input operand and a weight operand. Each multiplication in the sequence (also referred to as a cycle) is a multiplication of a different activation in the input operand with a different weight in the weight operand. The activation and weight in the same cycle may correspond to the same channel. The sequence of multiplication produces a product operand that includes a sequence of products. The MAC operation may also include accumulations in which multiple product operands are accumulated to produce an output operand of the MAC unit. The processing engine 370 may output multiple output operands at a time, each of which is generated by a different MAC unit. In a standard convolution, MAC operations may include accumulations across the channels. For instance, as opposed to generating an output operand, a MAC unit may accumulate products across different channels to generate a single output point.
[0098] In some embodiments, the processing engine 370 may perform MAC operations in quantized deep learning operations, such as MAC operations in a quantized convolution. In some embodiments, an MAC unit in the processing engine 370 may receive quantized activation and quantized weights and compute a quantized MAC result. The quantized MAC result may be a quantized value in an integer format and may be the output of the MAC unit. In some embodiments, the MAC unit may also include a quantization multiplier that can multiply a quantization scale with the quantized MAC result, and the output of the MAC unit may be a real value in a floating-point format. The MAC unit may include no quantization subtractors as zero-point offsetting is not needed for the MAC operations in quantized deep learning operations.
[0099] In some embodiments, the processing engine 370 may include sparsity acceleration logic for facilitating sparsity acceleration. For instance, each processing cell in the processing engine 370 may include one or more sparsity modules. In an example, each MAC column or each MAC row may have a corresponding sparsity module that accelerates MAC operations in the MAC column or MAC row. In some embodiments, a sparsity module accelerates computations in the processing engine 370 based on sparsity in activations, sparsity in weights, or both. The sparsity module may include a storage unit that stores a sparsity tensor, which may be loaded to the storage unit by the load module 360. The sparsity tensor may be an activation sparsity tensor, a weight sparsity tensor, or a combined sparsity tensor.
[0100] An activation sparsity tensor may be the sparsity tensor of an activation tensor and has the same number of elements as the activation tensor. An element in the activation sparsity tensor may indicate whether the corresponding element in the activation tensor is zero or not. For instance, a zero-valued in the activation sparsity tensor may indicate that the corresponding element in the activation tensor is zero. A one-valued in the activation sparsity tensor may indicate that the corresponding element in the activation tensor is nonzero. A weight sparsity tensor may be the sparsity tensor of a weight tensor and has the same number of elements as the weight tensor. An element in the weight sparsity tensor may indicate whether the corresponding element in the weight tensor is zero or not. For instance, a zero-valued in the weight sparsity tensor may indicate that the corresponding element in the weight tensor is zero. A one-valued in the weight sparsity tensor may indicate that the corresponding element in the weight tensor is nonzero. The sparsity module may generate a combined sparsity tensor using an activation sparsity tensor and a weight sparsity tensor. For instance, the sparsity module may multiply an element of the activation sparsity tensor with a corresponding element of the weight sparsity tensor to compute an element of the combined sparsity tensor. The positions of the three elements in their corresponding sparsity tensors may match. In some embodiments, each element in a sparsity tensor may be a bit, and the sparsity tensor may be referred to as a sparsity bitmap.
[0101] The sparsity module may use the sparsity tensor to identify activations and weights to be used in MAC operations by the MAC units. In an embodiment where the processing engine 370 operates in the combined sparsity mode, the sparsity module may identify activations and weights that correspond to nonzero valued elements of a combined sparsity tensor. In an embodiment where the processing engine 370 operates in the activation sparsity mode, the sparsity module may identify activations and weights that correspond to nonzero valued elements of an activation sparsity tensor. In an embodiment where the processing engine 370 operates in the weight sparsity mode, the sparsity module may identify activations and weights that correspond to nonzero valued elements of a weight sparsity tensor. The sparsity module may be bypassed in the dense mode as no sparsity acceleration would be conducted.
[0102] The post-processing engine 380 processes outputs of the processing engine 370. The post-processing engine 380 may include one or more post-processing elements. In some embodiments, the post-processing elements in the post-processing engine 380 may be arranged in an arrange that has rows and columns. In some embodiments, the post-processing engine 380 computes activation functions. The post-processing engine 380 may receive outputs of the processing engine 370 as inputs to the activation functions. In addition or alternative to activation functions, the post-processing engine 380 may perform other types of post processing on outputs of the processing engine 370. For instance, the post-processing engine 380 may apply a bias on an output of the processing engine 370. In some embodiments, the post-processing engine 380 may be bypassed for certain neural network operations.
[0103] The drain module 390 drains data from the processing engine 370 or from the post-processing engine 380. The drain module may write the data to the local memory 340. The drained data may be tensors, such as output tensors of neural network operations. In some embodiments, the drain module 390 may perform tensor permutation to change storage formats of tensors. For instance, the drain module 390 may permute tensors before writing the tensors to the local memory 340 so that the drain module 390 may write the tensors to the local memory 340 in the new formats. In some embodiments, the drain module 390 may perform tensor permutation to change Z-major formats to X-major formats or Y-major formats. For instance, a tensor drained by the drain module 390 from the processing engine 370 or from the post-processing engine 380 may be in a Z-major format. The drain module 390 may change the Z-major format to a X-major or Y-major format and write the tensor to the local memory 340 in the X-major or Y-major format.
[0104] In some embodiments, the drain module 390 may drain data on a cell level. For each processing cell, the drain module 390 may drain outputs of PEs in the processing cell based on a row index or column index of each PE. For instance, the drain module 390 may use a sequence of cycles to drain data from a processing cell. The drain module 390 may drain the output of some of the PE s in each cycle. The sequence of the cycles may be configured based on a configuration parameter indicating the operation mode of the load module 360.
[0105] In some embodiments, the drain module 390 includes sparsity encoding logic that can convert outputs of the processing engine 370 from a dense format to a sparse format. For instance, the drain module 390 may be implemented with one or more sparsity encoders. A sparsity encoder converts dense data to compressed data based on sparsity in the dense data. For instance, the sparsity encoder may remove zeros in an activation tensor computed by the processing engine 370 to convert the activation tensor to a compressed activation tensor. The sparsity encoder may also generate sparsity tensors, including activation sparsity tensors.
[0106] In some embodiments, the data drained from the processing engine 370 may be at least part of an output tensor (e.g., the output tensor 230 in FIG. 2A) of a deep learning operation. The sparsity encoder may generate a compressed version of the output tensor. The sparsity encoder may identify every zero-valued activation in the output tensor and remove these activations from the output tensor to generate a compressed activation tensor (aka “sparse activation tensor” ) . The sparsity encoder may also generate one or more sparsity tensors for the output tensor. A sparsity tensor may correspond to a portion of the output tensor (e.g., the vector 235 in FIG. 2A) . The sparsity tensor may include sparsity elements (e.g., bits) , each of which corresponds to a different activation in the vector and indicates whether the corresponding activation is zeroed or not.
[0107] The drain module 390 may write the compressed activation tensor and the one or more sparsity tensors into the local memory 340. The sparse activation tensor and the one or more sparsity tensors may be further loaded to the memory 310, e.g., through the DMA engine 320. Additionally or alternatively, the sparse activation tensor and the one or more sparsity tensors may be loaded by the load module 360 to the processing engine 370 for further computation, e.g., for performing a deep learning operation in the next layer. Certain aspects of the drain module 390 are described below in conjunction with FIG. 8.
[0108] FIG. 4 is a block diagram of a DNN module 400, in accordance with various embodiments. The DNN module 400 may be an embodiment of the DNN module 301 in FIG. 3. As shown in FIG. 4, the DNN module 400 includes an interface module 410, a training module 420, a compressing module 430, a compiler 440, an interpolation mapping module 450, and a datastore 460. In other embodiments, alternative configurations, different or additional components may be included in the DNN module 400. Further, functionality attributed to a component of the DNN module 400 may be accomplished by a different component included in the DNN module 400 or a different module or system. For instance, certain functionality attributed to the compiler 440 may be accomplished by the interpolation mapping module 450. Additionally or alternatively, certain functionality attributed to the interpolation mapping module 450 may be accomplished by the compiler 440.
[0109] The interface module 410 facilitates communications of the DNN module 400 with other modules or systems. For example, the interface module 410 establishes communications between the DNN module 400 with an external database to receive data that can be used to train DNNs or input into DNNs to perform tasks. As another example, the interface module 410 may distribute trained DNNs to other systems, e.g., computing devices configured to apply DNNs to perform tasks.
[0110] The training module 420 trains DNNs by using a training dataset. The training module 420 forms the training dataset. In an example where the training module 420 trains an DNN to recognize objects in images, the training dataset includes training images and training labels. The training labels describe ground-truth classifications of objects in the training images. In some embodiments, each label in the training dataset corresponds to an object in a training image. In some embodiments, a part of the training dataset may be used to initially train the DNN, and the rest of the training dataset may be held back as a validation subset used by the training module 420 to validate performance of a trained DNN. The data portion of the training dataset not including the tuning subset and the validation subset may be used to train the DNN.
[0111] The training module 420 also determines hyperparameters for training the DNN. Hyperparameters are variables specifying the DNN training process. Hyperparameters are different from parameters inside the DNN (e.g., weights of filters) . In some embodiments, hyperparameters include variables determining the architecture of the DNN, such as number of hidden layers, etc. Hyperparameters also include variables which determine how the DNN is trained, such as batch size, number of epochs, etc. A batch size defines the number of training samples to work through before updating the parameters of the DNN. The batch size is the same as or smaller than the number of samples in the training dataset. The training dataset can be divided into one or more batches. The number of epochs defines how many times the entire training dataset is passed forward and backwards through the entire network. The number of epochs defines the number of times that the deep learning algorithm works through the entire training dataset. One epoch means that each training sample in the training dataset has had an opportunity to update the parameters inside the DNN. An epoch may include one or more batches. The number of epochs may be 1, 5, 10, 50, 100, 500, 1000, or even larger.
