CUDA 12.6U1 NVCC Regression

Accelerated Computing CUDA CUDA NVCC Compiler
1

After an update from CUDA Toolkit 12.6 to 12.6U1, I am unable to compile previously working code. I get the following new error:

[build] nvcc warning : incompatible redefinition for option 'optimize', the last value of this option was used
[build] /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h(42): error: identifier "__builtin_ia32_ldtilecfg" is undefined
[build]     __builtin_ia32_ldtilecfg (__config);
[build]     ^
[build] 
[build] /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h(49): error: identifier "__builtin_ia32_sttilecfg" is undefined
[build]     __builtin_ia32_sttilecfg (__config);
[build]     ^
[build] 
[build] 2 errors detected in the compilation of "/home/bryce/dev/torch-discounted-cumsum-nd/torch_discounted_cumsum_nd/operator.cu"
  • OS = Ubuntu 24.04
  • NVCC = cuda_12.6.r12.6/compiler.34714021_0
  • GCC = 13.2 and 11.4

The source code can be found here. Requires pytorch >=2.4 as noted in the readme. I’m just on a Ryzen 5950X, I don’t even have AMX extensions.

1 Like