Patent Application
20190354833
Deep neural networks (DNNs) are used for many artificial intelligence and machine learning applications. These DNNs nominally include multiple hidden layers between an input layer and an output layer. Recently, DNNs have started to use an increasing number of layers which provide increased capacity and accuracy for various prediction problems in image, video, and speech recognition processing and analysis. However, deeper DNNs also result in increasingly greater performance challenges. For example, DNNs are extremely computationally expensive. It is not uncommon for a training task for a neural network to require days, weeks or even months to be performed.
To reduce the training time, multiple device architectures are used that enable the efficient distributed training of neural networks. However, these multiple device architectures require frequent communication between each of the processing elements to update the parameters of each neural network on each processing element with the updates from the other processing elements as each of the neural networks learn and converge to a solution. Consequently, there is significant traffic on the communication buses that connect the processing elements.
The speed at which updates can be sent between the processing elements is effectively the bottleneck to distributively training neural networks as the processing elements are becoming increasingly faster than the communication buses. The increased levels of communications traffic widens the imbalance between communication and computation and increases the overall time for neural network training.
A more detailed understanding may be had from the following description, given by way of example in conjunction with the accompanying drawings wherein:
Described herein is a method and system for reducing frequency of communication in neural network systems. The method includes computing the errors for samples run in a neural network, and sorting the samples after each epoch in a non-decreasing order of error, where the error is the difference between a ground truth and a prediction. In an illustrative definition, an epoch is the interval for completing a run through the neural network and the ground truth for a sample is the actual or given label for the sample/image. The updates (based on the errors) are then coalesced from multiple mini-batches of samples to reduce the ratio of communication to computation. Since the majority of samples have low errors, and increasingly so as the neural network converges to a solution, the method reduces the overall level of communications on the communications bus. The method addresses the imbalance between communication and computation, where processing elements, such as graphics processing units (GPUs), are becoming increasingly faster than the communication buses, such as Peripheral Component Interconnect Express (PCI-Ex). Communication buses such as the PCI-Ex can be more efficiently used without having to use higher bandwidth-based interconnects.
In various alternatives, the processor 102 includes a central processing unit (CPU), a graphics processing unit (GPU), a CPU and GPU located on the same die, or one or more processor cores, wherein each processor core can be a CPU or a GPU. In various alternatives, the memory 104 is located on the same die as the processor 102, or is located separately from the processor 102. The memory 104 includes a volatile or non-volatile memory, for example, random access memory (RAM), dynamic RAM, or a cache.
The storage 106 includes a fixed or removable storage, for example, a hard disk drive, a solid state drive, an optical disk, or a flash drive. The input devices 108 include, without limitation, a keyboard, a keypad, a touch screen, a touch pad, a detector, a microphone, an accelerometer, a gyroscope, a biometric scanner, or a network connection (e.g., a wireless local area network card for transmission and/or reception of wireless IEEE 802 signals). The output devices 110 include, without limitation, a display, a speaker, a printer, a haptic feedback device, one or more lights, an antenna, or a network connection (e.g., a wireless local area network card for transmission and/or reception of wireless IEEE 802 signals).
The input driver 112 communicates with the processor 102 and the input devices 108, and permits the processor 102 to receive input from the input devices 108. The output driver 114 communicates with the processor 102 and the output devices 110, and permits the processor 102 to send output to the output devices 110. It is noted that the input driver 112 and the output driver 114 are optional components, and that the device 100 will operate in the same manner if the input driver 112 and the output driver 114 are not present. The output driver 114 includes an accelerated processing device (“APD”) 116 which is coupled to a display device 118. The APD accepts compute commands and graphics rendering commands from processor 102, to processes those compute and graphics rendering commands, and provides pixel output to display device 118 for display.
As described in further detail below, the APD 116 includes one or more parallel processing units configured to perform computations in accordance with a single-instruction-multiple-data (“SIMD”) paradigm. Thus, although various functionality is described herein as being performed by or in conjunction with the APD 116, in various alternatives, the functionality described as being performed by the APD 116 is additionally or alternatively performed by other computing devices having similar capabilities that are not driven by a host processor (e.g., processor 102) and configured to provide (graphical) output to a display device 118. For example, it is contemplated that any processing system that performs processing tasks in accordance with a SIMD paradigm can perform the functionality described herein. Alternatively, it is contemplated that computing systems that do not perform processing tasks in accordance with a SIMD paradigm performs the functionality described herein.