[0112] The training module 420 defines the architecture of the DNN, e.g., based on some of the hyperparameters. The architecture of the DNN includes an input layer, an output layer, and a plurality of hidden layers. The input layer of an DNN may include tensors (e.g., a multidimensional array) specifying attributes of the input image, such as the height of the input image, the width of the input image, and the depth of the input image (e.g., the number of bits specifying the color of a pixel in the input image) . The output layer includes labels of objects in the input layer. The hidden layers are layers between the input layer and output layer. The hidden layers include one or more convolutional layers and one or more other types of layers, such as pooling layers, fully-connected layers, normalization layers, SoftMax or logistic layers, and so on. The convolutional layers of the DNN abstract the input image to a feature map that is represented by a tensor specifying the feature map height, the feature map width, and the feature map channels (e.g., red, green, blue images include 3 channels) . A pooling layer is used to reduce the spatial volume of input image after convolution. It is used between two convolution layers. A fully-connected layer involves weights, biases, and neurons. It connects neurons in one layer to neurons in another layer. It is used to classify images between different categories by training.
[0113] In the process of defining the architecture of the DNN, the training module 420 also adds an activation function to a hidden layer or the output layer. An activation function of a layer transforms the weighted sum of the input of the layer to an output of the layer. The activation function may be, for example, a ReLU activation function, a tangent activation function, or other types of activation functions.
[0114] After the training module 420 defines the architecture of the DNN, the training module 420 inputs a training dataset into the DNN. The training dataset includes a plurality of training samples. An example of a training sample includes an object in an image and a ground-truth label of the object. The training module 420 modifies the parameters inside the DNN ( “internal parameters of the DNN” ) to minimize the error between labels of the training objects that are generated by the DNN and the ground-truth labels of the objects. The internal parameters include weights of filters in the convolutional layers of the DNN. In some embodiments, the training module 420 uses a cost function to minimize the error.
[0115] The training module 420 may train the DNN for a predetermined number of epochs. The number of epochs is a hyperparameter that defines the number of times that the deep learning algorithm will work through the entire training dataset. One epoch means that each sample in the training dataset has had an opportunity to update internal parameters of the DNN. After the training module 420 finishes the predetermined number of epochs, the training module 420 may stop updating the parameters in the DNN. The DNN having the updated parameters is referred to as a trained DNN.
[0116] The training module 420 may also verify accuracy of DNNs after training. In some embodiments, the training module 420 inputs samples in a validation dataset into a trained DNN and uses the outputs of the DNN to determine the model accuracy. In some embodiments, a validation dataset may be formed of some or all the samples in the training dataset. Additionally or alternatively, the validation dataset includes additional samples, other than those in the training sets. In some embodiments, the training module 420 may determine an accuracy score measuring the precision, recall, or a combination of precision and recall of the DNN. The training module 420 may use the following metrics to determine the accuracy score: Precision = TP / (TP + FP) and Recall = TP / (TP + FN) , where precision may be how many the DNN correctly predicted (TP or true positives) out of the total it predicted (TP + FP or false positives) , and recall may be how many the DNN correctly predicted (TP) out of the total number of objects that did have the property in question (TP +FN or false negatives) . The F-score (F-score = 2 *PR / (P + R) ) unifies precision and recall into a single measure.
[0117] The training module 420 may compare the accuracy score with a threshold score. In an example where the training module 420 determines that the accuracy score of the DNN is less than the threshold score, the training module 420 instructs the training module 420 to re-train the DNN. In one embodiment, the training module 420 may iteratively re-train the DNN until the occurrence of a stopping condition, such as the accuracy measurement indication that the DNN may be sufficiently accurate, or a number of training rounds having taken place.
[0118] The compressing module 430 compresses DNNs. For instance, the compressing module 430 may add pruning operations to DNN layers to reduce computational complexity or memory usage. A pruning operation may prune weight tensors of a DNN layer by changing one or more nonzero valued weights of the layer to zeros. The modification may be done before, during, or after training. Weights may be pruned during training, during inference, or a combination of both. The compressing module 430 may determine a sparsity ratio for a DNN layer. The sparsity ratio may be a ratio of the number of zero-valued weight to the total number of weights in the layer. The compressing module 430 may perform the pruning operation till the sparsity ratio of the DNN layer meets a target sparsity ration, such as 10%, 20%, 30%, 40%, 50%, and so on.
[0119] In some embodiments, the compressing module 430 may select one or more layers in a DNN and modify each selected layer with a pruning operation. For instance, the compressing module 430 may select computationally complex layers, such as layers with large filters. For a pruning operation of a layer or of a type of layer, the compressing module 430 may determine a weight threshold that would not cause a loss of the accuracy of the DNN to exceed an accuracy loss constraint. A pruning operation may modify weights having absolute values above the weight threshold to zeros and leave the other weights unchanged. The weight pruning can reduce memory storage as zero-valued weights may not be stored. Also, the number of operations in the layer can be reduced as computations on zero-valued weights can be skipped without impacting the output of the layer. In some embodiments, the compressing module 430 may also measure energy saving, final DNN accuracy, or layer-wise sparsity caused by pruning operations.
[0120] After compressing a DNN, the compressing module 430 may fine tune the DNN, e.g., through a retraining process. The compressing module 430 may fine tunes DNNs after weights are pruned. In some embodiments, the fine-tuning process is a retraining or further training process. For instance, after weights in a DNN are pruned, the compressing module 430 may further train the DNN by inputting a training dataset into the DNN. The values of the unpruned weights in the DNN may be modified based on outputs of the DNN and ground-truth labels of the training samples in the training dataset. In some embodiments, the values of the pruned weights (i.e., zero) are not changed during the fine-tuning process. For instance, the compressing module 430 may place a mask over a pruned weight block and the mask can prevent values in the pruned weight blocks from being changed during the fine-tuning process. In other embodiments, the values of all weights, including the pruned weights, may be changed during the fine-tuning process. After one or more cycles of retraining and weight changing by the compressing module 430, the compressing module 430 may perform a new pruning process, e.g., by selecting weight blocks and pruning the selected weight blocks. In some embodiments, the weight pruning process may be repeated multiple times before the fine-tuning process is done. In some embodiments, the number of epochs in the fine-tuning process may be different from the number of epochs in the training process in which the pre-pruning values of the weights are determined. For instance, the fine-tuning process may have less epochs than the training process. In an example, the number of epochs in the fine-tuning process may be relatively small, such as 2, 3, 4, 5, and so on.
[0121] The compiler 440 compiles information of DNNs to executable instructions that can be executed, e.g., by the DNN accelerator 302, to carry out neural network operations in DNNs. In some embodiments, the compiler 405 may generate a graph representing a DNN. The graph may include nodes and edges. A node may represent a specific neural network operation in the DNN. An edge may connect two nodes and represent a connection between the two corresponding neural network operations. In an example, an edge may encode a tensor that flows from one of the neural network operations to the other neural network operation. The tensor may be an output tensor of the first neural network operation and an input tensor of the second neural network operation. The edge may encode one or more attributes of the tensor, such as size, shape, storage format, and so on. The compiler 440 may use the graph to generate executable DNNs. For instance, the compiler may generate computer program instructions (e.g., compilation descriptors) for executing DNNs. The instructions may be stored in registers associated with components of the DNN accelerator 302.
[0122] The interpolation mapping module 450 maps interpolation operations in DNNs to DPUs in DNN accelerators, e.g., the DPUs 330, as depthwise convolutions. The interpolation mapping module 450 may determine a horizontal scale factor and a vertical scale factor for an interpolation operation based on dimensions of the IFM and the OFM of the interpolation operation. The IFM has a spatial dimension of Hin×Win, where Hin is the height of the IFM, and Win is the width of the IFM. The IFM has a spatial dimension of Hout×Wout, where Hout is the height of the OFM, and Wout is the width of the OFM. The horizontal scale factor may be denoted as xscale=Win / Wout. The vertical scale factor may be denoted as yscale=Hin / Hout. In some embodiments, the horizontal scale factor or vertical scale factor may have a floating-point data format. For instance, or In some embodiments, the horizontal scale factor or vertical scale factor may be determined by the compiler 440 or the interpolation mapping module 450 in a compilation stage of the DNN.
[0123] The interpolation mapping module 450 may also determine vector coefficients for each (xout, yout) coordinate in the OFM. The vector coefficients for a (xout, yout) coordinate may include a vertical scaling coefficient and a horizonal scaling coefficient. In some embodiments (e.g., embodiments where scaling is static) , the interpolation mapping module 450 may determine vector coefficients offline, e.g., at compilation time. In other embodiments (e.g., embodiments where scaling is dynamic) , the interpolation mapping module 450 may determine vector coefficients at runtime, e.g., during DNN execution. The interpolation mapping module 450 may determine vector coefficients based on the scaling factors and coordinate transformation.
[0124] In some embodiments, the interpolation operation is a bi-linear interpolation. The interpolation mapping module 450 may decompose the bi-linear interpolation into a vertical interpolation and a horizonal interpolation. For the vertical interpolation, the interpolation mapping module 450 may generate a vertical scaling coefficient matrix α that has a spatial size of Hout×2. The interpolation mapping module 450 may initiate the values of all the elements in the vertical coefficient matrix to be zero. The interpolation mapping module 450 may also generate a vertical weight vector having a length of Hout and initiate the values of all the elements of the vertical vector to be zero. The interpolation mapping module 450 may identify every coordinate in the vertical dimension of the OFM, which may be denoted as yout and may be an integer in an inclusive range from 0 to (Hout-1) . For each yout in the OFM, the interpolation mapping module 450 may use the following functions to determine a vector coefficient α_yout: sy=math. floor (float ( (yout+0.5) ×yscale-0.5) ) , fy=float ( (yout+0.5) ×yscale-0.5) -sy, and α_youy= [1-fy, fy] ,
[0125] where math. floor () denotes a function that rounds a number down to the nearest integer. The vector coefficients of all the yout coordinates in the OFM constitute the vertical scaling coefficient matrix α.
