[DCU]Fix build error on rocm5.7

[onepick]

PR Category

Others

PR Types

Bug fixes

Description

fix build error and warning on rocm5.7

1. rocm5以上版本，头文件引用规则变化，要求从$ROCM_PATH/include进行引用
2. add patch to third_patry warprnnt/warpctc to fix build error
3. rocm5.7使用clang17, 编译器检测更加严格，编译rocm时，去掉Werror

[paddle-bot[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.

[onepick]

@luotao1 关于CI中没有通过的部分，都是在编译cuda的是时候，报出： nvcc fatal : Value '-Xcompiler' is not defined for option 'Werror'

但是我本地使用nvidia/cuda:11.6.1-cudnn8-devel-ubuntu20.04镜像编译的时候，并没有报出这个错误来，请问你们的环境有什么不一样的地方吗？

[onepick]

@qili93 有什么更新吗？

[qili93]

@onepick 当前代码的修改无法通过PR-CI-BUILD的CI，建议先修复下现有的CUDA编译问题，保障代码兼容性，谢谢！

[qili93]

nvidia/cuda:11.6.1-cudnn8-devel-ubuntu20.04

看了下日志里面显示的CUDA版本是

2024-04-29 15:51:37 ****************************************
2024-04-29 15:51:37 Paddle version: N/A
2024-04-29 15:51:37 Paddle With CUDA: N/A
2024-04-29 15:51:37 
2024-04-29 15:51:37 OS: ubuntu 20.04
2024-04-29 15:51:37 GCC version: (GCC) 8.2.0
2024-04-29 15:51:37 Clang version: N/A
2024-04-29 15:51:37 CMake version: version 3.18.0
2024-04-29 15:51:37 Libc version: glibc 2.31
2024-04-29 15:51:37 Python version: 3.10.13
2024-04-29 15:51:37 
2024-04-29 15:51:37 CUDA version: 11.8.89
2024-04-29 15:51:37 Build cuda_11.8.r11.8/compiler.31833905_0
2024-04-29 15:51:37 cuDNN version: 8.9.0
2024-04-29 15:51:37 Nvidia driver version: N/A
2024-04-29 15:51:37 Nvidia driver List: N/A
2024-04-29 15:51:37 ****************************************
2024-04-29 15:51:37 + bash /paddle/tools/get_cpu_info.sh

建议可以试试看 registry.baidubce.com/paddlepaddle/paddle:latest-dev-cuda11.8-cudnn8.6-trt8.5-gcc82 的镜像本地编译看下。也辛苦 @risemeup1 确认下 https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/detail/10589314/job/26072763 这流水线使用的基础编译镜像是什么。

[paddle-ci-bot[bot]]

Sorry to inform you that 475d533's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[onepick]

@qili93 fail的CI部分没任何错误提示？

[onecatcn]

2024-05-02 09:52:08 **************** 2024-05-02 09:52:08 0. Change compilation flag of warnings is not recommended. You must have one RD's (zhiqiu (Recommend), luotao1 or phlrain or Aurelius84) approval to use these methods. 2024-05-02 09:52:08 There are 1 approved errors. 2024-05-02 09:52:08 ****************

[onepick]

2024-05-02 09:52:08 **************** 2024-05-02 09:52:08 0. Change compilation flag of warnings is not recommended. You must have one RD's (zhiqiu (Recommend), luotao1 or phlrain or Aurelius84) approval to use these methods. 2024-05-02 09:52:08 There are 1 approved errors. 2024-05-02 09:52:08 ****************

got it. @zhiqiu @luotao1 编译中的warning过多，主要是因为rocm5.7升级编译器到clang17, clang17的检测更加智能/更加严格，所以更多的warning暴露了出来