GPU Texture Memory for Wavefield Data Loading
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current seismic data processing methods face challenges with computational capacity, especially in modeling large-scale seismic wave equation simulations, due to limitations in memory access and bandwidth, particularly when using GPUs for hybrid CPU/GPU approaches.
Innovation Solution
The method employs GPU texture memory to enhance memory access efficiency by mapping off-chip global memory to texture memory, allowing threads to read nearby memory addresses at higher bandwidth, thereby reducing loading and retrieving times by a factor of 10.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If off-chip global memory is used for storing wavefield data, then memory capacity is sufficient, but memory access bandwidth is limited and access time is long
Solution Approach 1:
The patent implements a hierarchical memory system where texture memory (faster, smaller) is nested within the off-chip global memory (slower, larger). The texture memory is configured to cache frequently accessed wavefield data from the off-chip global memory, creating a nested structure that combines the capacity advantages of off-chip memory with the speed advantages of on-chip texture memory.
Solution Approach 2:
The patent performs preliminary data transfer by copying wavefield data from off-chip global memory to texture memory before the main computation phase. This preliminary action ensures that frequently accessed data is already in the faster texture memory when needed during seismic wave propagation simulations, reducing access latency during critical computation phases.
2Power
If hybrid CPU/GPU processing is used, then computational capacity increases, but programming complexity and memory management difficulty increase
Solution Approach 1:
The patent introduces texture memory as an intermediary layer between CPU and GPU memory spaces. This intermediary simplifies memory management by providing a unified access interface that automatically handles data transfer and caching, reducing the programming complexity associated with direct CPU-GPU memory coordination.
Solution Approach 2:
The texture memory system serves multiple functions: it acts as high-speed cache for wavefield data, provides a unified memory interface for both CPU and GPU, and enables automatic data preprocessing. This multi-functionality reduces the overall system complexity by consolidating multiple memory management tasks into a single versatile component.
3Quantity of substance
If wavefield data is stored in off-chip memory, then storage capacity is adequate, but loading and retrieval time is excessive
Solution Approach 1:
The patent applies local quality optimization by identifying that not all wavefield data requires the same access speed. Frequently accessed data (such as wavefields near active sources or receivers) is cached in texture memory with high-speed access, while less frequently accessed data remains in off-chip global memory, creating differentiated access qualities based on local access patterns.
Solution Approach 2:
The system performs preliminary copying of anticipated wavefield data from off-chip global memory to texture memory before computation phases require this data. This preliminary action preemptively loads data into high-speed memory, eliminating loading delays during critical computation and retrieval operations.
Data Source
AI summary
A method for efficiently injecting sources and retrieving receiver wavefields using finite difference models for wavefield simulations on a GPU, when the source and receiver are not on numerical grid points, and therefore arbitrarily located. To accomplish that, the method employs GPU texture memory to increase the memory read bandwidth, and then positions at arbitrary or simulated locations the sources in a finite difference grid, as well as extends them over a newly generated number of grid points.