Skip to content

[libclc] Use CLC atomic functions for legacy OpenCL atom/atomic builtins #168325

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 4 commits into from
Nov 19, 2025
Merged

[libclc] Use CLC atomic functions for legacy OpenCL atom/atomic builtins #168325

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
d95e51c
[libclc] Use CLC atomic functions for legacy OpenCL atom/atomic builtins
wenju-he Nov 17, 2025
ae7fd06
update comment
wenju-he Nov 17, 2025
5976864
delete unused atom_int32_binary.inc
wenju-he Nov 18, 2025
fa1b9b0
change legacy atomic_dec/inc memory order to __ATOMIC_RELAXED
wenju-he Nov 18, 2025
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
3 changes: 0 additions & 3 deletions libclc/opencl/include/clc/opencl/atomic/atom_decl_int32.inc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,9 +6,6 @@
//
//===----------------------------------------------------------------------===//

#include <clc/clcfunc.h>
#include <clc/clctypes.h>

#define __CLC_DECLARE_ATOM(ADDRSPACE, TYPE) \
_CLC_OVERLOAD _CLC_DECL TYPE __CLC_FUNCTION(volatile ADDRSPACE TYPE *, TYPE);

Expand Down
3 changes: 0 additions & 3 deletions libclc/opencl/include/clc/opencl/atomic/atom_decl_int64.inc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,9 +6,6 @@
//
//===----------------------------------------------------------------------===//

#include <clc/clcfunc.h>
#include <clc/clctypes.h>

#define __CLC_DECLARE_ATOM(ADDRSPACE, TYPE) \
_CLC_OVERLOAD _CLC_DECL TYPE __CLC_FUNCTION(volatile ADDRSPACE TYPE *, TYPE);

Expand Down
1 change: 0 additions & 1 deletion libclc/opencl/lib/amdgcn/SOURCES
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,3 @@
cl_khr_int64_extended_atomics/minmax_helpers.ll
mem_fence/fence.cl
synchronization/barrier.cl
workitem/get_global_offset.cl
Expand Down
55 changes: 0 additions & 55 deletions libclc/opencl/lib/amdgcn/cl_khr_int64_extended_atomics/minmax_helpers.ll
View file
Open in desktop

This file was deleted.

29 changes: 16 additions & 13 deletions libclc/opencl/lib/generic/atomic/atom_add.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,32 +6,35 @@
//
//===----------------------------------------------------------------------===//

#include <clc/atomic/clc_atomic_fetch_add.h>
#include <clc/opencl/atomic/atom_add.h>
#include <clc/opencl/atomic/atomic_add.h>

// Non-volatile overloads are for backward compatibility with OpenCL 1.0.
Copy link
Contributor

@arsenm arsenm Nov 18, 2025

Choose a reason for hiding this comment

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

I don't remember the history here, but isn't this backwards? Even if the language is volatile qualified we probably shouldn't be emitting IR with volatile markers

Copy link
Contributor Author

@wenju-he wenju-he Nov 19, 2025

Choose a reason for hiding this comment

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

I don't remember the history here, but isn't this backwards?

OpenCL 1.0 atom_* builtins' pointer argument doesn't have volatile.

Even if the language is volatile qualified we probably shouldn't be emitting IR with volatile markers

Current implementation of CLC and OpenCL atomic function only have volatile in builtin function mangling. In LLVM IR, volatile is dropped, see the casting (ADDRSPACE __CLC_CASTTYPE *)Ptr that discards volatile at

llvm-project/libclc/clc/lib/generic/atomic/clc_atomic_def.inc

Line 42 in 9a15556

(ADDRSPACE __CLC_CASTTYPE *)Ptr, MemoryOrder, MemoryScope)); \

So it should match with your comment.


