failed build from source on NVIDIA Jetson AGX Xavier (ARM64)

[tmxklzp]

Describe the bug
Build from source on NVIDIA AGX Xavier TX2 (ARM64)
Ubuntu Ubuntu 18.04.6 LTS
CUDA 10.2
Python 3.8.15
cmake 3.20.5

CPU version of Jetson AGX Xavier is aarch64:

~$ uname -a
Linux surgicalai-xavier 4.9.201-tegra #1 SMP PREEMPT Fri Feb 19 08:42:04 PST 2021 aarch64 aarch64 aarch64 GNU/Linux

Error Log
when I followed https://docs.taichi-lang.org/docs/dev_install and this step errored:
Build Taichi from source

python3 setup.py develop --user

log:

[447/620] cd /home/build/taichi/taichi...+17 -D ARCH_cuda -I /home/build/taichi
FAILED: taichi/runtime/llvm/runtime_module/CMakeFiles/generate_llvm_runtime_cuda 
cd /home/build/taichi/taichi/runtime/llvm/runtime_module && /usr/bin/clang++-10 -c runtime.cpp -o runtime_cuda.bc -fno-exceptions -emit-llvm -std=c++17 -D ARCH_cuda -I /home/build/taichi
runtime.cpp:1121:23: warning: value size does not match register size specified by the constraint and modifier [-Wasm-operand-widths]
               : "=r"(ret)
                      ^
runtime.cpp:1120:37: note: use constraint modifier "w"
  asm volatile("match.all.sync.b32  %0, %1, %2;"
                                    ^~
                                    %w0
runtime.cpp:1122:22: warning: value size does not match register size specified by the constraint and modifier [-Wasm-operand-widths]
               : "r"(value), "r"(mask));
                     ^
runtime.cpp:1120:41: note: use constraint modifier "w"
  asm volatile("match.all.sync.b32  %0, %1, %2;"
                                        ^~
                                        %w1
runtime.cpp:1122:34: warning: value size does not match register size specified by the constraint and modifier [-Wasm-operand-widths]
               : "r"(value), "r"(mask));
                                 ^
runtime.cpp:1120:45: note: use constraint modifier "w"
  asm volatile("match.all.sync.b32  %0, %1, %2;"
                                            ^~
                                            %w2
runtime.cpp:1134:18: error: invalid input constraint 'l' in asm
               : "l"(value), "r"(mask));
                 ^
runtime.cpp:1144:44: warning: value size does not match register size specified by the constraint and modifier [-Wasm-operand-widths]
  asm volatile("activemask.b32 %0;" : "=r"(mask));
                                           ^
runtime.cpp:1144:32: note: use constraint modifier "w"
  asm volatile("activemask.b32 %0;" : "=r"(mask));
                               ^~
                               %w0
4 warnings and 1 error generated.

cmake arguments:

TAICHI_CMAKE_ARGS="-DTI_WITH_CUDA:BOOL=ON -DTI_WITH_LLVM:BOOL=ON -DTI_WITH_C_API:BOOL=ON "

and this command triggered error with same output:

cd /home/build/taichi/taichi/runtime/llvm/runtime_module && /usr/bin/clang++-10 -c runtime.cpp -o runtime_cuda.bc -fno-exceptions -emit-llvm -std=c++17 -D ARCH_cuda -I /home/build/taichi

[tmxklzp]

the error is throwed by this snippet:
~/taichi/taichi/runtime/llvm/runtime_module/runtime.cpp
L1129:

uint32 cuda_match_any_sync_i64(u32 mask, i64 value) {
#if ARCH_cuda
  u32 ret;
  asm volatile("match.any.sync.b64  %0, %1, %2;"
               : "=r"(ret)
               : "l"(value), "r"(mask));
  return ret;
#else
  return 0;
#endif
}

can I annotate this if condition and remain the "return 0;"?

[ailzhang]

@tmxklzp Hi, these warnings about register size can be suppressed following the suggestions by clang (adding w which means word register). You can find more details in https://stackoverflow.com/questions/54979550/value-size-does-not-match-register-size-specified-by-the-constraint-and-modifier. ;)
You're more than welcome to send a PR to fix it in taichi!

[tmxklzp]

@ailzhang Okay I'll check this answer about word register, but there is an error 1134:18: error: invalid input constraint 'l' in asm and this caused the failure of the building process(sorry I haven't paste all Cmake logs earlier):

runtime.cpp:1134:18: error: invalid input constraint 'l' in asm
               : "l"(value), "r"(mask));
                 ^
runtime.cpp:1144:44: warning: value size does not match register size specified by the constraint and modifier [-Wasm-operand-widths]
  asm volatile("activemask.b32 %0;" : "=r"(mask));
                                           ^
runtime.cpp:1144:32: note: use constraint modifier "w"
  asm volatile("activemask.b32 %0;" : "=r"(mask));
                               ^~
                               %w0
4 warnings and 1 error generated.
[450/620] Building C object taichi/rhi/opengl/C...pengl_rhi.dir/__/__/__/external/glad/src/gl.c.o
ninja: build stopped: subcommand failed.
Traceback (most recent call last):
  File "/home/username/.local/lib/python3.8/site-packages/skbuild/setuptools_wrap.py", line 640, in setup
    cmkr.make(make_args, install_target=cmake_install_target, env=env)
  File "/home/username/.local/lib/python3.8/site-packages/skbuild/cmaker.py", line 670, in make
    self.make_impl(clargs=clargs, config=config, source_dir=source_dir, install_target=install_target, env=env)
  File "/home/username/.local/lib/python3.8/site-packages/skbuild/cmaker.py", line 701, in make_impl
    raise SKBuildError(

An error occurred while building with CMake.
  Command:
    cmake --build . --target install --config Release -- -j8
  Install target:
    install
  Source directory:
    /home/username/taichi
  Working directory:
    /home/username/taichi/_skbuild/linux-aarch64-3.8/cmake-build
Please check the install target is valid and see CMake's output for more information.

[ailzhang]

that constraint might be x86 specific, try replacing it with "i"?

[tmxklzp]

@ailzhang yes I changed by "i" and it successly build! Thank you for your support. I'll check other warnings and try to make a PR to taichi

[johnnynunez]

@ailzhang @feisuzhu do you fix this permanently?
Digits, cuda arm laptops and jetson thor are coming #8629