836 results about "Graphics processing unit" patented technology

The invention discloses a large-scene three-dimensional reconstruction method based on three-dimensional Gaussian sputtering, and relates to computer graphics. The method comprises the following steps: collecting a multi-view image set of a large scene; obtaining a scene sparse point cloud according to the multi-view image set; performing monocular depth estimation on the multi-view image by using a pre-trained depth prediction network to obtain monocular depth estimation priori; the method comprises the following steps of: performing global training on a scene by utilizing scene sparse point cloud and monocular depth estimation prior to obtain an initial three-dimensional Gaussian model, and performing space grid division on the initial three-dimensional Gaussian model to obtain a plurality of scene blocks with axis alignment bounding boxes; setting image view angle data of each scene block; performing deep supervised training on the Gaussian ellipsoids in the plurality of scene blocks by using a parallel GPU (Graphics Processing Unit); combining the trained scene blocks to obtain a final three-dimensional Gaussian model; in view of low geometric structure reconstruction precision caused by only depending on color information of a multi-view image in large-scene three-dimensional rendering, the method improves the reconstruction precision of large-scene rendering.

The invention discloses a resource scheduling method for dynamic rendering and related equipment, and relates to the technical field of dynamic rendering, and the method comprises the following steps: obtaining time sequence data of a user touch event; based on the time sequence data, predicting a user behavior state through a behavior prediction model, and determining a current resource demand level; dynamically adjusting a resource allocation strategy and a video memory management strategy of the graphic processing unit according to the resource demand level; and executing a rendering task based on the adjusted resource allocation strategy, and performing picture output by adopting a double-buffer architecture so as to reduce delay caused by time sequence mismatch. According to the method and the device, by establishing closed-loop control of a user behavior-resource demand-supply strategy, accurate scheduling and efficient utilization of resources are realized, the problems of time sequence mismatch, resource waste and jamming in the traditional technology are solved, the rendering efficiency and the user experience are improved, and technical support is provided for high-concurrency and low-delay scenes.

The invention discloses a hybrid expert model reasoning method based on cooperation of a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit), and belongs to the field of deep learning. According to the method, a CPU-GPU computing framework of a hybrid expert model is constructed, heterogeneous computing resource loads are effectively balanced, and the hardware utilization rate is remarkably increased; an intelligent cache management mechanism based on dynamic priority scores is provided, high-demand experts are reserved preferentially, and the transmission overhead caused by cache missing is reduced; through pipeline parallel design for separating calculation and transmission tasks, CPU calculation and PCIe transmission are overlapped in the GPU execution period, and delay is effectively hidden. In addition, in combination with a multi-layer expert activation prediction prospective prefetching mechanism, the expert cache hit rate is improved. The method is compatible with hybrid expert models of different scales and structures, and stable and efficient reasoning acceleration is realized on a resource-limited heterogeneous platform.

The invention discloses a GPU (Graphics Processing Unit) and parallel IO (Input / Output) collaborative optimization method based on mode heterogeneous computing, and particularly relates to the field of heterogeneous computing optimizing.The method comprises the following steps: monitoring mode characteristics during execution of a target computing task in real time through an instruction analyzer, and synchronously collecting dynamic parameters of a heterogeneous computing environment to generate a multi-dimensional resource situation matrix; based on the mode and the resource situation matrix, collaborative optimization operation of the binding equipment dimension and the scheduling time dimension is executed; and constructing a three-level adaptive transmission mechanism and periodically implementing feedback optimization according to an execution state of the three-level adaptive transmission mechanism. According to the GPU and parallel IO collaborative optimization method based on mode heterogeneous computing, the data prefetching hit rate is increased by dynamically sensing the matching relation between a heterogeneous computing mode and an I / O access mode; end-to-end delay is reduced and task contention conflicts are reduced through a dynamic matching strategy of an equipment capability portrait and a task type; and the heterogeneous storage bandwidth peak value utilization rate and the I / O throughput are improved by differentially utilizing the hierarchical characteristics of the multi-stage storage.

