METHODS FOR PROCESSING HORIZONTAL STRIPES OF DATA IN AN EFFICIENT CONVOLUTIONAL ENGINE

Abstract

A convolutional engine is configured to process input data that is organized into horizontal stripes. The number of accumulators present in each convolver unit of the convolutional engine may equal a total number of rows of data in each of the horizontal stripes.

Description

FIELD OF THE INVENTION

The present invention relates to a hardware architecture for a convolutional engine, and more particularly relates to an efficient way to provide data values to compute units (called convolver units or functional units) of the convolutional engine.

BACKGROUND

Today, neural networks (in particular convolution neural networks) are widely used for performing image recognition/classification, object recognition/classification and image segmentation. While having numerous applications (e.g., object identification for self-driving cars, facial recognition for social networks, etc.), neural networks require intensive computational processing and frequent memory accesses. Described herein is an efficient hardware architecture for implementing a convolutional neural network.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 depicts a diagram providing an overview of model training and model application in a neural network.

FIG. 2 depicts a diagram of the input, model parameters and output of a convolution operation, the model parameters including a single 2-dimensional filter.

FIG. 3 depicts a diagram that explains the computation of a convolution operation using a 2-dimensional filter.

FIG. 4 depicts a diagram of the input, model parameters and output of a convolution operation, the model parameters including a plurality of 2-dimensional filters.

FIG. 5 depicts a diagram of the input, model parameters and output of a convolution operation, the model parameters including a single 3-dimensional filter.

FIG. 6 depicts a diagram that explains the computation of a convolution operation using a 3-dimensional filter.

FIG. 7 depicts a diagram of the input, model parameters and output of a convolution operation, the model parameters including a plurality of 3-dimensional filters.

FIG. 8 depicts a convolutional engine including a 2-D shift register and an array of convolver units, in accordance with one embodiment of the invention.

FIGS. 9A-9B depict the loading of data values into the convolutional engine, in accordance with one embodiment of the invention.

FIGS. 9C-9D depict the loading of filter weights into the convolutional engine, in accordance with one embodiment of the invention.

FIGS. 10A-10B depict the loading of a zero padding row into the 2-D shift register, in accordance with one embodiment of the invention.

FIGS. 10B-10D depict the loading of data values into the 2-D shift register, in accordance with one embodiment of the invention.

FIGS. 11A and 11B describe the processing of two convolver units for the spatial orientation of the data values depicted in FIG. 10D, in accordance with one embodiment of the invention.

FIG. 11C depicts the resulting partial sums following the processing of all active convolver units for the spatial orientation of the data values depicted in FIG. 10D, in accordance with one embodiment of the invention.

FIG. 12 depicts the data values after they have been shifted down one row of the 2-D shift register, as compared to the spatial orientation of the data values depicted in FIG. 10D.

FIGS. 13A-13D describe the processing of four convolver units for the spatial orientation of the data values depicted in FIG. 12, in accordance with one embodiment of the invention.

FIG. 13E depicts the resulting partial sums following the processing of all active convolver units for the spatial orientation of the data values depicted in FIG. 12, in accordance with one embodiment of the invention.

FIGS. 14A-14B depict the loading of data values into the convolutional engine, in accordance with one embodiment of the invention.

FIGS. 14C-14D depict the loading of filter weights into the convolutional engine, in accordance with one embodiment of the invention.

FIGS. 15A-15B depict the loading of a zero padding row into the 2-D shift register, in accordance with one embodiment of the invention.

FIGS. 15B-15D depict the loading of data values into the 2-D shift register, in accordance with one embodiment of the invention.

FIGS. 16A-16B describe the processing of two convolver units for the spatial orientation of the data values depicted in FIG. 15D, in accordance with one embodiment of the invention.

FIG. 16C depicts the resulting partial sums following the processing of all active convolver units for the spatial orientation of the data values depicted in FIG. 15D, in accordance with one embodiment of the invention.

FIG. 17 depicts the data values after they have been shifted down one row of the 2-D shift register as compared to the spatial orientation of the data values depicted in FIG. 15D.

FIGS. 18A-18B describe the processing of two convolver units for the spatial orientation of the data values depicted in FIG. 17, in accordance with one embodiment of the invention.

FIG. 18C depicts the resulting partial sums following the processing of all active convolver units for the spatial orientation of the data values depicted in FIG. 17, in accordance with one embodiment of the invention.

FIGS. 19A-19B depict the loading of bias values into the convolutional engine, in accordance with one embodiment of the invention.

FIG. 20 depicts the output of each of the convolver units, after the partial sums have been biased with bias values, in accordance with one embodiment of the invention.

FIG. 21 depicts internal components of a convolver unit, in accordance with one embodiment of the invention.

FIG. 22 depicts control circuitry for controlling the stride of a convolution operation, in accordance with one embodiment of the invention.

FIG. 23 depicts a generalized convolutional engine including a 2-D shift register and an array of functional units, in accordance with one embodiment of the invention.

FIG. 24 depicts internal components of a functional unit, in accordance with one embodiment of the invention.

FIG. 25 depicts three scenarios of data values being loaded from an input channel into a convolutional engine having m columns of convolver units, with scenario (a) illustrating the input channel having m columns of data values, scenario (b) illustrating the input channel having 3m−4 columns of data values, and scenario (c) illustrating the input channel having m/2 columns of data values, in accordance with one embodiment of the invention.

FIGS. 26A-26B depict the loading of data values into the convolutional engine for scenario (a), in accordance with one embodiment of the invention.

FIGS. 27A-27C depict the loading of data values into the convolutional engine for scenario (b), in accordance with one embodiment of the invention.

FIG. 28 depicts the loading of data values into the convolutional engine for scenario (c), in accordance with one embodiment of the invention.

FIGS. 29A-29B depict an alternate scheme for loading data values into the convolutional engine for scenario (c), in accordance with one embodiment of the invention.

FIG. 30 depicts a convolutional engine as one component of a larger system, in accordance with one embodiment of the invention.

FIG. 31 depicts a block diagram of a component for decompressing filter weights before the weights are provided to the convolver units, in accordance with one embodiment of the invention.

DETAILED DESCRIPTION OF THE INVENTION

In the following detailed description of the preferred embodiments, reference is made to the accompanying drawings that form a part hereof, and in which are shown by way of illustration specific embodiments in which the invention may be practiced. It is understood that other embodiments may be utilized and structural changes may be made without departing from the scope of the present invention. Description associated with any one of the figures may be applied to a different figure containing like or similar components/steps.

FIG. 1 depicts a diagram providing an overview of the training phase and the inference phase in a neural network. In the training phase, pairs of input and known (or desired) output may be provided to train model parameters (also called “weights”) of classification model 104. For conciseness, only one input and output pair (102, 106) is depicted in FIG. 1, but in practice many known input and output pairs will be used to train classification model 104. In the example of FIG. 1, input 102 is a matrix of numbers (which may represent the pixels of an image) and known output 106 is a vector of classification probabilities (e.g., the probability that the input image is a cat is 1, the probability that the input image is a dog is 0, and the probability that the input image is a human is 0). In one possible training process, the classification probabilities may be provided by a human (e.g., a human can recognize that the input image depicts a cat and assign the classification probabilities accordingly). At the conclusion of the model training process, the model parameters will have been estimated (e.g., W1=1.2, W2=3.8, W3=2.7). Sometimes, there may be intuitive ways to interpret the model parameters, but many times no intuition may be associated therewith, and the model parameters may simply be the parameters that minimize the error between the model's classification (or the model's classification probabilities) of a given set of input with the known classification (or known classification probabilities), while at the same time avoiding “model overfitting”.

In the inference (or prediction or feed-forward) phase, classification model 104 with trained parameters (i.e., parameters trained during the training phase) is used to classify a set of input. In the instant application, the trained classification model 104 provides the classification output 110 of a vector of probabilities (e.g., the probability that the input image is a cat is 0.3, the probability that the input image is a dog is 0.6, and the probability that the input image is a human is 0.1) in response to input 108.

One embodiment of classification model 104 is a convolutional neural network. A basic building block of a convolutional neural network is a convolution operation, which is described in FIGS. 2-7. As further described below, a convolution operation may refer to a 2-dimensional convolution operation with 2-dimensional input and a 2-dimensional filter, a 3-dimensional convolution operation with 3-dimensional input and a 3-dimensional filter, etc.

FIG. 2 depicts a diagram of the input, model parameters and output of a 2-dimensional convolution operation. In the example of FIG. 2, the input includes a 2-dimensional matrix of numerical values (each of the numerical values abstractly represented by “●”). The matrix in the example of FIG. 2 is a 4×4 matrix, but other input could have different dimensions (e.g., could be a 100×100 square matrix, a 20×70 rectangular matrix, etc.). Later presented examples will illustrate that the input may even be a 3-dimensional object. In fact, the input may be an object of any number of dimensions. The input may represent pixel values of an image or may represent the output of a previous convolution operation.

The model parameters may include a filter and a bias. In the example of FIG. 2, the filter is a 3×3 matrix of values (the values also called “weights”) and the bias is a scalar value. Typically, there is one bias associated with each filter. The example in FIG. 2 includes one filter, so there is one corresponding bias. However, in certain embodiments, if there were 5 filters, there would be 5 associated biases, one for each of the filters.

The convolution operator 208 (abbreviated “cony”) receives input 202 and the model parameters 204, 206, and generates output 210 called an activation map or a feature map. Each value of the activation map is generated as the sum of a dot product between of input 202 and filter 204 (at a certain spatial location relative to input 202) and bias 206. The computations to arrive at activation map 210 are described in more detail below in FIG. 3.

The first row of FIG. 3 describes the computation of the element at position (x=1, y=1) of activation map 210. As shown in the first row, the center of filter 204 is spatially aligned with the element at position (1, 1) of input 202. Such computation assumes the use of “zero padding” in which the input 202 is implicitly surrounded by a border of zeros. The advantage of using zero padding is that the dimensions of input 202 and output activation map 210 remain constant when using a 3×3 filter. A dot product is computed between filter 204 and the four values of input 202 that spatially align with filter 204. The dot product is then summed with bias b to arrive at the element at position (1, 1) of activation map 210.

The second row of FIG. 3 describes the computation of the element at position (1, 2) of activation map 210. As shown in the second row, the center of filter 204 is spatially aligned with the element at position (1, 2) of input 202. A dot product is computed between filter 204 and the six values of input 202 that spatially align with filter 204. The dot product is then summed with bias b to arrive at the element at position (1, 2) of activation map 210.

The third row of FIG. 3 describes the computation of the element at position (1, 3) of activation map 210. As shown in the third row, the center of filter 204 is spatially aligned with the element at position (1, 3) of input 202. A dot product is computed between filter 204 and the six values of input 202 that spatially align with filter 204. The dot product is then summed with bias b to arrive at the element at position (1, 3) of activation map 210.

The fourth row of FIG. 3 describes the computation of the element at position (4, 4) of activation map 210. As shown in the fourth row, the center of filter 204 is spatially aligned with the element at position (4, 4) of input 202. A dot product is computed between filter 204 and these four values of input 202 that spatially align with filter 204. The dot product is then summed with bias b to arrive at the element at position (4, 4) of activation map 210. In general, the convolution operation comprises a plurality of shift (or align), dot product and bias (or sum) steps. In the present example, the filter was shifted by 1 spatial position between dot product computations (called the step size or stride), but other step sizes of 2, 3, etc. are possible.

