Systems and methods for processing convolutional neural network operations using textures

Abstract

Convolutional neural network information may define a convolutional neural network including layers. The layers may define operations on an input to the convolutional neural network. The layers in the convolutional neural network information may be formatted as shaders. Input information defining the input to the convolutional neural network may be accessed. The input information may be formatted as an array of textures. The shaders may be applied to the textures to effectuate processing the input to the convolutional neural network through the layers of the convolutional neural network. One or more results may be obtained from applying the shaders to the array of textures.

Description

FIELD

This disclosure relates to systems and methods that processes convolutional neural network operations using textures, wherein layers of a convolutional neural network are formatted as shaders.

BACKGROUND

Processing convolutional neural network operations may be time consuming. Parallelizing threads of convolutional neural network operations may increase the speed of processing convolutional neural network operations. Such parallelizing methods may be limited by the use of one-dimensional weights and/or buffers.

SUMMARY

This disclosure relates to processing convolutional neural network operations using textures. Convolutional neural network information may define a convolutional neural network including layers. The layers may define operations on an input to the convolutional neural network. The layers in the convolutional neural network information may be formatted as shaders. Input information defining the input to the convolutional neural network may be accessed. The input information may be formatted as an array of textures. The shaders may be applied to the textures to effectuate processing the input to the convolutional neural network through the layers of the convolutional neural network. One or more results may be obtained from applying the shaders to the array of textures.

A system that processes convolutional neural network operations using textures may include one or more of physical storage media, processors, and/or other components. In some implementations, the system may include one or more graphics processing units. In some implementations, the graphics processing unit may be included in a mobile device. In some implementations, the system may include a graphics processing unit memory. The physical storage media may store convolutional neural network information and/or other information. The convolutional neural network information may define one or more convolutional neural networks. A convolutional neural network may include one or more layers. The layers may define one or more operations on an input to the convolutional neural network. The layers in the convolutional neural network information may be formatted as shaders. In some implementations, the layers of the convolutional neural network may include one or more convolution layers. The convolution layers may be processed using binary convolution.

The processor(s) may be configured by machine-readable instructions. Executing the machine-readable instructions may cause the processor(s) to facilitate processing convolutional neural network operations using textures. The machine-readable instructions may include one or more computer program components. The computer program components may include one or more of an access component, a shader component, a result component, and/or other computer program components. In some implementations, the computer program component may include an input conversion component. In some implementation, the computer program component may include a shader conversion component.

The shader conversion component may be configured to access layer information and/or other information. The layer information may defining one or more layers of a convolutional neural network. The shader conversion component may be configured to format one or more operations defined by the layer(s) as one or more of the shaders and/or other information. The shader conversion component may be configured to add the shaders and/or other information to the convolutional neural network information.

The access component may be configured to access input information and/or other information. The input information may define one or more inputs to the convolutional neural network. The input information may be formatted as an array of textures. The textures may be characterized by a width, a height, a number of channels, and/or other characteristics. In some implementations, the number of channels of the textures may be four. In some implementations, the array of textures may be stored using an int32 structure. The int32 structure may include four int8 structures. An individual int8 structure may store a four by two patch of a single channel of an individual texture.

The input conversion component may be configured to access one or more inputs and/or other information. The input conversion component may be configured to format the input as array of textures. The input conversion component may be configured to store the array of textures as the input information.

The shader component may be configured to apply the shaders to the array of textures. Applying the shaders to the array of textures may effectuate processing the input(s) to the convolutional neural network through the layers of the convolutional neural network. In some implementations, individual shaders may output information for four channels. In some implementation, the shaders may be applied to the array of textures by the graphics processing unit. In some implementations, one or more inputs to the shaders and one or more outputs of the shaders may be stored in the graphics processing unit memory.