The invention discloses a cross-platform GPU support-free real-time digital human rendering system and method, and relates to the technical field of digital human rendering, and the system comprises an offline modeling module which constructs a mouth base space model through principal component analysis based on video data of a target person; the real-time driving module is used for mapping real-time voice features into dynamic coefficients of the mouth base space model; the image synthesis module is used for reconstructing a mouth image according to the dynamic coefficient and fusing the mouth image with a pre-stored reference face image; the cross-platform adaptation layer is configured to realize real-time rendering without GPU dependence through compiling, memory mapping and multi-thread scheduling technologies; and the dynamic resource manager is used for optimizing resource occupation and ensuring stable operation in a multi-platform environment. According to the technical scheme, dependence on high-computing-power hardware and an ideal acoustic environment can be broken through, and a complete theoretical framework and an engineering implementation path are provided for large-scale popularization of a digital human technology in a lightweight terminal.

The invention discloses a cloud server physical GPU acceleration capability release method and system, and belongs to the technical field of computer GPU capability release, and the method specifically comprises the steps: collecting GPU operation parameters in real time, carrying out the preprocessing, storing the GPU operation parameters in a shared state database, constructing a lightweight virtualization abstraction layer on a physical GPU, dividing the physical GPU into a plurality of logic units, and storing the logic units in a shared state database; the method comprises the following steps: establishing a physical GPU, constructing an isolation layer, utilizing a three-dimensional resource segmentation strategy, adopting a priority queue, a jump mechanism and a delay tolerance mechanism in a scheduling process, dynamically allocating logic unit resources of the physical GPU, accelerating release of GPU resources, and realizing unification and real-time updating of physical GPU state information among nodes of a cloud server through a distributed state synchronization mechanism; through dynamic task scheduling, a jumping mechanism, a delay tolerance mechanism and distributed state synchronization, efficient GPU resource management is realized, and GPU capability release is accelerated.

Disclosed in the present application are a resource adjustment method and apparatus, an electronic device, a storage medium and a training platform, which are applied to the technical field of artificial intelligence. The method comprises: acquiring a container set of computing nodes executing a task to be trained and a container associated with same; on the basis of resource configuration information of the container set, resource occupation information of the computing nodes and user resource requirements, and in a mode that target graphics processing unit and target central processing unit cores executing said task are preferentially located in the same non-uniform memory access group, selecting target central processing unit cores needing to be bound for the container, and updating configuration information of a process group corresponding to the container, so as to complete binding of the central processing unit cores and the container. The present application can solve the problem that AI training tasks cannot be efficiently completed in the related art, and can maximally ensure the efficient completion of AI training tasks by means of resource adjustment during the execution of the AI training tasks.

A system (100) for optimizing the allocation of computing resources, comprising: a data acquisition unit (102) comprising a retrieval processor (102a) configured to retrieve historical performance data, which is then stored in a data storage module (104); a centralized processing unit (106) comprising a dedicated processor and a graphics processing unit (GPU) for performing a plurality of resource allocation functions, the centralized processing unit further comprising: a prediction module (106a) configured to implement machine learning algorithms to analyze historical performance data and generate workload forecasts; a performance monitoring module (106b) configured to track real-time system metrics, including CPU utilization, memory consumption, and network latency; a resource allocation module (106c) configured to dynamically adjust the computing resources based on the workload forecasts and real-time system metrics; a container orchestration module (106d) configured to manage the deployment and scaling of microservices across distributed containers; and an administrative user interface module (108) comprising a user display (110) and a visualization module (112), wherein the visualization module (112) generates representations of performance metrics and the user display (110) is configured to display system performance metrics and enable configuration of scaling policies.

The present disclosure relates to the technical field of Graphics Processing Units (GPUs), and in particular to a GPU scheduling method and apparatus, and a storage medium. The method comprises: acquiring a migration command, the migration command being used for migrating a workload of a source Virtual Machine (VM) from a source GPU core to a target GPU core; on the basis of the migration command, establishing an association relationship between the source VM and a corresponding hardware identifier on the target GPU core; and on the basis of the association relationship after the migration, using the target GPU core to process the workload from the source VM. According to embodiments of the present application, while the VM is not shut down, the target GPU core is used to process the workload from the source VM, and thus the workload of the source VM is migrated from a high-load source GPU core to an idle target GPU core, so that the thermal migration between multi-core GPU cores can be achieved, the load between a plurality of GPU cores can be balanced, the pressure of a high-load GPU core can be reduced, the response time can be shortened, and the throughput rate can be increased.

