Patent Application
20230140173
This disclosure relates generally to neural networks, and more specifically, to DNN accelerators with heterogeneous tiling.
DNNs are used extensively for a variety of artificial intelligence 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 each inference can require hundreds of millions of MAC (multiply-accumulate) operations as well as hundreds of millions of weight operand weights to be stored for classification or detection. Therefore, techniques to improve efficiency of DNNs are needed.
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.
Overview
The last decade has witnessed a rapid rise in AI(artificial intelligence) based data processing, particularly based on DNN. 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. 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. However, to extend the lifetime of these edge devices these platforms usually leverage DNN accelerators as the execution platform that are optimized for high performance and low power consumption for DNN workloads. However, DNNs come in a variety of shape and sizes, and the main challenge for these DNN accelerators is to ensure high MAC utilization and efficiency across this myriad of DNN workloads that will result in highest throughput and the lowest energy consumption for these resource constrained platforms.
A DNN accelerator is typically organized in the form of an array of PEs, where each PE consists of one or more of MAC units. An array of PEs may form one tile in a DNN accelerator. Multiple tiles are usually required to meet the top requirement as well as to schedule and map multiple layers of the same or different networks on each of the tiles.
Conventional DNN accelerators are made up of multiple tiles of the same size (for example, 16×16 PEs). These DNN accelerators typically have PE arrays, in which the number of rows and the number of columns are multiples of 16. These in the form of rows and columns that are almost always a multiple of 16 often rely on the compiler to maximize the PE utilization, i.e., to minimize unused PEs. However, there is a maximum limit till which the compiler can improve the utilization without changing the tile dimension. With a fixed tile size, compiler support provides limited improvements in utilization across a variety of DNN workloads that come in different shapes and sizes. Even with techniques such as workload splitting or workload tiling, it is difficult to improve the utilization significantly.
Clock gating or power gating of unused PEs have also been explored as means of reducing the power consumption for underutilized PE arrays. Clock gating and power gating require circuit additions leading to area, power, and performance overheads. Some solutions choose to apply clock gates without power gating to the unutilized PEs. However, such solutions can result in less energy savings compared to clock gating due to the presence of leakage. Therefore, improvement technology for improving PE utilization in DNN accelerators is needed.
Embodiments of the present disclosure may improve on at least some of the challenges and issues described above by providing an DNN accelerator that includes an ensemble of heterogeneous tiles to improve the utilization of PEs in the DNN accelerator for various DNN workloads. Heterogeneous tiles are tiles having different sizes. A tile includes a PE array including PEs arranged in columns and rows. The size (or dimensions) of a tile is determined by the number of PE columns and the number of PE rows in the tile and may be represented by “the number of PE columns×the number of PE rows.” The DNN accelerator can select tiles and assign workloads of running various DNN models to the tiles based on the dimensions of the tile and characteristics of DNN layers.
Using convolutional layers in DNNs as example, the DNN accelerator may search a tile set including a set of heterogenous tiles for running convolutional operations in a DNN. The DNN accelerator may use characteristics of some or all convolutional layers in the DNN layers to search the tile set. The characteristics of a convolutional layer may include dimensions of an output tensor of the convolutional layer. A convolutional operation includes MAC operations on an input tensor and a group of filters, the result of which is an output tensor. The dimensions of the output tensor may include a first dimension (“OX”) indicating a number of elements in a row in the matrix, a second dimension (“OY”) indicating a number of elements in a column in the matrix, and a third dimension (“OC”) indicating a number of output channels in the set of output channels. The dimensions of the output tensor can be determined based on dimensions of the input tensor, dimensions of the filters, and the number of filters used in the convolutional operations.
Within the selected tile set, a tile may be selected for an individual convolutional layer in the DNN in a way to maximize utilization of PEs in the tile set. The DNN accelerator may select the tile based on dimensions of the output tensor of the convolutional layer and the size of the tile. The DNN accelerator may select multiple tiles for a single layer or select a single tile for multiple layers. After the tile is selected, the DNN accelerator can distribute portions of the workload for running the convolutional operation of the layer to individual PEs through a partition of the output tensor. The DNN accelerator can partition the output tensor into segments based on dimensions of the PE array. For instance, the DNN accelerator may partition the output tensor in the OX and OY dimensions based on the number of PE rows in the tile and partition the output tensor in the OC dimension based on the number of PE columns in the tile. Further, the DNN accelerator can assign workloads of generating the output tensor segments to some or all of the PEs in the PE array. A PE can receive a workload of generating a respective output tensor segment and to perform a MAC operation for generating the respective output tensor segment.
Different from conventional DNN accelerators with homogeneous tiles (e.g., tiles having the same dimensions), the present disclosure provides a different type of DNN accelerator that includes homogeneous tiles. The DNN accelerator can also search for a tile set for a DNN, select tiles for layers in the DNN, and distribute workload to individual PEs in a way to maximize utilization of PEs in the tile set. Compared with conventional DNN accelerators, the DNN accelerator in the present disclosure can provide improvement in overall utilization of PEs. Such improvement is available in embodiments where a DNN layer is split over multiple tiles and embodiments where a DNN layer is run by a single tile.
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.
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.
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.
For the purposes of the present disclosure, the phrase “A and/or B” means (A), (B), or (A and B). For the purposes of the present disclosure, 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.
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 indicate 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.
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.
The terms “substantially,” “close,” “approximately,” “near,” and “about,” generally refer to being within +/−20% of a target value based on the input operand of a particular 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 based on the input operand of a particular value as described herein or as known in the art.
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.”
The DNN 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.
