[Fix] fix IndexElementwiseGet kernel CUDA error(700) on 0-size input

[DanielSun11]

PR Types

Operator Mechanism

PR Category

Bug fixes

Description

问题背景

当使用高级整数索引（list-of-list）对第一维为 0 的 Tensor 执行 __getitem__ 时，触发 CUDA error(700)（非法内存访问）或 CPU segfault：

import paddle
x = paddle.zeros([0, 5, 4, 3], dtype='complex128')
out = x[[[2, 3, 4], [1, 2, 5]]]  # GPU: CUDA error(700); CPU: SIGSEGV
out.backward(paddle.zeros_like(out))  # CPU backward: SIGSEGV

根因分析

调用链：__getitem__ → tensor__getitem_dygraph → ApplyGetitem → AdvancedIndex → index_elementwise_get_ad_func → IndexElementwiseGetKernel

AdvancedIndex 构造函数将被索引的维度用索引形状替换得到 src_sizes，如对 x.shape=[0,5,4,3] 用 [[2,3,4],[1,2,5]]（shape=[2,3]）索引维度 0，得到 src_sizes = [2, 3, 5, 4, 3]（numel=90）。

Forward kernel（GPU）：out->numel() = 90 != 0，原有 if (out->numel() == 0) return; 不触发；x.numel() = 0，x.data<T>() 返回 nullptr；kernel 访问 nullptr + offset → CUDA error(700)

Forward kernel（CPU）：同上，CPU 实现也缺少此检查

Backward kernel（GPU）：x_grad->numel() = 0，GpuMemsetAsync 填零后继续执行，访问空指针

Backward kernel（CPU）：x_grad->numel() = 0，dev_ctx.Alloc<T>(x_grad) 返回空指针，EigenVector<T>::Flatten(*x_grad) 对空指针操作 → SIGSEGV

修复方案

在四个 kernel 文件中增加对输入为空的早退检查：

1. GPU forward (index_elementwise_get_kernel.cu)：当 x.numel() == 0 时，用 GpuMemsetAsync 将输出填零并 return
2. CPU forward (index_elementwise_get_kernel.cc)：当 x.numel() == 0 时，用 memset 将输出填零并 return
3. GPU backward (index_elementwise_get_grad_kernel.cu)：当 x_grad->numel() == 0 时直接 return
4. CPU backward (index_elementwise_get_grad_kernel.cc)：在 Alloc<T> 之后、Eigen 操作之前加 if (x_grad->numel() == 0) return;

新增单测

• test/legacy_test/test_index_elementwise.py：新增 TestIndexElementwiseGet0SizeInput，覆盖 complex128、bool、float32、float64、int64、float16 等 dtype，包含正负索引及一维索引等场景（9 个测试方法）
• test/legacy_test/test_index_elementwise_grad.py：新增 TestIndexElementwiseGet0SizeInputGrad，覆盖 float32、float64 及负索引的反向场景（3 个测试方法）

所有 32 个新增测试方法均通过（CPU + GPU）。

是否引起精度变化

否

[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.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 25.00000% with 3 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@ae907b8). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...le/phi/kernels/cpu/index_elementwise_get_kernel.cc	33.33%	2 Missing ⚠️
...i/kernels/cpu/index_elementwise_get_grad_kernel.cc	0.00%	1 Missing ⚠️

❌ Your patch status has failed because the patch coverage (25.00%) is below the target coverage (90.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #78251   +/-   ##
==========================================
  Coverage           ?   25.00%           
==========================================
  Files              ?        2           
  Lines              ?        4           
  Branches           ?        0           
==========================================
  Hits               ?        1           
  Misses             ?        3           
  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.

[DanielSun11]

#78453 已有类似解决