The APD 116 executes commands and programs for selected functions, such as graphics operations and non-graphics operations that are suited for parallel processing and/or non-ordered processing. The APD 116 is used for executing graphics pipeline operations such as pixel operations, geometric computations, and rendering an image to display device 118 based on commands received from the processor 102. The APD 116 also executes compute processing operations that are not directly related to graphics operations, such as operations related to video, physics simulations, computational fluid dynamics, or other tasks, based on commands received from the processor 102.
The APD 116 includes compute units 132 that include one or more SIMD units 138 that perform operations at the request of the processor 102 in a parallel manner according to a SIMD paradigm. The SIMD paradigm is one in which multiple processing elements share a single program control flow unit and program counter and thus execute the same program but are able to execute that program with different data. In one example, each SIMD unit 138 includes sixteen lanes, where each lane executes the same instruction at the same time as the other lanes in the SIMD unit 138 but executes that instruction with different data. Lanes can be switched off with predication if not all lanes need to execute a given instruction. Predication can also be used to execute programs with divergent control flow. More specifically, for programs with conditional branches or other instructions where control flow is based on calculations performed by an individual lane, predication of lanes corresponding to control flow paths not currently being executed, and serial execution of different control flow paths allows for arbitrary control flow. In an implementation, each of the compute units 132 can have a local L1 cache. In an implementation, multiple compute units 132 share a L2 cache.
The basic unit of execution in compute units 132 is a work-item. Each work-item represents a single instantiation of a program that is to be executed in parallel in a particular lane. Work-items can be executed simultaneously as a “wavefront” on a single SIMD unit 138. One or more wavefronts are included in a “work group,” which includes a collection of work-items designated to execute the same program. A work group is executed by executing each of the wavefronts that make up the work group. In alternatives, the wavefronts are executed sequentially on a single SIMD unit 138 or partially or fully in parallel on different SIMD units 138. The wavefronts can be thought of as the largest collection of work-items that can be executed simultaneously on a single SIMD unit 138. Thus, if commands received from the processor 102 indicate that a particular program is to be parallelized to such a degree that the program cannot execute on a single SIMD unit 138 simultaneously, then that program is broken up into wavefronts which are parallelized on two or more SIMD units 138 or serialized on the same SIMD unit 138 (or both parallelized and serialized as needed). A scheduler 136 is configured to perform operations related to scheduling various wavefronts on different compute units 132 and SIMD units 138.
The parallelism afforded by the compute units 132 is suitable for graphics related operations such as pixel value calculations, vertex transformations, and other graphics operations. Thus in some instances, a graphics pipeline 134, which accepts graphics processing commands from the processor 102, provides computation tasks to the compute units 132 for execution in parallel.
The compute units 132 are also used to perform computation tasks not related to graphics or not performed as part of the “normal” operation of a graphics pipeline 134 (e.g., custom operations performed to supplement processing performed for operation of the graphics pipeline 134). An application 126 or other software executing on the processor 102 transmits programs that define such computation tasks to the APD 116 for execution.
The HSA 300 provides a unified view of fundamental computing elements. The HSA 300 allows a programmer to write applications that seamlessly integrate CPUs 320, also referred to as latency compute units, with GPUs 330, also referred to as throughput compute units, while benefiting from the best attributes of each. The HSA 300 allows the programmer to take advantage of the parallel processor in the GPU 330 as a peer to the traditional multi-core CPU 320, where each core can run one or more threads. A peer device is defined as an HSA device that shares the same memory coherency domain as another device.
The devices in the HSA 300 communicate with one another using queues as further explained with reference to
The CPU 405, GPU 420 and system memory 415 can be implemented as described with respect to
The GPU 420 provides graphics acceleration functionality and other compute functionality as described herein to system 400. The GPU 420 includes multiple command processors (CP) CP 1 . . . CP n 430, and multiple engines Engine 1 . . . Engine n 435, for example, 3D engines, unified video decoder (UVD) engines, digital rights management (DRM) direct memory access (DMA) engines and the like.
