CN112346836B - Preemption method, device, user equipment and storage medium for shared computing resources - Google Patents

Abstract

The present invention provides a preemption method, device, user equipment, and storage medium for GPU multi-core processors to share computing resources. The method includes: the management and scheduling thread receives the priority of the thread in the thread pool corresponding to the business core or thread from at least one business core or thread; the management and scheduling thread determines whether the priorities of all the threads are the same; if If yes, exit the preemption processing flow, otherwise execute step 3; the management scheduling thread obtains the lowest priority among the priorities corresponding to all currently executed threads; the management scheduling thread determines whether there is a thread with a higher priority than the lowest priority The threads are executed without obtaining computing resources; if not, exit the preemption processing flow; if yes, enter step 5; in order from low to high priority, select the thread corresponding to the lowest priority being executed The task of at least one business core or thread initiates a suspension operation on the selected task.

Description

Preemption method, device, user equipment and storage medium for shared computing resources

Technical field

The present invention relates to the technical field of graphics processors, and in particular to a preemption method, device, user equipment, and storage medium for GPU multi-core processors to share computing resources.

Background technique

GPU adopts streaming parallel computing mode. In order to increase the efficiency of parallel processing of data and improve the full utilization of underlying computing resources (SIMD/DIMD), a multi-core/multi-thread processor will be used to establish a data flow processing computing channel to perform computing tasks. Parse the command, configure relevant information, and let the computing resources (SIMD/DIMD) calculate and process the data.

Each core/thread corresponds to a thread pool (the dotted line in Figure 2). There are multiple threads in the thread pool (threads correspond to the priorities of the computing tasks performed). Each core/thread will manage it from its own Select a thread in the thread pool for execution. The scheduling method is mainly priority and time slice, so that the threads in each thread pool can achieve priority distribution of important computing tasks, and switch with the same priority according to the time slice to ensure that important tasks are executed after , ordinary tasks are then executed normally to maximize computing efficiency.

In the process of realizing the present invention, the inventor found that there are at least the following technical problems in the prior art:

Thread scheduling among multiple cores/threads is not considered in the prior art. Each core/thread executes the highest-priority computing task in its respective thread pool. If there are high-priority tasks in the current thread pool, the computing resources (SIMD/DIMD) will be exhausted, causing other cores/thread pools to have higher priority than the current thread. Priority tasks cannot be executed due to unavailability of computing resources.

Contents of the invention

The present invention provides a GPU multi-core processor shared computing resource preemption method, device, user equipment and storage medium, which can ensure that high-priority tasks in all thread pools can be executed in time.

On the one hand, a preemption method for shared computing resources of GPU multi-core processors is provided, including:

Step 1: The management scheduling thread receives the priority of the thread in the thread pool corresponding to the business core or thread from at least one business core or thread;

Step 2: The management scheduling thread determines whether the priorities of all the threads are the same; if yes, exit the preemption processing flow, otherwise execute step 3;

Step 3: The management scheduling thread obtains the lowest priority among the priorities corresponding to all currently executed threads;

Step 4: The management scheduling thread determines whether there is a thread with a higher priority than the lowest priority that has not obtained computing resources and is executed; if not, exit the preemption processing flow; if there is, enter step 5;

Step 5: In order of priority from low to high, the management scheduling thread selects the task of at least one business core or thread corresponding to the lowest priority being executed, initiates a suspension operation on the selected task, and saves the suspension. of the task and release the corresponding computing resources so that tasks with a higher priority than the lowest priority are executed.

Optionally, the method also includes:

Step 6: Jump to step 1.

Optionally, before step 6 of the method, the method further includes:

The management scheduling thread increases the priority of the suspended thread.

Optionally, before step 2, the method further includes:

The management scheduling thread sorts the threads according to the priority.

Optionally, before step 1, the method further includes:

GPU multi-core processors create and manage scheduling threads.

On the other hand, a preemption device for shared computing resources of GPU multi-core processors is provided, including:

A receiving unit, configured to receive from at least one business core or thread the priority of the thread in the thread pool corresponding to the business core or thread;

The first judgment unit is used to determine whether the priorities of all the threads are the same; if yes, exit the preemption processing flow, otherwise start the acquisition unit;

The acquisition unit is used to obtain the lowest priority among the priorities corresponding to all threads currently executing;

The second judgment unit is used to judge whether the thread with a higher priority than the lowest priority has not obtained the computing resource and is executed; if not, exit the preemption processing flow; if there is, start the selection unit;

A selection unit configured to select, in order of priority from low to high, the management scheduling thread to select at least one business core or thread task corresponding to the lowest priority being executed, and initiate a suspension operation on the selected task, The scene of the suspended task is saved, and the corresponding computing resources are released, so that tasks with a higher priority than the lowest priority are executed.