FIG. 4 is similar to FIG. 2, except that there are F filters 404, F biases 406 and F activation maps 410 instead of a single filter 204, a single bias 206 and a single activation map 210. The relation between the F filters 404, F biases 406 and F activation maps 410 is as follows: Filter f₁, bias b₁ and input 402 are used to compute activation map y₁ (in very much the same way that filter 204, bias 206 and input 202 were used to compute activation map 210 in FIG. 2); filter f₂, bias b₂ and input 402 are used to compute activation map y₂; and so on.

FIG. 5 is similar to FIG. 2, except that instead of a 2-dimensional input 202 and a 2-dimensional filter 204, a 3-dimensional input 502 and a 3-dimensional filter 504 are used. The computations to arrive at activation map 510 are described in more detail below in FIG. 6. While input 502 and filter 504 are 3-dimensional, activation map 510 is 2-dimensional, as will become clearer in the associated description of FIG. 6. Each “slice” of filter 504 (analogous to a “channel” of input 502) may be called a kernel. In FIG. 5, filter 504 is composed of five kernels, and input 502 is composed of five channels. If not already apparent, the number of kernels of filter 504 (or the size of the “z” dimension of filter 504) must match the number of channels of input 502 (or the size of the “z” dimension of input 502). During a convolution operation, channel 1 of input 502 aligns with kernel 1 of filter 504; channel 2 of input 502 aligns with kernel 2 of filter 504; and so on. Typically, there is no translation of filter 504 with respect to input 502 in the z-dimension during a convolution operation.

The first row of FIG. 6 describes the computation of the element at position (x=1, y=1) of activation map 510. As shown in the first row, the central axis 506 of filter 504 (with central axis drawn parallel to the z-axis) is aligned with the elements at positions (1, 1, z) for z∈{1, . . . , 5} of input 502. A dot product is computed between filter 504 and the twenty values of input 502 that spatially align with filter 504 (4 aligned values per channel×5 channels). The dot product is then summed with bias b to arrive at the element at position (1, 1) of activation map 510.

The second row of FIG. 6 describes the computation of the element at position (1, 2) of activation map 510. As shown in second first row, the central axis 506 of filter 504 is aligned with the elements at positions (1, 2, z) for z∈{1, . . . , 5} of input 502. A dot product is computed between filter 504 and the thirty values of input 502 that spatially align with filter 504 (6 aligned values per channel×5 channels). The dot product is then summed with bias b to arrive at the element at position (1, 2) of activation map 510.

The third row of FIG. 6 describes the computation of the element at position (1, 3) of activation map 510. As shown in the third row, the central axis 506 of filter 504 is aligned with the elements at positions (1, 3, z) for z∈{1, . . . , 5} of input 502. A dot product is computed between filter 504 and the thirty values of input 502 that spatially align with filter 504 (6 aligned values per channel×5 channels). The dot product is then summed with bias b to arrive at the element at position (1, 3) of activation map 510.

The fourth row of FIG. 6 describes the computation of the element at position (4, 4) of activation map 510. As shown in the fourth row, the central axis 506 of filter 504 is aligned with the elements at positions (4, 4, z) for z∈{1, . . . , 5} of input 502. A dot product is computed between filter 504 and the twenty values of input 502 that spatially align with filter 504 (4 aligned values per channel×5 channels). The dot product is then summed with bias b to arrive at the element at position (4, 4) of activation map 510.

FIG. 7 is similar to FIG. 5, except that there are F 3-dimensional filters 704, F biases 706 and F activation maps 710 (F>1), instead of a single 3-dimensional filter 504, a single bias b 505 and a single activation map 510. The relation between the F 3-dimensional filters 704, F biases 706 and F activation maps 710 is as follows: Filter f₁, bias b₁ and input 702 are used to compute activation map y₁ (in very much the same way that filter 504, bias b 505 and input 502 were used to compute activation map 510 in FIG. 5); filter f₂, bias b₂ and input 702 are used to compute activation map y₂; and so on.

The following figures describe a hardware architecture to perform the convolution operation of FIG. 7. Many of the examples assume the use of two filters, F=2, for simplicity. The examples further assume that the filters 704 are constructed using 3×3 kernels (i.e., each kernel being composed of 9 weights). It is understood, however, that the concepts/architectures described herein can be modified to accommodate kernels with other dimensions.

FIG. 8 depicts convolutional engine 708, in accordance with one embodiment of the invention. Convolutional engine 708 (depicted in FIG. 8) is a hardware architecture of the convolution operator (“cony”) 708 (depicted in FIG. 7). Convolutional engine 708 may include a 2-D shift register with an array of data storage elements:

$\left[\begin{array}{cccc}{d}_{1,1}& d_{1,2}& d_{1,3}& d_{1,4}\\ d_{2,1}& d_{2,2}& d_{2,3}& d_{2,4}\\ d_{3,1}& d_{3,2}& d_{3,3}& d_{3,4}\\ d_{4,1}& d_{4,2}& d_{4,3}& d_{4,4}\end{array}\right]$

In the simplified example of FIG. 8, the array is a four by four array. Each of the data storage elements may be formed by a plurality of D flip-flops (i.e., one D flip-flop to store each bit of a data signal). Therefore, if data storage element d_1,1 were to store eight bits, d_1,1 may be formed from eight D flip-flops. Each of the arrows between pairs of data storage elements represents an electrical connection (i.e., may be implemented as a wire). For example, data storage element d_1,1 (ref. num. 802) may be electrically coupled to storage element d_2,1 (ref. num. 802) via electrical connection 804. Further, the arrow may represent a one-directional flow of data (i.e., data being transmitted from data storage element d_1,1 to data storage element d_2,1, but not from d_2,1 to data storage element d_1,1). In the discussion that follows, the first row of data storage elements may be called a “header”, and the last row of data storage elements may be called a “footer”.

Convolutional engine 708 may further include an array of convolver units:

$\left[\begin{array}{cccc}C{U}_{1,1}& C{U}_{1,2}& C{U}_{1,3}& C{U}_{1,4}\\ C{U}_{2,1}& C{U}_{2,2}& C{U}_{2,3}& C{U}_{2,4}\end{array}\right]$

For conciseness, an array of convolver units may be called “a convolver array”. In the simplified example of FIG. 8, the convolver array is a two by four array. Convolver unit CU_1,2 has been labeled with reference numeral 806 (to facilitate later discussion). It is understood that a more typical embodiment will contain many more convolver units, such as in the example embodiment of FIG. 30. The operation of the 2-D shift register and the operation of the convolver units will be described in detail in the following figures.

FIG. 9A depicts the loading of data values into convolutional engine 708, in accordance with one embodiment of the invention. Each channel of input may be loaded into convolutional engine 708 in a serial fashion. FIG. 9A depicts the loading of the first channel 702a of input 702 into convolutional engine 708 (assuming that the channels are numbered from 1 to 5 in the left to right direction). As will be described in FIGS. 10B-10D, the rows of a particular channel may be loaded into convolutional engine 708 in a serial fashion. It is noted that terms such as a “row” and a “column” will be/are being used for convenience and with respect to how elements are depicted in the figures. Nevertheless, the meaning of such terms may or may not translate into how circuit elements are laid out on a chip, where a row could be interpreted as a column and vice versa, depending on the viewer's orientation with respect to the chip.

For simplicity, this first example describing the hardware architecture of a convolutional engine will handle the case in which the number of columns of an input channel is equal to the number of columns of the convolver array. In FIG. 9B, the number of columns of input channel 702a is assumed to equal the number of columns of the convolver array. For instance, input channel 702a may be a ten by four matrix of data values. FIGS. 27A-27C describe how to handle the scenario in which the number of columns of an input channel is greater than the number of columns of the convolver array. FIGS. 28, 29A and 29B describe two schemes for handling the case in which the number of columns of an input channel is less than the number of columns of the convolver array.

Typically, due to memory constraints of each convolver unit, convolutional engine 708 can only compute the convolution operation for a certain number of contiguous rows of the data values before the output needs to be saved (copied to a memory location separate from the convolver units—see memory 3002 in FIG. 30). Once the output is saved, the convolutional engine 708 can continue onto the next set of contiguous rows. In particular, if each convolver unit is constructed with n accumulators, convolutional engine 708 can compute the output of n contiguous input rows (plus two padding rows explained below). For simplicity of explanation, n contiguous input rows will be called a “horizontal stripe” of data. In the simplified example of FIG. 9B, there are two horizontal stripes 902a, 902b (while it is understood that in practice there could be any number of horizontal stripes). Due to the memory constraint of the convolver units, the convolutional engine 708 may process the horizontal stripes serially. In the example of FIG. 9B, horizontal stripe 902a is processed first, followed by horizontal stripe 902b.

For reasons that will be more apparent below, the loading of a leading row (i.e., first row of a horizontal stripe to be loaded) that is an external edge may be preceded by the loading of a zero padding row (as in row n of horizontal stripe 902a); the loading of a trailing row (i.e., last row of a horizontal stripe to be loaded) that is an external edge may be followed by the loading of a zero padding row (as in row 1 of horizontal stripe 902b); the loading of a leading row that is an internal edge may be preceded by the loading of a data padding row (as in row n of horizontal stripe 902b); and the loading of a trailing row that is an internal edge may be followed by the loading of a data padding row (as in row 1 of horizontal stripe 902a). If not already apparent, an “external edge” refers to a leading or trailing row of a horizontal stripe that forms an external boundary of an input channel, whereas an internal edge refers to a leading or trailing row of a horizontal stripe that is not part of an external boundary of an input channel. The reason for the zero or data padding row is tied to the 3×3 filter requiring data from a row above and a row below the row of interest to compute the convolution output. For a 5×5 filter, two padding rows (for the top row of a stripe) and two padding rows (for the bottom row of a stripe) or a total of four padding rows would have been needed.

In the particular example of FIG. 9B, the n+2 rows within the bolded and dashed rectangle are loaded into convolutional engine 708. The n+2 rows include a zero padding row, n rows of horizontal stripe 902a and a data padding row (equivalent to row n of horizontal stripe 902b).

FIGS. 9C-9D depict the loading of filter weights to convolutional engine 708, in accordance with one embodiment of the invention. More specifically, FIG. 9C depicts the loading of the nine weights of kernel 704a into each of the convolver units of the first row of the convolver array (i.e., CU_1,1, CU_1,2, CU_1,3 and CU_1,4), and FIG. 9D depicts the loading of the nine weights of kernel 704b into each of the convolver units of the second row of the convolver array (i.e., CU_2,1, CU_2,2, CU_2,3 and CU_2,4). Kernel 704a is the first kernel of filter f₁, and each of its weights is labeled with the superscript “1,1”, which is shorthand for (filter f₁, kernel 1). Kernel 704b is the first kernel of filter f₂, and each of its weights is labeled with the superscript “2,1”, which is shorthand for (filter f₁, kernel 1).

FIGS. 10A-10B depict the loading of a row of zero values into the 2-D shift register. FIGS. 10B-10D depict a row-by-row loading of data values from the first input channel 702a into the 2-D shift register and a row-to-row shifting of the data values through the 2-D shift register. Data values x_n,1, x_n,2, x_n,3 and x_n,4 may represent values from row n of horizontal stripe 902a of input channel 702a. Data values x_n-1,1, x_n-1,2, x_n-1,3 and x_n-1,4 may represent values from row n−1 of horizontal stripe 902a of input channel 702a. Data values x_n-2,1, x_n-2,2, x_n-2,3 and x_n-2,4 may represent values from row n−2 of horizontal stripe 902a of input channel 702a.

