GPGPU-sim with cuDNN

[bigwater]

Environment
Ubuntu 16.04.4 LTS
gcc/g++ 5.4.0 20160609
Python 2.7
CUDA 8.0
cuDNN 7.1.4
gpgpu-sim_distribution dev branch
pytorch-gpgpu-sim (removed the git dependencies of nervanagpu @ d4eefd5, since the repo is no longer there)

PYTORCH_BIN /usr/lib/x86_64-linux-gnu/libcudnn.so

The configurations used are from the configs folder of gpgpu-sim dev branch

MNIST
I use the MNIST sample from here. (Following deval281shah 's suggestions in another discussion. )
https://github.com/gpgpu-sim/gpgpu-sim_simulations

Config
I use TITANV config.
I also tried TITANX config, but a deadlock happened with that configuration.

The simulation runs for 39 minutes, and I checked it has stimulated a number of kernels, and it reported some related information such as IPC.

In the beginning, it generated a large amount of .. it cannot find all device function required.

Warning: cannot find deviceFun maxwell_zgemmBatched_32x32_raggedMn_ct
Warning: cannot find deviceFun maxwell_zgemmBatched_64x32_raggedMn_ct
Warning: cannot find deviceFun maxwell_zgemmBatched_32x32_raggedMn_cn
Warning: cannot find deviceFun maxwell_zgemmBatched_64x32_raggedMn_cn
Warning: cannot find deviceFun maxwell_zgemmBatched_32x32_raggedMn_tc
Warning: cannot find deviceFun maxwell_zgemmBatched_64x32_raggedMn_tc
Warning: cannot find deviceFun maxwell_zgemmBatched_32x32_raggedMn_tt
Warning: cannot find deviceFun maxwell_zgemmBatched_64x32_raggedMn_tt
Warning: cannot find deviceFun maxwell_zgemmBatched_32x32_raggedMn_tn
Warning: cannot find deviceFun maxwell_zgemmBatched_64x32_raggedMn_tn
Warning: cannot find deviceFun maxwell_zgemmBatched_32x32_raggedMn_nc

Therefore, cudaLaunchKernel fails to find the device function.

However, it ends up with
...

GPGPU-Sim PTX: Setting up arguments for 4 bytes starting at 0x7ffc0b96aa68..

GPGPU-Sim PTX: cudaLaunch for 0x0x4321f0 (mode=performance simulation) on stream 0
GPGPU-Sim PTX: ERROR launching kernel -- no PTX implementation found for 0x4321f0

Does anyone encounter this problem before? Any suggestions?

Thank you so much for your help.

[RedCarrottt]

Hi,
I did not met this issue. In my case, loading ptxinfo and extracting ptx files were going well.

Did you set CUDA_INSTALL_PATH? GPGPU-Sim may call the ptxas command based on the value.

[bigwater]

Hi,
I did not met this issue. In my case, loading ptxinfo and extracting ptx files were going well.

Did you set CUDA_INSTALL_PATH? GPGPU-Sim may call the ptxas command based on the value.

Thank you. I have fixed the ptxas problem. (Although I do not know how did it get fixed by only switching a computer. )

Have you tried the MNIST example in https://github.com/gpgpu-sim/gpgpu-sim_simulations? Seems the developers use this app for testing... But finally, I got a PTX not implementation error.

[RedCarrottt]

As you commented, I tried gpgpu-sim/gpgpu-sim_simulations.
When I run the mnistCUDNN, ptx extraction and some implicit_convolve_sgemm call work well. I could not find PTX not implementation error.
However, I've met a deadlock and the simulator has finally been dead as following message.

GPGPU-Sim uArch: Shader 32 bind to kernel 1 '_ZN5cudnn6detail23implicit_convolve_sgemmIffLi128ELi5ELi5ELi3ELi3ELi3ELi1ELb1ELb0ELb1EEEviiiPKT_iPT0_PS2_18kernel_conv_paramsiffiS6_S6_ii'
  <CTA alloc> : sm_idx=32 sid=32 max_cta_per_sm=16
GPGPU-Sim uArch: Shader 34 bind to kernel 1 '_ZN5cudnn6detail23implicit_convolve_sgemmIffLi128ELi5ELi5ELi3ELi3ELi3ELi1ELb1ELb0ELb1EEEviiiPKT_iPT0_PS2_18kernel_conv_paramsiffiS6_S6_ii'
  <CTA alloc> : sm_idx=34 sid=34 max_cta_per_sm=16
GPGPU-Sim uArch: ERROR ** deadlock detected: last writeback core 34 @ gpu_sim_cycle 12492 (+ gpu_tot_sim_cycle 4294867296) (87508 cycles ago)
GPGPU-Sim uArch: DEADLOCK  shader cores no longer committing instructions [core(# threads)]:
GPGPU-Sim uArch: DEADLOCK  0(64) 1(0) 2(64) 3(0) 4(64) 5(0) 6(64) 7(0)  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...
Re-run the simulator in gdb and use debug routines in .gdbinit to debug this
Aborted (core dumped)

[bigwater]

As you commented, I tried gpgpu-sim/gpgpu-sim_simulations.
When I run the mnistCUDNN, ptx extraction and some implicit_convolve_sgemm call work well. I could not find PTX not implementation error.
However, I've met a deadlock and the simulator has finally been dead as following message.

