[Danny On The Air (dannyota)]

Description

ConvBase::IsOnnxNodeSupported rejected any Conv/ConvTranspose whose C, H or W was unknown at partitioning time, so a graph with static channels and a symbolic width got no XNNPACK convolutions, although the kernel is created the same way for any width.

• Allow symbolic or unknown spatial dims when C is static. Kernels are still created once at prepack; the output shape and pads are computed per run from the actual input. Static shapes keep the cached path.
• ConvTranspose with SAME_* auto_pad or an output_shape attribute still falls back to CPU, since those derive padding from the input size.
• Reshaping an XNNPACK operator for a new spatial size reallocates its indirection buffer, but ORT's xnn_reallocate was ORT_NOT_IMPLEMENTED. Implement it with realloc semantics on ORT's allocator (size header, alignment kept, nullptr/zero handled) and cap request sizes at SIZE_MAX / 2.
• XNNPACK's deconvolution reshape (reshape_subconv2d_path) updates part of its state before it reallocates. After a failed reshape, a dynamic ConvTranspose now returns an error instead of running on inconsistent state. Conv is not affected: it reallocates before updating cached dims.

Testing (macOS arm64, --use_xnnpack --build_shared_lib)

• onnxruntime_provider_test --gtest_filter='*Conv*:*Xnnpack*': 443 pass, 0 fail (CUDA-only tests skipped).
• New: TestConvDynamicSpatial, TestConvTransposeDynamicSpatial and TestDynamicSpatialRepeatedRuns compare with CPU across widths 17, 64, 17, 18 and assert per-node XNNPACK placement; fallback tests for unknown C and the excluded ConvTranspose attributes; AllocatorReallocate, AllocatorSizeLimit; failed-reshape tests with an injected allocation failure and an allocator exception.

Motivation and Context

Fixes #32387.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🔵 Needs a closer look

The pinned XNNPACK sources are absent from this snapshot, so human review must confirm the reshape state-update ordering underlying Conv retry safety and ConvTranspose failure handling.

No open issues 🎉

What changed in this PR

Enables XNNPACK to handle Conv and ConvTranspose with symbolic spatial dimensions and static channels, addressing #32387’s streaming-inference use case while preserving the cached path for static shapes.

Changes:

• Computes dynamic output shapes at runtime and retains CPU fallback for unsupported ConvTranspose attributes.
• Implements allocator reallocation and prevents reuse of dynamic ConvTranspose operators after reshape or setup failures.
• Adds coverage for changing input sizes, provider placement, fallback behavior, allocation limits, and failure handling.

File	Description
onnxruntime/​test/​providers/​xnnpack/​xnnpack_basic_test.cc	Adds dynamic-shape, allocator, fallback, and failure tests.
onnxruntime/​core/​providers/​xnnpack/​xnnpack_init.h	Declares the test-only allocation-failure hook.
onnxruntime/​core/​providers/​xnnpack/​xnnpack_init.cc	Implements aligned reallocation semantics and allocation-size limits.
onnxruntime/​core/​providers/​xnnpack/​nn/​conv.cc	Uses runtime output-shape calculation.
onnxruntime/​core/​providers/​xnnpack/​nn/​conv_transpose.h	Tracks whether the operator is unusable after failure.
onnxruntime/​core/​providers/​xnnpack/​nn/​conv_transpose.cc	Calculates runtime shapes and guards against reuse after failure.
onnxruntime/​core/​providers/​xnnpack/​nn/​conv_base.h	Declares shared shape calculation and dynamic-dimension state.
onnxruntime/​core/​providers/​xnnpack/​nn/​conv_base.cc	Relaxes spatial-dimension checks and implements dynamic shape calculation.
cmake/​onnxruntime_providers_xnnpack.cmake	Enables allocation-failure injection for unit-test builds.

_{Review effort: Balanced}