[0126] For the horizontal interpolation, the interpolation mapping module 450 may generate a horizontal scaling coefficient matrix β that has a spatial size of Wout×2. The interpolation mapping module 450 may initiate the values of all the elements in the horizontal coefficient matrix to be zero. The interpolation mapping module 450 may also generate a horizontal weight vector having a length of Wout and initiate the values of all the elements of the horizontal weight vector to be zero. The interpolation mapping module 450 may then compute and update values of elements in the horizontal coefficient matrix and the horizontal weight vector. In an example, the interpolation mapping module 450 may identify every coordinate in the vertical dimension of the OFM, which may be denoted as xout and may be an integer in an inclusive range from 0 to (Wout-1) . For each xout in the OFM, the interpolation mapping module 450 may using the following functions to determine a vector coefficient β_xout: sx=math. floor (float ( (xout+0.5) ×xscale-0.5) ) , fx=float ( (xout+0.5) ×xscale-0.5) -sy, and β_xout= [1-fx, fx] .
[0127] The vector coefficients of all the xout coordinates in the OFM constitute the horizontal scaling coefficient matrix β.
[0128] In some embodiments, the interpolation mapping module 450 may map the vertical interpolation into a depthwise convolution and generate a vertical scaling kernel to be used to perform the depthwise convolution for the vertical interpolation. The interpolation mapping module 450 may define a vertical scaling kernel having a spatial size of 2×1× (Hout×C) ×1, where C is the number of channels. The data type of the weights may be FP16. The interpolation mapping module 450 may initiate the values of the weights in the kernel to be 0. The interpolation mapping module 450 may further select which element in the kernel to update. For each yout in the inclusive range from 0 to (Hout-1) , the interpolation mapping module 450 may take the vertical scaling coefficient, which is a 2×1 matrix and repeat it C times, which may result in a 3D tensor having a spatial size of 2×1×C. The interpolation mapping module 450 may further convert the 3D tensor into a 4D tensor having a spatial size of 2×1×C×1. The interpolation mapping module 450 may continue to update the elements in the kernel till the updating for all the yout coordinate and all the channels is complete. In some embodiments, the weights in different channels are the same. Depth convolution for vertical scaling may be performed on the IFM and the vertical scaling kernel.
[0129] The interpolation mapping module 450 may also map the horizontal interpolation into another depthwise convolution and generate a horizontal scaling kernel to be used to perform the depthwise convolution for the vertical interpolation. The horizontal scaling kernel may be a 4D tensor having a spatial size of 1×2× (Hout×C) ×1. The interpolation mapping module 450 may determine values of the elements in the horizontal scaling kernel based on the horizontal scaling coefficient. Depth convolution for horizontal scaling may be performed on the IFM and the horizontal scaling kernel.
[0130] In some embodiments, the interpolation mapping module 450 may change the storage format of input data. In an example, the interpolation mapping module 450 may change the storage format from NCHW to NHWC, where N represents the output channel dimension, C represents the input channel dimension, H represents the vertical dimension, and W represents the horizontal dimension. The interpolation mapping module 450 may perform a tensor permutation to change the order in which data elements in the tensor are stored.
[0131] In some embodiments (e.g., embodiments where the interpolation operation is a bi-cubic interpolation) , the interpolation mapping module 450 may determine a vertical scale factor and a horizontal scale factor as described above. Also, the interpolation mapping module 450 may compute scaling coefficients. In some embodiments (e.g., embodiments where scaling is static) , the interpolation mapping module 450 may determine vector coefficients offline, e.g., at compilation time. In other embodiments (e.g., embodiments where scaling is dynamic) , the interpolation mapping module 450 may determine vector coefficients at runtime, e.g., during DNN execution. The interpolation mapping module 450 may determine vector coefficients based on the scaling factors and coordinate transformation. Also, the interpolation mapping module 450 may perform additional compute of calculating weights (e.g., 1×4 or 4×1 weight tensor) at runtime based on the scaling factor and coordinate transformation. Each of the point map may be an integer mapping of the output over input coordinate, which can help in selecting the 4 rows or 4 columns needed for bi-cubic interpolation.
[0132] The interpolation mapping module 450 may compute a vertical scaling coefficient matrix α for the vertical interpolation in the bi-cubic interpolation. The spatial size of the vertical scaling coefficient matrix α may be Hout×4. The vertical scaling coefficient matrix αmay include a vector coefficient α_yout for each yout in the OFM. The interpolation mapping module 450 may use the following functions to determine a vector coefficient α_yout: sy=math. floor (float ( (yout+0.5) ×yscale-0.5) ) , fy=float ( (yout+0.5) ×yscale-0.5) -sy, α_yout= [α0, α1, α2, α3] , α0= ( (Af×(fy+1) -5×Af)×(fy+1)+8×Af)×(fy+1) -4×Sf, α1= ( (Sf+2)×fy- (Sf+3) )×fy×fy+1, α2= ( (Sf+2)×(1-fy) - (Af+3) )×(1-fy)×(1-fy)+1, and α3=1-α0-α1-α2
[0133] sy may be the weight value for yout.
[0134] The interpolation mapping module 450 may compute a horizontal scaling coefficient matrix β for the horizontal interpolation in the bi-cubic interpolation. The spatial size of the horizontal scaling coefficient matrix β may be Wout×4. The vertical scaling coefficient matrix β may include a vector coefficient β_yout for each yout in the OFM. The interpolation mapping module 450 may use the following functions to determine a vector coefficient β_xout: sx=math. floor (float ( (xout+0.5) ×xscale-0.5) ) , fx=float ( (xout+0.5) ×xscale-0.5) -sx, β_xout= [β0, β1, β2, β3] , β0= ( (Af× (fx+1) -5×Af)× (fx+1) +8×Af)× (fx+1) -4×Sf, β1= ( (Af+2)×fx- (Af+3) )×fx×fx+1, β2= ( (Af+2)× (1-fx) - (Af+3) )× (1-fx) × (1-fx) +1, and β3=1-β0-β1-β2
[0135] sx may be the weight value for xout.
[0136] The interpolation mapping module 450 may further generate a vertical scaling kernel for the vertical interpolation based on the vertical scaling coefficient matrix and generate a horizontal scaling kernel for the horizontal interpolation based on the horizontal scaling coefficient matrix. In some embodiments, the interpolation mapping module 450 may generate the vertical or horizontal scaling kernel by repeating the vertical or horizontal scaling coefficient matrix in one or more dimensions. In an example, the interpolation mapping module 450 may repeat the vertical or horizontal scaling coefficient matrix for multiple channels to obtain a tensor having a spatial size of 4× (Hout×C) , where C is the number of channels. The interpolation mapping module 450 may further reshape the tensor (e.g., adding one or more dimensions into the tensor) to generate the vertical or horizontal scaling kernel. In some embodiments, the vertical or horizontal scaling kernel may be a 4D tensor having a spatial size of 1×4× (Hout×C) ×1. Depth convolution for vertical scaling and horizontal scaling may be performed on the IFM and the horizontal scaling kernel.
[0137] In some embodiments, the interpolation mapping module 450 may change the storage format of input data. In an example, the interpolation mapping module 450 may change the storage format from NCHW to NHWC, where N represents the output channel dimension, C represents the input channel dimension, H represents the vertical dimension, and W represents the horizontal dimension. The interpolation mapping module 450 may perform a tensor permutation to change the storage format. The tensor permutation may include transposing one or more dimensions of the IFM or the filter. In some embodiments, the interpolation mapping module 450 may instruct the DNN accelerator 302 to perform vertical scaling first, then perform horizontal scaling on the result of the vertical scaling. In other embodiments, the interpolation mapping module 450 may instruct the DNN accelerator 302 to perform horizontal scaling first, then perform vertical scaling on the result of the horizontal scaling.
[0138] In some embodiments, the interpolation mapping module 450 may convert interpolation operations to depthwise convolutions by mapping interpolation parameters to depthwise convolution parameters. To generate an OFM of an interpolation operation (e.g., a bi-linear interpolation) , the interpolation mapping module 450 may prepare tensors on which a depthwise convolution is to be performed. The interpolation mapping module 450 may upsample (or instruct the DMA engine 320 to upsample) the IFM of the interpolation. The interpolation mapping module 450 may determine one or more duplicating scales, which may indicate how many times the IFM is duplicated. In some embodiments, interpolation mapping module 450 may determine a vertical duplicating scale, which may indicate how many times the IFM is duplicated in the vertical dimension, and a horizontal duplicating scale, which may indicate how many times the IFM is duplicated in the horizontal dimension.
[0139] The interpolation mapping module 450 may also determine one or more padding factors, based on which padding may be applied on the duplicated IFM. The one or more padding factors may include a vertical forward padding factor indicating the number of rows to be added above the first row of the IFM, a vertical back padding factor indicating the number of rows to be added below the last row of the IFM, a horizontal forward padding factor indicating the number of columns to be added to the left edge of the IFM, and a horizontal backward padding factor indicating the number of columns to be added to the right edge of the IFM.
[0140] The interpolation mapping module 450 may generate (or instruct the DMA engine 320) to generate an upsampled IFM by performing the duplication process (which may include one or more concatenation processes) based on the duplicating scales and performing the padding process based on the padding factors. The upsampled IFM has a larger size than the original IFM.
[0141] The upsampled IFM may be used as the input tensor of the depthwise convolution. The interpolation mapping module 450 may determine one or more intermediate parameters to convert interpolation parameters to depthwise convolution parameters. In an example, the intermediate parameters include a stride parameter, a padding parameter, and a kernel parameter, which may be computed using the following algorithm:
[0142] The intermediate parameters may be computed based on the mode and scale factor of the interpolation. For half pixel: In an embodiment of half pixel, the stride parameter equals 1, the padding parameter equals (1-Scale) ×0.5, and the kernel parameter equals 2×Scale. In another embodiment of half pixel, the stride parameter equals 2, the padding parameter equals (1-Scale) , and the kernel parameter equals Scale.For asymmetric and align corner: In an embodiment of asymmetric mode or aligned orders mode, the stride parameter equals 1, the padding parameter equals 0, and the kernel parameter equals Scale. Scale denotes the interpolation scale factor.
[0143] The intermediate parameters may be converted to parameters used for upsampling and depthwise convolution. In some embodiments (e.g., embodiments of upscale interpolation) , the duplicating scale may equal the kernel parameter. The forward padding factor (e.g., the horizontal forward padding factor or vertical forward padding factor) may equal the negation of the padding parameter. The back padding factor (e.g., the horizontal back padding factor or vertical back padding factor) may equal padding parameter+kernel parameter-stride parameter. The kernel shape (e.g., the vertical or horizontal dimension of the kernel) may equal the kernel parameter. The stride may equal the stride parameter.
