Skip to content

[flang][cuda] Add new entry points function for data transfer #108244

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Sep 16, 2024
Merged

[flang][cuda] Add new entry points function for data transfer #108244

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
968d2dd
[flang][cuda] Add runtime function for data transfer
clementval Sep 9, 2024
4c5bc93
clang-format
clementval Sep 11, 2024
b2ec76e
Add commont about pointer value
clementval Sep 11, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
47 changes: 47 additions & 0 deletions flang/include/flang/Runtime/CUDA/memory.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,47 @@
//===-- include/flang/Runtime/CUDA/memory.h ---------------------*- C++ -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//

#ifndef FORTRAN_RUNTIME_CUDA_MEMORY_H_
#define FORTRAN_RUNTIME_CUDA_MEMORY_H_

#include "flang/Runtime/descriptor.h"
#include "flang/Runtime/entry-names.h"
#include <cstddef>

static constexpr unsigned kHostToDevice = 0;
static constexpr unsigned kDeviceToHost = 1;
static constexpr unsigned kDeviceToDevice = 2;

namespace Fortran::runtime::cuda {

extern "C" {

/// Set value to the data hold by a descriptor. The \p value pointer must be
/// addressable to the same amount of bytes specified by the element size of
/// the descriptor \p desc.
void RTDECL(CUFMemsetDescriptor)(const Descriptor &desc, void *value,
const char *sourceFile = nullptr, int sourceLine = 0);

/// Data transfer from a pointer to a descriptor.
void RTDECL(CUFDataTransferDescPtr)(const Descriptor &dst, void *src,
std::size_t bytes, unsigned mode, const char *sourceFile = nullptr,
int sourceLine = 0);

/// Data transfer from a descriptor to a pointer.
void RTDECL(CUFDataTransferPtrDesc)(void *dst, const Descriptor &src,
std::size_t bytes, unsigned mode, const char *sourceFile = nullptr,
int sourceLine = 0);

/// Data transfer from a descriptor to a descriptor.
void RTDECL(CUFDataTransferDescDesc)(const Descriptor &dst,
const Descriptor &src, unsigned mode, const char *sourceFile = nullptr,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can this be used for entities (described by src/dst descriptors) of derived types with allocatable/pointer components or is this disallowed in CUF?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Do you mean smth like

module m
  type :: t1
    integer, allocatable :: a(:)
  end type
end

program p
  use m
  type(t1) :: h
  type(t1), device :: d

  h = d
end

This is allowed

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does it mean this API needs to basically reproduce the Assign runtime with regards to handling automatic reallocation of the components or is it supposed to make a shallow copy?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes. Basically it would be the same as Assign but with cuda memcpy instead of the standard copy. Hopefully we can adapt Assign to be called from here.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@vzakhari Do you have any more concerns about these new entry points?

int sourceLine = 0);

} // extern "C"
} // namespace Fortran::runtime::cuda
#endif // FORTRAN_RUNTIME_CUDA_MEMORY_H_
2 changes: 1 addition & 1 deletion flang/lib/Optimizer/Dialect/FIRType.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1467,4 +1467,4 @@ fir::getTypeSizeAndAlignmentOrCrash(mlir::Location loc, mlir::Type ty,
if (result)
return *result;
TODO(loc, "computing size of a component");
}
}
1 change: 1 addition & 0 deletions flang/runtime/CUDA/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,7 @@ set(CUFRT_LIBNAME CufRuntime_cuda_${CUDAToolkit_VERSION_MAJOR})
add_flang_library(${CUFRT_LIBNAME}
allocator.cpp
descriptor.cpp
memory.cpp
)

if (BUILD_SHARED_LIBS)
Expand Down
46 changes: 46 additions & 0 deletions flang/runtime/CUDA/memory.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,46 @@
//===-- runtime/CUDA/memory.cpp -------------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//

#include "flang/Runtime/CUDA/memory.h"
#include "../terminator.h"

#include "cuda_runtime.h"

namespace Fortran::runtime::cuda {
extern "C" {

void RTDEF(CUFMemsetDescriptor)(const Descriptor &desc, void *value,
const char *sourceFile, int sourceLine) {
Terminator terminator{sourceFile, sourceLine};
terminator.Crash("not yet implemented: CUDA data transfer from a scalar "
"value to a descriptor");
}

void RTDEF(CUFDataTransferDescPtr)(const Descriptor &desc, void *addr,
std::size_t bytes, unsigned mode, const char *sourceFile, int sourceLine) {
Terminator terminator{sourceFile, sourceLine};
terminator.Crash(
"not yet implemented: CUDA data transfer from a pointer to a descriptor");
}

void RTDEF(CUFDataTransferPtrDesc)(void *addr, const Descriptor &desc,
std::size_t bytes, unsigned mode, const char *sourceFile, int sourceLine) {
Terminator terminator{sourceFile, sourceLine};
terminator.Crash(
"not yet implemented: CUDA data transfer from a descriptor to a pointer");
}

void RTDECL(CUFDataTransferDescDesc)(const Descriptor &dstDesc,
const Descriptor &srcDesc, unsigned mode, const char *sourceFile,
int sourceLine) {
Terminator terminator{sourceFile, sourceLine};
terminator.Crash(
"not yet implemented: CUDA data transfer between two descriptors");
}
}
} // namespace Fortran::runtime::cuda
Loading