Question 8-byte alignment PTX code

[SimeonEhrig]

Hi,

I adapt your fatbinary code generation for the CUDA extension of the Cling. Thereby I found the following code:

llvm-project-cxxjit/clang/lib/CodeGen/JIT.cpp

Lines 1407 to 1412 in 87bbc36

	// The outer header of the fat binary is documented in the CUDA
	// fatbinary.h header. As mentioned there, the overall size must be a
	// multiple of eight, and so we must make sure that the PTX is.
	while (DevCD->DevAsm.size() % 7)
	DevCD->DevAsm += ' ';
	DevCD->DevAsm += '\0';

If I understand it correctly, the size of DevAsm should be a multiple of 8 and the last 8-byte block should end with '\0'.

Your implementation generates code with the size n*7+1+1.

I developed the following solution:

DevCD->DevAsm += std::string(7 - (DevCD->DevAsm.size() % 8), ' ') + '\0';

Did I find a bug or do I have a wrong idea?

[lxwithgod]

are you ok on cuda?
I can never succeed cuda jit，if global function is template ,the caller of this kernel is clangjit template,I use nvprof ,I can't see this kernel,but cudaLuanch,cudaSetupArgument.
I use cuda10.0/cuda9.2,gcc/g++ 7/8,clangjit-9.0 ,please tell me how you success .

[SimeonEhrig]

@lxwithgod
I'm not sure what do you mean with templated global functions, because the issue is about how to align PTX Code.

[hfinkel]

Thanks, @SimeonEhrig, I'll look at this soon.

[hfinkel]

Did I find a bug or do I have a wrong idea?

Yes, that logic seems incorrect. I've updated it just to use null padding (after always adding at least one null) and CUDA seems happy with that.