[Precision Depth Alignment] align paddle.nn.functional.interpolate forward

[zrr1999]

PR Category

Operator Mechanism

PR Types

Performance

Description

受影响的 API：

• paddle.nn.functional.interpolate ，新增对齐 case 3900/4769 个，现存精度问题case 620 个。

修改内容

• 规范 get_cubic_upsample_coefficients 重命名为 GetCubicUpsampleCoefficients，保持与其他函数命名风格一致。
• 添加 AreaPixelComputeScale 函数，提高代码复用性，并统一计算逻辑。AreaPixelComputeScale 部分与原本替代掉的逻辑不同，当 output_size > 1 时，原本逻辑是直接设置 ratio = 0，现在是根据align_corners的状态判断，为True时与原本逻辑一致，为 False时不进行该条件的判断。
• 移除 CubicConvolution2、CubicConvolution1和GetCubicUpsampleCoefficients 中不必要的static_cast。
• ratio_x 参数计算时使用与MT一致的数据类型，同时参数传递过程中使用double避免精度损失。
• 增加 InterpAntialiasKernel，使用抗锯齿算法，目前仅支持bilinear，bicubic算法。
• 修复单测问题，增加新单测

TODO

• 由于目前 Paddle 插值算法不同的 mode 共用同一个前处理，而nearest的T可能是整数，所以将
  funcs::AreaPixelComputeScale 统一写成了funcs::AreaPixelComputeScale，后续需要修改成funcs::AreaPixelComputeScale
• InferSymbolicShapeInterface 暂时注释，后续添加单测后再加回来。

注意

torch 的 AreaPixelComputeScale 实现有一些问题，

template <typename T>
HOSTDEVICE inline T AreaPixelComputeScale(int64_t input_size,
                                          int64_t output_size,
                                          bool align_corners,
                                          const T scale) {
  if (align_corners) {
    if (output_size > 1) {
      return static_cast<T>(input_size - 1) / (output_size - 1);
    } else {
      return static_cast<T>(0);
    }
  } else {
    return (scale > 0.) ? static_cast<T>(1.0) / scale
                        : (static_cast<T>(input_size) / output_size);
  }
}

上述 torch 等价实现中，当align_corners=False，scale=0, output_size=0 时，会出现x/0的情况，所以优化为下述实现（在合法情况下结果一致）。

template <typename T>
inline T AreaPixelComputeScale(int64_t input_size,
                               int64_t output_size,
                               bool align_corners,
                               const T scale) {
  if (align_corners) {
    if (output_size > 1) {
      return static_cast<T>(input_size - 1) / (output_size - 1);
    }
  } else {
    if (scale > 0.) {
      return static_cast<T>(1.0) / scale;
    }
    if (output_size > 0) {
      return static_cast<T>(input_size) / output_size;
    }
  }
  return static_cast<T>(0);
}

Pcard-67164

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[Copilot]

Pull Request Overview

This pull request aligns the precision and behavior of paddle.nn.functional.interpolate forward pass with reference implementations. The changes focus on improving numerical precision, adding anti-aliasing support for bilinear and bicubic interpolation modes, and refactoring the codebase for better maintainability.

Key Changes:

• Added anti-aliasing support for bilinear and bicubic interpolation modes with new interp_antialias kernel
• Unified ratio computation logic using new AreaPixelComputeScale and AreaPixelComputeSourceIndex helper functions
• Changed ratio parameters from float to double throughout the codebase to improve precision
• Renamed get_cubic_upsample_coefficients to GetCubicUpsampleCoefficients for naming consistency

Reviewed Changes

Copilot reviewed 15 out of 15 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
test/xpu/test_bilinear_interp_v2_op_xpu.py	Updated ratio computation logic to match new implementation
test/legacy_test/test_trilinear_interp_v2_op.py	Fixed compute_ratio function and corrected test parameters
test/legacy_test/test_nn_functional_interpolate.py	Removed anti-aliasing unsupported test (now supported)
test/legacy_test/test_bilinear_interp_v2_op.py	Added anti-aliasing tests and updated reference implementation
test/legacy_test/test_bicubic_interp_v2_op.py	Added anti-aliasing tests and updated reference implementation
python/paddle/nn/functional/common.py	Removed anti-aliasing restriction and added new operator integration
paddle/phi/ops/yaml/ops.yaml	Added interp_antialias operator definition
paddle/phi/ops/yaml/backward.yaml	Added interp_antialias_grad backward operator definition
paddle/phi/kernels/gpu/interpolate_kernel.cu	Implemented anti-aliasing kernels and improved precision with double
paddle/phi/kernels/gpu/interpolate_grad_kernel.cu	Updated gradient kernels to use double and fixed type casting
paddle/phi/kernels/funcs/interpolate_function.h	Added new helper functions and renamed existing functions
paddle/phi/kernels/cpu/interpolate_kernel.cc	Updated CPU kernels to use new helper functions
paddle/phi/kernels/cpu/interpolate_grad_kernel.cc	Updated CPU gradient kernels to use new helper functions
paddle/fluid/pir/dialect/operator/interface/infer_symbolic_shape/multiary_infer_sym.h	Added InterpAntialias symbolic shape inference declaration
paddle/fluid/pir/dialect/operator/interface/infer_symbolic_shape/multiary_infer_sym.cc	Added InterpAntialias symbolic shape inference implementation

---

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

[codecov-commenter]

Codecov Report

❌ Patch coverage is 38.29787% with 29 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@3569629). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
paddle/phi/kernels/cpu/interpolate_grad_kernel.cc	0.00%	16 Missing ⚠️
paddle/phi/kernels/funcs/interpolate_function.h	28.57%	10 Missing ⚠️
paddle/phi/kernels/cpu/interpolate_kernel.cc	70.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #76246   +/-   ##
==========================================
  Coverage           ?   38.29%           
==========================================
  Files              ?        4           
  Lines              ?       47           
  Branches           ?        0           
==========================================
  Hits               ?       18           
  Misses             ?       29           
  Partials           ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Copilot reviewed 15 out of 15 changed files in this pull request and generated 8 comments.

---

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

[zhangting2020]

LGTM for [Distribute-stable-test / Test

[wanghuancoder]

LGTM

[XiaoguangHu01]

LGTM