The result component may be configured to obtain one or more results from applying the shaders to the array of textures and/or other information. The result component may perform and/or obtain results from one or more processes that perform GPU to CPU conversion on the results of applying the shaders to the array of textures. The results component may obtain one or more results as one or more tensors, which may be used for additional operations

These and other objects, features, and characteristics of the system and/or method disclosed herein, as well as the methods of operation and functions of the related elements of structure and the combination of parts and economies of manufacture, will become more apparent upon consideration of the following description and the appended claims with reference to the accompanying drawings, all of which form a part of this specification, wherein like reference numerals designate corresponding parts in the various figures. It is to be expressly understood, however, that the drawings are for the purpose of illustration and description only and are not intended as a definition of the limits of the invention. As used in the specification and in the claims, the singular form of “a”, “an”, and “the” include plural referents unless the context clearly dictates otherwise.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 illustrates a system that processes convolutional neural network operations using textures.

FIG. 2 illustrates a method for processing convolutional neural network operations using textures.

FIG. 3A illustrates a formulation of convolutional layer output.

FIG. 3B illustrates a formulation of computing a weights average amplitude.

FIG. 3C illustrates a formulation of a convolution.

FIGS. 4A-4H illustrates operations of a convolutional neural network formatted as shader.

FIG. 5 illustrates example slices of a shader grid for a convolution operation.

FIG. 6 illustrates an example schematic showing weights storage in a 3D array.

FIG. 7 illustrates example int32 structure.

FIG. 8 illustrates an example processing flow using a shared memory.

DETAILED DESCRIPTION

FIG. 1 illustrates system 10 for processing convolutional neural network operations using textures. System 10 may include one or more of processor 11, storage media 12, interface 13 (e.g., bus, wireless interface), and/or other components. In some implementations, system 10 may include one or more graphics processing unit 14. In some implementations, graphics processing unit 14 may be included in a mobile device (e.g., smartphone, tablet, camera). In some implementations, system 10 may include graphics processing unit memory 15. Graphics processing unit memory 15 may be a part of graphics processing unit 14 (as shown in FIG. 1) and/or may be separate from graphics processing unit 14. Convolutional neural network information 20 may define a convolutional neural network including layers. The layers may define operations on an input to the convolutional neural network. The layers in convolutional neural network information 20 may be formatted as shaders. Input information defining the input to the convolutional neural network may be accessed. The input information may be formatted as an array of textures. The shaders may be applied to the textures to effectuate processing the input to the convolutional neural network through the layers of the convolutional neural network. One or more results may be obtained from applying the shaders to the array of textures.

Storage media 12 may be configured to include electronic storage medium that electronically stores information. Storage media 12 may store software algorithms, information determined by processor 11, information received remotely, and/or other information that enables system 10 to function properly. For example, electronic 12 may store information relating to convolutional neural network, layers, shaders, inputs, textures, and/or other information.

Storage media 12 may store convolutional neural network information 20 and/or other information. Convolutional neural network information 20 may define one or more convolutional neural networks. A convolutional neural network may refer to a neural network that receives an input and transforms the input through a series of layers. A convolutional neural network may include a series of linear and non-linear operations applied to an input tensor of data. A convolutional neural network may include one or more layers. For example, a convolutional neural network may include one or more of an input layer, an output layer, a convolution layer, a padding layer, a squeeze layer, an expand layer, a concatenation layer, a combine layer, a pooling layer, a normalization layer, a fully-connected layer, an activation layer, a drop layer, a flatten layer, and/or other layers. The layers may define one or more operations on an input to the convolutional neural network. The layers in the convolutional neural network information may be formatted as shaders. Shaders may refer programs/modules used to process shading—process color productions/visual transformation—within an image.

In some implementations, convolution layers may be processed using binary convolution. Binary convolution may provide for faster computational time of convolution at the expense of precision loss. For example, an input image (position (i, j), channel k) may be denoted as I_i,j,k. Convolutional layer output O_x,y,z may be defined as shown in FIG. 3A. In FIG. 3A, alpha may represent weights of a filter. Beta may represent a bias. Sigma may represent an activation function, such as hyperbolic tangent, rectified linear unit, leaky rectified linear unit, and/or other activation functions. W and H may denote the size of the filter, and D may denote the number of input channels. The principle of binary convolutions may include computing a weights average amplitude (as shown in FIG. 3B) and an input tensor subsampling amplitude I′ (same formula with/on the sampled patch of the formula for O in FIG. 3A). The convolution may be written as shown in FIG. 3C, where the binary values of the input tensor (+1 if positive; −1 if negative) is denoted as l_b and the binary version of the weights is denoted as a_b. The binary convolution (l_b star a_b) may be computed with bitwise operations (XOR). Uses of other convolutions are contemplated.