Example DNN
The convolutional layers 110 summarize the presence of features in the input image 105. 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 input feature map (IFM) 140) and a filter 150. As shown in
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 output feature map (OFM) 160). The OFM 160 is represented by a 5×5 2D array. The 5×5 2D array includes 5 output elements (also referred to as output points) in each row and 5 output elements in each column. For purpose of illustration, the standard convolution includes one filter in the embodiments of
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 array of output elements. As such, the 2D output array (i.e., the OFM 160) from the standard convolution 163 is referred to an OFM.
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
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 the rectified linear activation function (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 calculates 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 performs a convolution on the OFM 160 with new kernels and generates 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.
In some embodiments, a convolutional layer 110 has 4 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.
The pooling layers 120 down-sample feature maps generated by the convolutional layers, e.g., by summarizing the presents of features in the patches of the feature maps. A pooling layer 120 is placed between 2 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) has been applied to the OFM 160.
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 2 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.
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 receives 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 applies a linear combination and an activation function to the input operand and generates an individual partial sum. The individual partial sum 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 classify the input image 105 and returns an operand of size N, where N is the number of classes in the image classification problem. In the embodiments of
Example DNN Accelerator
The tile sets 210 include PEs that can run DNN models and function as neurons or nodes of DNNs. A tile set includes multiple PE arrays. A PE array includes PEs arranged in columns and rows. A PE may be a node of a DNN. The PE array may have a size indicating the number of columns, the number of rows, or a combination of both. For instance, for a PE array including 16 columns and 16 rows, the size of the PE array may be represented by “16×16” or “256.” The PE arrays in the same tile set 210 can have different sizes. A tile set 210 including PE arrays with different sizes is referred to as a heterogeneous tile. The tile sets 210 may be different combinations of PE arrays. For example, a tile set 210 may include at least one PE array that is different from all the PE arrays in another tile set 210. As another example, even though 2 tile sets 210 have the same PE arrays, the PE arrays may be arranged differently, e.g., locations of the PE arrays may be different.
In some embodiments, one or more tile sets 210 may be used for a single DNN. Within a tile set 210, one or more tiles may be used for a single layer, e.g., a convolutional layer. An example convolution layer is a convolutional layer 110 in
The workload manager 220 manages workloads of the tile sets 210. In some embodiments, the workload manager 220 manages workloads of the tile sets 210 in a way to maximize utilization of PEs in one or more tiles or one or more tile sets 210. The utilization of PEs in a tile or tile set may be measured based on a ratio of the number of active PEs to the total number of PEs in the tile or tile set 210. An active PE is a PE that performs one or more MAC operations during the execution of an DNN model. The workload manager 220 includes a tile set search module 240, a tile selection module 250, and a partitioning module 260. In other embodiments, alternative configurations, different or additional components may be included in the DNN accelerator 200. Further, functionality attributed to a component of the DNN accelerator 200 may be accomplished by a different component included in the DNN accelerator 200 or by a different system.
The tile set search module 240 searches a tile set 210 for a DNN model from the tile sets 210. For instance, the tile set search module 240 selects a tile set 210, that when running the DNN model, can achieve higher utilization of PEs than other tile sets 210. In some embodiments, the tile set search module 240 selects a tile set 210 for a DNN based on dimensions of output tensors of convolutional layers in the DNN. An example output tensor is the OFM 160 in
The tile set search module 240 may use a subset of all convolutional layers in the DNN to select a tile set 210 for the DNN. For instance, the tile set search module 240 may determine 2 factors for each convolutional layer in the DNN. The 2 factors include a first factor equal a product of multiplying OX with OY, which is denoted as OXOY, and a second factor equal OC. The tile set search module 240 may determine a first condition for the first factor, a second condition for the second factor, or a third condition for a combination of the first factor and the second factor. The combination may be a product of the first factor and the second factor. A condition may be a predetermined value range, e.g., 8 to 32, 8 to 64, 128 to 512, and so on. Then the tile set search module 240 can identify convolutional layers that can meet the first condition, the second condition, or both conditions. A condition is met if the value of the corresponding factor falls under the range specified in the condition.
After the tile set search module 240 identifies the convolutional layers in the subset, the tile set search module 240 may rank the convolutional layers in the subset. In an example, the tile set search module 240 finds the most common value of the third factor among all the convolutional layers in the subset. For instance, the tile set search module 240 may determine a frequency of a value of the third factor in the subset. The frequency indicating the number of convolutional layers whose third factors have the value. The tile set search module 240 may use the value of the third factor that has the highest frequency in the subset to determine which tile set 210 can achieve the highest PE utilization and then use the tile set 210 to run the DNN.
The tile selection module 250 selects a tile from a tile set 210 for one or more convolutional layers of a DNN. In some embodiments, a tile may perform a convolutional operation of a single convolution layer at a time. In other embodiments, a tile may perform convolutional operations of 2 or more convolution layers at a time. To select a tile for a convolutional layer, the tile selection module 250 may compare dimensions of the output tensor of the convolutional layer with dimensions of each tile in the tile set 210. For instance, the tile selection module 250 may compare the number of PE columns in each tile with OC of the output tensor and compare the number of PE rows in each tile with OXOY of the output tensor.
In some embodiments (such as embodiments where the convolutional operation can be executed by one tile), the tile selection module 250 may determine a first difference between the number of PE columns and OC of the output tensor and determines a second difference between the number of PE rows and OXOY. The tile selection module 250 may determine an aggregated difference for each tile by aggregating the first difference and second difference of the tile. The tile selection module 250 may select the tile having the smallest aggregated difference as the tile for the convolutional layer.