[0144] In other embodiments (e.g., embodiments of downscale interpolation) , the forward padding factor (e.g., the horizontal forward padding factor or vertical forward padding factor) may be computed as The back padding factor (e.g., the horizontal back padding factor or vertical back padding factor) may equal zero. The kernel shape (e.g., the vertical or horizontal dimension of the kernel) may equal the kernel parameter. The stride may equal the scale factor. In a half pixel mode, the stride parameter may be 2×Scale, the padding parameter may be Scale or Scale-1. The kernel parameter may be 2.
[0145] The datastore 460 stores data received, generated, used, or otherwise associated with the DNN module 400. For example, the datastore 460 stores the datasets used by the training module 420. The datastore 460 may also store data generated by the training module 420, such as the hyperparameters for training DNNs, internal parameters of trained DNNs (e.g., weights, etc. ) , data for sparsity acceleration (e.g., sparsity bitmap, etc. ) , and so on.The datastore 460 may store instructions, compilation descriptors, or other data generated by the compiler 440. The datastore 460 may further store data received, processed, or generated by the interpolation mapping module 450. The datastore 460 may include one or more memories. In the embodiment of FIG. 4, the datastore 460 is a component of the DNN module 400. In other embodiments, the datastore 460 may be external to the DNN module 400 and communicate with the DNN module 400 through a network.
[0146] FIG. 5 illustrates an example sparse cell 500, in accordance with various embodiments. The sparse cell 500 may be a processing cell in a processing engine, e.g., the processing engine 370 in FIG. 3. The sparse cell 500 includes 16 MAC units 510 (individually referred to as “MAC unit 510” ) , which constitutes a MAC array having four rows and four columns. The MAC array has a spatial shape of 5x4, meaning the height of the MAC array is four and the width of the MAC array is also 5. The sparse cell 500 also includes 16 weight register files 520 (individually referred to as “weight register file 520” ) , 16 activation register files 530 (individually referred to as “activation register file 530” ) , four row buffers 540 (individually referred to as “row buffer 540” ) , and sparsity modules 560 (individually referred to as “sparsity module 560” ) . In other embodiments, the sparse cell 500 may include fewer, more, or different components. For example, the sparse cell 500 may include a different number of MAC units 510, weight register files 520, activation register files 530, row buffers 540, or sparsity modules 560. As another example, the sparse cell 500 may include column buffers in lieu of or in addition to the row buffers 540. Also, the shape (e.g., the height or width) of the MAC array may be different.
[0147] The MAC units 510 are configured to perform MAC operations. Each MAC unit 510 may include one or more multipliers and one or more adders. A multiplier may multiply an activation with a weight at a time to compute a product. In some embodiments (e.g., embodiments where the MAC unit 510 includes multiple multipliers) , the multipliers may operate simultaneously to process multiple activation-weight pairs and compute multiple products in one cycle. An adder may accumulate products computed by the multipliers. Even though not shown in FIG. 5, the sparse cell may include an adder tree including a plurality of adder tiers. The first tier may receive outputs of a plurality of MAC units 510. The number of adders in the first tier may be half of the number of the MAC units 510, and each adder may accumulate the outputs of two MAC units 510. The second tier may receive outputs of adders in the first tier. The number of adders in the second tier may be half of the number of adders in the first tier, and each adder in the second tier may accumulate the outputs of two adders in the first tier. The adder tree may include one or more other tiers. The last tier may include a single adder that accumulates outputs of adders in the second last tier to compute a partial sum of the sparse cell 500.
[0148] The weight register files 520 store weights to be processed in MAC operations. In the embodiments of FIG. 5, four weight register files 520 are grouped into a storage set that stores data to be used by a column of MAC units 510. There are four storage sets corresponding to the four columns of MAC units 510. In some embodiments, a weight register file 520 may correspond to a MAC unit 510 and store data to be processed by the MAC unit. In some embodiments, all the 16 weight register files 520 constitute a weight storage unit.
[0149] The activation register files 530 stores activations to be processed in MAC operations. In the embodiments of FIG. 5, four activation register files 530 are grouped into a storage set that stores data to be used by a row of MAC units 510. There are four storage sets corresponding to the four rows of MAC units 510. In some embodiments, an activation register file 530 may correspond to a MAC unit 510 and store data to be processed by the MAC unit. In some embodiments, all the 16 activation register files 530 constitute an activation storage unit. The row buffers 540 store outputs of the MAC units 510. Each row buffer 540 may drain outputs of a single row of MAC units 510.
[0150] The sparsity module 560 facilitates dynamic sparsity-based acceleration in the sparse cell 500. In the embodiments of FIG. 5, each sparsity module 560 includes a sparsity tensor storage unit 565 and a control logic 567. The sparsity tensor storage unit 565 stores combined sparsity tensors. A combined sparsity tensor stored in the sparsity tensor storage unit 565 may correspond to an activation tensor and a weight tensor. A nonzero element in the combined sparsity tensor may correspond to a nonzero activation-weight pair that includes a nonzero activation and a nonzero weight. The position of the nonzero activation in the activation tensor may match the position of the nonzero weight in the weight tensor. The product of the nonzero activation and nonzero weight would be nonzero.
[0151] The control logic 567 may control transmission of activations and weights stored from the weight register files 520 and the activation register files 530 to the MAC units 510 based on sparsity tensors. For instance, the control logic 567 may select a subset of the weights stored in the weight register files 520 and select a subset of activations stored in the activation register files 530 based on a combined sparsity tensor. The selected weights and activations constitute nonzero activation-weight pairs. The control logic 567 may transmit the selected weights and activations to the MAC units 510 for performing MAC operations. The other weights stored in the weight register files 520 and the other activations stored in the activation register files 530 are skipped from computation. In the embodiments of FIG. 5, each sparsity module 560 controls sparsity acceleration in a respective MAC unit 510. As the sparsity acceleration is either based on both weight sparsity and activation sparsity, 16 sparsity modules 560 are used for acceleration computations in the 16 MAC units 510.
[0152] As shown in FIG. 5, the sparse cell 500 is associated with multiplexers (MUXs) 503, 504, 505, and 506. In other embodiments, the sparse cell 500 may be associated with a different number of MUXs or other devices. The MUX 503 facilitates loading weights, e.g., from the local memory 340, into the weight register files 520. The MUX 504 facilitates loading activations, e.g., from the local memory 340, into the activation register files 530. The MUX 505 facilitates loading sparsity tensors into the sparsity tensor storage unit 565. The MUX 506 may be a drain MUX that can facilitate draining outputs of the MAC units 510, e.g., to the local memory 340.
[0153] In some embodiments, the sparse cell 500 may also execute matrix multiplications converted from Fourier transform operations. For an example Fourier transform operation, the MAC units 510 may perform MAC operations in the two sequences of matrix multiplications converted from the Fourier transform operation. The weight register files 520 may be used to store data points in transformation tensor of the Fourier transform operation. The activation register file 530 may be used to store data points in the input tensor of the Fourier transform operation. The row buffers 540 may store data points in the output tensor of the Fourier transform operation.
[0154] FIG. 6 illustrates a sparse cell array 600, in accordance with various embodiments. The sparse cell array 600 may be an example of the processing engine 370 in FIG. 3. In FIG. 6, the sparse cell array 600 includes sparse cells 610 (individually referred to as “sparse cell 610” ) arranged in four columns and four rows, an activation memory 620, and a weight memory 630. In other embodiments, the sparse cell array 600 may include fewer, more, or different components. For instance, the sparse cell array 600 may include a different number of columns, rows, or sparse cells 610.
[0155] Each sparse cell 610 may perform sparsity accelerated MAC operations. The sparse cells 610 may facilitate dynamic sparsity mode. For instance, the sparsity modes of a sparse cell 610 may be dynamically changed between a combined sparsity mode, an activation sparsity mode, a weight sparsity mode, and a dense mode. An embodiment of a sparse cell 610 may be the sparse cell 500 in FIG. 5. The activation memory 620 stores activations, such as activations in input tensors of neural network operations. Activations may be loaded from the activation memory 620 to sparse cells 610. The weight memory 630 stores weights, such as weights in filters of neural network operations. Weights may be loaded from the weight memory 630 to sparse cells 610. The activation memory 620 or weight memory 630 may be a buffer. In other embodiments, the sparse cell array 600 may include a dense data memory and a sparse data memory in lieu of the activation memory 620 and weight memory 630. The dense data memory may store dense tensors. The sparse data memory may store sparse tensors.
[0156] The sparse cell array 600 may also execute matrix multiplications in Fourier transform operations. The activation memory 620 may be used to store input tensors of the Fourier transform operations. The weight memory 630 may be used to store transformation matrices of the Fourier transform operations.
[0157] FIG. 7 illustrates an example PE 700, in accordance with various embodiments. The PE 700 may be a unit component of a processing cell, e.g., a processing cell in the processing engine 370. In the embodiments of FIG. 7, the PE 700 includes an MAC unit 705, an activation register file 710, a weight register file 720, an output register file 750, and a sparsity accelerator 760. The MAC unit 705 includes a multiplier 730 and an adder 740. In other embodiments, the PE 700 may include fewer, more, or different components.
[0158] The activation register file 710 stores an activation operand, which may be a context. The activation register file 710 may be an example of the activation register files 530 in FIG. 5. The weight register file 720 stores a weight operand. The weight register file 720 may be an example of the weight register files 520 in FIG. 5. The activation operand and weight operand may be loaded from a memory (e.g., the memory 340) into the activation register file 710 and the weight register file 720, respectively. The sparsity accelerator 760 receives a sparsity bitmap 715 that corresponds to the sparse tensor in the weight register file 720. The sparsity bitmap 715 may be a combined sparsity bitmap when the MAC unit 705 operates in a combined sparsity mode. The sparsity bitmap 715 may be an activation sparsity bitmap when the MAC unit 705 operates in an activation sparsity mode. The sparsity bitmap 715 may be a weight sparsity bitmap when the MAC unit 705 operates in a weight sparsity mode. The sparsity bitmap 715 may have the same size (e.g., the same number of elements) as or a larger size than the activation operand or the weight operand.