Upon row n of horizontal stripe 902a being loaded into the second row of data storage elements (i.e., d_2,1, d_2,2, d_2,3 and d_2,4), the first row of convolver units (i.e., CU_1,1, CU_1,2, CU_1,3 and CU_1,4) corresponding to the second row of data storage elements may be activated. By “corresponding”, it is meant that there is a logical correspondence between convolver unit CU_1,1 and data storage element d_2,1, convolver unit CU_1,2 and data storage element d_2,2, and so on. The correspondences between the data storage element and convolver units are shown in the figures by the data storage element being drawn within the corresponding convolver unit. In a more typical embodiment with a high number of convolver units, most of the convolver units will receive data values from its corresponding data storage element and eight spatial neighbors (i.e., data storage element neighbors) of the corresponding data storage element. Such relationship is more difficult to appreciate from the example convolutional engine of FIG. 11A in which there is a small number of convolver units.

Active convolver units are drawn in FIG. 11A in bolded lines while non-active convolver units are drawn in FIG. 11A using non-bolded lines. In one embodiment, “active” means that a convolver unit is powered on, whereas “non-active” means that a convolver unit is powered off to save power. A controller (depicted as controller 2202 in FIG. 22 and controller 3006 in FIG. 30, but not depicted in other figures for conciseness of presentation) may be responsible for powering on and off convolver units. The controller may power on a row of convolver units once the data from row n of a horizontal stripe has been loaded into the data storage elements corresponding to the row of convolver units. The controller may power off a row of convolver units once data from row 1 of a horizontal stripe has been transferred out of the data storage elements corresponding to the row of convolver units.

FIGS. 11A and 11B describe the processing of two out of the four active convolver units for the spatial orientation of the data values depicted in FIG. 10D. While the processing of the two convolver units is described in two separate figures, it is understood that such processing typically occurs in parallel (i.e., at the same time) in order to increase the number of computations per clock cycle.

As depicted in FIG. 11A, convolver unit CU_1,1 (typical for convolver units located on the left and right edges of the convolver array) receives data and/or zero values from five neighboring data storage elements and one data value from the data storage element corresponding to convolver unit CU_1,1. More specifically, convolver unit CU_1,1 receives:

data value x_n-1,1 from data storage element d_1,1 via electrical connection 1100a,

data value x_n-1,2 from data storage element d_1,2 via electrical connection 1100b,

data value x_n,1 from data storage element d_2,1 via an electrical connection (not depicted)

data value x_n,2 from data storage element d_2,2 via electrical connection 1100c,

the zero value from data storage element d_3,1 via electrical connection 1100d, and

the zero value from data storage element d_3,2 via electrical connection 1100e.

For clarity of depiction, electrical interconnections (i.e., bolded arrows) between convolver units and data storage elements are only depicted when needed for discussion.

Once the data and/or zero values have been received, convolver unit CU_1,1 may compute the partial sum y₁ defined by w₂^1,1x_n-1,1+w₃^1,1x_n-1,2+w₅^1,1x_n,1+w₆^1,1x_n,2 (where w₂^1,1, w₃^1,2, w₅^1,2, and w₆^1,1 are four of the nine weights of kernel 704a depicted in FIG. 9C) and store the partial sum y₁ in accumulator 1102a of convolver unit CU_1,1. Accumulator 1102a may be part of a linear array of n accumulators, where n is the number of rows within horizontal stripe 902a. Accumulator 1102a may be configured to store the partial sums corresponding to row n of a horizontal stripe; accumulator 1102b may be configured to store the partial sums corresponding to row n−1 of a horizontal stripe; and so on. For clarity of explanation, it is noted that the bottom instance of convolver unit CU_1,1 and the top instance of convolver unit CU_1,1 are one and the same convolver unit, with the bottom instance showing additional details of the top instance.

As depicted in FIG. 11B, convolver unit CU_1,2 receives data and/or zero values from eight neighboring data storage elements and one data value from the data storage element corresponding to convolver unit CU_1,2. More specifically, convolver unit CU_1,2 receives:

data value x_n-1,1 from data storage element d_1,1 via electrical connection 1100f,

data value x_n-1,2 from data storage element d_1,2 via electrical connection 1100g,

data value x_n-1,3 from data storage element d_1,3 via electrical connection 1100h,

data value x_n,1 from data storage element d_2,1 via an electrical connection 1100i,

data value x_n,2 from data storage element d_2,2 via electrical connection (not depicted),

data value x_n,3 from data storage element d_2,3 via electrical connection 1100j,

the zero value from data storage element d_3,1 via electrical connection 1100k,

the zero value from data storage element d_3,2 via electrical connection 1100l, and

the zero value from data storage element d_3,3 via electrical connection 1100m.

Once the data values have been received, convolver unit CU_1,2 may compute the partial sum y₂ defined by w₁^1,1x_n-1,1+w₂^1,1x_n-1,2+w₃^1,1x_n-1,3+w₄^1,1x_n,1+w₅^1,1x_n,2+w₆^1,1x_n,3 (where w₁^1,1, w₂^1,1, w₃^1,1, w₄^1,1, w₅^1,1 and w₆^1,1 are six of the nine weights of kernel 704a depicted in FIG. 9C) and store the partial sum y₂ in accumulator 1104a of convolver unit CU_1,2.

Similar processing is performed by CU_1,3 and CU_1,4, so the details of these computations have been omitted for conciseness. At the conclusion of the processing by the four active convolver units for the spatial orientation of data values shown in FIG. 10D, four partial sums are computed and stored in accumulators 1102a, 1104a, 1106a and 1108a, as shown in FIG. 11C.

FIG. 12 depicts the 2-D shift register after the data and/or zero values have been shifted down one row of data storage elements, and data values x_n-2,1, x_n-2,2, x_n-2,3 and x_n-2,4 from the n−2 row of the horizontal stripe 902a have been loaded into the 2-D shift register. Once row n of horizontal stripe 902a has been loaded into data storage elements d_3,1, d_3,2, d_3,3, and d_3,4, the corresponding convolver units CU_2,1, CU_2,2, CU_2,3 and CU_2,4 are activated, in addition to CU_1,1, CU_1,2, CU_1,3 and CU_1,4 (as shown in FIG. 13A).

FIGS. 13A-13D describe the processing of four of the eight active convolver units, in accordance with one embodiment of the invention. While the processing of the four convolver units is described in four separate figures, it is understood that such processing typically occurs in parallel (i.e., at the same time) in order to increase the number of computations per clock cycle.

As depicted in FIG. 13A, convolver unit CU_1,1 may receive data values from the five neighboring data storage elements and the one corresponding data storage element. Convolver unit CU_1,1 may compute the partial sum y₅ defined by w₂^1,1x_n-2,1+w₃^1,1x_n-2,2+w₅^1,1x_n-1,1+w₆^1,1x_n-1,2+w₈^1,1x_n,1+w₉^1,1x_n,2 and store the partial sum y₅ in accumulator 1102b of convolver unit CU_1,1.

As depicted in FIG. 13B, convolver unit CU_1,2 may receive data values from the eight neighboring data storage elements and the one corresponding data storage element. Convolver unit CU_1,2 may compute the partial sum y₆ defined by w₁^1,1x_n-2,1+w₂^1,1x_n-2,2+w₃^1,1x_n-2,3+w₄^1,1x_n-1,1+w₅^1,1x_n-1,2+w₆^1,1x_n-1,3+w₇^1,1x_n,1+w₈^1,1x_n,2+w₉^1,1x_n,3 and store the partial sum y₆ in accumulator 1104b of convolver unit CU_1,2.

As depicted in FIG. 13C, convolver unit CU_1,3 may receive data values from the eight neighboring data storage elements and the one corresponding data storage element. Convolver unit CU_1,3 may compute the partial sum y₇ defined by w₁^1,1x_n-2,2+w₂^1,1x_n-2,3+w₃^1,1x_n-2,4+w₄^1,1x_n-1,2+w₅^1,1x_n-1,3+w₆^1,1x_n-1,4+w₇^1,1x_n,2+w₈^1,1x_n,3+w₉^1,1x_n,4 and store the partial sum y₇ in accumulator 1106b of convolver unit CU_1,3.

As depicted in FIG. 13D, convolver unit CU_2,1 may receive data and/or zero values from the five neighboring data storage elements and the one corresponding data storage element. Convolver unit CU_2,1 may then compute the partial sum y₉ defined by w₂^2,1x_n-1,1+w₃^2,1x_n-1,2+w₅^2,1x_n,1+w₆^2,1x_n,2 (where w₂^2,1, w₃^2,1, w₅^2,1, and w₆^2,1 are four of the nine weights of kernel 704b depicted in FIG. 9D) and store the partial sum y₉ in accumulator 1110a of convolver unit CU_2,1.

Similar processing may be performed by CU_1,4, CU_2,2, CU_2,3 and CU_2,4, so the details of these computations have been omitted for conciseness. At the conclusion of the processing by the active convolver units for the spatial orientation of data values shown in FIG. 12, eight (additional) partial sums have been computed and stored in accumulators 1102b, 1104b, 1106b, 1108b, 1110a, 1112a, 1114a and 1116a, as shown in FIG. 13E.

The processing of the 2-D shift register and the plurality of convolutional units continues in a similar fashion until row 1 of horizontal stripe 902a has been shifted through the 2-D shift register. At this point, data values of the next input channel and parameters (i.e., weights) of the kernels corresponding to the next input channel may be loaded into the convolutional engine, as depicted in FIGS. 14A-14D.

FIG. 14A depicts the loading of data values from the second input channel 702b into convolutional engine 708, in accordance with one embodiment of the invention. As depicted in greater detail in FIG. 14B, the second input channel 702b may include horizontal stripes 904a and 904b, and horizontal stripe 904a may be loaded into convolutional engine 708 in a similar manner as horizontal stripe 902a was loaded.

FIGS. 14C-14D depict the loading of filter weights into convolutional engine 708, in accordance with one embodiment of the invention. More specifically, FIG. 14C depicts the loading of the nine weights of kernel 704c into each of the convolver units of the first row of the convolver array (i.e., CU_1,1, CU_1,2, CU_1,3 and CU_1,4), and FIG. 14D depicts the loading of the nine weights of kernel 704b into each of the convolver units of the second row of the convolver array (i.e., CU_2,1, CU_2,2, CU_2,3 and CU_2,4). Kernel 704c is the second kernel of filter f₁, and each of its weights is labeled with the superscript “1,2”, which is shorthand for (filter f₁, kernel 2). Kernel 704d is the second kernel of filter f₂, and each of its weights is labeled with the superscript “2,2”, which is shorthand for (filter f₂, kernel 2).

