Missing Files in Repository for windows platform && verify issue && type defination issue in CUDA.h

[RioWu]

Hi , thanks for your great work,

But I couldn't find the files runtime/windows/rand48.h and three other files in the repository. These four files are defined in CMakeLists.txt as necessary for compilation on the Windows platform. Could you please add them to the repository, or let me know how these files are generated?

Looking forward to your reply.

[gongbell]

hi，the header files under the windows directory is not useful - they are the files from the clsmith tool. They should be removed from the project.

CUDASmith works only on linux, it is not supported to run on windows.

[RioWu]

Hi, thank you for your response, it has been very helpful.

If it's not too much trouble, I have a few more questions I'd like to ask:

I've made some modifications to the cuda_launcher.c.template from the repository and am using it as the host code. However, I noticed that the logic in this host code does not include any part for verifying test results. Was this omitted by mistake, or could you please tell me how I should design the verify function? That is, what kind of operations will the kernel code generated by cudaSmith perform, and how should I design the verify function to validate the test results?

I noticed that in the cuda.h file in the repository, all unsigned type definitions are defined as signed. Is this a typo, or is there a specific reason for this? I encounter compilation errors because of this when trying to compile the code with clang.

Thank you for your work, and I look forward to your next response.

[gongbell]

CUDASmith use differential testing to detect errors. The same kernel should result in the same output for two versions of nvcc or for nvcc and clang or for different optimization levels of the same compiler. Otherwise, a fault is detected.

cuda.h is used to compile with nvcc not clang.

[RioWu]

Hi,

Thank you for your reply. So the result verification of CudaSmith is determined by checking whether the results obtained from different compilers and different compilation parameters are consistent? There is no need to verify the correctness of the results by writing corresponding verify functions based on cases. I wonder if my understanding is correct.

In addition, if I want to use clang for compilation, how can I do it without using the cuda.h file in the repository?

Looking forward to your reply.