Hexagon Unveils Conv_2D Optimizations and New Conv_3D Support

Hexagon has made a significant leap in its processing capabilities with the introduction of optimizations for Conv_2D and the new support for Conv_3D. This update, tagged as b12179, was officially released on November 6, 2026, and is expected to enhance the performance of various applications relying on these convolutional operations.

The latest update includes a generalization of the Conv_2D implementation to support Conv_3D, which is crucial for applications requiring three-dimensional data processing. The new operation, HTP_OP_CONV_3D, has been integrated through the htp-ops.h and main.c files, sharing the existing op_conv_2d functionality. This integration allows Conv_3D to run on the established HMX tiled (implicit GEMM) Conv_2D path, effectively folding the kernel dimensions into a virtual kernel row. The output rows are now calculated as N*OD*OH, which optimizes the processing of multi-dimensional data.

Claude Code, a key contributor to this update, emphasized the importance of this enhancement, stating, "By generalizing the Conv_2D implementation, we are not only improving efficiency but also expanding the capabilities of our processing units to handle more complex tasks."

The update also includes several technical improvements aimed at enhancing the safety and efficiency of the processing operations. One notable change is the modification of checks to eliminate divide-by-zero hazards, which can lead to critical failures during processing. This adjustment ensures that the Conv_3D operations can proceed without the risk of encountering such errors.

Additionally, the Conv_2D tiled processing has been refined to compute the Conv_3D unfold divisors directly on the host, rather than on the device. This change reduces the computational load on the processing unit and improves overall performance. The fast division values for the kernel parameters have been optimized to fit within the existing struct size, ensuring that the update does not increase the memory footprint.

Another significant feature of this update is the support for arbitrary positive strides in Conv_2D operations. This enhancement allows for greater flexibility in processing different input formats, such as WHCN and CWHN. The update preserves load and shuffle paths for common strides while utilizing gather operations for larger strides, further optimizing the processing capabilities.

Moreover, Hexagon has improved the tile size selection for Conv_2D operations. The new approach prefers the largest feasible output-channel block, deriving spatial chunk capacity from VTCM usage. This method compares capacity and thread-aligned candidates, ensuring that the processing units operate at maximum efficiency.

The development of this update has been a collaborative effort, with contributions from various developers, including Yiwei Shao and Max Krasnyansky. Their involvement highlights the community-driven nature of Hexagon's advancements, which rely on collective expertise to push the boundaries of processing technology.

Looking ahead, Hexagon aims to continue refining its convolutional operations and expanding support for more complex data types. The introduction of Conv_3D is just the beginning, as the team explores further optimizations and enhancements that can be integrated into future updates.

The release of the b12179 update marks a pivotal moment for Hexagon, as it not only enhances the existing Conv_2D operations but also introduces robust support for Conv_3D. These advancements are set to significantly improve the performance of applications that rely on convolutional processing, paving the way for more sophisticated and efficient technologies in the future. As the demand for high-performance computing continues to grow, Hexagon's commitment to innovation remains steadfast, promising exciting developments in the realm of processing capabilities.