[flang-rt] Add Assign_omp RT call. #145465
Conversation
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
| allocatable.cpp | ||
| array-constructor.cpp | ||
| assign.cpp | ||
| assign_omp.cpp |
There was a problem hiding this comment.
I'm not sure if it's a good idea to make the Fortran runtime depend on the OpenMP runtime library. I think it makes more sense to have this routine live in the OpenMP offload runtime as a potentially generic API entry point.
There was a problem hiding this comment.
Fortran runtime function implementations exist in the OpenMP runtime library like this: https://github.com/llvm/llvm-project/blob/main/openmp/runtime/src/kmp_ftn_entry.h . If it is offload-only, it may also live in libomptarget.
There was a problem hiding this comment.
Context:
Assign_omp just does omp_target_memcpy between two device pointers.
void RTDEF(Assign_omp)(Descriptor &to, const Descriptor &from, const char *sourceFile, int sourceLine, omp::OMPDeviceTy omp_device)
This api is required, when hoisting "fir.call @_FortranAAssign(...)" from omp.target to the host in "lower-workdistribute" pass #140523
Descriptor struct is defined in flang-rt/runtime/descriptor.h.
Issue:
Now if I need to move this implementation to openmp runtime or libomptarget, I wouldn't have access to fortran-rt Descriptor structure there. Is there any solution to deal with such issue?
Probable solution:
Instead of creating new runtime api, may be, make call to "omp_target_memcpy" directly by extracting the ptrs from Descriptor structure at MLIR stage? Need to check if this is possible.
There was a problem hiding this comment.
I see two options:
- Use the CFI of Fortran to access the descriptor. It will sort of tie the OpenMP runtime to the Fortran runtime, but it will do so in an ISO-conforming way.
- Unpack the descriptor in the code-gen for the assign operation. It might be prudent to provide an accessor function for getting the pointers and size information from the descriptor.
| allocatable.cpp | ||
| array-constructor.cpp | ||
| assign.cpp | ||
| assign_omp.cpp |
There was a problem hiding this comment.
Fortran runtime function implementations exist in the OpenMP runtime library like this: https://github.com/llvm/llvm-project/blob/main/openmp/runtime/src/kmp_ftn_entry.h . If it is offload-only, it may also live in libomptarget.
| // If not present on the device it should already be a device ptr | ||
| if (!omp_target_is_present(voidAnyPtr, ompDevice)) | ||
| return anyPtr; | ||
| T *device_ptr = omp_get_mapped_ptr(anyPtr, ompDevice); |
There was a problem hiding this comment.
nit: use the same style for variables names: devicePtr.
|
|
||
| typedef int32_t OMPDeviceTy; | ||
|
|
||
| template <typename T> static T *getDevicePtr(T *anyPtr, OMPDeviceTy ompDevice) { |
There was a problem hiding this comment.
nit: for easier reasability:
| template <typename T> static T *getDevicePtr(T *anyPtr, OMPDeviceTy ompDevice) { | |
| template <typename T> static T *getDevicePtr(T *hostPtr, OMPDeviceTy ompDevice) { |
| std::size_t fromElements{from.Elements()}; | ||
|
|
||
| if (toElementBytes != fromElementBytes) | ||
| terminator.Crash("Assign: toElementBytes != fromElementBytes"); |
There was a problem hiding this comment.
Logging the number of element bytes (and elements below) in the crash might be helpful when debugging.
| terminator.Crash("Assign: toElements != fromElements"); | ||
|
|
||
| // Get base addresses and calculate length | ||
| void *to_base = to.raw().base_addr; |
There was a problem hiding this comment.
nit: Naming style is different in this function as well.
| } | ||
|
|
||
| RT_API_ATTRS static void Assign(Descriptor &to, const Descriptor &from, | ||
| Terminator &terminator, int flags, OMPDeviceTy omp_device) { |
There was a problem hiding this comment.
flags is not used. Do we need it here?
| std::size_t toElementBytes{to.ElementBytes()}; | ||
| std::size_t fromElementBytes{from.ElementBytes()}; | ||
| std::size_t toElements{to.Elements()}; | ||
| std::size_t fromElements{from.Elements()}; |
There was a problem hiding this comment.
You want to check also that descriptors are contiguous. You can have the same number of elements but the stride might be different.
There was a problem hiding this comment.
If you want it to work also on non contiguous descriptors, the Assign function as a mechanism to pass memmove function to use.
| //===-- lib/runtime/assign_omp.cpp ----------------------------------*- C++ | ||
| //-*-===// |
|
Will this add a hard OpenMP dependency for anyone building the flang runtime? This should probably be placed in a separate library just for OpenMP like it is done for CUDA Fortran. |
@clementval As discussed in #152756 and based on your feedback its better to not have openmp rt dependency on flang-rt. Closing this PR. |
Add runtime Assign_omp call which does memcpy of ptrs on device using "omp_target_memcpy".
void RTDEF(Assign_omp)(Descriptor &to, const Descriptor &from, const char *sourceFile, int sourceLine, omp::OMPDeviceTy omp_device)
This PR is pre-requisite for workdistribute lowering PR #140523