Fix paddle.vision.ops.yolo_loss for big tensor

[zrr1999]

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

非法配置已根据文档修改为以下内容

paddle.vision.ops.yolo_loss(Tensor([2863312, 30, 5, 5],"float64"), gt_box=Tensor([2863312, 5, 4],"float64"), gt_label=Tensor([2863312, 5],"int32"), anchors=list[10,13,16,30,33,23,30,61,62,45,59,119,116,90,156,198,373,326,], anchor_mask=list[0,1,2,], class_num=5, ignore_thresh=0.7, downsample_ratio=32, gt_score=None, use_label_smooth=True, scale_x_y=1.0, ) paddle.vision.ops.yolo_loss(Tensor([2863312, 30, 5, 5],"float64"), gt_box=Tensor([2863312, 5, 4],"float64"), gt_label=Tensor([2863312, 5],"int32"), anchors=list[10,13,16,30,33,23,30,61,62,45,59,119,116,90,156,198,373,326,], anchor_mask=list[0,1,2,], class_num=5, ignore_thresh=0.7, downsample_ratio=32, gt_score=Tensor([2863312, 5],"float64"), use_label_smooth=False, scale_x_y=1.0, ) paddle.vision.ops.yolo_loss(Tensor([2863312, 30, 5, 5],"float64"), gt_box=Tensor([2863312, 5, 4],"float64"), gt_label=Tensor([2863312, 5],"int32"), anchors=list[10,13,16,30,33,23,30,61,62,45,59,119,116,90,156,198,373,326,], anchor_mask=list[0,1,2,], class_num=5, ignore_thresh=0.7, downsample_ratio=32, gt_score=Tensor([2863312, 5],"float64"), use_label_smooth=True, scale_x_y=1.0, ) paddle.vision.ops.yolo_loss(Tensor([4793491, 14, 8, 8],"float32"), gt_box=Tensor([4793491, 10, 4],"float32"), gt_label=Tensor([4793491, 10],"int32"), anchors=list[10,13,16,30,], anchor_mask=list[0,1,], class_num=2, ignore_thresh=0.7, downsample_ratio=8, use_label_smooth=True, scale_x_y=1.0, ) paddle.vision.ops.yolo_loss(x=Tensor([11665, 255, 38, 38],"float32"), gt_box=Tensor([11665, 1, 4],"float32"), gt_label=Tensor([11665, 1],"int32"), gt_score=Tensor([11665, 1],"float32"), anchors=list[10,13,16,30,33,23,30,61,62,45,59,119,116,90,156,198,373,326,], anchor_mask=list[3,4,5,], class_num=80, ignore_thresh=0.7, downsample_ratio=16, use_label_smooth=True, ) paddle.vision.ops.yolo_loss(x=Tensor([2917, 255, 76, 76],"float32"), gt_box=Tensor([11665, 1, 4],"float32"), gt_label=Tensor([11665, 1],"int32"), gt_score=Tensor([11665, 1],"float32"), anchors=list[10,13,16,30,33,23,30,61,62,45,59,119,116,90,156,198,373,326,], anchor_mask=list[0,1,2,], class_num=80, ignore_thresh=0.7, downsample_ratio=8, use_label_smooth=True, )

问题出现在 GetEntryIndex 返回值的溢出，改为int64并修改其他对应位置即可

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 PR fixes integer overflow in the YOLO loss kernels by switching index calculations from 32-bit to 64-bit integers for very large tensors.

• Changed GetEntryIndex return type and computation to int64_t.
• Updated all kernel and gradient locals (index, box_idx, label_idx) to int64_t.
• Added <cstdint> include where needed.

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
paddle/phi/kernels/cpu/yolo_loss_kernel.cc	Switched index vars to int64_t and included <cstdint>
paddle/phi/kernels/cpu/yolo_loss_grad_kernel.cc	Updated grad kernel index vars to int64_t
paddle/phi/kernels/cpu/yolo_loss_functor.h	Changed GetEntryIndex signature/return to int64_t and logic

Comments suppressed due to low confidence (2)

paddle/phi/kernels/cpu/yolo_loss_kernel.cc:257

• [nitpick] Consider adding a unit test with tensors large enough (e.g. >2^31 elements) to validate that the int64_t indexing fully prevents overflow in real scenarios.

 int64_t box_idx = 

paddle/phi/kernels/cpu/yolo_loss_functor.h:39

• This header introduces int64_t but does not include <cstdint>. Add #include <cstdint> at the top to ensure int64_t is defined in all translation units.

static inline int64_t GetEntryIndex(int batch, 

[Copilot]

The expression entry * stride is evaluated as 32-bit int before being promoted to int64_t, which can still overflow. Cast one operand to int64_t (e.g. static_cast<int64_t>(entry) * stride) so the multiplication occurs in 64-bit.

Suggested change

	entry * stride + hw_idx;
	static_cast<int64_t>(entry) * stride + hw_idx;

[zrr1999]

代码中entry 取值只有5或0，stride 值较小，所以不需要int64

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Please upload report for BASE (develop@6026060). Learn more about missing BASE report.

Additional details and impacted files

@@ Coverage Diff @@ ## develop #73895 +/- ## =========================================== Coverage ? 100.00% =========================================== Files ? 3 Lines ? 7 Branches ? 0 =========================================== Hits ? 7 Misses ? 0 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.

[wanghuancoder]

LGTM 新的case麻烦提个PR同步到PaddleAPITest。case的numel需要超过int32上限

[zrr1999]

LGTM 新的case麻烦提个PR同步到PaddleAPITest。case的numel需要超过int32上限

好的，已在PFCCLab/PaddleAPITest#358 添加，第一个case，2863312305*5-2^31=352 满足numel超过int32上限