#define __CLC_IMPL(AS, TYPE) \
_CLC_OVERLOAD _CLC_DEF TYPE atom_add(volatile AS TYPE *p, TYPE val) { \
return __clc_atomic_fetch_add(p, val, __ATOMIC_RELAXED, \
__MEMORY_SCOPE_DEVICE); \
} \
_CLC_OVERLOAD _CLC_DEF TYPE atom_add(AS TYPE *p, TYPE val) { \
return atom_add((volatile AS TYPE *)p, val); \
}

#ifdef cl_khr_global_int32_base_atomics
#define __CLC_ATOMIC_OP add
#define __CLC_ATOMIC_ADDRESS_SPACE global
#include "atom_int32_binary.inc"
__CLC_IMPL(global, int)
__CLC_IMPL(global, unsigned int)
#endif // cl_khr_global_int32_base_atomics

#ifdef cl_khr_local_int32_base_atomics
#define __CLC_ATOMIC_OP add
#define __CLC_ATOMIC_ADDRESS_SPACE local
#include "atom_int32_binary.inc"
__CLC_IMPL(local, int)
__CLC_IMPL(local, unsigned int)
#endif // cl_khr_local_int32_base_atomics

#ifdef cl_khr_int64_base_atomics

#define __CLC_IMPL(AS, TYPE) \
_CLC_OVERLOAD _CLC_DEF TYPE atom_add(volatile AS TYPE *p, TYPE val) { \
return __sync_fetch_and_add_8(p, val); \
}

__CLC_IMPL(global, long)
__CLC_IMPL(global, unsigned long)
__CLC_IMPL(local, long)
__CLC_IMPL(local, unsigned long)
#undef __CLC_IMPL

#endif // cl_khr_int64_base_atomics
29 changes: 16 additions & 13 deletions libclc/opencl/lib/generic/atomic/atom_and.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,32 +6,35 @@
//
//===----------------------------------------------------------------------===//

#include <clc/atomic/clc_atomic_fetch_and.h>
#include <clc/opencl/atomic/atom_and.h>
#include <clc/opencl/atomic/atomic_and.h>

// Non-volatile overloads are for backward compatibility with OpenCL 1.0.

#define __CLC_IMPL(AS, TYPE) \
_CLC_OVERLOAD _CLC_DEF TYPE atom_and(volatile AS TYPE *p, TYPE val) { \
return __clc_atomic_fetch_and(p, val, __ATOMIC_RELAXED, \
__MEMORY_SCOPE_DEVICE); \
} \
_CLC_OVERLOAD _CLC_DEF TYPE atom_and(AS TYPE *p, TYPE val) { \
return atom_and((volatile AS TYPE *)p, val); \
}

#ifdef cl_khr_global_int32_extended_atomics
#define __CLC_ATOMIC_OP and
#define __CLC_ATOMIC_ADDRESS_SPACE global
#include "atom_int32_binary.inc"
__CLC_IMPL(global, int)
__CLC_IMPL(global, unsigned int)
#endif // cl_khr_global_int32_extended_atomics

#ifdef cl_khr_local_int32_extended_atomics
#define __CLC_ATOMIC_OP and
#define __CLC_ATOMIC_ADDRESS_SPACE local
#include "atom_int32_binary.inc"
__CLC_IMPL(local, int)
__CLC_IMPL(local, unsigned int)
#endif // cl_khr_local_int32_extended_atomics

#ifdef cl_khr_int64_extended_atomics

#define __CLC_IMPL(AS, TYPE) \
_CLC_OVERLOAD _CLC_DEF TYPE atom_and(volatile AS TYPE *p, TYPE val) { \
return __sync_fetch_and_and_8(p, val); \
}

__CLC_IMPL(global, long)
__CLC_IMPL(global, unsigned long)
__CLC_IMPL(local, long)
__CLC_IMPL(local, unsigned long)
#undef __CLC_IMPL

#endif // cl_khr_int64_extended_atomics
20 changes: 9 additions & 11 deletions libclc/opencl/lib/generic/atomic/atom_cmpxchg.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,13 +6,20 @@
//
//===----------------------------------------------------------------------===//

#include <clc/atomic/clc_atomic_compare_exchange.h>
#include <clc/opencl/atomic/atom_cmpxchg.h>
#include <clc/opencl/atomic/atomic_cmpxchg.h>

// Non-volatile overloads are for backward compatibility with OpenCL 1.0.