The invention relates to the technical field of GPUs, in particular to a GPU virtualization video memory management method and device, a storage medium and a program product. The method comprises the steps that a virtual machine transmits a page table root address corresponding to a target application program running in the virtual machine to a host machine; the virtual machine calls the non-memory operation type API in response to the target application program, intercepts calling of the non-memory operation type API and forwards the calling to the host machine; and in the host machine, accessing a video memory area corresponding to a target application program based on the page table root address through the drive of the non-memory operation type API, executing an instruction of the non-memory operation type API, and returning a processing result of the instruction of the non-memory operation type API to the virtual machine. According to the method and the device, the rear-end driver in the host machine can directly access the video memory resources allocated by the virtual machine, so that the resource utilization rate and the task execution efficiency are improved.

Approaches presented herein provide systems and methods for allocating streaming multiprocessors (SMs) to execute one or more tasks. A utilization for a given task may be determined by one or more parameters, such as a task execution time or a period. The utilization may then be used to assign a proportionate number of SMs associated with a processing unit, such as a graphics processing unit (GPU) or other type of processing unit, executing the task. The SMs may then be identified, allocated, and reserved until execution of the task is complete.

The invention relates to an application method of a GPU (Graphics Processing Unit) in design rule inspection in the field of integrated circuits, in particular to a layout design rule inspection method, system and equipment, a medium and a product. The layout design rule checking method comprises the following steps: inputting initial layout data, analyzing a design rule checking task to be executed, and marking the priority and complexity level of the task; the method comprises the following steps: dividing layout data into a plurality of data blocks according to regions on the basis of regional characteristics of the layout data, distributing tasks and corresponding data blocks to a GPU computing core according to task priorities and complexity levels, and distributing new data blocks according to the overall load condition of the GPU computing core after the GPU computing core completes data block processing, and repeating the steps until all tasks are executed, and outputting a layout design rule check result. The system, the computer equipment, the computer readable storage medium and the computer program product have the same beneficial effects as the layout design rule checking method.

The embodiment of the invention discloses a method for performing model reasoning by adopting a GPU (Graphics Processing Unit), electronic equipment, a computer readable storage medium and a computer software product. The method comprises the following steps: pre-loading and preheating a plurality of model instances in the GPU; setting a shared memory shared by at least two threads, and setting a state machine for the shared memory; the state machine can comprise a writable state and a non-writable state; when the state machine is in a writable state, writing a video frame into the shared memory, and setting the state machine to be in a non-writable state; calling the model instance loaded and preheated in the GPU to perform reasoning based on the video frames in the shared memory; and after the reasoning is completed, setting the state machine to be in a writable state. According to the embodiment of the invention, high efficiency and stability of model reasoning are realized, and calculation overhead and system complexity are reduced.

The invention relates to the technical field of GPGPU (General Purpose Graphics Processing Unit), and relates to an SIMT stack and thread scheduling optimization method and device for GPGPU (General Purpose Graphics Processing Unit) branch instructions, a medium and a GPGPU architecture. The method comprises the steps that in a GPGPU, an SIMT stack is optimized, an SM unit is expanded to be a double-instruction assembly line, and a main thread beam scheduler and an auxiliary thread beam scheduler are arranged. And in a non-branch state, half threads of each tube of the scheduler are executed in parallel by the assembly line. When a branch instruction is encountered, instruction addresses, thread masks and branch reunion points of the two branch paths are pressed into the SIMT stack. Judging conflicts according to the thread masks; if conflicts exist, performing serial execution; and if no conflict exists, executing on double pipelines in parallel. And when the branch is completed to a reunion point, the SIMT stack pops up an entry, and the execution is continued. The method has remarkable effects in the aspects of improving the thread resource utilization rate, improving the calculation efficiency, simplifying the branch control complexity, enhancing the GPGPU flexibility and the like.

The invention provides a method for model training, a host and a storage device. The method includes instructing, by a host device, a graphics processing unit (GPU) to load first intermediate data of a current layer of the model from a dynamic random access memory (DRAM) of a storage device into a memory of the GPU, where the current layer is a layer of the model that is performing computations; and based on the space capacity of the DRAM, notifying, by the host device, the storage device to pre-fetch second intermediate data of the layer to be calculated of the model from the NAND of the storage device to the DRAM of the storage device, the first intermediate data for the current layer is used by the GPU in back propagation of the model to perform computations for the current layer.