In other embodiments (such as embodiments where the convolutional operation needs multiple tiles, e.g., a subset of tiles in the tile set 210), the tile selection module 250 may determine a first difference between the total number of PE columns in the subset and OC of the output tensor and determines a second difference between the total number of PE rows in the subset and OXOY. The tile selection module 250 may determine an aggregated difference for each subset by aggregating the first difference and second difference of the subset. The tile selection module 250 may select the tiles in a subset having the smallest aggregated difference as the tiles for the convolutional layer.
The partitioning module 260 partitions the workload of a tile for a convolutional operation into workloads of individual PEs in the tile based on dimensions of an output tensor and assigns the workloads to individual PEs. As part of the mapping of workload on the PE array, the mapping of the dimensions OX, OY, and OC of the output tensor on the PE array can have a significant impact on the PE utilization (and hence performance) of the DNN accelerator 200. The partitioning module 260 can partition the output tensor to segments map the output tensor segments to individual PEs. The partitioning may be done for each layer of the DNN. The partitioning may determine the active PEs in the PE array while running the convolutional operation of the layer and thereby can determine the performance and power for the layer.
In some embodiments, the partitioning module 260 partitions the output tensor based on dimensions of the PE array. In some embodiments, the partitioning module 260 partitions OX and OY of the output tensor based on the number of PE rows in the tile, and partitions OC of the output tensor based on the number of PE columns in the tile. For instance, the partitioning module 260 may determine the OX and OY for each output tensor segment based on the number of PE rows and determine the OC of each output tensor segment based on the number of PE columns. Then the partitioning module 260 can identify the portions of the input tensor and filter for computing each output tensor segments, and transmit the portions of the input tensor and filter to the individual PEs for running MAC operations to computer the output tensor segments. Certain aspects of the workload manager 220 are described below in conjunction with
The memory 230 stores data associated with the DNN accelerator 200, such as data used by the DNN accelerator 200 for deep learning, data generated by the DNN accelerator 200, or data otherwise associated with the DNN accelerator 200. In some embodiments, the DNN accelerator 200 may be associated with multiple memories. In some embodiments, the memory 230 stores data associated with MAC operations. For instance, the memory stores some or all of the input, filters, and output of a DNN layer. In some embodiments, the memory 230 is a random-access memory (RAM), such as a static RAM (SRAM). The memory 230 may be byte-addressable, and each memory address identifies a single byte (8 bits) of storage. The memory 230 includes a plurality of storage units, each of which stores a single byte and has a memory address. Data larger than a single byte may be stored in storage units with consecutive memory addresses, i.e., adjacent storage units. For instance, 2 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 memory 230 in a single reading cycle. In other embodiments, 16 bits can be transferred from the memory 230 in multiple reading cycles, such as 2 cycles.
The tile set 300 is a set of heterogeneous PE array (also referred to as “heterogeneous tile set”), meaning at least one of the PE arrays 310, 320, 330, and 340 has at least one different dimension from the other PE arrays in the tile set 300. In an example, the PE array 310 may have more (or fewer) PE rows or columns than other PE arrays. Some of the PE arrays 310, 320, 330, and 340 may have same dimensions. For instance, 2 or 3 of the PE arrays 310, 320, 330, and 340 may have the same number of PE rows or PE columns. In some embodiments, each PE array has a different number of PE rows or PE columns from any of the other PE arrays. More information regarding PE array are described below in conjunction with
Each PE 410 performs an MAC operation on the input signals 450 and 460 and outputs the output signal 470, which is a result of the MAC operation. Some or all of the input signals 450 and 460 and the output signal 470 may be in an integer format, such as INT8, or FP format, such as FP16 or BF16. For purpose of simplicity and illustration, the input signals and output signal of all the PEs 410 have the same reference numbers, but the PEs 410 may receive different input signals and output different output signals from each other. Also, a PE 410 may be different from another PE 410, e.g., including more, fewer, or different components.
As shown in
In the embodiments of
As shown in
The input register file 540 temporarily stores input signals (e.g., contexts) received by the PE 410. The input feature data may include input feature data and output signals from other PEs 510. The weight register file 550 temporarily stores weights received by the PE 410. The output register file 560 temporarily stores output signals generated by the PE 410. For purpose of illustration and simplicity, the PE 410 in
The MAC unit 570 performs MAC operations on data in the input register file 540 and weight register file 550. The MAC unit 570 includes a multiply unit 580 and an accumulate unit 590. The multiply unit 580 performs multiply operations on input feature data in the input register file 540 and weights in the weight register file 550. The amount of time needed by the multiply unit 580 for a multiple operation depends on the sparsity level of the weights used in the multiple operation. If the weights are denser (i.e., the sparsity level is lower), the multiply unit 580 needs more time to perform the multiple operation. The accumulate unit 590 performs accumulate operations on the output of the multiply unit 580 and outputs signals from other PEs. The output of the accumulate unit 590 is the output signal of the PE 410. More details regarding MAC operations in PE are described below in conjunction with
Example MAC Operations
In the integer MAC operation 600, the bits in the input register file 617 and the weight register file 627 are fed sequentially into a multiplier 630, where the multiplier 630 performs a series of multiplication operations. Each multiplication operation is with a bit from the input register file 617 and a bit from the weight register file 627. The results of the multiplication operations are fed into an accumulator 635, which generates an individual partial sum of the PE. The individual partial sum of the PE can be stored in the output register file 640. The series of multiplication operations by the multiplier 630 and the accumulation operation by the accumulator 635 may constitute a MAC operation by the PE. The multiplier 630 and the accumulator 635 may operate with various integer formats or fixed-point formats.