#define __CLC_IMPL(AS, TYPE) \
_CLC_OVERLOAD _CLC_DEF TYPE atom_cmpxchg(volatile AS TYPE *p, TYPE cmp, \
TYPE val) { \
return atomic_cmpxchg(p, cmp, val); \
return __clc_atomic_compare_exchange(p, cmp, val, __ATOMIC_RELAXED, \
__ATOMIC_RELAXED, \
__MEMORY_SCOPE_DEVICE); \
} \
_CLC_OVERLOAD _CLC_DEF TYPE atom_cmpxchg(AS TYPE *p, TYPE cmp, TYPE val) { \
return atom_cmpxchg((volatile AS TYPE *)p, cmp, val); \
}

#ifdef cl_khr_global_int32_base_atomics
Expand All @@ -24,20 +31,11 @@ __CLC_IMPL(local, int)
__CLC_IMPL(local, unsigned int)
#endif // cl_khr_local_int32_base_atomics

#undef __CLC_IMPL

#ifdef cl_khr_int64_base_atomics

#define __CLC_IMPL(AS, TYPE) \
_CLC_OVERLOAD _CLC_DEF TYPE atom_cmpxchg(volatile AS TYPE *p, TYPE cmp, \
TYPE val) { \
return __sync_val_compare_and_swap_8(p, cmp, val); \
}

__CLC_IMPL(global, long)
__CLC_IMPL(global, unsigned long)
__CLC_IMPL(local, long)
__CLC_IMPL(local, unsigned long)
#undef __CLC_IMPL

#endif // cl_khr_int64_base_atomics
18 changes: 7 additions & 11 deletions libclc/opencl/lib/generic/atomic/atom_dec.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,13 +6,17 @@
//
//===----------------------------------------------------------------------===//

#include <clc/atomic/clc_atomic_dec.h>
#include <clc/opencl/atomic/atom_dec.h>
#include <clc/opencl/atomic/atom_sub.h>
#include <clc/opencl/atomic/atomic_dec.h>

// Non-volatile overloads are for backward compatibility with OpenCL 1.0.

#define __CLC_IMPL(AS, TYPE) \
_CLC_OVERLOAD _CLC_DEF TYPE atom_dec(volatile AS TYPE *p) { \
return atomic_dec(p); \
return __clc_atomic_dec(p, __ATOMIC_RELAXED, __MEMORY_SCOPE_DEVICE); \
} \
_CLC_OVERLOAD _CLC_DEF TYPE atom_dec(AS TYPE *p) { \
return atom_dec((volatile AS TYPE *)p); \
}

#ifdef cl_khr_global_int32_base_atomics
Expand All @@ -24,19 +28,11 @@ __CLC_IMPL(local, int)
__CLC_IMPL(local, unsigned int)
#endif // cl_khr_local_int32_base_atomics

#undef __CLC_IMPL

#ifdef cl_khr_int64_base_atomics

#define __CLC_IMPL(AS, TYPE) \
_CLC_OVERLOAD _CLC_DEF TYPE atom_dec(volatile AS TYPE *p) { \
return atom_sub(p, (TYPE)1); \
}

__CLC_IMPL(global, long)
__CLC_IMPL(global, unsigned long)
__CLC_IMPL(local, long)
__CLC_IMPL(local, unsigned long)
#undef __CLC_IMPL

#endif // cl_khr_int64_base_atomics
18 changes: 7 additions & 11 deletions libclc/opencl/lib/generic/atomic/atom_inc.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,13 +6,17 @@
//
//===----------------------------------------------------------------------===//

#include <clc/opencl/atomic/atom_add.h>
#include <clc/atomic/clc_atomic_inc.h>
#include <clc/opencl/atomic/atom_inc.h>
#include <clc/opencl/atomic/atomic_inc.h>

// Non-volatile overloads are for backward compatibility with OpenCL 1.0.

