CNN Shader Processing Using Texture Arrays on GPUs

The patent applies texture arrays instead of traditional one-dimensional buffers, utilizing multi-dimensional texture storage formats (2D, 3D, or 4D arrays) to organize weights and input data. This dimensional transformation enables GPUs to leverage their inherent parallel architecture more effectively, allowing simultaneous processing across multiple channels and spatial dimensions, thereby significantly improving processing speed and reducing computation time.

The patent replaces traditional CPU-based sequential processing with GPU-based parallel processing. By formatting convolutional layers as shaders and data as textures, the system leverages the GPU's massively parallel architecture, which can execute thousands of threads simultaneously, substituting the mechanical sequential execution model with a parallel computational model that dramatically accelerates convolution operations.

The patent uses a unified texture array format that can represent weights, biases, and input data across different convolutional layers and operations. This multi-functional approach allows the same GPU shader code and texture binding mechanism to handle various CNN operations (convolution, pooling, activation), reducing the need for separate processing paths and lowering overall system complexity despite high parallelization.

The patent changes the data representation parameters by using normalized floating-point values in texture arrays instead of integer formats, and by organizing data in channel-first or channel-last conventions. These parameter changes enable more efficient parallel computation on GPUs while maintaining compatibility with standard CNN architectures, balancing computational speed with implementation simplicity.