In the embodiments of
The bits in the input register file 717 and the weight register file 727 are fed sequentially into a multiplier 730, where the multiplier 730 performs a series of multiplication operations. Each multiplication operation is with a bit from the input register file 717 and a bit from the weight register file 727. The results of the multiplication operations are fed into an accumulator 735, which generates an individual partial sum of the PE. The individual partial sum of the PE can be stored in the output register file 740. The series of multiplication operations by the multiplier 730 and the accumulation operation by the accumulator 735 may constitute an floating-point MAC operation by the PE. The accumulator 735 may operate with a different floating-point bit precision from the multiplier 730. In an example, the multiplier 730 performs multiplications with FP16 or BF16 format, but the accumulator 735 performs accumulations with FP32 format.
Example Convolution Workload
The convolution operation has an output tensor 820 having dimensions OX, OY, and OC. The output tensor 820 is represented by cuboid in the coordinate system in
In the embodiments of
The partitioning module 260 may determine OC of the segment 825 based on the number of PE columns (i.e., 16) of the PE array 810 and the OC (i.e., 256) of the output tensor 820, e.g., the OC of the segment 825 is a result of dividing the OC of the output tensor 820 by the number of PE columns in the PE array 810. The partitioning module 260 may determine OX and OY of the segment 825 based on the number of PE rows (i.e., 16) of the PE array 810 and the OX (i.e., 14) and OY (i.e., 14) of the output tensor 820. To determine the OX and OY of the segment 825, the partitioning module 260 may determine 2 integer numbers, the product of which is no larger than the number of PE rows and use the 2 integer numbers as the OX and OY of the segment 825. IN some embodiments, the partitioning module 260 may select an integer number that is a divisor of the OX of the output tensor 820 as the OX of the segment 825, select an integer number that is a divisor of the OY of the output tensor 820 as the OY of the segment 825, or both. In an embodiment, the segment 825 has dimensions of OX=7, OY=2, and OC=16. In other embodiments, the segment 825 may have different dimensions.
Example Homogeneous Tile Set
Such a homogeneous tile set, when used to execute a DNN model, may not result in the most optimal PE utilization and performance for the DNN accelerator. Different workloads can be assigned to the PE arrays 910, 920, 930, and 940. Usually, the tile set 900 is selected for a DNN based on the heaviest workload in the DNN. For instance, each PE array may be used for a different convolutional layer, and the tile set 900 is selected so that a PE array can execute the convolutional operation of a layer having a biggest output tensor. The PE array receiving the heaviest workload (e.g., the PE array 910) may achieve the best PE utilization, i.e., the ratio of the number of active PEs to the number of all PEs in the PE array is the highest among the PE arrays 910, 920, 930, and 940. However, as different layers have different workloads, the other PE arrays 920, 930, and 940 may receive smaller workloads and the PE utilization would be lower.
As shown in
Example Heterogeneous Tile Sets
The heterogeneous tile sets 1010, 1020, and 1030 may be part of a DNN accelerator, such as the DNN accelerator in
Example Method of Deep Learning
The workload manager 220 identifies 1110 a tile set for executing tensor operations in a DNN. An example of the DNN is the DNN 100 in
In some embodiments, the workload manager 220 may select the tile set from a plurality of tile sets. The plurality of tile sets are combinations of different PE arrays. For instance, the workload manager 220 may determine dimensions of output tensors of a plurality of convolutional layers in the DNN. The workload manager 220 may identify a set of dimensions from the dimensions of the output tensors. The set of dimensions are dimensions of output tensors of multiple convolutional layers of the plurality of convolutional layers. The workload manager 220 may identify the tile set from a plurality of tile sets based on the set of dimensions. In some embodiments, an input tensor of a convolutional layer has a set of dimensions the set of dimensions, and the set of dimensions identified by the workload manager 220 has a higher frequency than any other set of dimensions in the DNN. For instance, the identified set of dimensions are dimensions of output tensors of more convolutional layers than any other sets of dimensions.
Then the workload manager 220 classifies the convolutional layers in the subset into a plurality of groups. Each group includes one or more convolution layers with output tensors having same dimensions. The workload manager 220 can rank the groups based on numbers of convolutional layers in the groups and select a group from the groups based on the ranking. The workload manager 220 can further identify the tile set, based on dimensions of a convolutional layer in the group, the tile set from a plurality of tile sets. Each of the plurality of tile sets is a combination of different PE arrays.
The workload manager 220 selects 1120 a PE array from the plurality of PE arrays for a convolutional layer in the DNN. The PE array can a part or a whole convolution operation in the convolutional layer. In some embodiments workload manager 220 selects a group of PE arrays from the plurality of PE arrays for the convolutional layer in the DNN. The group of PE arrays includes the PE array. Each PE array in the group may perform a portion of the convolutional operation. The workload manager 220 may assign workloads for different portions of the convolutional operation to the PE arrays. Each PE array may receive the workload of a different portion of the convolutional operation.
The workload manager 220 determines 1130 dimensions of an output tensor of the convolutional layer. The output tensor is a result of a convolutional operation to be performed by the PE array on an input tensor and a filter. In some embodiments, the output tensor includes a set of output channels. Each channel includes a matrix. The dimensions of the output tensor include a first dimension indicating a number of elements in a row in the matrix, a second dimension indicating a number of elements in a column in the matrix, and a third dimension indicating a number of output channels in the set of output channels. The workload manager 220 may determine the dimensions of the output tensor based on dimensions of the input tensor, a number of kernels in the filter, and dimensions of the kernels.