The CP 1 . . . CP n 430 controls the processing within GPU 420 and is connected to Engine 1 . . . Engine n 435. Each CP 1 . . . CP n 430 is associated with Engine 1 . . . Engine n 435 and each pair is an engine block (EB) EB 1 . . . EB n 437. In another embodiment, the CP 1 . . . CP n 430 is a single command processor. In general, the CP 1. . . CP n 430 receives instructions to be executed from the CPU 405, and coordinates the execution of those instructions on Engine 1 . . . Engine n 435 in GPU 420. In some instances, the CP 1 . . . CP n 430 generates one or more commands to be executed in GPU 420, that correspond to each command received from CPU 405. Logic instructions implementing the functionality of the CP 1 . . . CP n 430 is implemented in hardware, firmware, or software, or a combination thereof.
The memory 415 includes one or more memory devices and can be a dynamic random access memory (DRAM) or a similar memory device used for non-persistent storage of data. Memory 415 includes one or more memory buffers 445 through which CPU 405 communicates commands to GPU 420. The memory buffers 445 correspond to the engines 435 or the engine blocks 437, as appropriate. Memory buffers 445 are implemented as queues, ring buffers or other data structures suitable for efficient queuing of work items or command packets. In the instance of a queue, command packets are placed into and taken away from the memory buffers 445 in a circular manner. For purposes of illustration, memory buffers 445 are referred to as queue 1 . . . queue n 445 herein.
The memory 415 includes indirect buffers 455. The indirect buffers 455 hold the actual commands (e.g., instructions, data, pointers and non-pointers). For example, when the CPU 405 communicates a command packet to the GPU 420, the command packet is stored in the indirect buffer 455 and a pointer to that indirect buffer 455 is inserted in a queue 1 . . . queue n 445. As described herein below, certain of the indirect buffers 455 hold neuron data. That is, multiple indirect buffers are used for different purposes. The CPU 405, via driver 410, as a writer of the commands to queue 1 . . . queue n 445 and the GPU 420 as a reader of such commands, coordinate a write pointer and read pointer indicating the last item added and last item read, respectively, in queue 1 . . . queue n 445.
The above architecture provides a one-way communication from a host processor (the writer as represented by the driver 510) to the GPU 500 (the reader as represented by the command processor 505). Initially the read pointer 512 and the write pointer 514 point to the same location indicating that the queue 515 is empty. The queue 515 has free space 530 into which the driver 510 writes a command packet corresponding to a task. The driver 510 then updates the write pointer 514 to one position past the last command packet or the first available space. The write pointer 514 and read pointer 512 are now pointing to different locations. The command processor 505 fetches command packets at the read pointer 512 position and walks the read pointer 512 until it is equal to the write pointer 514.
Neural networks learn or train by comparing their classification of a sample with a ground truth classification of the sample. An error is defined as difference between what the neural network perceives and a ground truth. As each sample is run through the neural network, errors are determined and used to modify the weights, e.g., W1, W2, . . . , Wn, in the neural network. This is done iteratively for a given number of times. In general, updates based on the errors are communicated between each of the processing elements, e.g., processing elements 1 . . . MN, via a communications bus, e.g., the communications bus 425 or 605.
Frequent communication between each of the processing elements is necessary to update the weights of each neural network on each processing element with the updates from the other processing elements as each of the neural networks learn and converge to a solution. Consequently, there is significant traffic on the communication buses that connect the processing elements.
A commonly used method for adjusting the weights according to the error is gradient descent, which attempts to determine adjusted weights which results in lesser errors, where the reference to error means or relates to the order of the magnitude of the errors. One form of gradient descent implementation updates the neural network after each error determination and is computationally and communications extensive. Another form of gradient descent, batch gradient descent, updates the neural network after all samples in a training set are completed. Yet another form of gradient descent, mini-batch gradient descent, divides the training set into small batches (hereinafter mini-batches) and updates the neural network after the mini-batch is completed. Under multi-device or multi-node settings, the weights need to be synchronized. In the specific case of data parallelism, the synchronization is done in an all-to-all reduction. By default this is done after every batch or mini-batch. However, this limits parallelism by putting pressure on the communication bus.
