ggml/llama.cpp releases: b10664

TL;DR

The ggml-hexagon project added support for HTP unary operations (ABS and LOG) to accelerate GGML_OP_LOG and GGML_UNARY_OP_ABS on the Hexagon backend, enhancing performance. This update is significant as it improves computational efficiency in processing these mathematical functions using the Hexagon hardware.

Detailed Summary

The ggml-hexagon project, part of the llama.cpp suite, has updated its codebase with new unary operations for ABS and LOG. These updates include HVX-accelerated implementations specifically for the HTP backend, enhancing performance. The changes are registered in relevant support functions to ensure compatibility and efficiency across the system.

Key Points

• • Add HTP unary ops for ABS and LOG
• • Implement HVX-accelerated GGML_OP_LOG and GGML_UNARY_OP_ABS
• • Register HTP_OP_UNARY_ABS and HTP_OP_UNARY_LOG in op_remap_to_htp()
• • Update ggml_backend_hexagon_device_supports_op() to include ABS and LOG