The workload manager 220 partitions 1140 the output tensor into output tensor segments based on a size of the PE array. In some embodiments, the workload manager 220 may determine a fourth dimension and a fifth dimension of each output tensor segment based on the first number. The workload manager 220 may also determine a sixth dimension based on the second number. The fourth dimension indicates a number of elements in a row in the matrix. The fifth dimension indicates a number of elements in a column in the matrix. The sixth dimension indicates a number of output channels in the set of output channels
The workload manager 220 assigns 1150 workloads of generating the output tensor segments to a group of PEs in the PE array. Each PE in the group is to receive a workload of generating a respective output tensor segment and to perform a MAC operation for generating the respective output tensor segment. In some embodiments, for a workload of generating an output tensor segment, the workload manager 220 may identify a segment of the input tensor and a segment of the filter. The workload manager 220 may also transmit the segment of the input tensor and the segment of the filter into a PE in the group. The PE is to perform one or more MAC operations on the segment of the input tensor and the segment of the filter and to output the output tensor segment. The PE may include an input register file for storing the segment of the input tensor, a weight register file for storing the segment of the filter, an output register file for storing the output tensor segment, and a MAC unit for performing the one or more MAC operations. The input tensor may include one or more integer values or one or more floating-point values. A MAC operation may be an integer MAC operation (e.g., the integer MAC operation in
Example Deep Learning Environment
The deep learning server 1210 trains deep learning models using neural networks. A neural network is structured like the human brain and consists of artificial neurons, also known as nodes. These nodes are stacked next to each other in 3 types of layers: input layer, hidden layer(s), and output layer. Data provides each node with information in the form of inputs. The node multiplies the inputs with random weights, calculates them, and adds a bias. Finally, nonlinear functions, also known as activation functions, are applied to determine which neuron to fire. The deep learning server 1210 can use various types of neural networks, such as DNN, recurrent neural network (RNN), generative adversarial network (GAN), long short-term memory network (LSTMN), and so on. During the process of training the deep learning models, the neural networks use unknown elements in the input distribution to extract features, group objects, and discover useful data patterns. The deep learning models can be used to solve various problems, e.g., making predictions, classifying images, and so on. The deep learning server 1210 may build deep learning models specific to particular types of problems that need to be solved. A deep learning model is trained to receive an input and outputs the solution to the particular problem.
In
The database 1250 stores data received, used, generated, or otherwise associated with the deep learning server 1210. For example, the database 1250 stores a training dataset that the DNN system 1240 uses to train DNNs. In an embodiment, the training dataset is an image gallery that can be used to train a DNN for classifying images. The training dataset may include data received from the client devices 1220. As another example, the database 1250 stores hyperparameters of the neural networks built by the deep learning server 1210.
The distributer 1260 distributes deep learning models generated by the deep learning server 1210 to the client devices 1220. In some embodiments, the distributer 1260 receives a request for a DNN from a client device 1220 through the network 1230. The request may include a description of a problem that the client device 1220 needs to solve. The request may also include information of the client device 1220, such as information describing available computing resource on the client device. The information describing available computing resource on the client device 1220 can be information indicating network bandwidth, information indicating available memory size, information indicating processing power of the client device 1220, and so on. In an embodiment, the distributer may instruct the DNN system 1240 to generate a DNN in accordance with the request. The DNN system 1240 may generate a DNN based on the information in the request. For instance, the DNN system 1240 can determine the structure of the DNN and/or train the DNN in accordance with the request.
In another embodiment, the distributer 1260 may select the DNN from a group of pre-existing DNNs based on the request. The distributer 1260 may select a DNN for a particular client device 1220 based on the size of the DNN and available resources of the client device 1220. In embodiments where the distributer 1260 determines that the client device 1220 has limited memory or processing power, the distributer 1260 may select a compressed DNN for the client device 1220, as opposed to an uncompressed DNN that has a larger size. The distributer 1260 then transmits the DNN generated or selected for the client device 1220 to the client device 1220.
In some embodiments, the distributer 1260 may receive feedback from the client device 1220. For example, the distributer 1260 receives new training data from the client device 1220 and may send the new training data to the DNN system 1240 for further training the DNN. As another example, the feedback includes an update of the available computer resource on the client device 1220. The distributer 1260 may send a different DNN to the client device 1220 based on the update. For instance, after receiving the feedback indicating that the computing resources of the client device 1220 have been reduced, the distributer 1260 sends a DNN of a smaller size to the client device 1220.
The client devices 1220 receive DNNs from the distributer 1260 and applies the DNNs to perform machine learning tasks, e.g., to solve problems or answer questions. In various embodiments, the client devices 1220 input images into the DNNs and uses the output of the DNNs for various applications, e.g., visual reconstruction, augmented reality, robot localization and navigation, medical diagnosis, weather prediction, and so on. A client device 1220 may be one or more computing devices capable of receiving user input as well as transmitting and/or receiving data via the network 1230. In one embodiment, a client device 1220 is a conventional computer system, such as a desktop or a laptop computer. Alternatively, a client device 1220 may be a device having computer functionality, such as a personal digital assistant (PDA), a mobile telephone, a smartphone, an autonomous vehicle, or another suitable device. A client device 1220 is configured to communicate via the network 1230. In one embodiment, a client device 1220 executes an application allowing a user of the client device 1220 to interact with the deep learning server 1210 (e.g., the distributer 1260 of the deep learning server 1210). The client device 1220 may request DNNs or send feedback to the distributer 1260 through the application. For example, a client device 1220 executes a browser application to enable interaction between the client device 1220 and the deep learning server 1210 via the network 1230. In another embodiment, a client device 1220 interacts with the deep learning server 1210 through an application programming interface (API) running on a native operating system of the client device 1220, such as IOS® or ANDROID™.