A method and system addresses the bandwidth limitations of existing communications buses and cluster interconnects by reducing the communication frequency between processing elements. The method minimizes how fast the communication bus needs to be by minimizing the number of communication updates that need to be done. The communications pressure is alleviated by decreasing the communication frequency using the following heuristic: compute the error of each sample in a mini-batch and record it; sort the sample indices in non-decreasing order of errors; and coalesce the updates from several mini-batches to reduce the communication frequency. In an implementation, coalescing is done without sorting the errors first. For very large datasets, the coalescing reduces the communication frequency and improves the quality of the gradients without affecting accuracy. In an implementation, the method is realized by masking the communication updates and only conducting the communication updates when a locally accumulated error meets a significance threshold.
The method leverages the properties of deep learning or neural network algorithms and the datasets which are used for training the neural networks. The training datasets usually include a large distribution of samples, the majority of which are easily classified correctly, in addition to a few borderline samples, and another set of noisy samples. The distributed memory implementations usually implement mini-batch gradient descent, where the processing elements (also known as compute nodes) synchronize their neural networks after learning from each mini-batch. The described method reduces the communication frequency by computing the error for each sample at the end of each mini-batch and recording the computed error. The sample indices are then sorted in non-decreasing order of the error. In an implementation, the communication updates for low error samples are then coalesced since their updates are proportionally smaller in comparison to the high error samples. The coalescing fundamentally reduces the frequency of global synchronization for the mini-batch updates since the updates from a large number of small error samples is combined together. This reduces the overall pressure on the communication buses including system on chip buses or cluster interconnects.
In a next epoch or following epochs, mini-batches of samples are formed based on the sorted list (step 925). In an implementation, the mini-batches are formed starting from the sample which has the smallest error in the sorted list. In another implementation, the mini-batches alternate between samples with low errors and samples with high errors. That is, one mini-batch is formed starting from a sample which has the smallest error and a next mini-batch is formed starting from a sample which has the largest error. The parameters of the neural network are updated and the mini-batch is run (step 930). The updates related to a mini-batch or a set of coalesced mini-batches are communicated to the other processing elements if the updates have resulted in making a significant impact on the neural network (step 935). In this instance, significant impact is measured, for example, by determining if the errors or accumulated errors since the last communication update meet or exceed the significance threshold. The significance threshold is set, for example, based on desired features or characteristics such as, for example, by using an average of the maximum and minimum error as the significance threshold. The process starts over for further epochs until convergence is achieved (step 940).
In an implementation, the initial mini-batches in the next epoch(s) contain the samples with the smallest errors and the updates are likely to be communicated. The mini-batches which include the samples associated with larger errors are less likely to be communicated to the other processing elements as they do not have enough impact on the neural network. That is, the accumulated error from the later mini-batches is minimal as compared to the initial or earlier mini-batches and communication updates are not needed. In other words, as one progresses down the sorted list, there is less to learn and therefore less to communicate to the other processing elements. This, in turn, decreases the total pressure on the communications bus. In contrast, conventional methods communicate updates independent of any learning after each mini-batch.
It should be understood that many variations are possible based on the disclosure herein. Although features and elements are described above in particular combinations, each feature or element can be used alone without the other features and elements or in various combinations with or without other features and elements.
The methods provided can be implemented in a general purpose computer, a processor, or a processor core. Suitable processors include, by way of example, a general purpose processor, a special purpose processor, a conventional processor, a digital signal processor (DSP), a plurality of microprocessors, one or more microprocessors in association with a DSP core, a controller, a microcontroller, Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) circuits, any other type of integrated circuit (IC), and/or a state machine. Such processors can be manufactured by configuring a manufacturing process using the results of processed hardware description language (HDL) instructions and other intermediary data including netlists (such instructions capable of being stored on a computer readable media). The results of such processing can be maskworks that are then used in a semiconductor manufacturing process to manufacture a processor which implements aspects of the embodiments.
The methods or flow charts provided herein can be implemented in a computer program, software, or firmware incorporated in a non-transitory computer-readable storage medium for execution by a general purpose computer or a processor. Examples of non-transitory computer-readable storage mediums include a read only memory (ROM), a random access memory (RAM), a register, cache memory, semiconductor memory devices, magnetic media such as internal hard disks and removable disks, magneto-optical media, and optical media such as CD-ROM disks, and digital versatile disks (DVDs).
This application claims the benefit of U.S. provisional application No. 62/673,450 having a filing date of May 18, 2018, which is incorporated by reference as if fully set forth herein.
| Number | Date | Country | |
|---|---|---|---|
| 62673450 | May 2018 | US |