The invention relates to the technical field of organic light-emitting display data processing, in particular to a high-resolution organic display driving waveform generation method, which comprises the following steps of: dividing a pixel matrix into space blocks through distributed parameter analysis, extracting parasitic parameters in parallel by utilizing GPU (Graphics Processing Unit) nodes, and constructing a cross-block coupling model; generating a block parasitic parameter mapping table and a feature vector matrix; generating a segmented waveform parameter set containing curvature parameters based on the residual charge quantity data and dynamic gray scale segmented regulation and control; calling the brightness and voltage response data to execute nonlinear brightness correction, and generating a compensation pulse parameter in combination with a light intensity feedback loop; multi-mode driving strategies are classified and switched according to display content features, time sequence faults are eliminated through phase synchronization of a phase-locked loop, parasitic effect interference and brightness sudden change overshoot are effectively suppressed, gray scale compensation precision and refresh rate collaboration are improved, and display uniformity and dynamic picture smoothness are enhanced.

The invention provides a cuttable interactive software simulation method and system based on a GPU accelerated substance point method, and belongs to the field of computer graphics physical simulation, and the method comprises the steps: S1, carrying out the sampling and initialization of a to-be-simulated three-dimensional software model and a rigid body model interacting with the to-be-simulated three-dimensional software model; s2, based on a virtual scene to be simulated, constructing an initial GPU secondary sparse acceleration structure Ggrid; s3, carrying out expansion on the Ggrid; s4, updating the state of the rigid body model, and constructing a rigid body cutting field of Ggrid; s5, on the basis of the Ggrid and the rigid body cutting field, the physical explicit integral of the three-dimensional software model to be simulated is solved in parallel, and the position of the lagrangian particle at the next moment is obtained; and S6, during the last iteration of each frame, reconstructing a surface triangular mesh of the Lagrange particle cluster in real time based on Ggrid, and smoothing the surface triangular mesh. According to the method disclosed by the invention, multi-material soft deformation, cutting interaction, crushing and real-time surface reconstruction are realized.

The invention discloses a data processing method, a product, electronic equipment and a computer readable storage medium, relates to the technical field of computers, and aims to solve the problems of large GPU (Graphic Processing Unit) space occupation and high access delay in data processing in related technologies. Reading corresponding historical key value cache data from a key value cache of the memory extension equipment according to a historical key value cache data acquisition request sent by a host end; returning the historical key value cache data to the host end, and enabling the host end to process the current input data based on the historical key value cache data by adopting a local model to obtain key value data; and obtaining key value cache data corresponding to the key value data sent by the host end, and sending the key value cache data to a key value cache of the memory extension equipment for storage. The key value cache data is stored in the key value cache of the memory extension equipment, and the historical key value cache data is read from the key value cache, so that the occupation of the memory space of the GPU and the consumption of the memory bandwidth can be reduced, and the access delay is reduced.

The invention provides a parallel optimization method and system for a GPGPU (General Purpose Graphics Processing Unit) instruction execution stage, and the method comprises the steps: splitting an input instruction into a plurality of sub-instructions which can be independently scheduled based on an instruction type, operand dependency analysis and a hardware resource real-time state, and generating a dependency table to record the input and output association of the sub-instructions; acquiring real-time load data of the calculation unit, the storage unit and the communication unit, and distributing the sub-instructions to the calculation unit, the storage unit or the communication unit by adopting a priority dynamic adjustment algorithm and combining instruction key path analysis; according to the dependency relationship of the sub-instructions and the availability of hardware resources, the execution sequence of the sub-instructions is adjusted to maximize the degree of parallelism; and analyzing the dependency relationship among the sub-instructions in real time to maximize the degree of parallelism to dynamically adjust a parallel execution strategy so as to realize serial execution of dependent instructions and parallel execution of non-dependent instructions. According to the method, the complexity of instruction compiling is reduced, the efficiency of instruction compiling is improved, and meanwhile extra consumed time is reduced.

