[XPU] reduce_xxx and broadcast_xxx use int64_t shape

[dynamicheart]

PR Category

Custom Device

PR Types

Bug fixes

Description

reduce_xxx and broadcast_xxx use int64_t shape to prevent overflow

example:

import paddle

a = paddle.empty((3 * 1024 * 1024 * 1024,)).to("bool").any().item()
print(a)

>XPUAPI_DEBUG=0x1 python test.py
gtest_cast<float, bool>(api::kXPU3, "GM", "GM", 3221225472, 1024);
gtest_reduce_any<bool>(api::kXPU3, "GM", "GM", {-1073741824}, {0}, 1024);
[INVALID-SHAPE]{-1073741824}[src/wrapper/math_reduce_op.cpp:1379]
Traceback (most recent call last):
  File "/workspace/users/tmp_paddle/Paddle/test.py", line 3, in <module>
    a = paddle.empty((3 * 1024 * 1024 * 1024,)).to("bool").any().item()
  File "/usr/local/lib/python3.10/dist-packages/paddle/tensor/math.py", line 5144, in any
    return _C_ops.any(x, axis, keepdim)
OSError: (External) reduce_any XDNN Error, XDNN_INVALID_PARAM  (at /workspace/users/Paddle/paddle/phi/kernels/xpu/reduce_any_kernel.cc:46)

after fixed:

> XPUAPI_DEBUG=0x1 python test.py
gtest_cast<float, bool>(api::kXPU3, "GM", "GM", 3221225472, 1024);
gtest_reduce_any<bool>(api::kXPU3, "GM", "GM", {3221225472}, {0}, 1024);
    gtest_cast<int8_t, float>(api::kXPU3, "GM", "GM", 3221225472, 1024);
    gtest_cast<float, int8_t>(api::kXPU3, "GM", "GM", 1, 1024);
False

[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.

[skywalker2012]

LGTM

[skywalker2012]

43,44是否要改成int64_t?

[dynamicheart]

这个是xpu::plugin，之前的一套实验性算子接口，没有std::vector<int64_t>的参数形式，我们一般也不会编译也不会开这个编译选项PADDLE_WITH_XPU_PLUGIN，所以就保持其原状即可

[yongqiangma]

LGTM

[XiaoguangHu01]

LGTM