Processor 11 may be configured to provide information processing capabilities in system 10. As such, processor 11 may comprise one or more of a digital processor, an analog processor, a digital circuit designed to process information, a central processing unit, a graphics processing unit, a microcontroller, an analog circuit designed to process information, a state machine, and/or other mechanisms for electronically processing information. Processor 11 may be configured to execute one or more machine readable instructions 100 to facilitate processing convolutional neural network operations using textures. Machine readable instructions 100 may include one or more computer program components. Machine readable instructions 100 may include one or more of access component 102, shader component 104, result component 106, and/or other computer program components. In some implementations, machine readable instructions 100 may include input conversion component 108. In some implementation, machine readable instructions 100 may include shader conversion component 110.

Shader conversion component 110 may be configured to access layer information and/or other information. Layer information may defining one or more layers of a convolutional neural network. Shader conversion component 110 may access layer information from one or more storage locations. A storage location may include storage media 12, electronic storage of other device(s), and/or other locations. For example, shader conversion component 110 may access layer information stored in storage media 12. Shader conversion component 110 may access layer information associated with convolutional neural network information 20.

Shader conversion component 110 may be configured to format one or more operations defined by the layer(s) as one or more shaders and/or other information. One or more shaders may be characterized by parameters that define operations to be performed on one or more textures. For example, shaders may be characterized by biases parameters, weights parameters, input texture read parameters, output texture write parameters, and/or other parameters. FIGS. 4A-4H illustrates an example of operations of a convolutional neural network formatted as shader. Shader conversion component 110 may be configured to add the shaders and/or other information to convolutional neural network information 20.

Shader conversion component 110 may detect the layers (convolution layers and/or other layers) of the convolutional neural network and create shaders for the layers. In some implementations, the shaders may be optimized for the layer. Shader conversion component 110 may convert weights of the layers into texture. The weights of the convolutional operation may be stored in multichannel stack. The size of the shaders may be determined based on the size of the filter size. In some implementations, the size of the shaders may be constrained to a small area. Constraining the size of the shaders to a small area may provide for faster processing speed.

The shaders may write into output tensors based on input tensors. The shaders may compute multiple channels instead of a single channel. Use of shaders that processes on multiple channels may increase parallelization of processing convolutional neural network operations. For example, individual shaders may output information for four channels—individual shaders may compute four channels instead of a single channel. Computing on four channels may enable the shaders to sample approximately four times faster. Other numbers of channels are contemplated. Sampling may take longer than other aspects of the shaders operations in terms of time and increasing the speed of the sampling may result in faster processing of convolutional neural network operations.

For example, FIG. 5 illustrates example slices of a shader grid for a 1×1 convolution operation. Each shader may take all channels of each pixels to compute four channels of the output with four separate convolutional inside the same shader. The number of shaders may equal W×H×D/4, where W is the width of the input/output tensor, H the height of the input/output tensor, and D the number of channels of the output tensor. The number of parameters may equal D+C×D, where C is the number of input channels (weights and one bias per output filter). For convolutions on a bigger patch, for example for 3×3 patch, each shader may use all channels of 3×3 pixels to product four output channels. In such a case, the number of parameters may equal D+9+C×D. FIG. 6 illustrates an example schematic showing weights storage in a 3D array. Convolutional filter may be characterized by height (H), width (W), depth (C), and dimension (D). Biases vector may be characterized by dimension (D). In some implementations, Winograd type algorithms may be used to improve the computational time of convolutions.

Input information may define one or more inputs to a convolutional neural network. Inputs into a convolutional neural network may include visual content (e.g., one or more images, one or more videos). An input image may be characterized by an input tensor of three dimensions (three-dimensional data)—two spatial dimensions (e.g., x, y) and one depth/channel dimension (e.g., color space, such as RGB). Other types of input/dimensions are contemplated.

