Virtual Doorbell Mechanism for GPU-CPU Notification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional graphics processing unit (GPU) architectures face challenges in intra-device event notification, particularly when the GPU is a PCIe-based device or attached to a non-coherent fabric, preventing effective communication with the central processing unit (CPU) for scheduling work.
Innovation Solution
The implementation of hardware support for virtual doorbells in a memory-mapped I/O (MMIO) environment enables GPU notifications to be managed efficiently, allowing for effective communication and scheduling without requiring a coherent fabric.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional GPU architectures use coherent fabric for CPU-GPU communication, then notification mechanism works reliably, but the system cannot support PCIe-based GPUs or non-coherent fabric attachments
Solution Approach 1:
The patent introduces a doorbell register as an intermediary mechanism between CPU and GPU. The doorbell register acts as a mediator that enables notification functionality without requiring coherent fabric, allowing PCIe-based GPUs and non-coherent fabric attachments while maintaining reliable notification delivery through a dedicated notification register that CPU can write to signal GPU of pending work
2Adaptability or versatility
If GPU communicates notifications via coherent fabric, then notification delivery is reliable, but PCIe-based devices and non-coherent fabric attachments are prevented
Solution Approach 1:
The patent extracts the notification mechanism from the coherent fabric dependency. By separating the notification function from the coherent fabric requirement, the invention allows GPUs attached via PCIe or non-coherent fabric to operate independently of coherent fabric constraints, removing the blocking constraint while preserving notification functionality through dedicated doorbell and notification registers
3Productivity
If memory-mapped I/O environment is used without hardware support for virtual doorbells, then implementation is simpler, but notification and scheduling efficiency deteriorates
Solution Approach 1:
The patent implements self-service notification mechanism where the GPU autonomously monitors its own notification register and the CPU autonomously writes to the doorbell register. This self-service approach enables efficient notification and scheduling in memory-mapped I/O environments with hardware support for virtual doorbells, allowing each component to manage its own notification state without requiring complex external coordination mechanisms
Data Source
AI summary
An apparatus to facilitate doorbell notifications is disclosed. The apparatus includes memory-mapped I/O (MMIO) base address registers including a physical function (PF) and plurality of virtual functions (VF), wherein each function's base address register comprises a plurality of doorbell pages and doorbell hardware including doorbell registers, each having an assignable function identifier (ID) and offset, and comprising a plurality of doorbells to activate a doorbell notification in response to receiving a doorbell trigger from an associated doorbell page set upon detection of an access request.