The device also includes:

An increasing unit, used to increase the priority of the suspended thread.

On the other hand, a user equipment is provided, which includes the preemption device for shared computing resources of a GPU multi-core processor.

On the other hand, a preemption device for shared computing resources of GPU multi-core processors is provided, including:

memory;

and a processor coupled to the memory, the processor being configured to execute the preemption method for shared computing resources by a GPU multi-core processor based on instructions stored in the memory.

On the other hand, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a processor, the method for preempting shared computing resources of a GPU multi-core processor is implemented .

In the embodiment of the present invention, a preemption process is added between the multi-cores/multi-threads of the GPU stream processor to ensure that high-priority tasks in all thread pools can be executed in time.

Description of the drawings

Figure 1 is a flow chart of a preemption method for shared computing resources by a GPU multi-core processor according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of the system framework used in the preemption method for shared computing resources of a GPU multi-core processor according to another embodiment of the present invention;

Figure 3 is a flow chart of a preemption method for shared computing resources by a GPU multi-core processor according to another embodiment of the present invention;

Figure 4 is a schematic structural diagram of a preemption device for shared computing resources by a GPU multi-core processor according to an embodiment of the present invention;

FIG. 5 is a schematic structural diagram of a preemption device for shared computing resources by a GPU multi-core processor according to another embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments These are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

As shown in Figure 1, it is a preemption method for GPU multi-core processor shared computing resources according to the present invention, including:

Step 11: The management scheduling thread receives the priority of the thread in the thread pool corresponding to the business core or thread from at least one business core or thread; before this step, the method also includes: the GPU multi-core processor creates a management scheduling thread. When the business core/thread executes the computing task in the thread pool according to "time slice + priority", it informs the management core/thread of the priority of the currently executed task, so that the scheduled and managed core/thread can count the number of tasks executed by all cores/threads. priority.

Step 12: The management scheduling thread determines whether the priorities of all the threads are the same; if yes, exit the preemption processing flow, otherwise execute step 3; optionally, before step 12, the method also It includes: the management and scheduling thread sorts the threads according to the priority.

Step 13: The management scheduling thread obtains the lowest priority among the priorities corresponding to all currently executed threads;

Step 14: The management scheduling thread determines whether there is a thread with a higher priority than the lowest priority that has not obtained computing resources and is executed; if not, exit the preemption processing flow; if there is, enter step 15;

Step 15: In order of priority from low to high, the management scheduling thread selects the task of at least one business core or thread corresponding to the lowest priority being executed, initiates a suspension operation on the selected task, and releases the task. Corresponding computing resources are used so that tasks with a higher priority than the lowest priority are executed. At the same time, the scene of the suspended task is saved, so that execution can be reinitiated based on the saved scene next time.

Step 16: The management scheduling thread increases the priority of the suspended thread. The priority level can be raised by one or more levels, depending on the actual situation. Ensure that high-priority tasks are executed in a timely manner and low-priority tasks are not "starved to death."

Optionally, the method also includes: jumping to step 11 and continuing processing. Since multi-cores/multi-threads are executed in parallel, it may happen that high priority still cannot obtain computing resources, so you need to go back to step 11 and recheck.

The invention has the following beneficial effects:

1. The present invention is mainly aimed at increasing the preemption function between multi-cores/multi-threads in order of priority in GPU multi-core/multi-thread stream processors to ensure that high-priority tasks in all thread pools can obtain computing resources in time (SIMD /DIMD) and be executed promptly.

2) A "stop + increase priority" strategy is adopted for low-priority threads to ensure that low-priority threads will not be "starved to death".

