Add direct XYWH-CXCYWH conversion for better performance

[raimbekovm]

Summary

Implements direct conversion between XYWH and CXCYWH bounding box formats, removing the need for intermediate XYXY conversion. This resolves the TODO comment at line 307 in _meta.py.

Performance improvement: ~2.3x faster for XYWH↔CXCYWH conversions.

Changes

1. Added _xywh_to_cxcywh() function for direct XYWH → CXCYWH conversion
2. Added _cxcywh_to_xywh() function for direct CXCYWH → XYWH conversion
3. Updated _convert_bounding_box_format() to use direct conversion when possible
4. Removed the TODO comment

Both functions correctly handle integer tensors using the same rounding behavior as the two-step conversion path.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9326

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ You can merge normally! (1 Unrelated Failure)

As of commit 8517333 with merge base ec4f95a ():

FLAKY - The following job failed but was likely due to flakiness present on trunk:

• CMake / linux (linux.g5.4xlarge.nvidia.gpu, cuda, 12.6) / linux-job (gh) (detected as infra flaky with no log or failing log classifier)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[zy1git]

Thanks for the PR. I left a comment for a minor thing. Feel free to take a look.

[zy1git]

@raimbekovm Thanks a lot for this PR. Could you please add the parity test against reference implementation (the implementation which needs for intermediate XYXY conversion)?

Also, in your description, you mentioned that "Performance improvement: ~2.3x faster for XYWH↔CXCYWH conversions.". Could you please share benchmark script or benchmark configuration (shapes etc.) to show how you get 2.3x faster?

And I also notice that you submitted PR#9322. If that PR is correct (I am still reviewing it), the similar fix can be applied here for the function _cxcywh_to_xywh.

[raimbekovm]

@zy1git Thanks for the review! Here's the benchmark results and script.

Benchmark Results

num_boxes	dtype	conversion	direct	two-step	speedup
100	float32	xywh_to_cxcywh	0.0043ms	0.0095ms	2.24x
100	float32	cxcywh_to_xywh	0.0043ms	0.0088ms	2.04x
1,000	float32	xywh_to_cxcywh	0.0059ms	0.0135ms	2.29x
1,000	float32	cxcywh_to_xywh	0.0059ms	0.0126ms	2.15x
10,000	float32	xywh_to_cxcywh	0.0333ms	0.0778ms	2.34x
10,000	float32	cxcywh_to_xywh	0.0470ms	0.0885ms	1.88x
100,000	float32	xywh_to_cxcywh	0.2155ms	0.3804ms	1.77x
100,000	float32	cxcywh_to_xywh	0.1969ms	0.3562ms	1.81x
1,000	int64	xywh_to_cxcywh	0.0085ms	0.0176ms	2.06x
1,000	int64	cxcywh_to_xywh	0.0100ms	0.0177ms	1.76x
10,000	int64	xywh_to_cxcywh	0.0741ms	0.1192ms	1.61x
10,000	int64	cxcywh_to_xywh	0.0808ms	0.1229ms	1.52x

Average speedup: ~1.9x (ranges from 1.5x to 2.3x depending on tensor size and dtype)

The speedup comes from avoiding:

1. Intermediate XYXY tensor allocation
2. Extra arithmetic operations in the two-step path

Benchmark Script

import torch
import time

# Conversion functions from torchvision/transforms/v2/functional/_meta.py
def _xywh_to_xyxy(xywh, inplace):
    xyxy = xywh if inplace else xywh.clone()
    xyxy[..., 2:] += xyxy[..., :2]
    return xyxy

def _xyxy_to_cxcywh(xyxy, inplace):
    if not inplace:
        xyxy = xyxy.clone()
    xyxy[..., 2:].sub_(xyxy[..., :2])
    xyxy[..., :2].mul_(2).add_(xyxy[..., 2:]).div_(2, rounding_mode=None if xyxy.is_floating_point() else "floor")
    return xyxy

def _cxcywh_to_xyxy(cxcywh, inplace):
    if not inplace:
        cxcywh = cxcywh.clone()
    half_wh = cxcywh[..., 2:].div(-2, rounding_mode=None if cxcywh.is_floating_point() else "floor").abs_()
    cxcywh[..., :2].sub_(half_wh)
    cxcywh[..., 2:].add_(cxcywh[..., :2])
    return cxcywh

def _xyxy_to_xywh(xyxy, inplace):
    xywh = xyxy if inplace else xyxy.clone()
    xywh[..., 2:] -= xywh[..., :2]
    return xywh

# New direct conversion functions
def _xywh_to_cxcywh(xywh, inplace):
    if not inplace:
        xywh = xywh.clone()
    xywh[..., :2].add_(xywh[..., 2:].div(2, rounding_mode=None if xywh.is_floating_point() else "floor"))
    return xywh

def _cxcywh_to_xywh(cxcywh, inplace):
    if not inplace:
        cxcywh = cxcywh.clone()
    half_wh = cxcywh[..., 2:].div(-2, rounding_mode=None if cxcywh.is_floating_point() else "floor").abs_()
    cxcywh[..., :2].sub_(half_wh)
    return cxcywh

def benchmark(func, data, warmup=10, iterations=100):
    for _ in range(warmup):
        func(data.clone(), inplace=False)
    start = time.perf_counter()
    for _ in range(iterations):
        func(data.clone(), inplace=False)
    return (time.perf_counter() - start) / iterations * 1000

# Run benchmark
for num_boxes, dtype in [(100, torch.float32), (1000, torch.float32), (10000, torch.float32), (1000, torch.int64)]:
    data = torch.rand(num_boxes, 4, dtype=dtype) * 1000 if dtype.is_floating_point else torch.randint(0, 1000, (num_boxes, 4), dtype=dtype)
    direct = benchmark(lambda x, i: _xywh_to_cxcywh(x, i), data)
    two_step = benchmark(lambda x, i: _xyxy_to_cxcywh(_xywh_to_xyxy(x, False), False), data)
    print(f"{num_boxes} boxes, {dtype}: direct={direct:.4f}ms, two-step={two_step:.4f}ms, speedup={two_step/direct:.2f}x")

---

I've also added the parity test in the latest commit (992ef20).

Regarding PR #9322 - I'll apply the similar fix to _cxcywh_to_xywh once that PR is merged.

[zy1git]

left a comment on the test.

[zy1git]

The "new_format" is not passed to the test function. We can remove it, correct?

[raimbekovm]

Good point on both!

1. _xywh_to_cxcywh doesn't need half_wh since it's just a simple div(2).add() — no ceil trick like in _cxcywh_to_xywh. Let me know if you'd still prefer it for consistency.

2. You're right, new_format is unused — I'll remove it.

[zy1git]

@raimbekovm The new commit looks great! Thank you for your contribution!

However, it seems that the CI test fails because of a Lint problem. The error is that The ufmt formatter expects lines 184-186 of the _meta.py to be a single line instead of being split across three lines.

The thing is that you don't need to manually fix this. If you follow the contributing guide to set up your env, when you commit, the ufmt will do the linting fix automatically. Could you try that?

[raimbekovm]

Fixed! Thanks for the hint about ufmt.