#define __CLC_IMPL(AS, TYPE) \
_CLC_OVERLOAD _CLC_DEF TYPE atom_inc(volatile AS TYPE *p) { \
return atomic_inc(p); \
return __clc_atomic_inc(p, __ATOMIC_RELAXED, __MEMORY_SCOPE_DEVICE); \
} \
_CLC_OVERLOAD _CLC_DEF TYPE atom_inc(AS TYPE *p) { \
return atom_inc((volatile AS TYPE *)p); \
}

#ifdef cl_khr_global_int32_base_atomics
Expand All @@ -24,19 +28,11 @@ __CLC_IMPL(local, int)
__CLC_IMPL(local, unsigned int)
#endif // cl_khr_local_int32_base_atomics

#undef __CLC_IMPL

#ifdef cl_khr_int64_base_atomics

#define __CLC_IMPL(AS, TYPE) \
_CLC_OVERLOAD _CLC_DEF TYPE atom_inc(volatile AS TYPE *p) { \
return atom_add(p, (TYPE)1); \
}

__CLC_IMPL(global, long)
__CLC_IMPL(global, unsigned long)
__CLC_IMPL(local, long)
__CLC_IMPL(local, unsigned long)
#undef __CLC_IMPL

#endif // cl_khr_int64_base_atomics
23 changes: 0 additions & 23 deletions libclc/opencl/lib/generic/atomic/atom_int32_binary.inc
View file
Open in desktop

This file was deleted.

45 changes: 20 additions & 25 deletions libclc/opencl/lib/generic/atomic/atom_max.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,40 +6,35 @@
//
//===----------------------------------------------------------------------===//

#include <clc/atomic/clc_atomic_fetch_max.h>
#include <clc/opencl/atomic/atom_max.h>
#include <clc/opencl/atomic/atomic_max.h>

// Non-volatile overloads are for backward compatibility with OpenCL 1.0.

#define __CLC_IMPL(AS, TYPE) \
_CLC_OVERLOAD _CLC_DEF TYPE atom_max(volatile AS TYPE *p, TYPE val) { \
return __clc_atomic_fetch_max(p, val, __ATOMIC_RELAXED, \
__MEMORY_SCOPE_DEVICE); \
} \
_CLC_OVERLOAD _CLC_DEF TYPE atom_max(AS TYPE *p, TYPE val) { \
return atom_max((volatile AS TYPE *)p, val); \
}

#ifdef cl_khr_global_int32_extended_atomics
#define __CLC_ATOMIC_OP max
#define __CLC_ATOMIC_ADDRESS_SPACE global
#include "atom_int32_binary.inc"
__CLC_IMPL(global, int)
__CLC_IMPL(global, unsigned int)
#endif // cl_khr_global_int32_extended_atomics

#ifdef cl_khr_local_int32_extended_atomics
#define __CLC_ATOMIC_OP max
#define __CLC_ATOMIC_ADDRESS_SPACE local
#include "atom_int32_binary.inc"
__CLC_IMPL(local, int)
__CLC_IMPL(local, unsigned int)
#endif // cl_khr_local_int32_extended_atomics

#ifdef cl_khr_int64_extended_atomics

unsigned long __clc__sync_fetch_and_max_local_8(volatile local long *, long);
unsigned long __clc__sync_fetch_and_max_global_8(volatile global long *, long);
unsigned long __clc__sync_fetch_and_umax_local_8(volatile local unsigned long *,
unsigned long);
unsigned long
__clc__sync_fetch_and_umax_global_8(volatile global unsigned long *,
unsigned long);

#define __CLC_IMPL(AS, TYPE, OP) \
_CLC_OVERLOAD _CLC_DEF TYPE atom_max(volatile AS TYPE *p, TYPE val) { \
return __clc__sync_fetch_and_##OP##_##AS##_8(p, val); \
}

__CLC_IMPL(global, long, max)
__CLC_IMPL(global, unsigned long, umax)
__CLC_IMPL(local, long, max)
__CLC_IMPL(local, unsigned long, umax)
#undef __CLC_IMPL
__CLC_IMPL(global, long)
__CLC_IMPL(global, unsigned long)
__CLC_IMPL(local, long)
__CLC_IMPL(local, unsigned long)

#endif // cl_khr_int64_extended_atomics
Loading
Loading