Tutorial: deployment. An error occurred on arch x64 and cuda

[Routhleck]

Description

I tried the tutorial (Doc Home>>Deployment>>Tutorial: Run Taichi programs in C++ applications). And I found it can successfully run on vulkan backend, then I changed
app.py

ti.init(arch=ti.vulkan) -> ti.init(arch=ti.x64) or ti.init(arch=ti.cuda)
mod = ti.aot.Module(ti.vulkan) -> mod = ti.aot.Module(ti.x64) or mod = ti.aot.Module(ti.cuda)

app.cpp

ti::Runtime runtime(TI_ARCH_VULKAN) -> ti::Runtime runtime(TI_ARCH_X64) or ti::Runtime runtime(TI_ARCH_CUDA) 

After cmake compilation, ran ./build/TaichiAot , there was a segmentation fault on both x64 and cuda arch.

Thread 1 "TaichiAot" received signal SIGSEGV, Segmentation fault.
0x00007ffff5d31b50 in ?? () from /home/routhleck/miniconda3/envs/jax-taichi/lib/python3.8/site-packages/taichi/_lib/c_api/lib/libtaichi_c_api.so
(gdb) backtrace
#0  0x00007ffff5d31b50 in ?? () from /home/routhleck/miniconda3/envs/jax-taichi/lib/python3.8/site-packages/taichi/_lib/c_api/lib/libtaichi_c_api.so
#1  0x00007ffff5cc6534 in ?? () from /home/routhleck/miniconda3/envs/jax-taichi/lib/python3.8/site-packages/taichi/_lib/c_api/lib/libtaichi_c_api.so
#2  0x00007ffff5caaef7 in ti_load_aot_module ()
   from /home/routhleck/miniconda3/envs/jax-taichi/lib/python3.8/site-packages/taichi/_lib/c_api/lib/libtaichi_c_api.so
#3  0x0000555555557565 in ti::Runtime::load_aot_module(char const*) const ()
#4  0x000055555555686c in main ()

Then I noticed that "Before we start, it is recommended to install Taichi through taichi-nightly Python wheels using the following command. Be aware that there's no strong version compatibility enforced yet, so it's highly recommended to use the Taichi built from exactly the same commit." in tutorial. So I replaced taichi-nightly from 1.7.0.post20230804 to 1.6.0.post2023041. But I ran into a problem that I couldn't find libtaichi_c_api.so

-- Looking for Taichi components: Runtime
CMake Error at /home/routhleck/miniconda3/envs/jax-taichi/lib/python3.8/site-packages/taichi/_lib/c_api/taichi/lib/cmake/taichi/TaichiTargets.cmake:92 (message):
  The imported target "taichi_c_api" references the file

     "/home/routhleck/miniconda3/envs/jax-taichi/lib/python3.8/site-packages/taichi/_lib/c_api/lib/libtaichi_c_api.so"

  but this file does not exist.  Possible reasons include:

  * The file was deleted, renamed, or moved to another location.

  * An install or uninstall procedure did not complete successfully.

  * The installation package was faulty and contained

     "/home/routhleck/miniconda3/envs/jax-taichi/lib/python3.8/site-packages/taichi/_lib/c_api/taichi/lib/cmake/taichi/TaichiTargets.cmake"

  but not all the files it references.

Call Stack (most recent call first):
  /home/routhleck/miniconda3/envs/jax-taichi/lib/python3.8/site-packages/taichi/_lib/c_api/taichi/lib/cmake/taichi/TaichiConfig.cmake:27 (include)
  cmake/FindTaichi.cmake:101 (find_package)
  CMakeLists.txt:14 (find_package)

Development environment

hardware

• CPU: Intel Core i5-12490F
• GPU: Nvidia Geforce RTX 2070
• RAM: 32G DDR4 3200MHz

software

• OS: Ubuntu 22.04.2 LTS
• Python: 3.8
• taichi: 1.6.0
• taichi-nightly: 1.7.0.post20230804 or 1.6.0.post2023041

[bobcao3]

Can you try after installing taichi-nightly, remove the taichi package? (And regenerate the cmake files)

[Routhleck]

After removing the taichi package, and I reinstalled taichi-nightly. There still occured segmentation fault in ti_load_aot_module

[chaoming0625]

It seems that this feature in the doc is only supported in the Vulkan backend?

[Routhleck]

But the document say:
Currently ti.vulkan, ti.opengl, ti.x86 and ti.cuda are supported. ti.metal is not yet supported.

It seems that this feature in the doc is only supported in the Vulkan backend?

[jim19930609]

The Taichi AOT library shipped in the nightly-taichi does not have CUDA/x86 inside. You'll have to compile taichi from source code to enable C-API for CUDA/x86 or metal, etc.

Luckily, there's an easy way to do that. The build system for https://github.com/taichi-dev/taichi-aot-demo has options to turn on CUDA/x86 backends. Just turn on "TI_WITH_CPU" and "TI_WITH_CUDA" when you do cmake.

[Routhleck]

Thanks, I can compile taichi from source code!