[0159] Using the sparsity bitmap 715, the sparsity accelerator 760 selects four activations from the activation register file 710 and selects four weights from the weight register file 720. The sparsity accelerator 760 transmits the selected activations and weights to the multiplier 730. These selected data elements correspond to the nonzero valued elements of the sparsity bitmap 715. The four selected activations and the four selected weights may constitute four activation-weight pairs. The multiplier 730 may compute a product based on each activation-weight pair and therefore, compute four products in total. The four products may be provided to the adder 740. Even though FIG. 7 shows a single multiplier 730, the MAC unit 705 may include multiple multipliers that can perform multiple multiplication operations at the same time.
[0160] The adder 740 accumulates the four products and computes a unit-level internal partial sum. The four unselected elements of the dense tensor are not processed to save power and time, which would not impact the value of the unit-level internal partial sum. For instance, when the dense tensor is a dense activation tensor, the weights corresponding to the unselected activations are zeros so the products of the unselected activations and the weights would all be zero and have no contribution to the unit-level internal partial sum or other partial sums computed by the sparse cell. Similarly, when the dense tensor is a dense weight tensor, the activations corresponding to the unselected weights are zeros so the products of the unselected weights and the activations would all be zero and have no contribution to the unit-level internal partial sum or other partial sums computed by the sparse cell. In other embodiments, the MAC unit 705 may operate in a dense mode in which the sparsity bitmap 715 is not used and the sparsity accelerator 760 is inactive. The MAC unit 705 may process all the activations in the activation operand and all the weights in the weight operand.
[0161] The unit-level internal partial sum may be stored in the output register file 750. In some embodiments, the unit-level internal partial sum may be used multiple times. For instance, the activation operand may represent N data blocks in the input tensor of the convolution, where N is an integer greater than 1. Instead of processing all the N data blocks to compute N unit-level internal partial sums, the unit-level internal partial sum is computed once and used N times in the convolutional layers as N unit-level internal partial sums.
[0162] In some embodiments, the PE 700 receives one or more PE-level internal partial sums from one or more other PEs. The adder 740 or an accumulator (not shown in FIG. 7) can accumulate the one or more PE-level internal partial sums with the PE-level internal partial sum of the PE 700 and store the result of the accumulation (i.e., a multi-PE internal partial sum) in the output register file 750. The one or more other PEs may be in the same column as the PE 700 in a sparse cell. The multi-unit internal partial sum may be a column-level internal partial sum. In some embodiments, the PE-level internal partial sum of the PE 700 or the multi-unit internal partial sum may be sent to one or more other PEs for further accumulation.
[0163] Example Conversion of Interpolation to Depthwise Convolution
[0164] FIG. 8 illustrates an interpolation operation converted to an upsampling operation 801 and a depthwise convolution 802, in accordance with various embodiments. The interpolation operation is an operation to generate an OFM 820 from an IFM 810. For the purpose of illustration, the IFM 810 is a 3×3 matrix, shown by a 3×3 grid in FIG. 8. The numbers in the grid elements are the values of the pixels in the IFM 810. The OFM 820 is a 6×6 matrix, shown by a 6×6 grid in FIG. 8. The numbers in the grid elements are the values of the pixels in the OFM 820. In other embodiments, the IFM 810 or OFM 820 may have a different size. Also, the values of the pixels in the IFM 810 or OFM 820 may be different.
[0165] In the embodiments of FIG. 8, the interpolation operation is converted to upsampling operation 801 and depthwise convolution 802, which can be performed by the DNN accelerator 302 to compute the OFM 820. The parameters for the upsampling operation 801 and parameters for the depthwise convolution 802 are determined based on the scale factor of the interpolation operation. The scale factor of the interpolation operation is 2 in both the horizontal dimension and the vertical dimension. The mode of the interpolation operation may be half pixel. The strid parameter is determined to be 2. The padding parameter equal 1-scale=-1. The kernel parameter equals 2×scale=4. The parameters for the upsampling operation 801 and parameters for the depthwise convolution 802 are determined based on these three parameters. For instance, the upsampling scale equals the kernel parameter. The forward padding factor equals the negation of the padding parameter, which is 1. The back padding factor equals the padding parameter plus the kernel parameter minus the stride parameter, which is 1. The kernel dimension of the depthwise convolution 802 equals the kernel parameter, which is 4. The stride value of the depthwise convolution 802 equals the stride parameter, which is 2.
[0166] The upsampling operation 801 receives the IFM 810 as an input and outputs an upsampled IFM 830. In some embodiments, the upsampling operation 801 includes duplications of pixels in the IFM 810, followed by padding. For instance, every single pixel in the IFM 810 is duplicated to produce 14 pixels of the same value. Through the duplications, the 3×3 matrix is converted to a 12×12 matrix. The 12×12 matrix is then padded, e.g., by adding a row to the top, a row to the bottom, a column to the right, and a column to the left. The row added to the top has the same pixels as the top row of the 12×12 matrix. The row added to the bottom has the same pixels as the bottom row of the 12×12 matrix. The column added to the right has the same pixels as the right column of the 12×12 matrix. The column added to the left has the same pixels as the left column of the 12×12 matrix. The padding generates a 14×14 matrix.
[0167] A kernel 840 is generated for the depthwise convolution. The kernel 840 is a 4×4 matrix. The depthwise convolution is applied on the upsampled IFM 830 and the kernel 840 and generates the OFM 820. Even though the IFM 810, upsampled IFM 830, kernel 840, and OFM 820 are 2D tensors in FIG. 8, the IFM 810, upsampled IFM 830, kernel 840, or OFM 820 may be a 3D tensor in other embodiments. For instance, the IFM 810 may include a number N channels. The upsampling operation 801 may be performed for each of the channels so that the upsampled IFM 810 may also include N channels. The depthwise convolution 802 may be applied on each of the channels using the kernel 840 so that the OFM 820 also has Nchannels. Different from standard convolution, depthwise convolution does not cause collapse of input channels so that the channels in the IFM can be preserved in the OFM.
[0168] FIGS. 9A and 9B illustrate a depthwise convolution performed for vertical scaling, in accordance with various embodiments. In some embodiments, the vertical scaling is in a bi-cubic interpolation operation. FIG. 9A shows an IFM 910 and a kernel 920. The IFM 910 is the IFM of the bi-cubic interpolation operation and is also used as the IFM of the depthwise convolution. The IFM 910 has a spatial size Hin×Win, where Hin is the height of the IFM 910 and Win is the width of the IFM 910. The kernel 920 has a spatial size Hout×4, where Hout is the height of the OFM of the interpolation operation and 4 is the width of the kernel 920. The kernel 920 may be determined by the interpolation mapping module 450 based on one or more scale factors of the bi-cubic interpolation.
[0169] The first row of the kernel 920, which is highlighted with a dot pattern in FIG. 9A, corresponds to the first four rows (i.e., rows 0-3) of the IFM 910, which are also highlighted with the dot pattern in FIG. 9A. The second row of the kernel 920, which is highlighted with a diagonal line pattern in FIG. 9A, corresponds to rows 5-8 of the IFM 910, which are also highlighted with the diagonal line pattern in FIG. 9A. The depthwise operation may be performed using the IFM 910 and the kernel 920.
[0170] The output of the depthwise convolution is an intermediate feature map 930 shown in FIG. 9B. In some embodiments, the first row of the intermediate feature map 930, which is highlighted by the dot pattern in FIG. 9B, may be computed from rows 0-3 of the IFM 910 and the first row of the kernel 920. The second row of the intermediate feature map 930, which is highlighted by the diagonal line pattern in FIG. 9B, may be computed from rows 5-8 of the IFM 910 and the second row of the kernel 920. A depthwise convolution having a kernel shape 1×1×1×4 and an input tensor shape 1×1×Win×4 is performed to compute each row of the intermediate feature map 930. This continues till all the rows of the intermediate feature map 930 are computed. The intermediate feature map 930 has a spatial size Hout×Win. Even though the IFM 910, kernel 920, and intermediate feature map 930 are all 2D tensors, the IFM 910, kernel 920 or intermediate feature map 930 may be a 3D tensor in some embodiments. The third dimension may be the channel dimension. A depthwise convolution may be performed in each of the channels. The intermediate feature map 930 may have the same channels as the IFM 910.
[0171] FIGS. 10A and 10B illustrate a depthwise convolution performed for horizontal scaling, in accordance with various embodiments. In some embodiments, the horizontal scaling is in a bi-cubic interpolation operation. The horizontal scaling may be performed after a vertical scaling. In some embodiments, the depthwise convolution in FIGS. 10A and 10B is performed on an intermediate feature map 1010 and a kernel 1020, which are shown in FIG. 10A, to compute an OFM 1030, which is shown in FIG. 10B. The intermediate feature map 1010 may be an output of a previously performed depthwise convolution that is converted from the vertical scaling in the bi-cubic interpolation operation. For instance, the intermediate feature map 1010 may be the intermediate feature map 930 in FIG. 9B. The intermediate feature map 1010 has a spatial size Hout×Win, where Hout is the height of the OFM 1030 and Win is the width of the IFM of the bi-cubic interpolation operation. The kernel 1020 has a spatial size Hout×4, where Hout is the height of the OFM 1030 and 4 is the width of the kernel 1020. The kernel 1020 may be determined by the interpolation mapping module 450 based on one or more scale factors of the bi-cubic interpolation.
[0172] In some embodiments, the first column of the OFM 1030, which is highlighted by a dot pattern in FIG. 10B, may be computed from columns 2-5 of the intermediate feature map 1010 and the first column of the kernel 1020, which are highlighted by the dot pattern in FIG. 10A. The second column of the OFM 1030, which is highlighted by the diagonal line pattern in FIG. 10B, may be computed from columns 8-11 of the intermediate feature map 1010 and the second column of the kernel 1020, which are highlighted by the diagonal line pattern in FIG. 10A. A depthwise convolution having a kernel shape 1×1×4×1 and an input tensor shape 1×1×4×Hout may be performed to compute each column of the OFM 1030. This continues till all the columns of the OFM 1030 are computed. The OFM 1030 has a spatial size Hout×Wout. Even though the intermediate feature map 1010, kernel 1020, and OFM 1030 are all 2D tensors, the intermediate feature map 1010, kernel 1020 or OFM 1030 may be a 3D tensor in some embodiments. The third dimension may be the channel dimension. A depthwise convolution may be performed in each of the channels. The OFM 1030 may have the same channels as the intermediate feature map 1010.
