perf: reduce TensorView metadata allocations

[voltjia]

Summary

• Add exact by-value TensorView::Shape / TensorView::Strides constructor overloads so owned metadata can be moved into TensorView without redundant copies.
• Construct initializer-list metadata directly, and add Linux allocation regression coverage plus focused construction, indexing, and transpose microbenchmarks.

Motivation

TensorView is intended to be a lightweight framework-neutral tensor metadata layer, but common rvalue, initializer-list, operator[], and T() paths created temporary vectors and then copied them into the result. Non-empty rank-2 views therefore performed three or four metadata allocations where two owned buffers are sufficient.

The existing public Shape = std::vector<Size> and Strides = std::vector<Stride> aliases, generic constructors, object data members, and behavior are preserved in this first phase.

Related issue: N/A - this performance work is not linked to an existing issue.

Type of Change

• feat - new feature / new backend capability / new public API
• fix - bug fix
• perf - performance improvement without behavior 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 (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• Hygon (WITH_HYGON)
• MetaX (WITH_METAX)
• Moore (WITH_MOORE)
• Cambricon (WITH_CAMBRICON)
• Ascend (WITH_ASCEND)
• Build system / CMake / generated headers
• Public headers / installed consumer API
• Documentation only

Smoke Build and Test Result

CPU Release build on the Linux x86_64 NVIDIA validation host (GCC 11.4.0):

cmake -S . -B build-cpu \
  -DCMAKE_BUILD_TYPE=Release \
  -DWITH_CPU=ON \
  -DINFINI_RT_BUILD_TESTING=ON \
  -DINFINI_RT_BUILD_PERFORMANCE_TESTING=ON
cmake --build build-cpu -j2
ctest --test-dir build-cpu --output-on-failure

100% tests passed, 0 tests failed out of 10

NVIDIA Release build in the existing accelerator-dev/nvidia:latest image on 8x A100-SXM4-80GB (CUDA 13.1.80, GCC 13.3.0):

docker run --rm --gpus all \
  -v "$PWD:/workspace/InfiniRT" -w /workspace/InfiniRT \
  accelerator-dev/nvidia:latest \
  cmake -S . -B build-nvidia \
    -DCMAKE_BUILD_TYPE=Release \
    -DWITH_NVIDIA=ON \
    -DINFINI_RT_BUILD_TESTING=ON \
    -DINFINI_RT_BUILD_PERFORMANCE_TESTING=ON
docker run --rm --gpus all \
  -v "$PWD:/workspace/InfiniRT" -w /workspace/InfiniRT \
  accelerator-dev/nvidia:latest \
  cmake --build build-nvidia -j2
docker run --rm --gpus all --entrypoint ctest \
  -v "$PWD:/workspace/InfiniRT" -w /workspace/InfiniRT \
  accelerator-dev/nvidia:latest \
  --test-dir build-nvidia -E '^perf_' --output-on-failure

100% tests passed, 0 tests failed out of 8

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
CPU	Yes	Full build and tests passed	10/10 passed, including allocation, performance, install, and installed-consumer tests
NVIDIA	Yes	NVIDIA build and non-performance smoke tests passed	8/8 smoke tests passed; the final clang-format-only amend rebuilt successfully in the same image
Iluvatar	Yes	Not run - SDK/host unavailable in this pass	Backend-specific code is unchanged; maintainer smoke validation requested
Hygon	Yes	Not run - SDK/host unavailable in this pass	Backend-specific code is unchanged; maintainer smoke validation requested
MetaX	Yes	Not run - SDK/host unavailable in this pass	Backend-specific code is unchanged; maintainer smoke validation requested
Moore	Yes	Not run - SDK/host unavailable in this pass	Backend-specific code is unchanged; maintainer smoke validation requested
Cambricon	Yes	Not run - SDK/host unavailable in this pass	Backend-specific code is unchanged; maintainer smoke validation requested
Ascend	Yes	Not run - SDK/host unavailable in this pass	Backend-specific code is unchanged; maintainer smoke validation requested

Full `ctest` output

CPU:
100% tests passed, 0 tests failed out of 10
Total Test time (real) = 2.55 sec

NVIDIA smoke (performance executables excluded; run before the final clang-format-only amend):
100% tests passed, 0 tests failed out of 8
Total Test time (real) = 11.75 sec

The amended head rebuilt successfully in the same NVIDIA image.

Benchmark / Performance Impact

Harness: tests/performance/perf_tensor_view.cc, Release CPU backend on the same Linux x86_64 validation host, GCC 11.4.0, rank 2, shape [32, 64], float32 where applicable, 1,000 warmup iterations, 200,000 measured iterations per sample, and 7 samples. Values below are mean nanoseconds per operation.

Benchmark	Baseline	Candidate	Delta
rvalue full metadata construction	74.86 ns	37.27 ns	-50.2%
rvalue default-stride construction	64.61 ns	41.99 ns	-35.0%
initializer-list construction	92.58 ns	48.24 ns	-47.9%
operator[]	91.90 ns	48.88 ns	-46.8%
T()	95.51 ns	49.70 ns	-48.0%
numel() control	3.92 ns	3.92 ns	0.0%

The allocation regression test verifies the corresponding metadata allocation reductions: rvalue full metadata, initializer-list, operator[], and T() paths decrease from four allocations to two; rvalue default-stride construction decreases from three to two. Copy and move construction remain at two and zero allocations, respectively.

Notes for Reviewers

• This is intentionally a first-phase optimization: Shape and Strides remain std::vector, generic tensor-like/range constructors remain available, and TensorView data member types and order are unchanged.
• The allocation-count test is registered only on Linux because it relies on ELF cross-DSO operator new interposition.
• Small-vector storage, borrowed metadata views, downstream pass-by-reference changes, and unrelated TensorView correctness fixes are deliberately out of scope.
• git diff --check and both clang-format 21 CI checks pass.