FIGS. 15A-15B depict the loading of a row of zero values into the 2-D shift register. FIGS. 15B-15D depict a row-by-row loading of data values from the second input channel 702b into the 2-D shift register and a row-to-row shifting of the data values through the 2-D shift register. Data values x_n,1′, x_n,2′, x_n,3′ and x_n,4′ may represent values from row n of horizontal stripe 904a of input channel 702b. Data values x_n-1,1′, x_n-1,2′, x_n-1,3′ and x_n-1,4′ may represent values from row n−1 of horizontal stripe 904a of input channel 702b. Data values x_n-2,1′, x_n-2,2′, x_n-2,3′ and x_n-2,4′ may represent values from row n−2 of horizontal stripe 904a of input channel 702b. Upon row n of horizontal stripe 904a being loaded into the second row of data storage elements, the first row of convolver units may be activated (as shown in FIG. 16A).

FIGS. 16A and 16B describe the processing of two out of the four active convolver units for the spatial orientation of the data values depicted in FIG. 15D. As depicted in FIG. 16A, convolver unit CU_1,1 may receive data and/or zero values from the five neighboring data storage elements and one data value from the data storage element corresponding to convolver unit CU_1,1. Once the data values have been received, convolver unit CU_1,1 may compute the partial sum y₁₃ defined by w₂^1,2x_n-1,1′+w₃^1,2x_n-1,2′+w₅^1,2x_n,1′+w₆^1,2x_n,2′ (where w₂^1,2, w₃^1,2, w₅^1,2, and w₆^1,2 are four of the nine weights of kernel 704c depicted in FIG. 14C). The partial sum y₁₃ may be summed with y₁ (the partial sum previously computed by convolver unit CU_1,1 for row n) and the new partial sum y₁+y₁₃ may be stored in accumulator 1102a.

As depicted in FIG. 16B, convolver unit CU_1,2 may receive data and/or zero values from the eight neighboring data storage elements and one data value from the data storage element corresponding to convolver unit CU_1,2. Once the data and/or zero values have been received, convolver unit CU_1,2 may compute the partial sum y₁₄ defined by w₁^1,2x_n-1,1′+w₂^1,2x_n-1,2′+w₃^1,2x_n-1,3′+w₄^1,2x_n,1′+w₂^1,2x_n,2′+w₆^1,2x_n,3′ (where w₁^1,2, w₂^1,2, w₃^1,2, w₄^1,2, w₅^1,2 and w₆^1,2 are six of the nine weights of kernel 704c depicted in FIG. 14C). The partial sum y₁₄ may be summed with y₂ (the partial sum previously computed by convolver unit CU_1,2 for row n) and the new partial sum y₂+y₁₄ may be stored in accumulator 1104a.

Similar processing is performed by CU_1,3 and CU_1,4, so the details of these computations have been omitted for conciseness. At the conclusion of the processing by the four active convolver units for the spatial orientation of data values shown in FIG. 15D, four partial sums have been updated and stored in accumulators 1102a, 1104a, 1106a and 1108a, as shown in FIG. 16C.

FIG. 17 depicts the 2-D shift register after the data and/or zero values have been shifted down one row of data storage elements, and data values and x_n-2,1′, x_n-2,2′, x_n-2,3′ and x_n-2,4′ from the n−2 row of the horizontal stripe 904a have been loaded into the 2-D shift register. Once row n of horizontal stripe 904a has been loaded into data storage elements d_3,1, d_3,2, d_3,3 and d_3,4, the corresponding convolver units CU_2,1, CU_2,2, CU_2,3 and CU_2,4 are activated, in addition to CU_1,1, CU_1,2, CU_1,3 and CU_1,4 (as shown in FIG. 18A).

FIGS. 18A-18B describe the processing of two of the eight active convolver units, in accordance with one embodiment of the invention. As depicted in FIG. 18A, convolver unit CU_1,1 may receive data values from the five neighboring data storage elements and the one corresponding data storage element. Convolver unit CU_1,1 may then compute the partial sum y₁₇ defined by w₂^1,2x_n-2,1′+w₃^1,2x_n-2,2′+w₅^1,2x_n-1,1′+w₆^1,2x_n-1,2′+w₈^1,2x_n,1′+w₉^1,2x_n,2′. The partial sum y₁₇ may be summed with y₅ (the partial sum previously computed by convolver unit CU_1,1 for row n−1) and the new partial sum y₅+y₁₇ may be stored in accumulator 1102b.

As depicted in FIG. 18B, convolver unit CU_1,2 may receive data values from the eight neighboring data storage elements and the one corresponding data storage element. Convolver unit CU₁₂ may then compute the partial sum y₁₈ defined by w₁^1,2x_n-2,1′+w₂^1,2x_n-2,2′+w₃^1,2x_n-2,3′+w₄^1,2x_n-1,1′+w₅^1,2x_n-1,2′+w₆^1,2x_n-1,3′+w₇^1,2x_n,1′+w₈^1,2x_n,2′+w₉^1,2x_n,3′. The partial sum y₁₈ may be summed with y₆ (the partial sum previously computed by convolver unit CU_1,2 for row n−1) and the new partial sum y₆+y₁₈ may be stored in accumulator 1104b.

Similar processing is performed by convolver units CU_1,3, CU_1,4, CU_2,1, CU_2,2, CU_2,3 and CU_2,4, so the details of these computations have been omitted for conciseness. At the conclusion of the processing by the active convolver units for the spatial orientation of data values shown in FIG. 17, eight (additional) partial sums have been updated and stored in accumulators 1102b, 1104b, 1106b, 1108b, 1110a, 1112a, 1114a and 1116a, as shown in FIG. 18C.

The processing of the 2-D shift register and the plurality of convolutional units continues in a similar fashion until row 1 of horizontal stripe 904a has been shifted through the 2-D shift register. The processing of the 2-D shift register and the plurality of convolutional units then continues until all of the remaining input channels have been processed in a manner similar to the processing of the first two input channels.

At this point (or earlier in the process), bias values may be loaded into the convolutional units. More specifically, FIG. 19A depicts the loading of bias value b₁ into the first row of convolver units (CU_1,1, CU_1,2, CU_1,3 and CU_1,4) and FIG. 19B depicts the loading of bias value b₂ into the second row of convolver units (CU_2,1, CU_2,2, CU_2,3 and CU_2,4). The partial sums computed by the first row of convolver units may be biased by bias value b₁, and the partial sums computed by the second row of convolver units may be biased by bias value b₂ (as depicted in FIG. 20) to yield the output of the convolution operation.

In the examples so far, it was assumed that the number of rows of the convolver array equals the number filters. This relationship, however, does not always hold. If the number of filters were less than the number of rows of the convolver array, unused rows of the convolver array could be deactivated. If the number of filters were more than the number of rows of the convolver array, the convolution operations would essentially need to be repeated. For instance, if there were six filters and only three rows of convolver units, then the convolution operations could be performed for filters 1-3, and the same convolution operations would be repeated, except that filters 1-3 would be substituted with filters 4-6.

Some motivation is now provided for the above-described architecture of the convolutional engine. The architecture essentially attempts to strike a balance between the fan-out of data storage elements (related to the sizing of circuit components) and the number of computations per clock cycle (related to the speed of computation). At one extreme of solely maximizing the computations per clock cycle, the 2-D shift register could have been reduced to three rows of data storage elements, with CU_1,1, CU_2,1, CU_3,1, . . . wired to the same six data storage elements; CU_1,2, CU_2,2, CU_3,2, . . . wired to the same nine data storage elements, etc. While the computations per clock cycle would be greater than the above-described architecture, the fan-out of the data storage elements would be much greater (requiring larger circuit components to drive the increased output capacitance). At the other extreme of solely minimizing the fan-out, three contiguous rows of the 2-D shift register could have been used exclusively for filter 1, three contiguous rows of the 2-D shift register could have been used exclusively for filter 2, and so on. While the fan-out would be lower than the above-described architecture, the number of computations per clock cycle would essentially be reduced by two-thirds, as compared to the above-described architecture. In light of this explanation, the motivation for the above-described architecture should now be more apparent as one which strikes a balance between the fan-out of data storage elements and the number of computations per clock cycle.

FIG. 21 depicts internal components of convolver unit 806 (i.e., CU_1,2), in accordance with one embodiment of the invention. Convolver unit 806 may include nine multipliers (2102a, . . . , 2102i). Each of the multipliers may be electrically coupled to a data storage element (i.e., one of the data storage elements of the 2-D shift register) and may be configured to receive a data value stored in the corresponding data storage element. In particular, multipliers 2102a, 2102b, 2102c, 2102d, 2102e, 2102f, 2102g, 2102h, and 2102i are electrically coupled to data storage elements d_1,1, d_1,2, d_1,3, d_2,1, d_2,2, d_2,3, d_3,1, d_3,2, and d_3,3, and are configured to receive data values x₁, x₂, x₃, x₄, x₅, x₆, x₇, x₈, and x₉, from data storage elements 2102a, 2102b, 2102c, 2102d, 2102e, 2102f, 2102g, 2102h, and 2102i, respectively. The data value stored in a data storage element typically changes with each clock cycle. For example, in the context of FIG. 10C, x₁ would equal x_n,1; in FIG. 10D, x₁ would equal x_n-1,1; and so on. The same comment applies for the other data values.

Each of the multipliers is further configured to receive a weight. In particular, multipliers 2102a, 2102b, 2102c, 2102d, 2102e, 2102f, 2102g, 2102h, and 2102i are configured to receive weights w₁, w₂, w₃, w₄, w₅, w₆, w₇, w₈, and w₉, respectively. A different set of weights may be loaded for each channel of input data 702. For example, in the context of FIG. 9C, w₁ would equal w₁^1,1; in the context of FIG. 14C, w₁ would equal w₁^1,2; and so on.

Each of the multipliers may multiply two values so as to generate the product of the two values. In particular, multipliers 2102a, 2102b, 2102c, 2102d, 2102e, 2102f, 2102g, 2102h, and 2102i may multiply data values x₁, x₂, x₃, x₄, x₅, x₆, x₇, x₈, and x₉ with weights w₁, w₂, w₃, w₄, w₅, w₆, w₇, w₈, and w₉ so as to generate the products w₁x₁, w₂x₂, w₃x₃, w₄x₄, w₅x₅, w₆x₆, w₇x₇, w₈x₈, and w₉x₉, respectively. In an embodiment in which signal values (including data values and weights) are represented in the log domain, a specialized multiplier may be implemented using a bit-shifter and an adder (the specialized multiplier further performing a log-to-linear transformation). For more details on such an implementation, see, e.g., Daisuke Miyashita et al. “Convolutional Neural Networks using Logarithmic Data Representation” arXiv preprint arXiv:1603.01025, 2016. Therefore, for clarity, when stated that a product is generated, it is understood that such a computation can be implemented using a multiplier, or the combination of a bit-shifter and an adder.

Convolver unit 806 may further include a plurality of adders and the values that are summed by the adders may depend on control signal s1. When the data values x₁, . . . , x₉ are from the first input channel 702a, control signal s1 may be set to 0, causing output selector 2106 to deliver the zero value to adder 2104h. In this mode of operation, the partial sum w₁x₁+w₂x₂+w₃x₃+w₄x₄+w₅x₅+w₆x₆+w₇x₇+w₈x₈+w₉x₉ is computed, and is not based on any previous partial sums. The partial sum is then stored in one of the accumulators 1104a, 1104b, etc. depending on which row of a horizontal stripe the data values are from. If the data values are from row n, the partial sum would be stored in accumulator 1104a; if the data values are from row n−1, the partial sum would be stored in accumulator 1104b; and so on.