That is to say, the present invention avoids the situation that when computing resources (SIDM/DIMD) are exhausted by low-priority tasks, high-priority computing tasks distributed by cores/threads cannot be executed, and at the same time eliminates the problem of low-priority computing caused by preemption. The situation where the task is "starved to death". In GPU multi-core processors that support priority, the highest priority tasks executed by all cores/threads can be given priority to computing resources (SIMD/MIMD) for computing processing. Not only can threads be scheduled according to priority and time slice within the thread pool, but threads can also be scheduled according to priority between thread pools, ensuring that the computing tasks of high-priority threads are executed while maximizing the computing efficiency of the GPU.

The present invention discloses a preemption method for realizing GPU multi-core processor shared computing resources (SIMD/MIMD), aiming at scheduling between multi-cores/multi-threads within the dotted line range of Figure 2. The main design concept of the present invention is: by counting the priorities of all cores/threads, sorting them from high to low, and then checking whether the computing task of the highest priority core/thread is being executed; if it is not being executed, then it will be The execution of the computing task of the lowest priority core/thread stops execution, saves the scene, and enters the ready state; then the computing task of the highest priority core/thread will be executed and will be replaced by the priority of the lower priority core/thread. Raise one level to ensure that high-priority tasks are executed in time and low-priority tasks will not be "starved to death"; if the highest-priority core/thread task is already executing, continue to check the next-highest priority based on the sorting result. Perform the same operations as above until the computing tasks of all high-priority cores/threads are executed, or all cores/threads have the same priority, then exit the preemption process. A priority scheduling mechanism is added between the multi-cores/multi-threads of the GPU multi-core/multi-thread processor. By scheduling the priority of tasks executed by the multi-core/multi-thread, it ensures that high-priority tasks in all parallel execution cores/threads can be obtained first. Computing resources (SIMD/MIMD) execution.

In the design of the present invention, multi-core/multi-thread respectively adopt the strategy of "priority + time slice" for the tasks of their respective thread pools unchanged, and a priority-based preemption strategy is added for multi-core/multi-thread. Track the running status of high-priority computing tasks of execution threads to ensure that high-priority tasks can be executed in a timely manner.

Another embodiment of the present invention is described below. As shown in Figure 3, the method includes the following steps:

Step 1: Create a scheduling management core/thread to receive the priority information sent by all business threads in the thread pool, and manage and schedule threads according to the priority based on the received information. When the business core/thread executes the computing task in the thread pool according to "time slice + priority", it informs the management core/thread of the priority of the currently executed task. The core/thread managed by the scheduler counts the priorities of all core/thread execution tasks.

Step 2: The management scheduling thread sorts the reported thread priority information from high to low priority, and then checks the sorted priority information. If the priorities of all business threads are the same, it means that there is no need to preempt the thread and exit the preemption process. Otherwise, proceed to step 3;

Step 3: Obtain high-priority tasks in sequence;

Step 4: According to the priority order from high to low, check in turn whether the business core/thread executing the next highest priority task has obtained the computing resources for execution. If yes, proceed to step 5. If not, proceed to step 6.

Step 5: Check whether the cores/threads of all high-priority tasks have obtained computing resources. If yes, exit the preemption process; if not, proceed to step 4, select the next highest priority thread and repeat the above process;

Step 6: In order of priority from low to high, select the low-priority and executing core/thread, initiate a stop operation, save the scene of the computing task corresponding to the low-priority thread, and release the corresponding computing resources.

Step 7: Increase the priority of the suspended thread to prevent low-priority threads from never being able to execute due to task preemption and "starving to death". After stopping the low-priority core/thread tasks, proceed from step 3. Since multi-cores/multi-threads are executed in parallel, it may happen that high priority still cannot obtain computing resources, so you need to go back to step 3 and recheck.

As shown in Figure 4, it is a preemption device for GPU multi-core processor shared computing resources according to an embodiment of the present invention, including:

A receiving unit, configured to receive from at least one business core or thread the priority of the thread in the thread pool corresponding to the business core or thread;

The first judgment unit is used to determine whether the priorities of all the threads are the same; if yes, exit the preemption processing flow, otherwise start the acquisition unit;

The acquisition unit is used to obtain the lowest priority among the priorities corresponding to all threads currently executing;

The second judgment unit is used to judge whether the thread with a higher priority than the lowest priority has not obtained the computing resource and is executed; if not, exit the preemption processing flow; if there is, start the selection unit;

A selection unit configured to select, in order of priority from low to high, the management scheduling thread to select at least one business core or thread task corresponding to the lowest priority being executed, and initiate a suspension operation on the selected task, The scene of the suspended task is saved, and the corresponding computing resources are released, so that tasks with a higher priority than the lowest priority are executed.

