[zzhfz]

Summary

• Add an optional, off-by-default detailed dispatch trace in src/operator.h.
• Emit each tensor's index, device, dtype, shape, and strides at operator call entry.
• Keep the existing INFINI_OPS_TRACE_CALLS output unchanged.

Motivation

The existing operator-call trace reports the operator name, device, and
implementation index, but does not expose tensor metadata needed for
per-operator cost analysis.

The new INFINI_OPS_TRACE_CALLS_DETAILED mode provides this metadata without
reading tensor values, adding device synchronization, or changing operator
computation and scheduling.

No related issue.

Type of Change

• feat — new feature / new operator / new platform
• fix — bug fix
• perf — performance improvement (no behavioral change)
• refactor — code restructuring without behavior change
• test — adding or fixing tests only
• docs — documentation only
• build / ci — build system or CI configuration
• chore — tooling, formatting, or other non-code changes
• Breaking change

Platforms Affected

This change is in the shared operator dispatch path. It is active only when
INFINI_OPS_TRACE_CALLS_DETAILED is enabled.

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

Targeted CPU validation was run with a local pytest script against a CPU smoke
build. The validation script is not included in this commit.

tests/test_trace_calls.py
7 passed in 15.91s

##Validated:

• trace disabled, basic-only, detailed-only, and both enabled
• float32, float16, and bfloat16 metadata
• contiguous and non-contiguous strides
• concurrent JSON record parsing

Moore / MUSA

Tested on an MTT S5000 with a Moore-only build (WITH_CPU=OFF).

[INFINI_OPS_TRACE_CALLS_DETAILED] {"operator_name": "Add", "implementation": 0, "tensors": [{"index": 0, "device": "moore", "dtype": "float16", "shape": [8], "strides": [1]}, ...]}
ADD_OK=True

[INFINI_OPS_TRACE_CALLS_DETAILED] {"operator_name": "Gemm", "implementation": 0, "tensors": [{"index": 0, "device": "moore", "dtype": "float16", "shape": [4,8], "strides": [8,1]}, ...]}
GEMM_OK=True