When the data values x₁, . . . , x₉ are from one of the subsequent input channels (e.g., 702b, etc.), control signal s1 may be set to 1, causing output selector 2106 to deliver a previously computed partial sum to adder 2104h. In particular, if the data values are from row n of a horizontal stripe, the previously computed partial sum stored in accumulator 1104a would be provided to adder 2104h; if the data values are from row n−1, the previously computed partial sum stored in accumulator 1104b would be provided to adder 2104h; and so on.

When control signal s1 is set to 2, output selector 2106 may be configured to deliver a partial sum from an accumulator to adder 2104j, which sums the partial sum with bias b_k. The resulting sum may be stored back into the accumulator from which the partial sum was read. For an efficient implementation, an entire vector of partial sums may be read from the accumulator array (1104a, 1104b, . . . ), summed with bias b_k, and the vector (now with biasing) may be stored back into the accumulator array. Such computation may implement the biasing operation described for CU_1,2 in FIG. 20.

It is further noted that in an embodiment in which signal values are represented in the log domain, specialized adders (built using comparators, bit-shifters and adders) may receive two values in the linear domain (since the preceding specialized multipliers performed a log-to-linear transformation) and return the resulting sum in the log domain. Details of such specialized adders may also be found in Daisuke Miyashita et al. “Convolutional Neural Networks using Logarithmic Data Representation” arXiv preprint arXiv:1603.01025, 2016.

Any of the convolver units that receive nine data values (and nine weights) may have a similar hardware architecture as convolver unit CU_1,2, and hence will not be described for conciseness. For convolver units that receive less than nine data values, the hardware architecture could still be similar to the hardware architecture of convolver unit CU_1,2, except that some of the inputs to the multipliers could be hardwired to the zero value (data input or weight could be set to the zero value). For example, since CU_1,1 does not receive data values x₁, x₄ and x₇, weights w₁, w₄ and w₇ could be set to zero. In another embodiment, some of the multipliers could even be omitted. For example, since CU_1,1 does not receive data values x₁, x₄ and x₇, multipliers 2102a, 2102d and 2102g could be omitted.

In one embodiment of the invention, the computations of all nine multipliers (or their equivalents in the log domain) and nine adders (or their equivalents in the log domain) take place all within one clock cycle. That is, if data values are stored in the nine data storage elements at clock cycle n, the partial sum is stored in the accumulators at clock cycle n+1. Further, for increased throughput, new data values may be stored in the nine data storage elements at clock cycle n+1 while the partial sum is stored. Therefore the computation of a new partial sum may be performed during every clock cycle.

Details are now provided as to how the stride of the convolution operation can be set using the hardware architecture. Recall, the stride (or the step size) is the number of pixels or data values that the filter is shifted between dot product operations. FIG. 22 illustrates that by setting every odd row and every odd column of convolver units to be active and setting every even row and every even column of convolver units to be non-active (by means of control signals provided by controller 2202), a stride of 2 may be achieved. It should be apparent how other stride values can be set. For a stride of 3, rows 3x+1 for x∈{0, 1, 2, . . . } of convolver units and columns 3x+1 for x∈{0, 1, 2, . . . } of convolver units may be set to be active and all other rows and columns may be set to be non-active. Even strides of less than 1 are possible. For example, for a stride of ½, input 702 can be interpolated before it is loaded into convolutional engine 708. For a 2×2 input matrix of

$\left[\begin{array}{cc}a& b\\ c& d\end{array}\right]$

the following 3×3 interpolated matrix can be provided as input to convolutional engine 708 in order to achieve a stride of ½:

$\left[\begin{array}{ccc}a& \frac{a+b}{2}& b\\ \frac{a+c}{2}& \frac{a+b+c+d}{4}& \frac{b+d}{2}\\ c& \frac{c+d}{2}& d\end{array}\right]$

While a linear interpolation was used in the present example, it is understood that other forms of interpolation (e.g., polynomial interpolation, spline interpolation, etc.) are also possible.

While the discussion so far has focused on the convolution operation, a convolutional neural network typically involves other types of operations, such as the max pool and rectification operators. The convolver unit was presented first for ease of understanding, but now a more generalized form of a convolver unit, called a “functional unit” will now be described for handling other types of operations common in a convolutional neural network in addition to the convolution operation.

FIG. 23 depicts convolutional engine 2300 including a 2-D shift register and an array of functional units, in accordance with one embodiment of the invention. Convolutional engine 2300 is similar to the above-described convolutional engine 708, except that the convolver units have been replaced with functional units. One of the functional units, FU_1,2, is labeled as 2302 and its hardware architecture is described below in FIG. 23.

FIG. 24 depicts internal components of functional unit 2302, in accordance with one embodiment of the invention. There are two main differences between functional unit 2302 and convolver unit 806. First, functional unit 2302 has the ability to compute the maximum of a sum (needed to perform the max pool operation). Second, functional unit 2302 has the ability to compute the rectification of a value. In order to compute the maximum of a sum, each of the nine adders (2104a, . . . , 2104i) of the convolver unit may be replaced with a function selector (2404a, . . . , 2404i). The function selector receives control signal s2, allowing the selection between an adder and a comparator (see inset in FIG. 24). With the adder selected, the functional unit, for the most part, is transformed back into the hardware architecture of convolver unit 806, and functional unit 2302 is configured to perform the above-described convolution operation. With the comparator selected, functional unit 2302 is configured to compute max(w₁x₁, w₂x₂, w₃x₃, w₄x₄, w₅x₅, w₆x₆, w₇x₇, w₈x₈, w₉x₉) when control signal s1 is set to 0, and max(w₁x₁, w₂x₂, w₃x₃, w₄x₄, w₅x₅, w₆x₆, w₇x₇, w₈x₈, w₉x₉, previous partial sum) when control signal s1 is set to 1. Therefore, when operating the convolutional engine 2302 in a manner similar to FIGS. 8-18C, except with the comparator selected, the maximum of the pointwise multiplication of a three dimensional filter (e.g., f₁) with a three dimensional volume of input (i.e., a volume of input that aligns with the filter—as described in FIG. 6) may be computed. It should now be apparent that the max pool operator may be implemented with the comparators of a functional unit selected and the stride set equal to the magnitude of one dimension of a kernel of the filter (e.g., for a 3×3 kernel, the stride would be set to be 3).

When the control signal s1 is set to 2, functional unit is configured to perform the rectification operation. Control signal s1 being set to 2 causes output selector 2406 to provide the value stored in one or more of the accumulators 1104a, 1104b, . . . to rectifier 2408, which performs the following rectification operation:

$\mathrm{rect}\left(x\right)=\{\begin{array}{c}x\mathrm{for}x>0\\ 0\mathrm{for}x\le 0\end{array}$

In a data representation in which a data value is represented by a zero bit (indicating whether the data value is 0), a sign bit (indicating the whether the data value is positive or negative) and the magnitude (indicating the magnitude of the data value), rectifier 2408 can be configured to return 0 whenever the sign bit indicates a negative number or if the zero bit is set, and return the magnitude otherwise.

When the control signal s1 is set to 3, functional unit is configured to add a bias value to the data stored in accumulators 1104a, 1104b, etc. similar to the operation of convolver unit 806.

FIG. 25 depicts three scenarios of data values being loaded from input channel 702a into convolutional engine 708 having m columns of convolver units, with scenario (a) illustrating input channel 702a having m columns of data values, scenario (b) illustrating input channel 702a having 3m−4 columns of data values, and scenario (c) illustrating input channel 702a having m/2 columns of data values, in accordance with one embodiment of the invention. Scenario (a) was previously described in FIG. 9B, but will be more fully discussed in FIGS. 26A-26B. Scenario (b) discusses an example in which the number of columns of input channel 702a is greater than the number of columns of the convolver array. Scenario (c) discusses an example in which the number of columns of input channel 702a is less than the number of columns of the convolver array. While a convolutional engine is more abstractly depicted, it should be understood that the architecture of a convolutional engine may be similar to earlier described examples, with a 2-D shift register and a convolver array.

FIG. 26A depicts the loading of a zero padding row, horizontal stripe 902a and a data padding row (corresponding to row n of horizontal stripe 902b) into convolutional engine 708. (If not already apparent, the bolded dashed rectangle denotes the portion of input channel 702a being loaded into convolutional engine 708.) More specifically, the zero padding row is first loaded into the 2-D shift register of convolutional engine 708, followed by row n of horizontal stripe 902a, followed by row n−1 of horizontal stripe 902a, . . . followed by row 1 of horizontal stripe 902a, and followed by the data padding row. As described above, each time a row of data storage elements stores row n of a horizontal stripe, the convolver units corresponding to that row of data storage elements are activated. Each time row 1 of a horizontal stripe is shifted out of a row of data storage elements, the convolver units corresponding to that row of data storage elements are de-activated.

FIG. 26B depicts the loading of one data padding row (corresponding to row 1 of horizontal stripe 902a), horizontal stripe 902b and a zero padding row into convolutional engine 708. More specifically, the data padding row is first loaded into the 2-D shift register of convolutional engine 708, followed by row n of horizontal stripe 902b, followed by row n−1 of horizontal stripe 902b, . . . followed by row 1 of horizontal stripe 902b, and followed by the zero padding row.

While input channel 702a included two horizontal stripes to illustrate the concept of a single “horizontal cut line” through the input data (conceptually located at the boundary of horizontal stripes 902a and 902b), it is understood that an input channel would have more horizontal stripes if there were more horizontal cut lines. For a horizontal stripe that is bordered above and below by other horizontal stripes, the loading of that horizontal stripe would be preceded by a data padding row and followed by another data padding row.

FIGS. 27A-27C illustrate a scenario in which “vertical cut lines” through input channel 702a are needed, and how to handle the vertical cut lines. Generally, a vertical cut line is needed whenever the number of columns of the input channel is greater than the number of columns of the convolver array. The present example discusses the scenario in which the number of columns of the input channel is equal to 3m−4, where in is the number of columns of the convolver array. Whenever the number of columns of the input channel is equal to two more than a multiple of m−2 (as is true in the present example), the convolver array is utilized in an efficient manner (no unused convolver units), but if this relationship does not hold, the concepts described below still apply, but the convolver array will be utilized in a less efficient manner (will have unused convolver units). Further, for the sake of clarity of illustration and explanation, horizontal cut lines, zero padding rows, and data padding rows are not discussed in the example of FIGS. 27A-27C. Nevertheless, it is expected that one of ordinary skill in the art will be able to combine concepts from FIGS. 26A-26B and 27A-27B in order to handle scenarios in which there are both horizontal and vertical cut lines.

In FIG. 27A, input channel 702a is divided into vertical stripes 906a, 906b and 906c. Conceptually, one can imagine a first vertical cut line separating vertical stripe 906a from vertical stripe 906b, and a second vertical cut line separating vertical stripe 906b from 906c. In an efficient use of the convolutional engine, interior vertical stripes (such as 906b) contain m−2 columns, whereas exterior vertical stripes (such as 906a and 906c) contain m−1 columns. FIG. 27A depicts in columns (including the m−1 columns of vertical stripe 906a and one data padding column) being loaded into convolutional engine 708. The right most column of convolver units (which aligns with the data padding column) is non-active, as the output of these convolver units would have produced a convolution output treating the data padding column as an external column (which is not true in the current scenario). The remaining m−1 columns of the convolver units operate in a similar manner as the convolver units that have been previously described.

