UPSTREAM PR #17042: ggml-hexagon: fix test-backend-ops failures on specific binary ops

[DajanaV]

Mirrored from ggml-org/llama.cpp#17042

Summary

Fixes test-backend-ops failures in ggml-hexagon by correcting the index calculation for binary operations.

Changes

• Fixed index calculation in binary ops to align with the cpu implementation in ggml/src/ggml-cpu/ops.cpp

Testing

TODO:

[loci-review]

Access the complete analysis in the LOCI Dashboard

Performance Analysis Summary

Overview

Analysis of version 274a4670-5520-4a90-8221-9afd5d526898 compared to base version 0211d925-b37b-4bdf-9ef8-7fce0422fc7b reveals minimal performance variations with no impact on core inference functions.

Key Findings

Performance Metrics:

• Highest Response Time change: can_reuse function (+0.096%, +0.06 ns) in build.bin.libllama.so
• Highest Throughput change: common_sampler_accept function (-0.11%, -0.07 ns improvement) in build.bin.llama-tts
• Both functions are non-core utilities with negligible absolute impact

Core Function Impact:
No changes detected in critical inference functions (llama_decode, llama_encode, llama_tokenize). The modified functions operate in auxiliary roles:

• can_reuse: Graph parameter validation utility
• common_sampler_accept: Sampling mechanism helper

Inference Performance Impact:
Token throughput remains unaffected as no core tokenization or inference functions show performance changes. The reference baseline (7% tokens/second reduction per 2ms llama_decode slowdown) does not apply since llama_decode shows no measurable change.

Power Consumption Analysis:
System-wide power consumption remains stable across all binaries with changes <0.001%. Affected binaries show negligible decreases:

• build.bin.libllama.so: -0.0002% power reduction
• build.bin.llama-tts: -0.0002% power reduction

Technical Analysis:

• Flame Graph: can_reuse exhibits simple leaf function behavior with 65 ns self-contained execution
• CFG Comparison: Identical assembly code between versions indicates performance variation stems from external factors (memory layout, cache alignment) rather than algorithmic changes
• Code Review: PR UPSTREAM PR #17042: ggml-hexagon: fix test-backend-ops failures on specific binary ops #100 modifies GGML Hexagon backend binary operations, unrelated to the measured performance changes

Conclusion:
The analysis reveals stable performance with minor variations in non-critical functions. No regressions affect core inference capabilities, and the system maintains consistent energy efficiency and computational performance.