Input information may be formatted as one or more arrays of textures. The textures may be characterized by a width, a height, a number of channels, and/or other characteristics. In some implementations, the number of channels of the textures may be four. For example, three-dimensional data of an image input tensor may be stored in a Texture2DArray structure and/or other structures. The type of data inside the structure may be RGBA floating point 32 (four values) or half-floating point 16 (e.g., IEEE 754 half-precision binary floating-point format). Each point (x, y) may store a number of c channels and each channel may contain four values. Because each channel contains four values, the depth of the tensor may be the number of channels of the input image divided by four, and sampling may be approximately four times faster.

In some implementations, the array of textures may be stored using an int32 structure. The int32 structure may include four int8 structures. An individual int8 structure may store a four by two patch of a single channel of an individual texture. FIG. 7 illustrates two int32 structures. Int32 structures may be characterized by a width, a height, and a depth. FIG. 7 shows four int8 structures of the first int32 structure. First int32 structure may store 8 pixels and four channels. Each square may store a value in binary (e.g., carry a value of one or zero). Thus, each int32 may store eight spatial values for four channels.

Int32 structure may be used to store input, output, filters, and/or other values. Storage of values in spatial binary storage may facilitate computation of binary convolutions. Using a XOR operation, the output values for a convolution may be computed (e.g., summing the components of resulting int8 after product with the amplitudes). The binary format for an int8 may be “abcdefgh,” where each letter belongs to {0,1}. A transformation 2*(x−½) may be performed before and after the calculation to recenter the data after the XOR operation. The int32 structure may be used for 4×4 convolutions. Other size convolutions may use the int32 structure through padding (e.g., for 3×3 convolutions), through duplication (e.g., for multiples of 4×4 convolutions), or through both padding and duplication.

Access component 102 may be configured to access input information and/or other information. Access component 102 may access input information from one or more storage locations. A storage location may include storage media 12, electronic storage of one or more image sensors (not shown in FIG. 1), and/or other locations. For example, access component 102 may access input information stored in storage media 12. In some implementations, input information may define one or more visual content (e.g., image, video). Access component 102 may access input information defining one or more visual content during acquisition of the visual content and/or after acquisition of the visual content by one or more image sensors. For example, access component 102 may access input information defining an image/video while the image/video is being captured by one or more image sensors. Access component 102 may access input information defining an image/video after the image/video has been captured and stored in memory (e.g., storage media 12).

Input conversion component 108 may be configured to access one or more inputs and/or other information. For example, input conversion component 108 may be configured to access one or more of images, videos, and/or other information as input into a convolutional neural network. Input conversion component 108 may be configured to access inputs and/or other information during acquisition of the input/information and/or after acquisition of the input/information (e.g., by one or more image sensors). For example, input conversion component 108 may access input defining an image/video while the image/video is being captured by one or more image sensors. Input conversion component 108 may access input defining an image/video after the image/video has been captured and stored in memory (e.g., storage media 12).

Input conversion component 108 may be configured to format the input as array of textures. Input conversion component 108 may format the input using the Texture2DArray structure and/or the int32 structure. Inputs conversion component 108 may split input tensor into stacks of textures. Individual textures may contain information regarding the input on up to a certain number of channels. For example, individual textures may contain information regarding the input on up to four channels. For example, an input tensor with four channels may be stored in one texture. An input tensor with eight channels may be stored in two textures. An input tensor with five hundred channels may be stored in a hundred and twenty five textures. The depth of the number may equal the number of channels of the input divided by four. Other numbers of channels are contemplated. Input conversion component 108 may be configured to store the array of textures as input information. Input information stored by input conversion component 108 may be accessed by access component 102

Shader component 104 may be configured to apply the shaders to the array of textures. Shaders may be applied to the textures by sampling (extracting data from certain positions) and writing on output textures after a computation (a layer) is applied to the sampled data. In some implementations, sampling may include one-to-one sampling. In some implementations, sampling may include other operations, such as application of filters (e.g., bilateral filtering) and/or other operations.

Individual shaders may write to a set number of channels (e.g., four channels) of the output. For example, for writing to sixty-four channels in an output, sixteen shaders writing to four channels may be used. Applying the shaders to the array of textures may effectuate processing the input(s) to the convolutional neural network through the layers of the convolutional neural network. Parallelization of the shaders/textures in multiple channels may increase the speed of processing convolutional neural network operations.