GPGPU-Sim uArch: Shader 32 bind to kernel 1 '_ZN5cudnn6detail23implicit_convolve_sgemmIffLi128ELi5ELi5ELi3ELi3ELi3ELi1ELb1ELb0ELb1EEEviiiPKT_iPT0_PS2_18kernel_conv_paramsiffiS6_S6_ii'
  <CTA alloc> : sm_idx=32 sid=32 max_cta_per_sm=16
GPGPU-Sim uArch: Shader 34 bind to kernel 1 '_ZN5cudnn6detail23implicit_convolve_sgemmIffLi128ELi5ELi5ELi3ELi3ELi3ELi1ELb1ELb0ELb1EEEviiiPKT_iPT0_PS2_18kernel_conv_paramsiffiS6_S6_ii'
  <CTA alloc> : sm_idx=34 sid=34 max_cta_per_sm=16
GPGPU-Sim uArch: ERROR ** deadlock detected: last writeback core 34 @ gpu_sim_cycle 12492 (+ gpu_tot_sim_cycle 4294867296) (87508 cycles ago)
GPGPU-Sim uArch: DEADLOCK  shader cores no longer committing instructions [core(# threads)]:
GPGPU-Sim uArch: DEADLOCK  0(64) 1(0) 2(64) 3(0) 4(64) 5(0) 6(64) 7(0)  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...  + others ...
Re-run the simulator in gdb and use debug routines in .gdbinit to debug this
Aborted (core dumped)

Which configuration files are you using? When I used TITANX (Maxwell), I also got a deadlock. Could you try TITANV config instead?

Seems in your log file, you also got "Warning: cannot find deviceFun" --- I guess the problem is gpgpusim failed to find some device functions. (may be they do not belong to cuDNN? will check later. )

Thank you.

[RedCarrottt]

Hello, I've found that mnistCUDNN require both libcudnn and libcublas, but GPGPU-Sim support extracting PTX files from only one library(its path is defined at PYTORCH_BIN variable).

After I modified GPGPU-Sim so that it supports extracting PTX files from multiple libraries.

I uploaded my pull request(PR #116).

Although mnistCUDNN calls several kernels for fermi or maxwell architecture(ex. cudnn7maxwell4gemm) with volta architecture config file, anyway it works well.

[cng123]

Was mnistCUDNN linked statically with libcudnn and libcublas? I believe that was a requirement in order to account for gpgpu-sim only being able to extract PTX files from 1 file. The Makefile in deval281shah's mnistCUDNN example should be passing flags to statically link those libraries.

[masa-laboratory]

@RedCarrottt @bigwater
Hello,

when I test minist/main.py, the output GPGPU-Sim PTX: Parsing libcaffe2_ gpu.1.sm_ 61.ptx lasted a long time and hasn't stopped yet. Have you ever encountered this problem when running this? Is it a mistake or does it take so long to handle PTX itself.

Thanks.

[jidle123]

你好

当我测试时，输出持续了很长时间并且还没有停止。您在运行此软件时是否遇到过此问题？这是一个错误还是处理 PTX 本身需要很长时间。minist/main.py``GPGPU-Sim PTX: Parsing libcaffe2_ gpu.1.sm_ 61.ptx

谢谢。

您好，请问这个问题解决了吗？我遇到了跟你一样的问题，一直在输出libcaffe2_gpu.xx.sm_61.ptx，根本停不下来

[jidle123]

@RedCarrottt @bigwater Hello,

when I test minist/main.py, the output GPGPU-Sim PTX: Parsing libcaffe2_ gpu.1.sm_ 61.ptx lasted a long time and hasn't stopped yet. Have you ever encountered this problem when running this? Is it a mistake or does it take so long to handle PTX itself.

Thanks.

Hi, Have you deal it yet?

[itsMaoMao]

你好
当我测试时，输出持续了很长时间并且还没有停止。您在运行此软件时是否遇到过此问题？这是一个错误还是处理 PTX 本身需要很长时间。 minist/main.pyGPGPU-Sim PTX: Parsing libcaffe2_ gpu.1.sm_ 61.ptx ``
谢谢。

您好，请问这个问题解决了吗？我遇到了跟你一样的问题，一直在输出libcaffe2_gpu.xx.sm_61.ptx，根本停不下来

ptx文件生成等待一点时间（我是10分钟-20分钟之间）他会完成的，虽然我之后的运行中遇到一些问题，但是正确的代码段ptx文件这个时间内可以完成生成。btw你成功运行了吗？

[jidle123]

你好 我后面没跑这个了，这个问题我也不知道该怎么处理，但我初步猜测是在安装基于gpgpusim的pytorch拓展时，pytorch将库不仅仅替换了最开始的库，还应该自动改了最开始的指向库的链接。我这么说的原因是，如果你尝试回去重跑vecadd.cu，发现会报一样的错，目前应该处于一个进退两难的地方。 我尝试向论文作者发邮件，但论文作者并没有给任何有用反馈。

…

---- 回复的原邮件 ---- | 发件人 | ***@***.***> | | 日期 | 2024年10月30日 10:18 | | 收件人 | ***@***.***> | | 抄送至 | ***@***.***>***@***.***> | | 主题 | Re: [gpgpu-sim/gpgpu-sim_distribution] GPGPU-sim with cuDNN (#112) | 你好 当我测试时，输出持续了很长时间并且还没有停止。您在运行此软件时是否遇到过此问题？这是一个错误还是处理 PTX 本身需要很长时间。 minist/main.pyGPGPU-Sim PTX: Parsing libcaffe2_ gpu.1.sm_ 61.ptx `` 谢谢。 您好，请问这个问题解决了吗？我遇到了跟你一样的问题，一直在输出libcaffe2_gpu.xx.sm_61.ptx，根本停不下来 ptx文件生成等待一点时间（我是10分钟-20分钟之间）他会完成的，虽然我之后的运行中遇到一些问题，但是正确的代码段ptx文件这个时间内可以完成生成。btw你成功运行了吗？ — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>