[0173] FIG. 11A is a flowchart of a method 1100 of executing an interpolation operation in a DNN, in accordance with various embodiments. The method 1100 may be performed by the DNN system 300 in FIG. 3. Although the method 1100 is described with reference to the flowchart illustrated in FIG. 11A, many other methods for executing interpolation operations in DNNs may alternatively be used. For example, the order of execution of the steps in FIG. 11A may be changed. As another example, some of the steps may be changed, eliminated, or combined.
[0174] The DNN system 300 determines 1110 a scale factor of the interpolation operation based on a dimension of an IFM of the interpolation operation and a dimension of an OFM of the interpolation operation.
[0175] The DNN system 300 computes 1120 a scaling coefficient based on the scale factor and a position of the pixel in the OFM for a pixel in the OFM. In some embodiments, the interpolation operation is a bi-linear interpolation, and the scaling coefficient is a vector comprising two elements. In other embodiments, the interpolation operation is a bi-cubic interpolation, and the scaling coefficient is a vector comprising four elements.
[0176] The DNN system 300 generates 1130 a kernel from one or more scaling coefficients of one or more pixels in the OFM. In some embodiments, the kernel comprises weights, and values of at least some of the weights are values in the one or more scaling coefficients. In some embodiments, the interpolation operation is a bi-linear interpolation, and a dimension of the kernel is two. In other embodiments, the interpolation operation is a bi-cubic interpolation, and a dimension of the kernel is four. In some embodiments, the OFM has a number of channels, and a dimension of the kernel is the number or a multiple of the number.
[0177] The DNN system 300 performs 1140 a depthwise convolution on the IFM and the kernel. In some embodiments, the DNN system 300 performs a subsequent depthwise convolution on the output of the depthwise convolution, wherein the output of the subsequent depthwise convolution is the OFM of the interpolation operation.
[0178] The DNN system 300 determines 1150 the value of the pixel in the OFM using an output of the depthwise convolution. In some embodiments, the DNN system 300 determines another scale factor of the interpolation operation based on another dimension of an IFM of the interpolation operation and another dimension of an OFM of the interpolation operation. The dimension of the IFM and the dimension of the OFM are widths. The another dimension of the IFM and the another dimension of the OFM are heights. The subsequent depthwise convolution is performed based on the another scale factor.
[0179] In some embodiments, the position of the pixel is a position of the pixel in the dimension of the OFM. For the pixel in the OFM, the DNN system 300 computes another scaling coefficient based on the another scale factor and a position of the pixel in the another dimension of the OFM. The DNN system 300 generates a kernel for the subsequent depthwise convolution using the another scaling coefficient.
[0180] FIG. 11B is a flowchart of another method 1105 of executing an interpolation operation in a DNN, in accordance with various embodiments. The method 1105 may be performed by the DNN system 300 in FIG. 3. Although the method 1105 is described with reference to the flowchart illustrated in FIG. 11B, many other methods for executing interpolation operations in DNNs may alternatively be used. For example, the order of execution of the steps in FIG. 11B may be changed. As another example, some of the steps may be changed, eliminated, or combined.
[0181] The DNN system 300 upsamples 1115 an IFM of the interpolation operation to generate an upsampled feature map. In some embodiments, the DNN system 300 forms an intermediate feature map by duplicating one or more pixels in the IFM. The intermediate feature map comprises a matrix formed by duplicating a single pixel in the IFM. In some embodiments, the DNN system 300 determines a number of times the single pixel is duplicated based on the scale factor of the interpolation factor. In some embodiments, the DNN system 300 adds one or more sequences of new pixels to an edge of the intermediate feature map. In some embodiments, the DNN system 300 determines a number of sequences of new pixels to be added to the edge of the intermediate feature map based on the scale factor of the interpolation factor.
[0182] The DNN system 300 determines 1125 a scale factor of the interpolation operation based on a dimension of the IFM of the interpolation operation and a dimension of an OFM of the interpolation operation.
[0183] The DNN system 300 generates 1135 a kernel based on the scale factor of the interpolation operation. In some embodiments, the DNN system 300 determines one or more dimensions of the kernel based on the scale factor of the interpolation operation.
[0184] The DNN system 300 performs 1145 a depthwise convolution on the IFM and the kernel to compute the OFM of the interpolation operation. In some embodiments, the DNN system 300 determines a stride value based on the scale factor of the interpolation operation. The stride value indicates a number of pixels that kernel moves at a time in the depthwise convolution.
[0185] Example Computing Device
[0186] FIG. 12 is a block diagram of an example computing device 2000, in accordance with various embodiments. In some embodiments, the computing device 2000 can be used as at least part of the DNN system 300. A number of components are illustrated in FIG. 12 as included in the computing device 2000, but any one or more of these components may be omitted or duplicated, as suitable for the application. In some embodiments, some or all of the components included in the computing device 2000 may be attached to one or more motherboards. In some embodiments, some or all of these components are fabricated onto a single system on a chip (SoC) die. Additionally, in various embodiments, the computing device 2000 may not include one or more of the components illustrated in FIG. 12, but the computing device 2000 may include interface circuitry for coupling to the one or more components. For example, the computing device 2000 may not include a display device 2006, but may include display device interface circuitry (e.g., a connector and driver circuitry) to which a display device 2006 may be coupled. In another set of examples, the computing device 2000 may not include an audio input device 2018 or an audio output device 2008 but may include audio input or output device interface circuitry to which an audio input device 2018 or audio output device 2008 may be coupled.
[0187] The computing device 2000 may include a processing device 2002 (e.g., one or more processing devices) . The processing device 2002 processes electronic data from registers and / or memory to transform that electronic data into other electronic data that may be stored in registers and / or memory. The computing device 2000 may include a memory 2004, which may itself include one or more memory devices such as volatile memory (e.g., DRAM) , nonvolatile memory (e.g., read-only memory (ROM) ) , high bandwidth memory (HBM) , flash memory, solid state memory, and / or a hard drive. In some embodiments, the memory 2004 may include memory that shares a die with the processing device 2002. In some embodiments, the memory 2004 includes one or more non-transitory computer-readable media storing instructions executable to perform interpolation operations in DNNs (e.g., the method 1100 described in conjunction with FIG. 11A or the method 1105 described in conjunction with FIG. 11B) or some operations performed by one or more components of the DNN system 300. The instructions stored in the one or more non-transitory computer-readable media may be executed by the processing device 2002.
[0188] In some embodiments, the computing device 2000 may include a communication chip 2012 (e.g., one or more communication chips) . For example, the communication chip 2012 may be configured for managing wireless communications for the transfer of data to and from the computing device 2000. The term "wireless" and its derivatives may be used to describe circuits, devices, systems, methods, techniques, communications channels, etc., that may communicate data through the use of modulated electromagnetic radiation through a nonsolid medium. The term does not imply that the associated devices do not contain any wires, although in some embodiments they might not.
[0189] The communication chip 2012 may implement any of a number of wireless standards or protocols, including but not limited to Institute for Electrical and Electronic Engineers (IEEE) standards including Wi-Fi (IEEE 802.10 family) , IEEE 802.16 standards (e.g., IEEE 802.16-2005 Amendment) , Long-Term Evolution (LTE) project along with any amendments, updates, and / or revisions (e.g., advanced LTE project, ultramobile broadband (UMB) project (also referred to as "3GPP2" ) , etc. ) . IEEE 802.16 compatible Broadband Wireless Access (BWA) networks are generally referred to as WiMAX networks, an acronym that stands for worldwide interoperability for microwave access, which is a certification mark for products that pass conformity and interoperability tests for the IEEE 802.16 standards. The communication chip 2012 may operate in accordance with a Global System for Mobile Communication (GSM) , General Packet Radio Service (GPRS) , Universal Mobile Telecommunications System (UMTS) , High Speed Packet Access (HSPA) , Evolved HSPA (E-HSPA) , or LTE network. The communication chip 2012 may operate in accordance with Enhanced Data for GSM Evolution (EDGE) , GSM EDGE Radio Access Network (GERAN) , Universal Terrestrial Radio Access Network (UTRAN) , or Evolved UTRAN (E-UTRAN) . The communication chip 2012 may operate in accordance with Code-division Multiple Access (CDMA) , Time Division Multiple Access (TDMA) , Digital Enhanced Cordless Telecommunications (DECT) , Evolution-Data Optimized (EV-DO) , and derivatives thereof, as well as any other wireless protocols that are designated as 3G, 4G, 5G, and beyond. The communication chip 2012 may operate in accordance with other wireless protocols in other embodiments. The computing device 2000 may include an antenna 2022 to facilitate wireless communications and / or to receive other wireless communications (such as AM or FM radio transmissions) .
[0190] In some embodiments, the communication chip 2012 may manage wired communications, such as electrical, optical, or any other suitable communication protocols (e.g., the Ethernet) . As noted above, the communication chip 2012 may include multiple communication chips. For instance, a first communication chip 2012 may be dedicated to shorter-range wireless communications such as Wi-Fi or Bluetooth, and a second communication chip 2012 may be dedicated to longer-range wireless communications such as global positioning system (GPS) , EDGE, GPRS, CDMA, WiMAX, LTE, EV-DO, or others. In some embodiments, a first communication chip 2012 may be dedicated to wireless communications, and a second communication chip 2012 may be dedicated to wired communications.
[0191] The computing device 2000 may include battery / power circuitry 2014. The battery / power circuitry 2014 may include one or more energy storage devices (e.g., batteries or capacitors) and / or circuitry for coupling components of the computing device 2000 to an energy source separate from the computing device 2000 (e.g., AC line power) .