In an embodiment, a client device 1220 is an integrated computing device that operates as a standalone network-enabled device. For example, the client device 1220 includes display, speakers, microphone, camera, and input device. In another embodiment, a client device 1220 is a computing device for coupling to an external media device such as a television or other external display and/or audio output system. In this embodiment, the client device 1220 may couple to the external media device via a wireless interface or wired interface (e.g., an HDMI (High-Definition Multimedia Interface) cable) and may utilize various functions of the external media device such as its display, speakers, microphone, camera, and input devices. Here, the client device 1220 may be configured to be compatible with a generic external media device that does not have specialized software, firmware, or hardware specifically for interacting with the client device 1220.
The network 1230 supports communications between the deep learning server 1210 and client devices 1220. The network 1230 may comprise any combination of local area and/or wide area networks, using both wired and/or wireless communication systems. In one embodiment, the network 1230 may use standard communications technologies and/or protocols. For example, the network 1230 may include communication links using technologies such as Ethernet, 12010.11, worldwide interoperability for microwave access (WiMAX), 3G, 4G, code division multiple access (CDMA), digital subscriber line (DSL), etc. Examples of networking protocols used for communicating via the network 1230 may include multiprotocol label switching (MPLS), transmission control protocol/Internet protocol (TCP/IP), hypertext transport protocol (HTTP), simple mail transfer protocol (SMTP), and file transfer protocol (FTP). Data exchanged over the network 1230 may be represented using any suitable format, such as hypertext markup language (HTML) or extensible markup language (XML). In some embodiments, all or some of the communication links of the network 1230 may be encrypted using any suitable technique or techniques.
Example DNN System
The interface module 1310 facilitates communications of the DNN system 1300 with other systems. For example, the interface module 1310 establishes communications between the DNN system 1300 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 1310 supports the DNN system 1300 to distribute DNNs to other systems, e.g., computing devices configured to apply DNNs to perform tasks.
The training module 1320 trains DNNs by using a training dataset. The training module 1320 forms the training dataset. In an embodiment where the training module 1320 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 validation module 1330 to validate performance of a trained DNN. The portion of the training dataset not including the tuning subset and the validation subset may be used to train the DNN.
The training module 1320 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 13, 130, 500, 1300, or even larger.
The training module 1320 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 2 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 category by training.
In the process of defining the architecture of the DNN, the training module 1320 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 rectified linear unit activation function, a tangent activation function, or other types of activation functions.
After the training module 1320 defines the architecture of the DNN, the training module 1320 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 1320 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 1320 uses a cost function to minimize the error.
The training module 1320 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 1320 finishes the predetermined number of epochs, the training module 1320 may stop updating the parameters in the DNN. The DNN having the updated parameters is referred to as a trained DNN.
The validation module 1330 verifies accuracy of trained DNNs. In some embodiments, the validation module 1330 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 validation module 1330 determines may determine an accuracy score measuring the precision, recall, or a combination of precision and recall of the DNN. The validation module 1330 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 reference classification model correctly predicted (TP or true positives) out of the total it predicted (TP+FP or false positives), and recall may be how many the reference classification model 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.
The validation module 1330 may compare the accuracy score with a threshold score. In an example where the validation module 1330 determines that the accuracy score of the augmented model is lower than the threshold score, the validation module 1330 instructs the training module 1320 to re-train the DNN. In one embodiment, the training module 1320 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.
The inference module 1340 applies the trained or validated DNN to perform tasks. For instance, the inference module 1340 inputs images into the DNN. The DNN outputs classifications of objects in the images. As an example, the DNN may be provisioned in a security setting to detect malicious or hazardous objects in images captured by security cameras. As another example, the DNN may be provisioned to detect objects (e.g., road signs, hazards, humans, pets, etc.) in images captured by cameras of an autonomous vehicle. The input to the DNN may be formatted according to a predefined input structure mirroring the way that the training dataset was provided to the DNN. The DNN may generate an output structure which may be, for example, a classification of the image, a listing of detected objects, a boundary of detected objects, or the like. In some embodiments, the inference module 1340 distributes the DNN to other systems, e.g., computing devices in communication with the DNN system 1300, for the other systems to apply the DNN to perform the tasks.
The memory 1350 stores data received, generated, used, or otherwise associated with the DNN system 1300. For example, the memory 1350 stores the datasets used by the training module 1320 and validation module 1330. The memory 1350 may also store data generated by the training module 1320 and validation module 1330, such as the hyperparameters for training DNNs, internal parameters of trained DNNs (e.g., values of tunable parameters of FALUs), etc. In the embodiment of
Example Computing Device
The computing device 1400 may include a processing device 1402 (e.g., one or more processing devices). The processing device 1402 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 1400 may include a memory 1404, 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 1404 may include memory that shares a die with the processing device 1402. In some embodiments, the memory 1404 includes one or more non-transitory computer-readable media storing instructions executable to perform operations for deep learning, e.g., the method 1100 described above in conjunction with
In some embodiments, the computing device 1400 may include a communication chip 1412 (e.g., one or more communication chips). For example, the communication chip 1412 may be configured for managing wireless communications for the transfer of data to and from the computing device 1400. 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.
