Fix round_prefer_ceil nearest mode for negative halfway values in Resize op

[Copilot]

Description

ROUND_PREFER_CEIL in the Resize operator used bare std::round/roundf, which rounds away from zero. This is correct for positive halfway values (e.g., round(0.5) = 1 = ceil(0.5)) but wrong for negative halfway values (e.g., round(-0.5) = -1, but ceil(-0.5) = 0).

Negative coordinates occur naturally with the half_pixel coordinate transformation mode for the first output pixels when upsampling.

Added an explicit negative-halfway check, mirroring the existing positive-halfway check in ROUND_PREFER_FLOOR:

// CPU (upsamplebase.h)
case ROUND_PREFER_CEIL:
  return [](float x_original, bool) {
    if (x_original == static_cast<int64_t>(x_original) - 0.5f) {
      return static_cast<int64_t>(std::ceil(x_original));
    }
    return static_cast<int64_t>(std::round(x_original));
  };

Same fix applied to the CUDA implementation (resize_impl.cu).

Added two test cases in resize_op_test.cc:

1. ResizeOpNearestUpSample_RoundPreferCeil_HalfPixel — exercises non-integer scale (26→64) from the original issue report, verifying correct source pixel selection at fractional boundaries.
2. ResizeOpNearestUpSample_RoundPreferCeil_HalfPixel_2x2to7x8 — exercises a positive 0.5 boundary where round_prefer_ceil selects ceiling.

Motivation and Context

The round_prefer_floor path already had an explicit halfway-case override (for positive values where std::round disagrees with floor). The round_prefer_ceil path was missing the symmetric fix for negative values, violating the ONNX spec semantics of "at ties, prefer ceiling."

[tianleiwu]

@copilot, please add a test case for this.

[Copilot]

@copilot, please add a test case for this.

Added two test cases in commit 076bea0:

1. ResizeOpNearestUpSample_RoundPreferCeil_HalfPixel — exercises non-integer scale (26→64) from the original issue
2. ResizeOpNearestUpSample_RoundPreferCeil_HalfPixel_2x2to7x8 — exercises a positive 0.5 boundary where round_prefer_ceil selects ceiling

Both pass on the CPU provider.

[github-actions]

You can commit the suggested changes from lintrunner.

[tianleiwu]

Summary

Clean, well-scoped bug fix that adds the missing negative-halfway-value handling to ROUND_PREFER_CEIL in both CPU and CUDA Resize implementations. The fix mirrors the existing ROUND_PREFER_FLOOR positive-halfway check symmetrically. Two new test cases exercise the fix with different scale factors. Low risk.

CPU & CUDA Fix

The fix correctly mirrors the ROUND_PREFER_FLOOR pattern: FLOOR checks + 0.5f for positive halfway values where std::round/roundf (rounds away from zero) disagrees with floor; CEIL now checks - 0.5f for negative halfway values where std::round/roundf disagrees with ceil. The symmetry is clean and easy to reason about.

CUDA implementation uses _Ceil consistently with the existing _Floor usage in NearestPixel_ROUND_PREFER_FLOOR, and static_cast<int> matching the CUDA convention.

Tests

Two complementary test cases: one exercises a non-integer scale (26→64) with negative intermediate coordinates where the bug manifests, the other exercises positive 0.5 boundaries confirming round_prefer_ceil selects ceiling there too. Both follow existing patterns.