The device of this embodiment can be used to execute the technical solutions of the above method embodiments. Its implementation principles and technical effects are similar and will not be described again here.

An embodiment of the present invention also provides a user equipment, which includes the above-mentioned preemption device for GPU multi-core processor shared computing resources.

As shown in Figure 5, an embodiment of the present invention also provides a preemption device for GPU multi-core processors to share computing resources, including:

memory 602;

and a processor 601 coupled to the memory, the processor being configured to execute the preemption method for shared computing resources of a GPU multi-core processor based on instructions stored in the memory.

Embodiments of the present invention also provide a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions. When the computer instructions are executed by the processor, the preemption of the shared computing resources of the GPU multi-core processor is realized. method.

Those of ordinary skill in the art can understand that all or part of the processes in the above method embodiments can be implemented by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. The program can be stored in a computer-readable storage medium. During execution, the process may include the processes of the embodiments of each of the above methods. The storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM) or a random access memory (Random Access Memory, RAM), etc.

The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present invention. All are covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims (10)

1. A preemption method for shared computing resources of GPU multi-core processors, which is characterized by including: Step 1: The management scheduling thread receives the priority of the thread in the thread pool corresponding to the business core or thread from multiple business cores or threads, wherein thread scheduling is performed in each thread pool according to priority and time slice; Step 2: The management scheduling thread determines whether the priorities of all the threads are the same; if yes, exit the preemption processing flow, otherwise execute step 3; Step 3: The management scheduling thread obtains the lowest priority among the priorities corresponding to all currently executed threads; Step 4: The management scheduling thread determines whether there is a thread with a higher priority than the lowest priority that has not obtained computing resources and is executed; if not, exit the preemption processing flow; if there is, enter step 5; Step 5: In order of priority from low to high, the management scheduling thread selects the task of at least one business core or thread corresponding to the lowest priority being executed, initiates a suspension operation on the selected task, and releases the task. Corresponding computing resources are used to enable tasks with a higher priority than the lowest priority to be executed; and the scene of the suspended task is saved so that execution can be reinitiated based on the saved scene next time. 2. The method according to claim 1, characterized in that, the method further comprises: Step 6: Jump to step 1. 3. The method according to claim 2, characterized in that, after the step 5 and before the method step 6, the method further includes: The management scheduling thread increases the priority of the suspended thread. 4. The method according to claim 1, characterized in that, before step 2, the method further includes: The management scheduling thread sorts the threads according to the priority. 5. The method according to claim 1, characterized in that, before step 1, the method further includes: GPU multi-core processors create and manage scheduling threads. 6. A preemption device for shared computing resources of GPU multi-core processors, characterized by including: The receiving unit is used to receive the priorities of threads in the thread pools corresponding to the business cores or threads from multiple business cores or threads, wherein thread scheduling is performed in each thread pool according to the priority and time slice; The first judgment unit is used to judge whether the priorities of all the threads are the same; if yes, exit the preemption processing flow, otherwise start the acquisition unit; The acquisition unit is used to obtain the lowest priority among the priorities corresponding to all threads currently executing; The second judgment unit is used to judge whether the thread with a higher priority than the lowest priority has not obtained the computing resource and is executed; if not, exit the preemption processing flow; if there is, start the selection unit; The selection unit is configured to manage and schedule the thread to select the task of at least one business core or thread corresponding to the lowest priority being executed in order from low to high priority, initiate a suspension operation on the selected task, and save the suspension. The scene of the task, and release the corresponding computing resources, so that the task with a higher priority than the lowest priority is executed; and save the scene of the suspended task, so that the execution can be reinitiated according to the saved scene next time. 7. The device of claim 6, further comprising: An increasing unit, used to increase the priority of the suspended thread. 8. A user equipment, characterized in that the user equipment includes the preemption device for GPU multi-core processor shared computing resources according to any one of claims 6 to 7. 9. A preemption device for shared computing resources of GPU multi-core processors, including: memory; and a processor coupled to the memory, the processor configured to perform the GPU multi-core processor sharing computing resources of any one of claims 1 to 5 based on instructions stored in the memory. Preemption method. 10. A computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, which when executed by a processor implement the GPU multi-core processing according to any one of claims 1 to 5 Preemption method for shared computing resources between servers.