The shaders may be applied to the array of textures by graphics processing unit 14. Graphics processing unit 14 may process textures/images in parallel. For example, graphics processing unit 14 may apply shaders to textures/images with RGBA channels in parallel. Graphics processing unit 14 may apply shaders to multiple textures/images in parallel. Parallelization of the processes in both space and depth may enable faster processing of the operations of the convolutional neural network. In some implementations, processing/parallelization by graphics processing unit 14 may be effectuated via one or more application programming interfaces/frameworks (e.g., Apple Metal).

In some implementations, one or more inputs to the shaders and one or more outputs of the shaders may be stored in graphics processing unit memory 15. Graphics processing unit memory 15 may be a part of graphics processing unit 14 (as shown in FIG. 1) and/or may be separate from graphics processing unit 14. Graphics processing unit memory 15 may provide a common buffer for inputs and outputs of one or more shaders. Use of graphics processing unit memory 15 to store and retrieve operations performed by the shaders may enable more efficient and faster storage/retrieval operations than if the inputs and/or outputs of the shaders are stored in other locations and/or processed by a CPU (e.g., processor 11). The shaders may take as input what is stored in graphics processing unit memory 15 (output of the prior shader) and may write output to graphics processing unit memory (for the next shader to take as input). Such use of graphics processing unit memory 15 as a common buffer may enable processing of one layer of a convolutional neural network to another layer of the convolutional neural network without returning to a CPU (e.g., processor 11).

For example, FIG. 8 illustrates an example processing flow using shared memory 850. In FIG. 8, CPU 800 may pass to shader A 810 an input. Shader A 810 may perform operations on the input and write its output in shared memory 850. Shader B 812 may take as its input the output written in shared memory 850 by shader A 810. Shader B 812 may perform operations on the input and write its output in shared memory 850. Shader C 814 may take as its input the output written in shared memory 850 by shader B 812. Shader C 814 may perform operations on the input and write its output in shared memory 850. Operations of the shaders may proceed in such fashion until the output of the final shader is sent back to CPU 800.

Result component 106 may be configured to obtain one or more results from applying the shaders to the array of textures and/or other information. Result component 106 may obtain one or more results from the application of the last shader to the array of textures. Result component 106 may be configured to obtain one or more results from one or more earlier applications of the shaders to the array of textures. Result component 106 may perform and/or obtain results from one or more processes that perform GPU to CPU conversion on the results of applying the shaders to the array of textures. Results component 106 may obtain one or more results as one or more tensors, which may be used for additional operations, such as calculation, display, printing, and/or other operations.

Although processor 11 is shown in FIG. 1 as a single entity, this is for illustrative purposes only. In some implementations, processor 11 may comprise a plurality of processing units. These processing units may be physically located within the same device, or processor 11 may represent processing functionality of a plurality of devices operating in coordination. Processor 11 may be configured to execute one or more components by software; hardware; firmware; some combination of software, hardware, and/or firmware; and/or other mechanisms for configuring processing capabilities on processor 11.

Although graphics processing unit 14 is shown in FIG. 1 as a single entity, this is for illustrative purposes only. In some implementations, graphics processing unit 14 may comprise a plurality of graphics processing units. These graphics processing units may be physically located within the same device, or graphics processing unit 14 may represent graphic processing functionality of a plurality of devices operating in coordination. Graphics processing unit 14 may be configured to execute one or more components by software; hardware; firmware; some combination of software, hardware, and/or firmware; and/or other mechanisms for configuring graphics processing capabilities on graphics processing unit 14

It should be appreciated that although computer components are illustrated in FIG. 1 as being co-located within a single processing unit, in implementations in which processor 11 comprises multiple processing units, one or more of computer program components may be located remotely from the other computer program components.

The description of the functionality provided by the different computer program components described herein is for illustrative purposes, and is not intended to be limiting, as any of computer program components may provide more or less functionality than is described. For example, one or more of computer program components 102, 104, 106, 108, and/or 110 may be eliminated, and some or all of its functionality may be provided by other computer program components. As another example, processor 11 may be configured to execute one or more additional computer program components that may perform some or all of the functionality attributed to one or more of computer program components 102, 104, 106, 108, and/or 110 described herein.