[0192] The computing device 2000 may include a display device 2006 (or corresponding interface circuitry, as discussed above) . The display device 2006 may include any visual indicators, such as a heads-up display, a computer monitor, a projector, a touchscreen display, a liquid crystal display (LCD) , a light-emitting diode display, or a flat panel display, for example.
[0193] The computing device 2000 may include an audio output device 2008 (or corresponding interface circuitry, as discussed above) . The audio output device 2008 may include any device that generates an audible indicator, such as speakers, headsets, or earbuds, for example.
[0194] The computing device 2000 may include an audio input device 2018 (or corresponding interface circuitry, as discussed above) . The audio input device 2018 may include any device that generates a signal representative of a sound, such as microphones, microphone arrays, or digital instruments (e.g., instruments having a musical instrument digital interface (MIDI) output) .
[0195] The computing device 2000 may include a GPS device 2016 (or corresponding interface circuitry, as discussed above) . The GPS device 2016 may be in communication with a satellite-based system and may receive a location of the computing device 2000, as known in the art.
[0196] The computing device 2000 may include another output device 2010 (or corresponding interface circuitry, as discussed above) . Examples of the other output device 2010 may include an audio codec, a video codec, a printer, a wired or wireless transmitter for providing information to other devices, or an additional storage device.
[0197] The computing device 2000 may include another input device 2020 (or corresponding interface circuitry, as discussed above) . Examples of the other input device 2020 may include an accelerometer, a gyroscope, a compass, an image capture device, a keyboard, a cursor control device such as a mouse, a stylus, a touchpad, a bar code reader, a Quick Response (QR) code reader, any sensor, or a radio frequency identification (RFID) reader.
[0198] The computing device 2000 may have any desired form factor, such as a handheld or mobile computer system (e.g., a cell phone, a smart phone, a mobile internet device, a music player, a tablet computer, a laptop computer, a netbook computer, an ultrabook computer, a personal digital assistant (PDA) , an ultramobile personal computer, etc. ) , a desktop computer system, a server or other networked computing component, a printer, a scanner, a monitor, a set-top box, an entertainment control unit, a vehicle control unit, a digital camera, a digital video recorder, or a wearable computer system. In some embodiments, the computing device 2000 may be any other electronic device that processes data.
[0199] Select Examples
[0200] The following paragraphs provide various examples of the embodiments disclosed herein.
[0201] Example 1 provides a method of executing an interpolation operation in a neural network, including determining a scale factor of the interpolation operation based on a dimension of an IFM of the interpolation operation and a dimension of an OFM of the interpolation operation; for a pixel in the OFM, computing a scaling coefficient based on the scale factor and a position of the pixel in the OFM; generating a kernel from one or more scaling coefficients of one or more pixels in the OFM; performing a depthwise convolution on the IFM and the kernel; and determining a value of the pixel in the OFM using an output of the depthwise convolution.
[0202] Example 2 provides the method of example 1, in which the interpolation operation is a bi-linear interpolation, and the scaling coefficient is a vector including two elements.
[0203] Example 3 provides the method of example 1 or 2, in which the interpolation operation is a bi-cubic interpolation, and the scaling coefficient is a vector including four elements.
[0204] Example 4 provides the method of any one of examples 1-3, in which the kernel includes weights, and values of at least some of the weights are values in the one or more scaling coefficients.
[0205] Example 5 provides the method of any one of examples 1-4, in which the interpolation operation is a bi-linear interpolation, and a dimension of the kernel is two.
[0206] Example 6 provides the method of any one of examples 1-5, in which the interpolation operation is a bi-cubic interpolation, and a dimension of the kernel is four.
[0207] Example 7 provides the method of any one of examples 1-6, in which the OFM has a number of channels, and a dimension of the kernel is the number or a multiple of the number.
[0208] Example 8 provides the method of any one of examples 1-7, further including performing a subsequent depthwise convolution on the output of the depthwise convolution, in which the output of the subsequent depthwise convolution is the OFM of the interpolation operation.
[0209] Example 9 provides the method of example 8, further including determining another scale factor of the interpolation operation based on another dimension of an IFM of the interpolation operation and another dimension of an OFM of the interpolation operation, in which the dimension of the IFM and the dimension of the OFM are widths, and the another dimension of the IFM and the another dimension of the OFM are heights, and the subsequent depthwise convolution is performed based on the another scale factor.
[0210] Example 10 provides the method of example 9, in which the position of the pixel is a position of the pixel in the dimension of the OFM, and the method further includes for the pixel in the OFM, computing another scaling coefficient based on the another scale factor and a position of the pixel in the another dimension of the OFM; and generating a kernel for the subsequent depthwise convolution using the another scaling coefficient.
[0211] Example 11 provides one or more non-transitory computer-readable media storing instructions executable to perform operations for executing an interpolation operation in a neural network, the operations including determining a scale factor of the interpolation operation based on a dimension of an IFM of the interpolation operation and a dimension of an OFM of the interpolation operation; for a pixel in the OFM, computing a scaling coefficient based on the scale factor and a position of the pixel in the OFM; generating a kernel from one or more scaling coefficients of one or more pixels in the OFM; performing a depthwise convolution on the IFM and the kernel; and determining a value of the pixel in the OFM using an output of the depthwise convolution.
[0212] Example 12 provides the one or more non-transitory computer-readable media of example 11, in which the interpolation operation is a bi-linear interpolation, and the scaling coefficient is a vector including two elements.
[0213] Example 13 provides the one or more non-transitory computer-readable media of example 11 or 12, in which the interpolation operation is a bi-cubic interpolation, and the scaling coefficient is a vector including four elements.
[0214] Example 14 provides the one or more non-transitory computer-readable media of any one of examples 11-13, in which the kernel includes weights, and values of at least some of the weights are values in the one or more scaling coefficients.
[0215] Example 15 provides the one or more non-transitory computer-readable media of any one of examples 11-14, in which the interpolation operation is a bi-linear interpolation, and a dimension of the kernel is two.
[0216] Example 16 provides the one or more non-transitory computer-readable media of any one of examples 11-15, in which the interpolation operation is a bi-cubic interpolation, and a dimension of the kernel is four.
[0217] Example 17 provides the one or more non-transitory computer-readable media of any one of examples 11-16, in which the OFM has a number of channels, and a dimension of the kernel is the number or a multiple of the number.
[0218] Example 18 provides the one or more non-transitory computer-readable media of any one of examples 11-17, in which the operations further include performing a subsequent depthwise convolution on the output of the depthwise convolution, in which the output of the subsequent depthwise convolution is the OFM of the interpolation operation.
[0219] Example 19 provides the one or more non-transitory computer-readable media of example 18, in which the operations further include determining another scale factor of the interpolation operation based on another dimension of an IFM of the interpolation operation and another dimension of an OFM of the interpolation operation, in which the dimension of the IFM and the dimension of the OFM are widths, and the another dimension of the IFM and the another dimension of the OFM are heights, and the subsequent depthwise convolution is performed based on the another scale factor.
[0220] Example 20 provides the one or more non-transitory computer-readable media of example 19, in which the position of the pixel is a position of the pixel in the dimension of the OFM, and the one or more non-transitory computer-readable media further includes for the pixel in the OFM, computing another scaling coefficient based on the another scale factor and a position of the pixel in the another dimension of the OFM; and generating a kernel for the subsequent depthwise convolution using the another scaling coefficient.
[0221] Example 21 provides an apparatus, including a computer processor for executing computer program instructions; and a non-transitory computer-readable memory storing computer program instructions executable by the computer processor to perform operations for executing an interpolation operation in a neural network, the operations including determining a scale factor of the interpolation operation based on a dimension of an IFM of the interpolation operation and a dimension of an OFM of the interpolation operation, for a pixel in the OFM, computing a scaling coefficient based on the scale factor and a position of the pixel in the OFM, generating a kernel from one or more scaling coefficients of one or more pixels in the OFM, performing a depthwise convolution on the IFM and the kernel, and determining a value of the pixel in the OFM using an output of the depthwise convolution.
[0222] Example 22 provides the apparatus of example 21, in which the interpolation operation is a bi-linear interpolation, and the scaling coefficient is a vector including two elements.
[0223] Example 23 provides the apparatus of example 21 or 22, in which the interpolation operation is a bi-cubic interpolation, and the scaling coefficient is a vector including four elements.
[0224] Example 24 provides the apparatus of any one of examples 21-23, in which the kernel includes weights, and values of at least some of the weights are values in the one or more scaling coefficients.
[0225] Example 25 provides the apparatus of any one of examples 21-24, in which the OFM has a number of channels, and a dimension of the kernel is the number or a multiple of the number.
[0226] Example 26 provides the apparatus of any one of examples 21-25, in which the operations further include performing a subsequent depthwise convolution on the output of the depthwise convolution, in which the output of the subsequent depthwise convolution is the OFM of the interpolation operation.
[0227] Additional Select Examples
[0228] The following paragraphs provide various examples of the embodiments disclosed herein.
[0229] Example 1 provides a method of executing an interpolation operation in a neural network, including upsampling an IFM of the interpolation operation to generate an upsampled feature map; determining a scale factor of the interpolation operation based on a dimension of the IFM of the interpolation operation and a dimension of an OFM of the interpolation operation; generating a kernel based on the scale factor of the interpolation operation; and performing a depthwise convolution on the IFM and the kernel to compute the OFM of the interpolation operation.
[0230] Example 2 provides the method of example 1, in which upsampling the IFM includes forming an intermediate feature map by duplicating one or more pixels in the IFM, the intermediate feature map including a matrix formed by duplicating a single pixel in the IFM.
[0231] Example 3 provides the method of example 2, further including determining a number of times the single pixel is duplicated based on the scale factor of the interpolation factor.
[0232] Example 4 provides the method of example 2 or 3, in which upsampling the IFM further includes adding one or more sequences of new pixels to an edge of the intermediate feature map.
[0233] Example 5 provides the method of example 4, further including determining a number of sequences of new pixels to be added to the edge of the intermediate feature map based on the scale factor of the interpolation factor.
[0234] Example 6 provides the method of any one of examples 1-5, in which generating the kernel includes determining one or more dimensions of the kernel based on the scale factor of the interpolation operation.
