Skip to content

Explicit gpu global copies #2260

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

Open
ThrudPrimrose wants to merge 18 commits into
base: main
Choose a base branch
from
Open

Explicit gpu global copies #2260

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
18 commits
Select commit Hold shift + click to select a range
f519b26
Extend cases supported by the explicit copy transformations
ThrudPrimrose Jan 6, 2026
e483a25
Refactor, use views+
ThrudPrimrose Jan 6, 2026
4d7156f
Refactor
ThrudPrimrose Jan 6, 2026
09e29e6
Refactor
ThrudPrimrose Jan 6, 2026
8a754c3
Extensions
ThrudPrimrose Jan 7, 2026
dad01d3
Fix bug
ThrudPrimrose Jan 7, 2026
e38016c
Fix
ThrudPrimrose Jan 7, 2026
944db27
Check for GPU outputs in current stream generation
ThrudPrimrose Jan 7, 2026
c77ea55
Fix cpp codegen
ThrudPrimrose Jan 7, 2026
0ed6406
Precommit
ThrudPrimrose Jan 9, 2026
1dac422
Merge branch 'main' into explicit-gpu-global-copies
ThrudPrimrose Jan 11, 2026
f762d5f
Merge branch 'main' into explicit-gpu-global-copies
ThrudPrimrose Jan 23, 2026
02b4182
Merge branch 'main' into explicit-gpu-global-copies
ThrudPrimrose Jan 23, 2026
bca117c
Add dtype
ThrudPrimrose Jan 24, 2026
9dd1605
Rm gpu def
ThrudPrimrose Jan 24, 2026
3e83c74
Merge branch 'main' into explicit-gpu-global-copies
ThrudPrimrose Jan 28, 2026
c9cc32d
Update dace/transformation/passes/gpu_specialization/helpers/copy_str…
ThrudPrimrose Jan 28, 2026
6aa4c27
Update dace/transformation/passes/gpu_specialization/helpers/copy_str…
ThrudPrimrose Jan 28, 2026
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
7 changes: 5 additions & 2 deletions dace/codegen/targets/cpp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -807,9 +807,12 @@ def unparse_cr(sdfg, wcr_ast, dtype):
def connected_to_gpu_memory(node: nodes.Node, state: SDFGState, sdfg: SDFG):
for e in state.all_edges(node):
path = state.memlet_path(e)
if ((isinstance(path[0].src, nodes.AccessNode)
and path[0].src.desc(sdfg).storage is dtypes.StorageType.GPU_Global)):
if (((isinstance(path[0].src, nodes.AccessNode)
and path[0].src.desc(sdfg).storage is dtypes.StorageType.GPU_Global))
or ((isinstance(path[-1].dst, nodes.AccessNode)
and path[-1].dst.desc(sdfg).storage is dtypes.StorageType.GPU_Global))):
return True

return False


Expand Down
18 changes: 16 additions & 2 deletions dace/dtypes.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,6 +71,7 @@ class ScheduleType(AutoNumberEnum):
GPU_ThreadBlock = () #: Thread-block code
GPU_ThreadBlock_Dynamic = () #: Allows rescheduling work within a block
GPU_Persistent = ()
GPU_Warp = ()
Copy link
Contributor

@alexnick83 alexnick83 Jan 28, 2026

Choose a reason for hiding this comment

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

I believe we need this, but is it actually used anywhere in this PR?

Copy link
Collaborator Author

@ThrudPrimrose ThrudPrimrose Jan 28, 2026

Choose a reason for hiding this comment

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

It is not utilized for the copies, but it is part of the new GPU codegen, I potentially had error when porting dtypes from the GPU codegen branch to this branch.


