Batched DMA Tile Retrieval for Random Frame Region Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional processing accelerators face inefficiencies due to system latencies in memory access and inefficient configuration, particularly in handling sparse pixel processing and random regions of interest, leading to performance loss and resource wastage.
Innovation Solution
A system and method that configures a DMA system to identify and obtain tiles of interest in batches, performing DMA transfers and updates concurrently, reducing memory reads/writes and conserving resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If separate transactions are used to retrieve data for random tiles, then each tile can be obtained individually, but performance loss occurs due to latency bubbles associated with gaps in requests
Solution Approach 1:
The patent merges multiple separate tile retrieval transactions into a single batched DMA transfer operation. The system collects multiple tile identifiers and their corresponding memory addresses, then initiates a single DMA transfer that retrieves all tiles simultaneously, eliminating the latency bubbles that occur between sequential requests.
Solution Approach 2:
The system performs preliminary actions by collecting and batching tile identifiers and memory addresses before initiating the DMA transfer. This preparation phase allows the system to organize all necessary retrieval information in advance, enabling efficient bulk transfer without repeated setup overhead.
2Adaptability or versatility
If multiple channels are used to address deficiencies in random tile retrieval, then additional computing resources are dedicated, but latency bubbles still occur
Solution Approach 1:
The DMA system is configured to handle both sequential and random access patterns through a universal batched transfer mechanism. The same DMA infrastructure processes different access patterns by simply varying the input parameters (tile identifiers and addresses), eliminating the need for separate specialized channels while maintaining adaptability.
Solution Approach 2:
The system adapts to different access patterns by changing parameters such as the number of tiles to retrieve, their memory addresses, and the batch size, rather than changing the fundamental retrieval mechanism. This allows flexible handling of random access patterns using the same efficient batched DMA transfer approach.
3Ease of operation
If tiles are obtained in separate instances rather than batches, then each retrieval can be processed independently, but the number of memory reads/writes increases
Solution Approach 1:
The patent combines multiple independent tile retrieval operations into a single batched DMA transfer. Instead of performing separate memory read operations for each tile, the system consolidates all retrieval requests into one unified transfer operation, dramatically reducing the total number of memory reads/writes while maintaining the ability to process different tiles independently through parameter specification.
Data Source
AI summary
In various examples, systems and methods are disclosed that relate to processing data based at least on random regions in a frame. For example, an accelerator such as a vector processing unit (VPU) can determine one or more regions within a frame that that are involved in operations performed using, or to be performed using, the VPU. The VPU can generate descriptors that cause direct memory access (DMA) transfers to be performed such that the VPU obtains the data associated with the regions involved in the operations without necessarily obtaining the data associated with the entire frame.