[CIR][CUDA] Generate device stubs #1332
Conversation
clang/lib/CIR/CodeGen/CIRGenCUDA.cpp
Outdated
| // Now emit the call to cudaLaunchKernel | ||
| // cudaError_t cudaLaunchKernel(const void *func, dim3 gridDim, dim3 blockDim, | ||
| // void **args, size_t sharedMem, | ||
| // cudaStream_t stream); |
There was a problem hiding this comment.
Seems like we could have a ... = cir.cuda.setup_device_stub <name>, args, dim3_ty = <some_type>, stream_ty = <some_type2> ... that will hide both __cudaPopCallConfiguration and cudaLaunchKernel calls. This will then be expanded in LoweringPrepare to these calls (so we don't have to postpone this to LLVMLowering).
However, I'd rather see you adding this as-is first (after the other comment about OG is addressed) and in a follow up PR we can raise the representation and move it to LoweringPrepare.
There was a problem hiding this comment.
Similar things also happen at call site. Shall we also generate a cir.cuda.call_kernel for that and expand in LoweringPrepare?
There was a problem hiding this comment.
which callsite do you mean? not the call happening in the device stub? if not in the device host, does it not call __cudaPopCallConfiguration to retrieve the dims? Perhaps we should have a bit more of direct CIRGen to have a better grasp of uses of these internal functions before we raise them.
There was a problem hiding this comment.
I mean the place where we invoke the kernel, for example in main we can write global_fn<<<1, 1>>>(a, b, c). This is the place where we call __cudaPushCallConfiguration for device stub to pop. I guess I'll directly generate them and adjust according to review.
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
AdUhTkJm
force-pushed
the
main
branch
2 times, most recently
from
0832402 to
fe4f3c5
Compare
Now we're able to generate device stubs.
A simple explanation:
We first store function arguments inside a
void* args[], which shall be passed intocudaLaunchKernel.Then we retrieve configuration using
__cudaPopCallConfiguration, popping the config pushed by callsite. (We can't generate calls to kernels currently.)Now we have enough arguments. Invoke
cudaLaunchKerneland we're OK.