The communication chip 1412 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 1412 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 1412 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 1412 may operate in accordance with 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 1412 may operate in accordance with other wireless protocols in other embodiments. The computing device 1400 may include an antenna 1422 to facilitate wireless communications and/or to receive other wireless communications (such as AM or FM radio transmissions).
In some embodiments, the communication chip 1412 may manage wired communications, such as electrical, optical, or any other suitable communication protocols (e.g., the Ethernet). As noted above, the communication chip 1412 may include multiple communication chips. For instance, a first communication chip 1412 may be dedicated to shorter-range wireless communications such as Wi-Fi or Bluetooth, and a second communication chip 1412 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 1412 may be dedicated to wireless communications, and a second communication chip 1412 may be dedicated to wired communications.
The computing device 1400 may include battery/power circuitry 1414. The battery/power circuitry 1414 may include one or more energy storage devices (e.g., batteries or capacitors) and/or circuitry for coupling components of the computing device 1400 to an energy source separate from the computing device 1400 (e.g., AC line power).
The computing device 1400 may include a display device 1406 (or corresponding interface circuitry, as discussed above). The display device 1406 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.
The computing device 1400 may include an audio output device 1408 (or corresponding interface circuitry, as discussed above). The audio output device 1408 may include any device that generates an audible indicator, such as speakers, headsets, or earbuds, for example.
The computing device 1400 may include an audio input device 1418 (or corresponding interface circuitry, as discussed above). The audio input device 1418 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).
The computing device 1400 may include a GPS device 1416 (or corresponding interface circuitry, as discussed above). The GPS device 1416 may be in communication with a satellite-based system and may receive a location of the computing device 1400, as known in the art.
The computing device 1400 may include an other output device 1410 (or corresponding interface circuitry, as discussed above). Examples of the other output device 1410 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.
The computing device 1400 may include an other input device 1420 (or corresponding interface circuitry, as discussed above). Examples of the other input device 1420 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 (register filelD) reader.
The computing device 1400 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 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 1400 may be any other electronic device that processes data.
Select Examples
The following paragraphs provide various examples of the embodiments disclosed herein.
Example 1 provides a method of deep learning, the method including identifying a tile set for executing tensor operations in a DNN, the tile set including a plurality of PE arrays having different sizes, each PE array including PEs arranged in a first number of columns and a second number of rows and having a size determined by the first number and the second number; selecting a PE array from the plurality of PE arrays for a convolutional layer in the DNN; determining dimensions of an output tensor of the convolutional layer, the output tensor being a result of a convolutional operation to be performed by the PE array on an input tensor and a filter; partitioning the output tensor into output tensor segments based on a size of the PE array; and assigning workloads of generating the output tensor segments to a group of PEs in the PE array, where each PE in the group is to receive a workload of generating a respective output tensor segment and to perform a multiply-accumulation (MAC) operation for generating the respective output tensor segment.
Example 2 provides the method of example 1, where identifying the tile set for executing the tensor operations in the DNN includes determining dimensions of output tensors of a plurality of convolutional layers in the DNN; identifying a set of dimensions from the dimensions of the output tensors, wherein the set of dimensions are dimensions of output tensors of multiple convolutional layers of the plurality of convolutional layers; and identifying the tile set from a plurality of tile sets based on the set of dimensions, where each of the plurality of tile sets is a combination of different PE arrays.
Example 3 provides the method of example 2, where identifying the tile set for executing the tensor operations in the DNN further includes identifying the plurality of convolutional layers from all convolutional layers in the DNN, where dimensions of the plurality of convolutional layers are within one or more predetermined dimension ranges.
Example 4 The method of any of the preceding examples, where selecting the PE array from the plurality of PE arrays for the convolutional layer in the DNN includes selecting a group of PE arrays from the plurality of PE arrays for the convolutional layer in the DNN, where the group of PE arrays includes the PE array.
Example 5 provides the method of any of the preceding examples, where determining dimensions of the output tensor of the convolutional layer includes determining the dimensions of the output tensor based on dimensions of the input tensor, a number of kernels in the filter, and dimensions of the kernels.
Example 6 provides the method of any of the preceding examples, where the output tensor includes a set of output channels, each output channel including a matrix, and the dimensions of the output tensor include a first dimension indicating a number of elements in a row in the matrix, a second dimension indicating a number of elements in a column in the matrix, and a third dimension indicating a number of output channels in the set of output channels.
Example 7 provides the method of example 6, where partitioning the output tensor into output tensor segments based on a size of the PE array includes determining a fourth dimension and a fifth dimension of each output tensor segment based on the first number; and determining a sixth dimension based on the second number, where the fourth dimension indicates a number of elements in a row in the matrix, the fifth dimension indicates a number of elements in a column in the matrix, and the sixth dimension indicates a number of output channels in the set of output channels.
Example 8 provides the method of any of the preceding examples, where assigning the workloads of generate the output tensor segments to the group of PEs in the PE array includes for a workload of generating an output tensor segment, identifying a segment of the input tensor and a segment of the filter; and transmitting the segment of the input tensor and the segment of the filter into a PE in the group, where the PE is to perform one or more MAC operations on the segment of the input tensor and the segment of the filter and to output the output tensor segment.
Example 9 provides the method of example 8, where the PE includes an input register file for storing the segment of the input tensor; a weight register file for storing the segment of the filter; an output register file for storing the output tensor segment; and a MAC unit for performing the one or more MAC operations.
Example 10 provides the method of example 8 or 9, where the input tensor includes one or more integer values or one or more floating-point values.
