CUDA Resize: add optimized 3D nearest resize kernel for 5D up/down sa…

[johannes-rehm-snkeos]

Summary

This PR adds CUDA support for optimized nearest-neighbor 3D resize mapping/execution in the Resize operator path, and adds targeted regression coverage.

The implementation introduces a dedicated 3D fast path for nearest resize to handle the last three spatial dimensions (D/H/W) efficiently when outer dimensions are unchanged.

What Changed

CUDA Resize implementation

File: onnxruntime/core/providers/cuda/tensor/resize_impl.cu

• Added 3D nearest mapping kernel:
  • _ResizeNearestMappingKernel3D
• Added 3D nearest compute kernel:
  • _ResizeNearestKernel3D
• Added optimized 3D dispatch path in ResizeNearestImpl:
  • Enabled when:
    • rank >= 3
    • coordinate_transformation_mode != tf_crop_and_resize
    • all outer scales (except last 3 dims) are 1.0

This keeps existing behavior unchanged for other cases while using the optimized path for true 3D nearest resize workloads.

Regression tests

File: onnxruntime/test/providers/cpu/tensor/resize_op_test.cc

Added CUDA-targeted regression tests:

• ResizeOpNearestUpSampleTest_5D_CudaRegression_Optimized3DMapping
• ResizeOpNearestDownSampleTest_5D_CudaRegression_Optimized3DMapping

Why

The previous nearest implementation relied on the generic path for these 3D scenarios. This change introduces a dedicated CUDA 3D path to improve performance for 5D nearest resize workloads.

Fixes #14596

[johannes-rehm-snkeos]

@microsoft-github-policy-service agree

[tianleiwu]

@johannes-rehm-snkeos, could you share benchmark/profiling results that show that the new kernel is better?

[johannes-rehm-snkeos]

@tianleiwu I used your script from here: #14596 (comment) and got the following results:

Profiling of Torch:

           Type  Time(%)      Time     Calls       Avg       Min       Max  Name
 GPU activities:   54.73%  4.98039s       503  9.9014ms  5.6153ms  39.858ms  void at::native::_GLOBAL__N__b9911c4e_20_UpSampleNearest3d_cu_2b4cf812::upsample_nearest3d_out_frame<float, __operator_&__(at::native::nearest_neighbor_compute_source_index(float, int, int))>(float const *, __int64, __int64, __int64, __int64, __int64, __int64, __int64, __int64, at::native::nearest_neighbor_compute_source_index*, float, float, float)

Profiling of onnxruntime-gpu==1.24.3:

           Type  Time(%)      Time     Calls       Avg       Min       Max  Name
 GPU activities:   90.56%  40.4292s       503  80.376ms  76.639ms  98.755ms  void onnxruntime::cuda::_ResizeNearestKernel<float>(int, onnxruntime::cuda::TArray<__int64, int=8>, onnxruntime::cuda::_ResizeNearestKernel<float, onnxruntime::cuda::DivMod<int>, int=8>, float const *, onnxruntime::cuda::_ResizeNearestKernel<float, onnxruntime::cuda::DivMod<int>, int=8>*, __int64, onnxruntime::cuda::_ResizeNearestKernel<float, onnxruntime::cuda::DivMod<int>, int=8>, __int64 const *, onnxruntime::cuda::NearestMappingInfo const *)

Profiling of johannes-rehm-snkeos:cuda-resize-nearest-3d-kernel:

           Type  Time(%)      Time     Calls       Avg       Min       Max  Name
 GPU activities:   43.42%  3.75516s       503  7.4655ms  1.9272ms  37.947ms  void onnxruntime::cuda::_ResizeNearestKernel3D<float, bool=0>(__int64, __int64, __int64, __int64, __int64, int, onnxruntime::cuda::DivMod<int>, onnxruntime::cuda::DivMod, onnxruntime::cuda::DivMod, float const *, onnxruntime::cuda::DivMod<int>*, __int64, onnxruntime::cuda::DivMod<int>, onnxruntime::cuda::NearestMappingInfo const *)

[Copilot]

Pull request overview

This PR adds a CUDA optimized fast-path for nearest-neighbor 3D resize (mapping + execution) to improve performance on rank≥3 tensors where only the last three dimensions are resized and all outer-dimension scales are 1.0, and introduces CUDA-targeted regression tests to validate the new path.

Changes:

• Added CUDA nearest-neighbor 3D mapping and compute kernels and a dispatch fast-path in ResizeNearestImpl.
• Enabled the new 3D optimized path when coordinate_transformation_mode != tf_crop_and_resize and all outer scales (except last 3 dims) are exactly 1.0.
• Added CUDA regression tests covering 5D nearest upsample and downsample scenarios intended to hit the optimized 3D mapping path.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
onnxruntime/core/providers/cuda/tensor/resize_impl.cu	Introduces optimized nearest-neighbor 3D mapping/compute CUDA kernels and a conditional fast-path in ResizeNearestImpl.
onnxruntime/test/providers/cpu/tensor/resize_op_test.cc	Adds CUDA-targeted regression tests for 5D nearest resize upsample/downsample intended to exercise the optimized 3D path.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

[tianleiwu]

/azp run Linux QNN CI Pipeline, Win_TRT_Minimal_CUDA_Test_CI, Windows ARM64 QNN CI Pipeline, Windows GPU Doc Gen CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 4 pipeline(s).