UPSTREAM PR #17883: ggml : remove redundant src in ggml_cast

[loci-dev]

Mirrored from ggml-org/llama.cpp#17883

Don't think this src[1] is needed

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary

Overview

The code change removes a self-referential assignment (result->src[1] = result) from the ggml_cast function in ggml.c. This modification eliminates an unnecessary pointer assignment during tensor type casting operations.

Key Findings

Performance-Critical Area Impact:

The change affects the ggml_cpy_impl function in build.bin.libggml-base.so, which shows:

• Response time increased by 21 ns (from 2236 ns to 2257 ns)
• Throughput decreased by 1 operation (from 92 to 91 operations)

Inference Impact:

No measurable impact on tokens per second is expected. The affected function ggml_cpy_impl is a tensor utility operation used during graph construction, not during the primary inference path. The 21 ns response time increase is negligible compared to the millisecond-scale operations in llama_decode that directly affect tokenization throughput. Core inference functions (llama_decode, llama_encode, llama_tokenize) remain unmodified.

Code Change Analysis:

The removed line result->src[1] = result was creating a circular reference where the result tensor pointed to itself as a source operand. This self-reference was semantically incorrect for the CPY operation, as the actual source tensor is already assigned to result->src[0]. The fix improves code correctness by eliminating this redundant assignment, with minimal performance overhead during tensor creation.