The invention discloses a thread management optimization method and device based on GPU sharing, and the method comprises the steps: inputting the attribute information of a target task and a current GPU state into a target prediction model when the target task is obtained, so as to obtain a prediction result outputted by the target prediction model; under the condition that the performance parameters reach the standard, calling thread block configuration and a shared memory to execute the target task so as to improve the data processing efficiency of each thread in the execution process of the target task; when it is monitored that the load of any thread is abnormal, the load of any thread is distributed to other threads with the fragment rate meeting the requirement, so that the execution efficiency of the target task is guaranteed. According to the method, the target prediction model is utilized, the thread block configuration of the target task is determined, the thread block configuration is called and the shared memory is shared, the data processing efficiency of the target task is improved, the abnormal load is distributed to other threads for execution, and the execution efficiency of the target task in the GPU shared environment is effectively guaranteed.

The invention discloses a whole machine hot standby switching method of a network system, electronic equipment, a medium and a product, and relates to the technical field of computers, and the method comprises the following steps: when a first target host device in a fault state exists in a plurality of host devices, deleting the first target host device, and generating corresponding fault information, and receiving a first topology configuration file of a new graphic processing unit updated by the multi-host control processing unit based on the fault information, so as to update a routing table of the PCIe switch based on the first topology configuration file, and mapping resources of the graphic processing unit of the first target host device to a second target host device in a normal state, the problems that the port mapping of the PCIe switch is relatively fixed and the data is difficult to switch to the standby equipment quickly are solved, and the flexible connection of the GPUs and the continuous updating of the routing table are realized through the PCIe switch, so that the efficient data interaction and resource mapping among the GPUs are realized.

This disclosure describes supporting distributed graphics and compute engines in a multi-dielet processor, such as, for example, a multi-dielet graphics processing unit (GPU), architectures and synchronization in such architectures. Each multi-dielet processor includes a hardware-implemented remapping capability and / or a hardware-implemented memory barrier capability.

The embodiments of the present application relate to the technical field of terminals. Provided are a rendering processing method and an electronic device. The method comprises: determining a plurality of rendering instructions corresponding to first graphic data to be rendered of a first application program, wherein the plurality of rendering instructions include a first rendering instruction; and then reading, from a graphics processing unit (GPU) memory, a first rendering resource corresponding to the first rendering instruction, and executing the plurality of rendering instructions on the basis of rendering resources respectively corresponding to the plurality of rendering instructions, so as to render the first graphic data. In this way, a rendering resource is read from a GPU memory, so as to reduce data interaction between a GPU and a main memory, thereby reducing the bandwidth consumption of a terminal device.

The invention provides a thread block scheduling module, a general-purpose computing graphic processing unit, equipment and a product, and relates to the technical field of data process.The method comprises the steps that an operation data analysis unit conducts data analysis on all first thread blocks of a computing task issued by a host, and second thread blocks sharing data with all the first thread blocks are determined; the mapping relation between the first thread block and the second thread block is written into a thread block data mapping table of the queue management unit; the resource management unit is used for recording available resources and used resources of each streaming multiprocessor and information of a thread block currently allocated to each streaming multiprocessor; and the thread block allocation unit is used for allocating the thread blocks of the first target thread block to the corresponding stream multiprocessors according to the thread block data mapping table of the first target thread block under the condition of detecting that any stream multiprocessor recorded by the resource management unit has idle available resources.

The invention discloses a model reasoning acceleration method and device based on heterogeneous reasoning, equipment and a medium, and relates to the technical field of artificial intelligence. The method comprises the steps that cue words are input into a target reasoning model located in a central processing unit; inputting the obtained first token generation information and the target token sequence into a target draft model located in a graphics processor so as to generate a to-be-determined draft token generation tree and perform iterative updating; generating a target draft token set based on the obtained target draft token spanning tree, and inputting spanning tree information and the target draft token set into a target inference model; and verifying the target draft token set based on the obtained target verification probability value to determine a next round of target token sequence and first token generation information, and then skipping to the step of inputting to a target draft model located in the graphics processor until a preset stop condition is met. Therefore, the reasoning speed of the model can be further improved to meet the requirement of efficient reasoning.

