PR Category

Operator Mechanism

PR Types

Bug fixes

Description

• Add negative index detection and conversion in DealWithIndex function
• Support negative indexing in advanced indexing like tensor[[-1]]
• Fix issue75574: negative indexing in strided slice operations
• Maintain backward compatibility with existing positive indexing
• Add comprehensive test cases for negative indexing scenarios

Fixes:

• paddle dev 到对矩阵取索引结果异常 #75574

修复PaddlePaddle中fancy indexing操作中负索引处理的问题。当使用tensor[[-1]]这种高级索引时，负索引没有被正确转换为正索引，导致索引错误。

复现步骤

import paddle

edge_index = paddle.arange(17758, dtype='int64')
strided_edge_index = edge_index[::32]
print(strided_edge_index[[-1]])  # 期望: 17728，实际可能出错
assert strided_edge_index[[-1]] == 17728  # 应该通过但实际失败

解决方案

在paddle/fluid/pybind/slice_utils.h的DealWithIndex函数中添加负索引处理逻辑：

// Handle negative indices for advanced indexing
if (indice.defined() && (indice.dtype() == paddle::DataType::INT64 || 
                        indice.dtype() == paddle::DataType::INT32)) {
  auto dims = transed_sub_tensor->dims();
  if (dims.size() > 0) {
    int64_t dim_size = dims[0];
    auto dim_size_tensor = full_ad_func({1}, dim_size, paddle::DataType::INT64, indice.place());
    
    auto negative_mask = less_than_ad_func(indice, zeros_like_ad_func(indice));
    auto positive_indices = where_ad_func(negative_mask, 
                                        add_ad_func(indice, dim_size_tensor), 
                                        indice);
    indice = positive_indices;
  }
}

修复内容

• 修复tensor[[-1]]负索引问题
• 支持步长切片后的负索引：strided_tensor[[-1]]
• 支持混合索引：正索引和负索引混合使用
• 向后兼容，不影响现有功能

测试用例

# 基本负索引
edge_index = paddle.arange(100, dtype='int64')
assert edge_index[[-1]].item() == 99

# 步长切片负索引
strided = edge_index[::10]
assert strided[[-1]].item() == 90

# 多负索引
result = edge_index[[-1, -2, -3]]
assert (result.numpy() == [99, 98, 97]).all()

影响范围

• 仅影响fancy indexing中的负索引处理
• 向后兼容，不影响现有正索引功能
• 最小性能影响，只在检测到负索引时进行转换

Fixes: #issue75574