Snitch = ()
Snitch_Multicore = ()
Expand All @@ -84,6 +85,11 @@ class ScheduleType(AutoNumberEnum):
ScheduleType.GPU_Persistent,
]

# A subset of GPU schedule types for ExperimentalCUDACodeGen
EXPERIMENTAL_GPU_SCHEDULES = [
ScheduleType.GPU_Warp,
]

# A subset of CPU schedule types
CPU_SCHEDULES = [
ScheduleType.CPU_Multicore,
Expand All @@ -95,6 +101,8 @@ class ScheduleType(AutoNumberEnum):
StorageType.GPU_Shared,
]

GPU_KERNEL_ACCESSIBLE_STORAGES = [StorageType.GPU_Global, StorageType.GPU_Shared, StorageType.Register]


@undefined_safe_enum
class ReductionType(AutoNumberEnum):
Expand Down Expand Up @@ -192,7 +200,8 @@ class TilingType(AutoNumberEnum):
ScheduleType.GPU_ThreadBlock: StorageType.Register,
ScheduleType.GPU_ThreadBlock_Dynamic: StorageType.Register,
ScheduleType.SVE_Map: StorageType.CPU_Heap,
ScheduleType.Snitch: StorageType.Snitch_TCDM
ScheduleType.Snitch: StorageType.Snitch_TCDM,
ScheduleType.GPU_Warp: StorageType.Register,
}

# Maps from ScheduleType to default ScheduleType for sub-scopes
Expand All @@ -207,9 +216,10 @@ class TilingType(AutoNumberEnum):
ScheduleType.GPU_Device: ScheduleType.GPU_ThreadBlock,
ScheduleType.GPU_ThreadBlock: ScheduleType.Sequential,
ScheduleType.GPU_ThreadBlock_Dynamic: ScheduleType.Sequential,
ScheduleType.GPU_Warp: ScheduleType.Sequential,
ScheduleType.SVE_Map: ScheduleType.Sequential,
ScheduleType.Snitch: ScheduleType.Snitch,
ScheduleType.Snitch_Multicore: ScheduleType.Snitch_Multicore
ScheduleType.Snitch_Multicore: ScheduleType.Snitch_Multicore,
}

# Maps from StorageType to a preferred ScheduleType for helping determine schedules.
Expand Down Expand Up @@ -1240,6 +1250,7 @@ class string(_DaCeArray, npt.NDArray[numpy.str_]): ...
class vector(_DaCeArray, npt.NDArray[numpy.void]): ...
class MPI_Request(_DaCeArray, npt.NDArray[numpy.void]): ...
class float32sr(_DaCeArray, npt.NDArray[numpy.float32]): ...
class gpuStream_t(_DaCeArray, npt.NDArray[numpy.void]): ...
# yapf: enable
else:
# Runtime definitions
Expand All @@ -1260,6 +1271,7 @@ class float32sr(_DaCeArray, npt.NDArray[numpy.float32]): ...
complex128 = typeclass(numpy.complex128)
string = stringtype()
MPI_Request = opaque('MPI_Request')
gpuStream_t = opaque('gpuStream_t')
float32sr = Float32sr()


Expand All @@ -1281,6 +1293,7 @@ class Typeclasses(AutoNumberEnum):
float64 = float64
complex64 = complex64
complex128 = complex128
gpuStream_t = gpuStream_t


_bool = bool
Expand Down Expand Up @@ -1508,6 +1521,7 @@ def can_access(schedule: ScheduleType, storage: StorageType):
ScheduleType.GPU_Persistent,
ScheduleType.GPU_ThreadBlock,
ScheduleType.GPU_ThreadBlock_Dynamic,
ScheduleType.GPU_Warp,
]:
return storage in [StorageType.GPU_Global, StorageType.GPU_Shared, StorageType.CPU_Pinned]
elif schedule in [ScheduleType.Default, ScheduleType.CPU_Multicore, ScheduleType.CPU_Persistent]:
Expand Down
Empty file added dace/transformation/passes/gpu_specialization/__init__.py
View file
Open in desktop
Empty file.
Empty file added dace/transformation/passes/gpu_specialization/helpers/__init__.py
View file
Open in desktop
Empty file.
Loading
Loading