Interleaved Multi-Channel Input for SIMD Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current electronic devices face challenges in efficiently processing images due to high resource demands, particularly in mobile platforms with limited processing, memory, and energy resources, leading to slow and inefficient image processing tasks.
Innovation Solution
The method involves interleaving input image channels to produce a multi-channel input, loading it into a SIMD processor, and convolving it with a multi-channel filter, optimizing cache access and parallel processing to accelerate Deep Convolution Network (DCN) convolution operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional image processing methods are used, then processing can be performed with standard hardware, but processing time is long and efficiency is low
Solution Approach 1:
The patent segments the convolution operation into distinct phases: loading input image channels, loading filter coefficients into registers, performing parallel multiplication operations, and accumulating results. This segmentation allows each phase to be optimized independently, with filter coefficients residing in fast registers and input channels being loaded and processed in parallel batches, thereby reducing total processing time while maintaining standard hardware architecture
Solution Approach 2:
The patent applies preliminary action by pre-loading filter coefficients into processor registers before the actual convolution operation begins. This preliminary placement of frequently accessed data in high-speed register memory eliminates repeated cache accesses during processing, significantly reducing convolution time and improving overall image processing efficiency without requiring specialized hardware
2Measurement precision
If complex processing tasks are implemented, then more sophisticated image analysis is achieved, but resource consumption increases and processing speed decreases
Solution Approach 1:
The patent changes the parameter of data organization by interleaving multiple input image channels into a single continuous data stream. This parameter change allows the SIMD processor to process multiple channels simultaneously using vector instructions, achieving sophisticated multi-channel image analysis while reducing the computational overhead and energy consumption associated with handling separate channels individually
3Productivity
If multiple input channels are processed separately, then each channel can be handled individually, but processing time increases and efficiency decreases
Solution Approach 1:
The patent merges multiple input image channels into a single interleaved data structure that can be processed simultaneously by the SIMD processor. This combining approach allows parallel processing of all channels through vector operations, dramatically increasing processing speed while the actual device complexity remains moderate since it leverages standard SIMD capabilities rather than requiring complex custom hardware architectures
Data Source
AI summary
A method performed by an electronic device is described. The method includes interleaving multiple input image channels to produce an interleaved multi-channel input. The method also includes loading the interleaved multi-channel input to a single-instruction multiple data (SIMD) processor. The method further includes convolving the interleaved multi-channel input with a multi-channel filter.