FIG. 27B depicts in columns (including the m−2 columns of vertical stripe 906b bordered on the right and left sides by a data padding column) being loaded into convolutional engine 708. The left most and right most columns of convolver units (which align with the data padding columns) are non-active, for reasons similar to those provided above. The remaining m−2 columns of the convolver units operate in a similar manner as the convolver units that have been previously described.

FIG. 27C depicts in columns (including one data padding column and the m−1 columns of vertical stripe 906c) being loaded into convolutional engine 708. The left most column of convolver units (which aligns with the data padding column) is non-active, for reasons similar to those provided above. The remaining m−1 columns of the convolver units operate in a similar manner as the convolver units that have been previously described.

FIG. 28 describes the scenario in which the number of columns of the input channel 702a is equal to m/2, in which in is the number of columns of the convolutional engine. The variable in is assumed to be an even number for the example of FIG. 28, but need not be an even number in general. Whenever the number of columns of the input channel is equal to a divisor of in (as is true in the present example), the convolver array is utilized in an efficient manner (i.e., will have no unused convolver units), but if this relationship does not hold, the concepts described below still apply, but the convolver array will be utilized in a less efficient manner (i.e., will have unused convolver units).

The example of FIG. 28 illustrates the concept of a “vertical cut line” through the convolutional engine 708, in which there is no transfer of data between region 708a (which includes the first half of the “columns” of the convolutional engine) and region 708b (which includes the second half of the “columns” of the convolutional engine). The term column, when used in the context of a convolutional engine, includes a column of the 2-D shift register and the corresponding column of convolutional units. Conceptually, one can imagine a vertical cut line that separates region 708a from region 708b. Region 708a essentially functions independently from region 708b, allowing region 708a to be configured to perform a convolution with a first set of filters (e.g., filters 1 through 10), and region 708b to be configured to perform the convolution with a second set of filters (e.g., filters 11-20). The number of filters (10 in each region) was chosen for clarity of explanation, and it is understood that there could have been a different number of filters in one or both of the two regions. The partitioning of the convolutional engine into independent regions (2 regions in this case, but could be more regions in other cases), allows for an increased throughput for the convolutional engine (in this case a doubling of the throughput).

As a concrete example, suppose that convolutional engine has 14 columns of convolver units. The weights of filter 1 would be loaded in each of convolver units 1-7 of the first row of convolver units, and the weights of filter 11 would be loaded in each of convolver units 8-14 of the first row of convolver units.

To configure convolutional engine 708 to operate with a “vertical cut line”, convolver units in the right most column of region 708a have weights w₃, w₆ and w₉ set to zero (regardless of what those weights might be from the filter kernels), and convolver units in the left most column of region 708b have weights w₁, w₄ and w₇ set to zero (regardless of what those weights might be from the filter kernels). Such setting of weights results in the data flow shown in the inset of FIG. 28, in which convolver units in the right most column of region 708a do not receive any data values from its “right neighbors”, and the convolver units in the left most column of region 708b do not receive any data values from its “left neighbors”.

When input channel 702a is loaded into convolutional engine 708, it is loaded into region 708a row-by-row, and at the same time, it is loaded into region 708b row-by-row. If the propagation of data through convolutional engine 708 could conceptually be viewed as a ticker tape traversing in the vertical direction, there would be one ticker tape traversing down region 708a, and there would be a mirror image of that ticker tape traversing down region 708b.

While FIG. 28 illustrated an example with one vertical cut line through the convolutional engine, it should be apparent how a convolutional engine could be modified to have multiple vertical cut lines. Further, for the sake of clarity of illustration and explanation, horizontal cut lines, zero padding rows, and data padding rows are not discussed in the example of FIG. 28. Nevertheless, it is expected that one of ordinary skill in the art will be able to combine concepts from FIGS. 26A-26B and 28 together to handle scenarios in which there are both horizontal and vertical cut lines.

FIGS. 29A-29B illustrate another scheme for handling the scenario in which the number of columns of the input channel 702a is equal to m/2, in which m is the number of columns of convolutional engine 708. The scheme involves combining the concept of a horizontal cut line through the input data (described in FIGS. 26A-26B) and the concept of a vertical cut line through the convolver array (described in FIG. 28). In the FIGS. 26A-26B, the two horizontal stripes were processed one after another (i.e., serially). However, in the example of FIGS. 29A-29B, the horizontal stripes 908a and 908b are processed in parallel, with horizontal stripe 908a processed in region 708a, and horizontal stripe 908b processed in region 708b. The same filters are populated in regions 708a and 708b, in contrast to the scheme of FIG. 28.

Since there are several overlapping rectangles in FIG. 29A, the scheme is conceptually redrawn in FIG. 29B, which more clearly shows the data loaded into region 708a and region 708b. If not already apparent, it is noted that row 1 of horizontal stripe 908a is identical to the data padding row that precedes horizontal stripe 908b, and the data padding row that follows horizontal stripe 908a is identical to row n of horizontal stripe 908b.

Similar to the scheme of FIG. 28, the scheme of FIGS. 29A-29B also has the effect of doubling the throughput. At this point, since there are two possible schemes for handling the m/2 scenario, one might wonder which scheme is preferable. One consideration between the scheme of FIG. 28 and the scheme of FIGS. 29A-29B is the number of filters versus the number of rows of the input channel. If there are many more filters than the number of rows of the input channel, then the scheme of FIG. 28 might be preferred, whereas if there are many more rows of the input channel than the number of filters, then the scheme of FIGS. 29A-29B might be preferred. Intuitively, the former case would be analogous to a long skinny column of filters, in which it would be advantageous to cut the long skinny column of filters in half (place one half in region 708a and the other half in region 708b), whereas the latter case would be analogous to a long skinny column of input data, in which it would be advantageous to cut the long skinny column of input data in half and process the two halves of input data in parallel.

Other considerations for favoring one scheme over the another might also include the number of filters relative to the number of rows of convolver units. If the number of filters were less than the number of rows of convolver units, then the scheme of FIG. 29A-29B might be preferred, whereas if the number of filters were more than the number of rows of convolver units, then the scheme of FIG. 28 might be preferred.

FIG. 30 depicts convolutional engine 708 as one component of system 3000, in accordance with one embodiment of the invention. System 3000 may include memory 3002, shift and format module 3004, convolutional engine 708 and controller 3006.

Memory 3002 may be implemented using static random-access memory (SRAM), and may store input data 702, and the output of the convolutional engine 708 (e.g., convolution output, max pool output, rectified output, etc.).

Shift and format module 3004 is an interface between memory 3002 and convolutional engine 708 and is configured to shift and format the data. For instance, in the example of FIG. 29A, providing horizontal stripe 908b to region 708b of the convolutional engine would be one task performed by shift and format module 3004. Achieving a stride of ½ (or a stride less than one) could also involve shift and format module 3004, in which the above-described interpolation could be performed by the shift and format module 3004.

In the embodiment of FIG. 30, convolutional engine 708 contains a more typical number of data storage elements and convolver units. FIG. 30 depicts a convolutional engine with a 64 by 256 array of convolver units 806 and a 66 by 256 array of data storage elements configured as a 2-D shift register. Similar to the previously-described embodiments, the first row of convolver units logically corresponds with the second row of data storage elements, and the last row of convolver units logically corresponds with the second to last row of data storage elements.

Controller 3006 may be responsible for performing many of the above-described control operations. For example, controller 3006 may provide the control signals that set convolver units to be active and non-active (and hence, the above-described controller 2202 may be part of controller 3006). Controller 3006 may be responsible for providing control signal s1 (described in FIGS. 21 and 24) for controlling the output of output selectors 2106 and 2406. Controller 3006 may be responsible for providing control signal s2 (described in FIG. 24) for controlling whether a functional unit is programmed to output a convolution output or a max pool output. Controller 3006 may logically partition an input channel into horizontal stripes, and/or vertical stripes (more appropriately called chunks when there are vertical and horizontal cut lines) based on the dimensions of the input channel relative to the dimensions of the convolver array. Controller 3006 may control shift and format module 3004 to perform the necessary shift and format operations. Controller 3006 may determine which weights are to be loaded to which convolutional units. Controller 3006 may determine whether to override filter weights with zero values in order to logically partition the convolutional engine into multiple independent regions (as depicted in FIGS. 28, 29A and 29B). Controller 3006 may also contain the logic that determines, for the loading of a horizontal stripe into the convolutional engine, whether the horizontal stripe is to be preceded by a zero padding row or a data padding row, or whether the horizontal stripe is to be followed by a zero padding row or a data padding row. These are merely some examples of the functions that may be performed by controller 3006.

FIG. 31 depicts a block diagram of weight decompressor 3100 for decompressing filter weights before the weights are provided to the convolver units, in accordance with one embodiment of the invention. Weight decompressor 3100 may utilize dictionary 3102 to decompress weights. In one embodiment, compressed weights are keys to a look-up table (an embodiment of the dictionary), and the records corresponding to the keys in the look-up table are the decompressed weights. The 256 convolver units may be logically and/or physically grouped into 16 groups, each group including 16 convolver units. The decompressed weights may be provided to each of the 16 groups of convolver units.

Thus, an efficient convolutional engine has been described. In one embodiment, the convolutional engine includes a two-dimensional shift register having a three by four array of data storage elements:

$\left[\begin{array}{cccc}{d}_{1,1}& d_{1,2}& d_{1,3}& d_{1,4}\\ d_{2,1}& d_{2,2}& d_{2,3}& d_{2,4}\\ d_{3,1}& d_{3,2}& d_{3,3}& d_{3,4}\end{array}\right]$

wherein, at a first moment in time,

data storage element d_1,1 stores data value x_1,1,

data storage element d_1,2 stores data value x_1,2,

data storage element d_1,3 stores data value x_1,3,

data storage element d_1,4 stores data value x_1,4,

data storage element d_2,1 stores data value x_2,1,

data storage element d_2,2 stores data value x_2,2,

data storage element d_2,3 stores data value x_2,3,

data storage element d_2,4 stores data value x_2,4,

data storage element d_3,1 stores data value x_3,1,

data storage element d_3,2 stores data value x_3,2,

data storage element d_3,3 stores data value x_3,3, and

data storage element d_3,4 stores data value x_3,4.