Example 11 provides one or more non-transitory computer-readable media storing instructions executable to perform operations for deep learning, the operations including identifying a tile set for executing tensor operations in a DNN, the tile set including a plurality of PE arrays having different sizes, each PE array including PEs arranged in a first number of columns and a second number of rows and having a size determined by the first number and the second number; selecting a PE array from the plurality of PE arrays for a convolutional layer in the DNN; determining dimensions of an output tensor of the convolutional layer, the output tensor being a result of a convolutional operation to be performed by the PE array on an input tensor and a filter; partitioning the output tensor into output tensor segments based on a size of the PE array; and assigning workloads of generating the output tensor segments to a group of PEs in the PE array, where each PE in the group is to receive a workload of generating a respective output tensor segment and to perform a multiply-accumulation (MAC) operation for generating the respective output tensor segment.
Example 12 provides the one or more non-transitory computer-readable media of example 11, where identifying the tile set for executing the tensor operations in the DNN includes determining dimensions of output tensors of a plurality of convolutional layers in the DNN; identifying a set of dimensions from the dimensions of the output tensors, wherein the set of dimensions are dimensions of output tensors of multiple convolutional layers of the plurality of convolutional layers; and identifying the tile set from a plurality of tile sets based on the set of dimensions, where each of the plurality of tile sets is a combination of different PE arrays.
Example 13 provides the one or more non-transitory computer-readable media of example 12, where identifying the tile set for executing the tensor operations in the DNN further includes identifying the plurality of convolutional layers from all convolutional layers in the DNN, where dimensions of the plurality of convolutional layers are within one or more predetermined dimension ranges.
Example 14 provides the one or more non-transitory computer-readable media of any one of examples 11-13, where selecting the PE array from the plurality of PE arrays for the convolutional layer in the DNN includes selecting a group of PE arrays from the plurality of PE arrays for the convolutional layer in the DNN, where the group of PE arrays includes the PE array.
Example 15 provides the one or more non-transitory computer-readable media of any one of examples 11-14, where determining dimensions of the output tensor of the convolutional layer includes determining the dimensions of the output tensor based on dimensions of the input tensor, a number of kernels in the filter, and dimensions of the kernels.
Example 16 provides the one or more non-transitory computer-readable media of any one of examples 11-15, where the output tensor includes a set of output channels, each output channel including a matrix, and the dimensions of the output tensor include a first dimension indicating a number of elements in a row in the matrix, a second dimension indicating a number of elements in a column in the matrix, and a third dimension indicating a number of output channels in the set of output channels.
Example 17 provides the one or more non-transitory computer-readable media of example 16, where partitioning the output tensor into output tensor segments based on a size of the PE array includes determining a fourth dimension and a fifth dimension of each output tensor segment based on the first number; and determining a sixth dimension based on the second number, where the fourth dimension indicates a number of elements in a row in the matrix, the fifth dimension indicates a number of elements in a column in the matrix, and the sixth dimension indicates a number of output channels in the set of output channels.
Example 18 provides the one or more non-transitory computer-readable media of any one of examples 11-17, where assigning the workloads of generate the output tensor segments to the group of PEs in the PE array includes for a workload of generating an output tensor segment, identifying a segment of the input tensor and a segment of the filter; and transmitting the segment of the input tensor and the segment of the filter into a PE in the group, where the PE is to perform one or more MAC operations on the segment of the input tensor and the segment of the filter and to output the output tensor segment.
Example 19 provides the one or more non-transitory computer-readable media of example 18, where the PE includes an input register file for storing the segment of the input tensor; a weight register file for storing the segment of the filter; an output register file for storing the output tensor segment; and a MAC unit for performing the one or more MAC operations.
Example 20 provides the one or more non-transitory computer-readable media of any one of examples 11-19, where the input tensor includes one or more integer values or one or more floating-point values.
Example 21 provides a DNN accelerator, the DNN accelerator including a tile set including a plurality of PE arrays having different sizes, each PE array including PEs arranged in a first number of columns and a second number of rows and having a size determined by the first number and the second number; a workload manager configured to manage workloads of the tile set by: selecting a PE array from the plurality of PE arrays for a convolutional layer in the DNN, determining dimensions of an output tensor of the convolutional layer, the output tensor being a result of a convolutional operation to be performed by the PE array on an input tensor and a filter, partitioning the output tensor into output tensor segments based on a size of the PE array, and assigning workloads of generating the output tensor segments to a group of PEs in the PE array, where each PE in the group is to receive a workload of generating a respective output tensor segment and to perform a multiply-accumulation (MAC) operation for generating the respective output tensor segment; and a memory configured to store the input tensor, the filter, and the output tensor.
Example 22 provides the DNN accelerator of example 21, where the DNN accelerators further includes a plurality of tile sets that includes the tile set, and each of the plurality of tile sets is a combination of different PE arrays.
Example 23 provides the DNN accelerator of example 21 or 22, where the DNN includes a plurality of convolutional layers, and the tile set is selected from the plurality of tile sets based on one or more of the plurality of convolutional layers.
Example 24 provides the DNN accelerator of any one of examples 21-23, where the output tensor includes a set of output channels, each output channel including a matrix, and the dimensions of the output tensor include a first dimension indicating a number of elements in a row in the matrix, a second dimension indicating a number of elements in a column in the matrix, and a third dimension indicating a number of output channels in the set of output channels.
Example 25 provides the DNN accelerator of any one of examples 21-24, where the PE includes an input register file for storing a segment of the input tensor; a weight register file for storing a segment of the filter; an output register file for storing the output tensor segment; and a MAC unit for performing the MAC operation on the segment of the input tensor and the segment of the filter to generate the output tensor segment.
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.