The electronic storage media of storage media 12 may be provided integrally (i.e., substantially non-removable) with one or more components of system 10 and/or removable storage that is connectable to one or more components of system 10 via, for example, a port (e.g., a USB port, a Firewire port, etc.) or a drive (e.g., a disk drive, etc.). Storage media 12 may include one or more of optically readable storage media (e.g., optical disks, etc.), magnetically readable storage media (e.g., magnetic tape, magnetic hard drive, floppy drive, etc.), electrical charge-based storage media (e.g., EPROM, EEPROM, RAM, etc.), solid-state storage media (e.g., flash drive, etc.), and/or other electronically readable storage media. Storage media 12 may be a separate component within system 10, or storage media 12 may be provided integrally with one or more other components of system 10 (e.g., processor 11). Although storage media 12 is shown in FIG. 1 as a single entity, this is for illustrative purposes only. In some implementations, storage media 12 may comprise a plurality of storage units. These storage units may be physically located within the same device, or storage media 12 may represent storage functionality of a plurality of devices operating in coordination.

The electronic storage media of graphics processing unit memory 15 may be provided integrally (i.e., substantially non-removable) with one or more components of system 10 and/or removable storage that is connectable to one or more components of system 10 via, for example, a port (e.g., a USB port, a Firewire port, etc.) or a drive (e.g., a disk drive, etc.). Graphics processing unit memory 15 may include one or more of optically readable storage media (e.g., optical disks, etc.), magnetically readable storage media (e.g., magnetic tape, magnetic hard drive, floppy drive, etc.), electrical charge-based storage media (e.g., EPROM, EEPROM, RAM, etc.), solid-state storage media (e.g., flash drive, etc.), and/or other electronically readable storage media. Graphics processing unit memory 15 may be a separate component within system 10, or graphics processing unit memory 15 may be provided integrally with one or more other components of system 10 (e.g., graphics processing unit 14). Although graphics processing unit memory 15 is shown in FIG. 1 as a single entity, this is for illustrative purposes only. In some implementations, graphics processing unit memory 15 may comprise a plurality of storage units. These storage units may be physically located within the same device, or graphics processing unit memory 15 may represent storage functionality of a plurality of devices operating in coordination.

FIG. 2 illustrates method 200 for processing convolutional neural network operations using textures. The operations of method 200 presented below are intended to be illustrative. In some implementations, method 200 may be accomplished with one or more additional operations not described, and/or without one or more of the operations discussed. In some implementations, two or more of the operations may occur substantially simultaneously.

In some implementations, method 200 may be implemented in one or more processing devices (e.g., a digital processor, an analog processor, a digital circuit designed to process information, a central processing unit, a graphics processing unit, a microcontroller, an analog circuit designed to process information, a state machine, and/or other mechanisms for electronically processing information). The one or more processing devices may include one or more devices executing some or all of the operation of method 200 in response to instructions stored electronically on one or more electronic storage mediums. The one or more processing devices may include one or more devices configured through hardware, firmware, and/or software to be specifically designed for execution of one or more of the operation of method 200.

Referring to FIG. 2 and method 200, at operation 201, input information may be accessed. The input information may define an input to a convolutional neural network. The input information may be formatted as an array of textures. The textures may be characterized by a width, a height, and a number of channels. The convolutional neural network may be defined by convolutional neural network information. The convolutional neural network may include layers. The layers may define operations on the input to the convolutional neural network. The layers in the convolutional neural network may be formatted as shaders. In some implementation, operation 201 may be performed by a processor component the same as or similar to access component 102 (Shown in FIG. 1 and described herein).

At operation 202, the shaders may be applied to the array of textures. Applying the shaders to the array of textures may effectuate processing the input to the convolutional neural network through the layers of the convolutional neural network. In some implementations, operation 202 may be performed by a processor component the same as or similar to shader component 104 (Shown in FIG. 1 and described herein).

At operation 203, one or more results may be obtained from applying the shaders to the array of textures. In some implementations, operation 203 may be performed by a processor component the same as or similar to result component 106 (Shown in FIG. 1 and described herein).