The convolutional engine further includes a first convolver unit having a first plurality of multipliers, m_1,1¹, m_1,2¹, m_1,3¹, m_2,1¹, m_2,2¹, m_2,3¹, m_3,1¹, m_3,2¹, and m_3,3¹, wherein:

• • multiplier m_1,1¹ is electrically coupled to data storage element d_1,1, and is configured to multiply data value x_1,1 with weight w₁ so as to generate a product w₁x_1,1,
  • multiplier m_1,2¹ is electrically coupled to data storage element d_1,2, and is configured to multiply data value x_1,2 with weight w₂ so as to generate a product w₂x_1,2,
  • multiplier m_1,3¹ is electrically coupled to data storage element d_1,3, and is configured to multiply data value x_1,3 with weight w₃ so as to generate a product w₃x_1,3,
  • multiplier m_2,1¹ is electrically coupled to data storage element d_2,1, and is configured to multiply data value x_2,1 with weight w₄ so as to generate a product w₄x_2,1,
  • multiplier m_2,2¹ is electrically coupled to data storage element d_2,2, and is configured to multiply data value x_2,2 with weight w₅ so as to generate a product w₅x_2,2,
  • multiplier m_2,3¹ is electrically coupled to data storage element d_2,3, and is configured to multiply data value x_2,3 with weight w₆ so as to generate a product w₆x_2,3,
  • multiplier m_3,1¹ is electrically coupled to data storage element d_3,1, and is configured to multiply data value x_3,1 with weight w₇ so as to generate a product w₇x_3,1,
  • multiplier m_3,2¹ is electrically coupled to data storage element d_3,2, and is configured to multiply data value x_3,2 with weight w₈ so as to generate a product w₈x_3,2, and
  • multiplier m_3,3¹ is electrically coupled to data storage element d_3,3, and is configured to multiply data value x_3,3 with weight w₉ so as to generate a product w₉x_3,3.The convolutional engine further includes a second convolver unit comprising a second plurality of multipliers, m_1,1², m_1,2², m_2,1², m_2,2², m_2,3², m_3,1², m_3,2², and m_3,3², wherein:
  • multiplier m_1,1² is electrically coupled to data storage element d_1,2, and is configured to multiply data value x_1,2 with weight w, so as to generate a product w₁x_1,2,
  • multiplier m_1,2² is electrically coupled to data storage element d_1,3, and is configured to multiply data value x_1,3 with weight w₂ so as to generate a product w₂x_1,3,
  • multiplier m_1,3² is electrically coupled to data storage element d_1,4, and is configured to multiply data value x_1,4 with weight w₃ so as to generate a product w₃x_1,4,
  • multiplier m_2,1² electrically coupled to data storage element d_2,2, and is configured to multiply data value x_2,2 with weight w₄ so as to generate a product w₄x_2,2,
  • multiplier m_2,2² is electrically coupled to data storage element d_2,3, and is configured to multiply data value x_2,3 with weight w₅ so as to generate a product w₅x_2,3,
  • multiplier m_2,3² is electrically coupled to data storage element d_2,4, and is configured to multiply data value x_2,4 with weight w₆ so as to generate a product w₆x_2,4,
  • multiplier m_3,1² is electrically coupled to data storage element d_3,2, and is configured to multiply data value x_3,2 with weight w₇ so as to generate a product w₇x_3,2,
  • multiplier m_3,2² is electrically coupled to data storage element d_3,3, and is configured to multiply data value x_3,3 with weight w₈ so as to generate a product w₈x_3,3, and
  • multiplier m_3,3² is electrically coupled to data storage element d_3,4, and is configured to multiply data value x_3,4 with weight w₉ so as to generate a product w₉x_3,4.

In various embodiments, the first convolver unit may be configured to generate a sum of terms, the terms including at least the product w₁x_1,1, the product w₂x_1,2, the product w₃x_1,3, the product w₄x_2,1, the product w₅x_2,2, the product w₆x_2,3, the product w₇x_3,1, the product w₈x_3,2, the product w₉x_3,3 and b₁, wherein b₁ is a bias value. Further, the second convolver unit may be configured to compute a sum of terms, the terms including at least the product w₁x_1,2, the product w₂x_1,3, the product w₃x_1,4, the product w₄x_2,2, the product w₅x_2,3, the product w₆x_2,4, the product w₇x_3,2, the product w₈x_3,3, the product w₉x_3,4 and b₁, wherein b₁ is a bias value.

In some instances:

data storage element d_1,1 is electrically coupled to data storage element d_2,1,

data storage element d_2,1 is electrically coupled to data storage element d_3,1,

data storage element d_1,2 is electrically coupled to data storage element d_2,2,

data storage element d_2,2 is electrically coupled to data storage element d_3,2,

data storage element d_1,3 is electrically coupled to data storage element d_2,3,

data storage element d_2,3 is electrically coupled to data storage element d_3,3,

data storage element d_1,4 is electrically coupled to data storage element d_2,4, and

data storage element d_2,4 is electrically coupled to data storage element d_3,4.

Further embodiments of the invention provide an apparatus, that includes a two-dimensional shift register having a four by four array of data storage elements:

$\left[\begin{array}{cccc}{d}_{1,1}& d_{1,2}& d_{1,3}& d_{1,4}\\ d_{2,1}& d_{2,2}& d_{2,3}& d_{2,4}\\ d_{3,1}& d_{3,2}& d_{3,3}& d_{3,4}\\ d_{4,1}& d_{4,2}& d_{4,3}& d_{4,4}\end{array}\right]$

wherein, at a first moment in time,

• • data storage element d_1,1 stores data value x_1,1,
  • data storage element d_1,2 stores data value x_1,2,
  • data storage element d_1,3 stores data value x_1,3,
  • data storage element d_1,4 stores data value x_1,4,
  • data storage element d_2,1 stores data value x_2,1,
  • data storage element d_2,2 stores data value x_2,2,
  • data storage element d_2,3 stores data value x_2,3,
  • data storage element d_2,4 stores data value x_2,4,
  • data storage element d_3,1 stores data value x_3,1,
  • data storage element d_3,2 stores data value x_3,2,
  • data storage element d_3,3 stores data value x_3,3,
  • data storage element d_3,4 stores data value x_3,4,
  • data storage element d_4,1 stores data value x_4,1,
  • data storage element d_4,2 stores data value x_4,2,
  • data storage element d_4,3 stores data value x_4,3, and
  • data storage element d_4,4 stores data value x_4,4.The apparatus also includes a first convolver unit comprising a first plurality of multipliers, m_1,1¹, m_1,2¹, m_1,3¹, m_2,1¹, m_2,2¹, m_2,3¹, m_3,1¹, m_3,2¹, and m_3,3¹, wherein:
  • multiplier m_1,1¹ is electrically coupled to data storage element d_1,1, and is configured to multiply data value x_1,1 with weight w, so as to generate a product w₁x_1,1,
  • multiplier m_1,2¹ is electrically coupled to data storage element d_1,2, and is configured to multiply data value x_1,2 with weight w₂ so as to generate a product w₂x_1,2,
  • multiplier m_1,3¹ is electrically coupled to data storage element d_1,3, and is configured to multiply data value x_1,3 with weight w₃ so as to generate a product w₃x_1,3,
  • multiplier m_2,1¹ is electrically coupled to data storage element d_2,1, and is configured to multiply data value x_2,1 with weight w₄ so as to generate a product w₄x_2,1,
  • multiplier m_2,2¹ is electrically coupled to data storage element d_2,2, and is configured to multiply data value x_2,2 with weight w₅ so as to generate a product w₅x_2,2,
  • multiplier m_2,3¹ is electrically coupled to data storage element d_2,3, and is configured to multiply data value x_2,3 with weight w₆ so as to generate a product w₆x_2,3,
  • multiplier m_3,1¹ is electrically coupled to data storage element d_3,1, and is configured to multiply data value x_3,1 with weight w₇ so as to generate a product w₇x_3,1,
  • multiplier m_3,2¹ is electrically coupled to data storage element d_3,2, and is configured to multiply data value x_3,2 with weight w₈ so as to generate a product w₈x_3,2,
  • and
  • multiplier m_3,3¹ is electrically coupled to data storage element d_3,3, and is configured to multiply data value x_3,3 with weight w₉ so as to generate a product w₉x_3,3.The apparatus also includes a second convolver unit comprising a second plurality of multipliers, m_1,1², m_1,2², m_1,3², m_2,1², m_2,2², m_2,3², m_3,1², m_3,2², and m_3,3², wherein:
  • multiplier m_1,1² is electrically coupled to data storage element d_1,2, and is configured to multiply data value x_1,2 with weight w₁ so as to generate a product w₁x_1,2,
  • multiplier m_1,2² is electrically coupled to data storage element d_1,3, and is configured to multiply data value x_1,3 with weight w₂ so as to generate a product w₂x_1,3,
  • multiplier m_1,3² is electrically coupled to data storage element d_1,4, and is configured to multiply data value x_1,4 with weight w₃ so as to generate a product w₃x_1,4,
  • multiplier m_2,1² is electrically coupled to data storage element d_2,2, and is configured to multiply data value x_2,2 with weight w₄ so as to generate a product w₄x_2,2,
  • multiplier m_2,2² is electrically coupled to data storage element d_2,3, and is configured to multiply data value x_2,3 with weight w₅ so as to generate a product w₅x_2,3,
  • multiplier m_2,3² is electrically coupled to data storage element d_2,4, and is configured to multiply data value x_2,4 with weight w₆ so as to generate a product w₆x_2,4,
  • multiplier m_3,1² is electrically coupled to data storage element d_3,2, and is configured to multiply data value x_3,2 with weight w₇ so as to generate a product w₇x_3,2,
  • multiplier m_3,2² is electrically coupled to data storage element d_3,3, and is configured to multiply data value x_3,3 with weight w₈ so as to generate a product w₈x_3,3, and
  • multiplier m_3,3² is electrically coupled to data storage element d_3,4, and is configured to multiply data value x_3,4 with weight w₉ so as to generate a product w₉x_3,4,
  • a third convolver unit comprising a third plurality of multipliers, m_1,1³, m_1,2³, m_1,3³, m_2,1³, m_2,2³, m_2,3³, m_3,1³, m_2,2³, and m_3,3³, wherein:
  • multiplier m_1,1³ is electrically coupled to data storage element d_2,1, and is configured to multiply data value x_2,1 with weight w₁₀ so as to generate a product w₁₀x_2,1,
  • multiplier m_1,2³ is electrically coupled to data storage element d_2,2, and is configured to multiply data value x_2,2 with weight w₁₁ so as to generate a product w₁₁x_2,2,
  • multiplier m_1,3³ is electrically coupled to data storage element d_2,3, and is configured to multiply data value x_2,3 with weight w₁₂ so as to generate a product w₁₂x_2,3,
  • multiplier m_2,1³ is electrically coupled to data storage element d_3,1, and is configured to multiply data value x_3,1 with weight w₁₃ so as to generate a product w₁₃x_3,1,
  • multiplier m_2,2³ is electrically coupled to data storage element d_3,2, and is configured to multiply data value x_3,2 with weight w₁₄ so as to generate a product w₁₄x_3,2,
  • multiplier m_2,3³ is electrically coupled to data storage element d_3,3, and is configured to multiply data value x_3,3 with weight w₁₅ so as to generate a product w₁₅x_3,3,
  • multiplier m_3,1³ is electrically coupled to data storage element d_4,1, and is configured to multiply data value x_4,1 with weight w₁₆ so as to generate a product w₁₆x_4,1,
  • multiplier m_3,2³ is electrically coupled to data storage element d_4,2, and is configured to multiply data value x_4,2 with weight w₁₇ so as to generate a product w₁₇x_4,2, and
  • multiplier m_3,3³ is electrically coupled to data storage element d_4,3, and is configured to multiply data value x_4,3 with weight w₁₈ so as to generate a product w₁₈x_4,3.And the apparatus includes a fourth convolver unit comprising a fourth plurality of multipliers, m_1,4⁴, m_1,2⁴, m_1,3⁴, m_2,1⁴, m_2,2⁴, m_2,3⁴, m_3,1⁴, m_3,2⁴, and m_3,3⁴, wherein:
  • multiplier m_1,1⁴ is electrically coupled to data storage element d_2,2, and is configured to multiply data value x_2,2 with weight w₁₀ so as to generate a product w₁₀x_2,2,
  • multiplier m_1,2⁴ is electrically coupled to data storage element d_2,3, and is configured to multiply data value x_2,3 with weight w₁₁ so as to generate a product w₁₁x_2,3,
  • multiplier m_1,3⁴ is electrically coupled to data storage element d_2,4, and is configured to multiply data value x_2,4 with weight w₁₂ so as to generate a product w₁₂x_2,4,
  • multiplier m_2,1⁴ is electrically coupled to data storage element d_3,2, and is configured to multiply data value x_3,2 with weight w₁₃ so as to generate a product w₁₃x_3,2,
  • multiplier m_2,2⁴ is electrically coupled to data storage element d_3,3, and is configured to multiply data value x_3,3 with weight w₁₄ so as to generate a product w₁₄x_3,3,
  • multiplier m_2,3⁴ is electrically coupled to data storage element d_3,4, and is configured to multiply data value x_3,4 with weight w₁₅ so as to generate a product w₁₅x_3,4,
  • multiplier m_3,1⁴ is electrically coupled to data storage element d_4,2, and is configured to multiply data value x_4,2 with weight w₁₆ so as to generate a product w₁₆x_4,2,
  • multiplier m_3,2⁴ is electrically coupled to data storage element d_4,3, and is configured to multiply data value x_4,3 with weight w₁₇ so as to generate a product w₁₇x_4,3, and
  • multiplier m_3,3⁴ is electrically coupled to data storage element d_4,4, and is configured to multiply data value x_4,4 with weight w₁₈ so as to generate a product w₁₈x_4,4.