[0235] Example 7 provides the method of any one of examples 1-6, further including determining a stride value based on the scale factor of the interpolation operation, in which the stride value indicates a number of pixels that kernel moves at a time in the depthwise convolution.
[0236] Example 8 provides one or more non-transitory computer-readable media storing instructions executable to perform operations for executing an interpolation operation in a neural network, the operations including upsampling an IFM of the interpolation operation to generate an upsampled feature map; determining a scale factor of the interpolation operation based on a dimension of the IFM of the interpolation operation and a dimension of an OFM of the interpolation operation; generating a kernel based on the scale factor of the interpolation operation; and performing a depthwise convolution on the IFM and the kernel to compute the OFM of the interpolation operation.
[0237] Example 9 provides the one or more non-transitory computer-readable media of example 8, in which upsampling the IFM includes forming an intermediate feature map by duplicating one or more pixels in the IFM, the intermediate feature map including a matrix formed by duplicating a single pixel in the IFM.
[0238] Example 10 provides the one or more non-transitory computer-readable media of example 9, in which the operations further include determining a number of times the single pixel is duplicated based on the scale factor of the interpolation factor.
[0239] Example 11 provides the one or more non-transitory computer-readable media of example 9 or 10, in which upsampling the IFM further includes adding one or more sequences of new pixels to an edge of the intermediate feature map.
[0240] Example 12 provides the one or more non-transitory computer-readable media of example 11, in which the operations further include determining a number of sequences of new pixels to be added to the edge of the intermediate feature map based on the scale factor of the interpolation factor.
[0241] Example 13 provides the one or more non-transitory computer-readable media of any one of examples 8-12, in which generating the kernel includes determining one or more dimensions of the kernel based on the scale factor of the interpolation operation.
[0242] Example 14 provides the one or more non-transitory computer-readable media of any one of examples 8-13, in which the operations further include determining a stride value based on the scale factor of the interpolation operation, in which the stride value indicates a number of pixels that kernel moves at a time in the depthwise convolution.
[0243] Example 15 provides an apparatus, including a computer processor for executing computer program instructions; and a non-transitory computer-readable memory storing computer program instructions executable by the computer processor to perform operations for executing an interpolation operation in a neural network, the operations including upsampling an IFM of the interpolation operation to generate an upsampled feature map, determining a scale factor of the interpolation operation based on a dimension of the IFM of the interpolation operation and a dimension of an OFM of the interpolation operation, generating a kernel based on the scale factor of the interpolation operation, and performing a depthwise convolution on the IFM and the kernel to compute the OFM of the interpolation operation.
[0244] Example 16 provides the apparatus of example 15, in which upsampling the IFM includes forming an intermediate feature map by duplicating one or more pixels in the IFM, the intermediate feature map including a matrix formed by duplicating a single pixel in the IFM.
[0245] Example 17 provides the apparatus of example 16, further including determining a number of times the single pixel is duplicated based on the scale factor of the interpolation factor.
[0246] Example 18 provides the apparatus of example 16 or 17, in which upsampling the IFM further includes adding one or more sequences of new pixels to an edge of the intermediate feature map.
[0247] Example 19 provides the apparatus of example 18, further including determining a number of sequences of new pixels to be added to the edge of the intermediate feature map based on the scale factor of the interpolation factor.
[0248] Example 20 provides the apparatus of any one of examples 15-19, in which generating the kernel includes determining one or more dimensions of the kernel based on the scale factor of the interpolation operation.
[0249] The above description of illustrated implementations of the disclosure, including what is described in the Abstract, is not intended to be exhaustive or to limit the disclosure to the precise forms disclosed. While specific implementations of, and examples for, the disclosure are described herein for illustrative purposes, various equivalent modifications are possible within the scope of the disclosure, as those skilled in the relevant art will recognize. These modifications may be made to the disclosure in light of the above detailed description.
Claims
1.A method of executing an interpolation operation in a neural network, comprising:determining a scale factor of the interpolation operation based on a dimension of an input feature map of the interpolation operation and a dimension of an output feature map of the interpolation operation;for a pixel in the output feature map, computing a scaling coefficient based on the scale factor and a position of the pixel in the output feature map;generating a kernel from one or more scaling coefficients of one or more pixels in the output feature map;performing a depthwise convolution on the input feature map and the kernel; anddetermining a value of the pixel in the output feature map using an output of the depthwise convolution.2.The method of claim 1, wherein the interpolation operation is a bi-linear interpolation, and the scaling coefficient is a vector comprising two elements.3.The method of claim 1 or 2, wherein the interpolation operation is a bi-cubic interpolation, and the scaling coefficient is a vector comprising four elements.4.The method of any one of claims 1-3, wherein the kernel comprises weights, and values of at least some of the weights are values in the one or more scaling coefficients.5.The method of any one of claims 1-4, wherein the interpolation operation is a bi-linear interpolation, and a dimension of the kernel is two.6.The method of any one of claims 1-5, wherein the interpolation operation is a bi-cubic interpolation, and a dimension of the kernel is four.7.The method of any one of claims 1-6, wherein the output feature map has a number of channels, and a dimension of the kernel is the number or a multiple of the number.8.The method of any one of claims 1-7, further comprising:performing a subsequent depthwise convolution on the output of the depthwise convolution, wherein the output of the subsequent depthwise convolution is the output feature map of the interpolation operation.9.The method of claim 8, further comprising:determining another scale factor of the interpolation operation based on another dimension of an input feature map of the interpolation operation and another dimension of an output feature map of the interpolation operation,wherein the dimension of the input feature map and the dimension of the output feature map are widths, and the another dimension of the input feature map and the another dimension of the output feature map are heights, and the subsequent depthwise convolution is performed based on the another scale factor.10.The method of claim 9, wherein the position of the pixel is a position of the pixel in the dimension of the output feature map, and the method further comprises:for the pixel in the output feature map, computing another scaling coefficient based on the another scale factor and a position of the pixel in the another dimension of the output feature map; andgenerating a kernel for the subsequent depthwise convolution using the another scaling coefficient.11.One or more non-transitory computer-readable media storing instructions executable to perform operations for executing an interpolation operation in a neural network, the operations comprising:determining a scale factor of the interpolation operation based on a dimension of an input feature map of the interpolation operation and a dimension of an output feature map of the interpolation operation;for a pixel in the output feature map, computing a scaling coefficient based on the scale factor and a position of the pixel in the output feature map;generating a kernel from one or more scaling coefficients of one or more pixels in the output feature map;performing a depthwise convolution on the input feature map and the kernel; anddetermining a value of the pixel in the output feature map using an output of the depthwise convolution.12.The one or more non-transitory computer-readable media of claim 11, wherein the interpolation operation is a bi-linear interpolation, and the scaling coefficient is a vector comprising two elements.13.The one or more non-transitory computer-readable media of claim 11 or 12, wherein the interpolation operation is a bi-cubic interpolation, and the scaling coefficient is a vector comprising four elements.14.The one or more non-transitory computer-readable media of any one of claims 11-13, wherein the kernel comprises weights, and values of at least some of the weights are values in the one or more scaling coefficients.15.The one or more non-transitory computer-readable media of any one of claims 11-14, wherein the interpolation operation is a bi-linear interpolation, and a dimension of the kernel is two.16.The one or more non-transitory computer-readable media of any one of claims 11-15, wherein the interpolation operation is a bi-cubic interpolation, and a dimension of the kernel is four.17.The one or more non-transitory computer-readable media of any one of claims 11-16, wherein the output feature map has a number of channels, and a dimension of the kernel is the number or a multiple of the number.18.The one or more non-transitory computer-readable media of any one of claims 11-17, wherein the operations further comprise:performing a subsequent depthwise convolution on the output of the depthwise convolution, wherein the output of the subsequent depthwise convolution is the output feature map of the interpolation operation.19.The one or more non-transitory computer-readable media of claim 18, wherein the operations further comprise:determining another scale factor of the interpolation operation based on another dimension of an input feature map of the interpolation operation and another dimension of an output feature map of the interpolation operation,wherein the dimension of the input feature map and the dimension of the output feature map are widths, and the another dimension of the input feature map and the another dimension of the output feature map are heights, and the subsequent depthwise convolution is performed based on the another scale factor.20.The one or more non-transitory computer-readable media of claim 19, wherein the position of the pixel is a position of the pixel in the dimension of the output feature map, and the one or more non-transitory computer-readable media further comprises:for the pixel in the output feature map, computing another scaling coefficient based on the another scale factor and a position of the pixel in the another dimension of the output feature map; andgenerating a kernel for the subsequent depthwise convolution using the another scaling coefficient.21.An apparatus, comprising:a computer processor for executing computer program instructions; anda non-transitory computer-readable memory storing computer program instructions executable by the computer processor to perform operations for executing an interpolation operation in a neural network, the operations comprising:determining a scale factor of the interpolation operation based on a dimension of an input feature map of the interpolation operation and a dimension of an output feature map of the interpolation operation,for a pixel in the output feature map, computing a scaling coefficient based on the scale factor and a position of the pixel in the output feature map,generating a kernel from one or more scaling coefficients of one or more pixels in the output feature map,performing a depthwise convolution on the input feature map and the kernel, anddetermining a value of the pixel in the output feature map using an output of the depthwise convolution.22.The apparatus of claim 21, wherein the interpolation operation is a bi-linear interpolation, and the scaling coefficient is a vector comprising two elements.23.The apparatus of claim 21 or 22, wherein the interpolation operation is a bi-cubic interpolation, and the scaling coefficient is a vector comprising four elements.24.The apparatus of any one of claims 21-23, wherein the kernel comprises weights, and values of at least some of the weights are values in the one or more scaling coefficients.25.The apparatus of any one of claims 21-24, wherein the operations further comprise:performing a subsequent depthwise convolution on the output of the depthwise convolution, wherein the output of the subsequent depthwise convolution is the output feature map of the interpolation operation.
Citation Information
Patent Citations
Multilayer convolution neural network optimization system and method
CN105844653A
Method for processing image based on super-resolution
CN117934273A
Video Frame Interpolation Via Feature Pyramid Flows
US20220400226A1
Method and apparatus for efficient non-integer scaling in neural network accelerators
US20220405881A1