Although the system(s) and/or method(s) of this disclosure have been described in detail for the purpose of illustration based on what is currently considered to be the most practical and preferred implementations, it is to be understood that such detail is solely for that purpose and that the disclosure is not limited to the disclosed implementations, but, on the contrary, is intended to cover modifications and equivalent arrangements that are within the spirit and scope of the appended claims. For example, it is to be understood that the present disclosure contemplates that, to the extent possible, one or more features of any implementation can be combined with one or more features of any other implementation.

Claims

1. A system for processing convolutional neural network operations using textures, the system comprising: physical storage media storing convolutional neural network information, the convolutional neural network information defining a convolutional neural network, the convolutional neural network including one or more layers, the one or more layers defining one or more operations on an input to the convolutional neural network; andone or more physical processors configured by machine-readable instructions to: format the one or more operations defined by the one or more of the layers as one or more shaders;access input information, the input information defining the input to the convolutional neural network, the input formatted as an array of textures, the textures characterized by a width, a height, and a number of channels;apply the one or more shaders to the array of textures, wherein applying the one or more shaders to the array of textures effectuates processing the input to the convolutional neural network through the one or more layers of the convolutional neural network; andobtain one or more results from applying the one or more shaders to the array of textures.

2. The system of claim 1, wherein the one or more shaders are characterized by one or more of a biases parameter, a weights parameter, an input texture read parameter, and/or an output texture write parameter.

3. The system of claim 1, wherein the one or more physical processors are further configured by machine-readable instructions to: access the input;format the input as the array of textures; andstoring the array of textures as the input information.

4. The system of claim 1, wherein the number of channels of the textures is four.

5. The system of claim 4, wherein individual shaders output information for four channels.

6. The system of claim 1, further comprising a graphics processing unit, wherein the one or more shaders are applied to the array of textures by the graphics processing unit.

7. The system of claim 6, further comprising a graphics processing unit memory, wherein one or more inputs to the one or more shaders and one or more outputs of the one or more shaders are stored in the graphics processing unit memory.

8. The system of claim 6, wherein the graphics processing unit is included in a mobile device.

9. The system of claim 1, wherein the layers of the convolutional neural network include one or more convolution layers, the convolution layers processed using binary convolution.

10. The system of claim 1, wherein the array of textures is stored using an int32 structure, the int32 structure including four int8 structures, an individual int8 structure storing a four by two patch of a single channel of an individual texture.

11. A method for processing convolutional neural network operations using textures, the method comprising: accessing convolutional neural network information, the convolutional neural network information defining a convolutional neural network, the convolutional neural network including one or more layers, the one or more layers defining one or more operations on an input to the convolutional neural network;formatting the one or more operations defined by the one or more of the layers as one or more shaders;accessing input information, the input information defining the input to the convolutional neural network, the input formatted as an array of textures, the textures characterized by a width, a height, and a number of channels;applying the one or more shaders to the array of textures, wherein applying the one or more shaders to the array of textures effectuates processing the input to the convolutional neural network through the one or more layers of the convolutional neural network; andobtaining one or more results from applying the one or more shaders to the array of textures.

12. The method of claim 11, wherein the one or more shaders are characterized by one or more of a biases parameter, a weights parameter, an input texture read parameter, and/or an output texture write parameter.

13. The method of claim 11, further comprising: accessing the input;formatting the input as the array of textures; andstoring the array of textures as the input information.

14. The method of claim 11, wherein the number of channels of the textures is four.

15. The method of claim 14, wherein individual shaders output information for four channels.

16. The method of claim 11, wherein the one or more shaders are applied to the array of textures by a graphics processing unit.

17. The method of claim 16, wherein one or more inputs to the one or more shaders and one or more outputs of the one or more shaders are stored in a graphics processing unit memory.

18. The method of claim 16, wherein the graphics processing unit is included in a mobile device.

19. The method of claim 11, wherein the layers of the convolutional neural network include one or more convolution layers, the convolution layers processed using binary convolution.

20. The method of claim 11, wherein the array of textures is stored using an int32 structure, the int32 structure including four int8 structures, an individual int8 structure storing a four by two patch of a single channel of an individual texture.