In some embodiments, the first convolver unit of this apparatus may be configured to generate a sum of terms, the terms including at least the product w₁x_1,1, the product w₂x_1,2, the product w₃x_1,3, the product w₄x_2,1, the product w₅x_2,2, the product w₆x_2,3, the product w₇x_3,1, the product w₈x_3,2, the product w₉x_3,3 and b₁, wherein b₁ is a bias value. Also, the second convolver unit may be configured to compute a sum of terms, the terms including at least the product w₁x_1,2, the product w₂x_1,3, the product w₃x_1,4, the product w₄x_2,2, the product w₅x_2,3, the product w₆ x_2,4, the product w₇ x_3,2, the product w₈ x_3,3, the product w₉x_3,4 and b₁, wherein b₁ is a bias value.

In still further embodiments, the third convolver unit may be configured to generate a sum of terms, the terms including at least the product w₁₀x_2,1, the product w₁₁x_2,2, the product w₁₂x_2,3, the product w₁₃ x_3,1, the product w₁₄x_3,2, the product w₁₅x_3,3, the product w₁₆x_4,1, the product w₁₇x_4,2, the product w₁₈x_4,3 and b₂, wherein b₂ is a bias value. Also, the fourth convolver unit may be configured to compute a sum of terms, the terms including at least the product w₁₀x_2,2, the product w₁₁x_2,3, the product w₁₂x_2,4, the product w₁₃x_3,2, the product w₁₄x_3,3, the product w₁₅x_3,4, the product w₁₆x_4,2, the product w₁₇x_4,3, the product w₁₈x_4,4 and b₂, wherein b₂ is a bias value.

In various embodiments:

data storage element d_1,1 is electrically coupled to data storage element d_2,1,

data storage element d_2,1 is electrically coupled to data storage element d_3,1,

data storage element d_3,1 is electrically coupled to data storage element d_4,1,

data storage element d_1,2 is electrically coupled to data storage element d_2,2,

data storage element d_2,2 is electrically coupled to data storage element d_3,2,

data storage element d_3,2 is electrically coupled to data storage element d_4,2,

data storage element d_1,3 is electrically coupled to data storage element d_2,3,

data storage element d_2,3 is electrically coupled to data storage element d_3,3,

data storage element d_3,3 is electrically coupled to data storage element d_4,3,

data storage element d_1,4 is electrically coupled to data storage element d_2,4,

data storage element d_2,4 is electrically coupled to data storage element d_3,4, and

data storage element d_3,4 is electrically coupled to data storage element d_4,4.

Still another embodiment of the invention provides an apparatus that includes a two-dimensional synchronous shift register comprising a p by q array of data storage elements:

$\left[\begin{array}{ccc}{d}_{1,1}& \dots & d_{1,q}\\ ⋮& \ddots & ⋮\\ d_{p,1}& \dots & d_{p,q}\end{array}\right]$

wherein a first row of data storage elements d_1,1, . . . , d_1,q receives q data values on each clock cycle and each row of data storage elements d_k,1, . . . , d_k,q receives q data values from a previous row of data storage elements d_k-1,1, . . . , d_k-1,q on each clock cycle, for 1<k≤p; and a convolver array configured to process the data values stored in the two-dimensional synchronous shift register, wherein the convolver array comprises a p−2 by q array of convolver units, wherein for convolver units CU_i,j, 1≤i≤p−2 and 2≤j≤q−1:

• • (i) a first input of CU_i,j is electrically coupled to data storage element d_i,j−1,
  • (ii) a second input of CU_i,j is electrically coupled to data storage element d_i+1,j−1,
  • (iii) a third input of CU_i,j is electrically coupled to data storage element d_i+2,j−1,
  • (iv) a fourth input of CU_i,j is electrically coupled to data storage element d_i,j,
  • (v) a fifth input of CU_i,j is electrically coupled to data storage element d_i+1,j,
  • (vi) a sixth input of CU_i,j is electrically coupled to data storage element d_i+2,j,
  • (vii) a seventh input of CU_i,j is electrically coupled to data storage element d_i,j+1,
  • (viii) an eighth input of CU_i,j is electrically coupled to data storage element d_i+1,j+1, and
  • (ix) a ninth input of CU_i,j is electrically coupled to data storage element d_i+2,j+1.

In some embodiments, for convolver units CU_i,1, 1≤i≤p−2,

• • (i) at least one of a first input of CU_i,1 or a weight associated with the first input is set to logical zero,
  • (ii) at least one of a second input of CU_i,1 or a weight associated with the second input is set to logical zero,
  • (iii) at least one of a third input of CU_i,1 or a weight associated with the third input is set to logical zero,
  • (iv) a fourth input of CU_i,1 is electrically coupled to data storage element d_i,1,
  • (v) a fifth input of CU_i,1 is electrically coupled to data storage element d_i+1,1,
  • (vi) a sixth input of CU_i,1 is electrically coupled to data storage element d_i+2,1,
  • (vii) a seventh input of CU_i,1 is electrically coupled to data storage element d_i,2,
  • (viii) an eighth input of CU_i,1 is electrically coupled to data storage element d_i+1,2, and
  • (ix) a ninth input of CU_i,1 is electrically coupled to data storage element d_i+2,2.

Further, in some embodiments, for convolver units CU_i,q, 1≤i≤p−2,

• • (i) a first input of CU_i,q is electrically coupled to data storage element d_i,q−1,
  • (ii) a second input of CU_i,q is electrically coupled to data storage element d_i+1,q−1,
  • (iii) a third input of CU_i,q is electrically coupled to data storage element d_i+2,q−1,
  • (iv) a fourth input of CU_i,q is electrically coupled to data storage element d_i,q,
  • (v) a fifth input of CU_i,q is electrically coupled to data storage element d_i+1,q,
  • (vi) a sixth input of CU_i,q is electrically coupled to data storage element d_i+2,q,
  • (vii) at least one of a seventh input of CU_i,q or a weight associated with the seventh input is set to logical zero,
  • (viii) at least one of an eighth input of CU_i,q or a weight associated with the eighth input is set to logical zero, and
  • (ix) at least one of a ninth input of CU_i,q or a weight associated with the ninth input is set to logical zero.

It is to be understood that the above-description is intended to be illustrative, and not restrictive. Many other embodiments will be apparent to those of skill in the art upon reviewing the above description. The scope of the invention should, therefore, be determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled.

Claims

1. A method for processing input data with a first input channel and a second input channel, the first input channel comprising a first horizontal stripe and a second horizontal stripe, the second input channel comprising a third horizontal stripe and a fourth horizontal stripe, the method comprising: processing, by a convolutional engine comprising a plurality of convolver units, the first horizontal stripe, the processing of the first horizontal stripe comprising computing a first plurality of dot products based on the first horizontal stripe and providing the first plurality of dot products to accumulators of the convolver units, wherein a total number of rows of the first horizontal stripe equals a total number of accumulators present in each of the convolver units;processing, by the convolutional engine, the third horizontal stripe, the processing of the third horizontal stripe comprising computing a second plurality of dot products based on the third horizontal stripe and summing the second plurality of dot products and the first plurality of dot products with the accumulators of the convolver units, wherein a total number of rows of the third horizontal stripe equals the total number of accumulators present in each of the convolver units;processing, by the convolutional engine, the second horizontal stripe, the processing of the second horizontal stripe comprising computing a third plurality of dot products based on the second horizontal stripe and providing the third plurality of dot products to the accumulators of the convolver units, wherein a total number of rows of the third horizontal stripe equals the total number of accumulators present in each of the convolver units; andprocessing, by the convolutional engine, the fourth horizontal stripe, the processing of the fourth horizontal stripe comprising computing a fourth plurality of dot products based on the fourth horizontal stripe and summing the third plurality of dot products and the fourth plurality of dot products with the accumulators of the convolver units, wherein a total number of rows of the fourth horizontal stripe equals the total number of accumulators present in each of the convolver units.

2. The method of claim 1, further comprising loading a zero padding row into the convolutional engine, followed by loading the first horizontal stripe into the convolutional engine, wherein computing the first plurality of dot products is based on the first horizontal stripe and the zero padding row.

3. The method of claim 1, further comprising loading the first horizontal stripe into the convolutional engine, followed by loading a data padding row into the convolutional engine, wherein computing the first plurality of dot products is based on the first horizontal stripe and the data padding row.

4. The method of claim 3, wherein the data padding row is identical to a leading row of the second horizontal stripe.

5. The method of claim 1, further comprising loading a zero padding row into the convolutional engine, followed by loading the third horizontal stripe into the convolutional engine, wherein computing the second plurality of dot products is based on the third horizontal stripe and the zero padding row.

6. The method of claim 1, further comprising loading the third horizontal stripe into the convolutional engine, followed by loading a data padding row into the convolutional engine, wherein computing the second plurality of dot products is based on the third horizontal stripe and the data padding row.

7. The method of claim 6, wherein the data padding row is identical to a leading row of the fourth horizontal stripe.

8. The method of claim 1, further comprising loading a data padding row into the convolutional engine, followed by loading the second horizontal stripe into the convolutional engine, wherein computing the third plurality of dot products is based on the second horizontal stripe and the data padding row.

9. The method of claim 8, wherein the data padding row is identical to a trailing row of the first horizontal stripe.

10. The method of claim 1, further comprising loading the second horizontal stripe into the convolutional engine, followed by loading a zero padding row into the convolutional engine, wherein computing the third plurality of dot products is based on the second horizontal stripe and the zero padding row.

11. The method of claim 1, further comprising loading a data padding row into the convolutional engine, followed by loading the fourth horizontal stripe into the convolutional engine, wherein computing the fourth plurality of dot products is based on the fourth horizontal stripe and the data padding row.

12. The method of claim 11, wherein the data padding row is identical to a trailing row of the third horizontal stripe.

13. The method of claim 1, further comprising loading the fourth horizontal stripe into the convolutional engine, followed by loading a zero padding row into the convolutional engine, wherein computing the fourth plurality of dot products is based on the fourth horizontal stripe and the zero padding row.