The invention discloses a model performance test method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence. The theoretical maximum lexical throughput of a target large language model is calculated based on the video memory bandwidth of a graphics processor, the model parameter quantity, the byte number corresponding to the quantization precision and the video memory bandwidth utilization rate; meanwhile, the benchmark performance throughput is obtained, a theoretical corresponding first concurrency number is calculated in combination with the theoretical maximum lexical unit throughput and the concurrency competition loss coefficient, then the model test is executed based on the first concurrency number to obtain the actual maximum lexical unit throughput and a corresponding second concurrency number, and a model performance test result is generated. The problems that in the prior art, due to the fact that manual testing is conducted depending on manual intervention, a continuous approaching attempt mode is adopted, a reasonable test starting point is not deduced in combination with hardware core bottlenecks and key parameters, evaluation is time-consuming and labor-consuming, the result is prone to being affected by artificial factors, and accuracy and consistency are poor can be solved.

The invention discloses a graphics processor simulation method, a simulator, a device, electronic equipment and a storage medium, and belongs to the technical field of simulation. Loading a target system call simulator for simulating the target graphics processor; the target system call simulator comprises a user space simulation layer, a graphics processor simulation layer and a graphics drive simulation layer connected with the graphics processor simulation layer and the user space simulation layer; and running a target program in the user space simulation layer, and directly sending a graphic system call generated in the running process of the target program to the graphic drive simulation layer, so that the graphic drive simulation layer transmits a corresponding GPU instruction set to the graphic processor simulation layer based on the graphic system call, and executes the GPU instruction set through the graphic drive simulation layer. Therefore, the software and hardware of the corresponding graphics processor can be simulated by calling the simulator through the lightweight target system, and the simulation precision of the graphics processor can be improved under the condition of ensuring relatively high simulation efficiency.

The invention discloses a high-performance CPU-GPU (Central Processing Unit-Graphics Processing Unit) cooperative processing architecture of an audio frequency integrated signal processor, belonging to the technical field of audio frequency integrated signal processing, the architecture is based on a dynamic priority scheduling model and realizes efficient collaboration of a CPU and a GPU through hierarchical resource management and protocol level optimization, a hardware layer adopts a multi-GPU cluster and a distributed storage node, and the CPU-GPU cooperative processing architecture has the advantages that the multi-GPU cluster and the distributed storage node are integrated; high-concurrency task processing is supported; the transmission layer fuses RapidIO and an Ethernet protocol, and adapts to a short frame control signal and a long packet data stream respectively; and the application layer calculates task resource demands through dynamic priority weights, and ensures low time delay of key tasks in combination with a normalized allocation algorithm. According to the architecture, in an audio signal processing scene, the task scheduling efficiency is improved by 40%, the short frame transmission delay is as low as 0.5, the throughput of a long data stream reaches 100 Gbps, and the requirements for real-time performance and calculation precision in a complex acoustic environment can be met.

The invention discloses a server and a server monitoring method, and relates to the technical field of servers, the server comprises a graphics processor board and a management board, a first connector is additionally arranged on the graphics processor board, the management board independent of the graphics processor board is additionally arranged, and the management board is connected with the graphics processor board through the first connector. The information monitoring device is used for monitoring the running state of the graphics processor board, so that on the basis of not changing the structure of the graphics processor board, comprehensive coverage and quick response of information monitoring of a plurality of computing devices on the graphics processor board are realized, and the problems of untimely and incomplete information monitoring caused by increase of the devices are solved; and the stability and the safety of the graphics processor board are improved.

The invention discloses a ray tracing method and device and an image rendering system, and relates to the technical field of image rendering. The image rendering system provided by the invention comprises a plurality of graphics processors, a first graphics processor determines a target bounding box intersected with a first light ray, and if geometry data in the target bounding box is not stored in the first graphics processor, a first cross-processor intersection request is sent, and then executing a light intersection task aiming at the second light. The first graphics processing unit can initiate the cross-processor intersection request after determining that the geometry data in the target bounding box intersected with the first light is not stored, so that the overall time consumption of parallel execution of a light intersection task by multiple GPUs can be greatly reduced, and the efficiency of parallel execution of image rendering by multiple GPUs can be improved; and cross-processor light intersection can meet the requirement of real-time image rendering.