https://github.com/wenju-he updated https://github.com/llvm/llvm-project/pull/168318
>From 5965c4bd306c1a227188ea6eeb9c3d247e9ab2f3 Mon Sep 17 00:00:00 2001 From: Wenju He <[email protected]> Date: Mon, 17 Nov 2025 06:52:38 +0100 Subject: [PATCH 1/4] [libclc] Add OpenCL atomic_*_explicit builtins Implement atomic_*_explicit (e.g. atomic_store_explicit) with memory_order plus optional memory_scope. OpenCL memory_order maps 1:1 to Clang (e.g. OpenCL memory_order_relaxed == Clang __ATOMIC_RELAXED), so we pass it unchanged to clc_atomic_* function which forwards to Clang _scoped_atomic* builtins. Other changes: * Add __opencl_get_clang_memory_scope helper in opencl/utils.h (OpenCL scope -> Clang scope). * Correct atomic_compare_exchange return type to bool. * Fix atomic_compare_exchange to return true when value stored in the pointer equals expected value. --- .../atomic/atomic_compare_exchange_strong.h | 3 + .../atomic/atomic_compare_exchange_weak.h | 3 + .../include/clc/opencl/atomic/atomic_decl.inc | 91 ++++++++++-- .../clc/opencl/atomic/atomic_exchange.h | 3 + .../clc/opencl/atomic/atomic_fetch_add.h | 3 + .../clc/opencl/atomic/atomic_fetch_and.h | 3 + .../clc/opencl/atomic/atomic_fetch_max.h | 3 + .../clc/opencl/atomic/atomic_fetch_min.h | 3 + .../clc/opencl/atomic/atomic_fetch_or.h | 3 + .../clc/opencl/atomic/atomic_fetch_sub.h | 3 + .../clc/opencl/atomic/atomic_fetch_xor.h | 3 + .../include/clc/opencl/atomic/atomic_load.h | 3 + .../include/clc/opencl/atomic/atomic_store.h | 3 + libclc/opencl/include/clc/opencl/types.h | 48 +++++++ libclc/opencl/include/clc/opencl/utils.h | 33 +++++ .../atomic/atomic_compare_exchange_strong.cl | 7 +- .../atomic/atomic_compare_exchange_weak.cl | 7 +- .../opencl/lib/generic/atomic/atomic_def.inc | 131 +++++++++++++++--- .../lib/generic/atomic/atomic_exchange.cl | 7 +- .../lib/generic/atomic/atomic_fetch_add.cl | 7 +- .../lib/generic/atomic/atomic_fetch_and.cl | 7 +- .../lib/generic/atomic/atomic_fetch_max.cl | 7 +- .../lib/generic/atomic/atomic_fetch_min.cl | 7 +- .../lib/generic/atomic/atomic_fetch_or.cl | 7 +- .../lib/generic/atomic/atomic_fetch_sub.cl | 7 +- .../lib/generic/atomic/atomic_fetch_xor.cl | 7 +- .../opencl/lib/generic/atomic/atomic_load.cl | 7 +- .../opencl/lib/generic/atomic/atomic_store.cl | 7 +- 28 files changed, 325 insertions(+), 98 deletions(-) create mode 100644 libclc/opencl/include/clc/opencl/types.h create mode 100644 libclc/opencl/include/clc/opencl/utils.h diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_compare_exchange_strong.h b/libclc/opencl/include/clc/opencl/atomic/atomic_compare_exchange_strong.h index 59bfa0e87dd8f..4870b13329e4f 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_compare_exchange_strong.h +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_compare_exchange_strong.h @@ -9,6 +9,9 @@ #ifndef __CLC_OPENCL_ATOMIC_ATOMIC_COMPARE_EXCHANGE_STRONG_H__ #define __CLC_OPENCL_ATOMIC_ATOMIC_COMPARE_EXCHANGE_STRONG_H__ +#include <clc/opencl/opencl-base.h> +#include <clc/opencl/types.h> + #define __CLC_FUNCTION atomic_compare_exchange_strong #define __CLC_COMPARE_EXCHANGE diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_compare_exchange_weak.h b/libclc/opencl/include/clc/opencl/atomic/atomic_compare_exchange_weak.h index 7106c3e061d65..103d4f5504d71 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_compare_exchange_weak.h +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_compare_exchange_weak.h @@ -9,6 +9,9 @@ #ifndef __CLC_OPENCL_ATOMIC_ATOMIC_COMPARE_EXCHANGE_WEAK_H__ #define __CLC_OPENCL_ATOMIC_ATOMIC_COMPARE_EXCHANGE_WEAK_H__ +#include <clc/opencl/opencl-base.h> +#include <clc/opencl/types.h> + #define __CLC_FUNCTION atomic_compare_exchange_weak #define __CLC_COMPARE_EXCHANGE diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_decl.inc b/libclc/opencl/include/clc/opencl/atomic/atomic_decl.inc index 38d250f0693f7..a36e68bca86a2 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_decl.inc +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_decl.inc @@ -25,32 +25,105 @@ #define __CLC_ATOMIC_GENTYPE __CLC_XCONCAT(atomic_, __CLC_GENTYPE) +#define __CLC_FUNCTION_EXPLICIT __CLC_XCONCAT(__CLC_FUNCTION, _explicit) + +#ifdef __CLC_NO_VALUE_ARG +#define __CLC_DECL_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, memory_order Order, \ + memory_scope Scope); +#elif defined(__CLC_RETURN_VOID) +#define __CLC_DECL_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DECL void __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ + memory_order Order, memory_scope Scope); +#elif defined(__CLC_COMPARE_EXCHANGE) +#define __CLC_DECL_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DECL bool __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, \ + ADDRSPACE __CLC_GENTYPE *Expected, __CLC_GENTYPE Desired, \ + memory_order Order, memory_scope Scope); +#else +#define __CLC_DECL_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ + memory_order Order, memory_scope Scope); +#endif + +__CLC_DECL_ATOMIC(global) +__CLC_DECL_ATOMIC(local) +#if _CLC_GENERIC_AS_SUPPORTED +__CLC_DECL_ATOMIC() +#endif + +#undef __CLC_DECL_ATOMIC + +#if defined(__opencl_c_atomic_scope_device) + +#ifdef __CLC_NO_VALUE_ARG +#define __CLC_DECL_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, memory_order Order); +#elif defined(__CLC_RETURN_VOID) +#define __CLC_DECL_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DECL void __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ + memory_order Order); +#elif defined(__CLC_COMPARE_EXCHANGE) +#define __CLC_DECL_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DECL bool __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, \ + ADDRSPACE __CLC_GENTYPE *Expected, __CLC_GENTYPE Desired, \ + memory_order Success, memory_order Failure); +#else +#define __CLC_DECL_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ + memory_order Order); +#endif + +__CLC_DECL_ATOMIC(global) +__CLC_DECL_ATOMIC(local) +#if _CLC_GENERIC_AS_SUPPORTED +__CLC_DECL_ATOMIC() +#endif + +#undef __CLC_DECL_ATOMIC + +#endif // defined(__opencl_c_atomic_scope_device) + +#if defined(__opencl_c_atomic_order_seq_cst) && \ + defined(__opencl_c_atomic_scope_device) + #ifdef __CLC_NO_VALUE_ARG -#define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ +#define __CLC_DECL_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr); #elif defined(__CLC_RETURN_VOID) -#define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ +#define __CLC_DECL_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DECL void __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value); #elif defined(__CLC_COMPARE_EXCHANGE) -#define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ - _CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE __CLC_FUNCTION( \ +#define __CLC_DECL_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DECL bool __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, \ ADDRSPACE __CLC_GENTYPE *Expected, __CLC_GENTYPE Desired); #else -#define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ +#define __CLC_DECL_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value); #endif -__CLC_DEFINE_ATOMIC(global) -__CLC_DEFINE_ATOMIC(local) +__CLC_DECL_ATOMIC(global) +__CLC_DECL_ATOMIC(local) #if _CLC_GENERIC_AS_SUPPORTED -__CLC_DEFINE_ATOMIC() +__CLC_DECL_ATOMIC() #endif -#undef __CLC_DEFINE_ATOMIC +#undef __CLC_DECL_ATOMIC + +#endif // defined(__opencl_c_atomic_order_seq_cst) && + // defined(__opencl_c_atomic_scope_device) #endif // __CLC_HAVE_FP_ATOMIC || __CLC_HAVE_INT_ATOMIC diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_exchange.h b/libclc/opencl/include/clc/opencl/atomic/atomic_exchange.h index 9d949825b58c3..d47691b373eb0 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_exchange.h +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_exchange.h @@ -9,6 +9,9 @@ #ifndef __CLC_OPENCL_ATOMIC_ATOMIC_EXCHANGE_H__ #define __CLC_OPENCL_ATOMIC_ATOMIC_EXCHANGE_H__ +#include <clc/opencl/opencl-base.h> +#include <clc/opencl/types.h> + #define __CLC_FUNCTION atomic_exchange #define __CLC_BODY <clc/opencl/atomic/atomic_decl.inc> diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_add.h b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_add.h index bae5a7a7e19bb..9ec29e1a553da 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_add.h +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_add.h @@ -9,6 +9,9 @@ #ifndef __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_ADD_H__ #define __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_ADD_H__ +#include <clc/opencl/opencl-base.h> +#include <clc/opencl/types.h> + #define __CLC_FUNCTION atomic_fetch_add #define __CLC_BODY <clc/opencl/atomic/atomic_decl.inc> diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_and.h b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_and.h index 9f9d2225f910e..fb51102911228 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_and.h +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_and.h @@ -9,6 +9,9 @@ #ifndef __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_AND_H__ #define __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_AND_H__ +#include <clc/opencl/opencl-base.h> +#include <clc/opencl/types.h> + #define __CLC_FUNCTION atomic_fetch_and #define __CLC_BODY <clc/opencl/atomic/atomic_decl.inc> diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_max.h b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_max.h index bef102dc82f48..8902e000a1024 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_max.h +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_max.h @@ -9,6 +9,9 @@ #ifndef __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_MAX_H__ #define __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_MAX_H__ +#include <clc/opencl/opencl-base.h> +#include <clc/opencl/types.h> + #define __CLC_FUNCTION atomic_fetch_max #define __CLC_BODY <clc/opencl/atomic/atomic_decl.inc> diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_min.h b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_min.h index d7e346dc44368..0b79b5d9f9d18 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_min.h +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_min.h @@ -9,6 +9,9 @@ #ifndef __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_MIN_H__ #define __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_MIN_H__ +#include <clc/opencl/opencl-base.h> +#include <clc/opencl/types.h> + #define __CLC_FUNCTION atomic_fetch_min #define __CLC_BODY <clc/opencl/atomic/atomic_decl.inc> diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_or.h b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_or.h index aa00982e15a56..5928e15cc3f53 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_or.h +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_or.h @@ -9,6 +9,9 @@ #ifndef __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_OR_H__ #define __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_OR_H__ +#include <clc/opencl/opencl-base.h> +#include <clc/opencl/types.h> + #define __CLC_FUNCTION atomic_fetch_or #define __CLC_BODY <clc/opencl/atomic/atomic_decl.inc> diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_sub.h b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_sub.h index 3d04ed7ba34f8..76e519f933121 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_sub.h +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_sub.h @@ -9,6 +9,9 @@ #ifndef __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_SUB_H__ #define __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_SUB_H__ +#include <clc/opencl/opencl-base.h> +#include <clc/opencl/types.h> + #define __CLC_FUNCTION atomic_fetch_sub #define __CLC_BODY <clc/opencl/atomic/atomic_decl.inc> diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_xor.h b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_xor.h index 2cdff08069025..c0befd44eae20 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_xor.h +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_fetch_xor.h @@ -9,6 +9,9 @@ #ifndef __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_XOR_H__ #define __CLC_OPENCL_ATOMIC_ATOMIC_FETCH_XOR_H__ +#include <clc/opencl/opencl-base.h> +#include <clc/opencl/types.h> + #define __CLC_FUNCTION atomic_fetch_xor #define __CLC_BODY <clc/opencl/atomic/atomic_decl.inc> diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_load.h b/libclc/opencl/include/clc/opencl/atomic/atomic_load.h index 7db259b136ec8..1aaa26bdecc9e 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_load.h +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_load.h @@ -9,6 +9,9 @@ #ifndef __CLC_OPENCL_ATOMIC_ATOMIC_LOAD_H__ #define __CLC_OPENCL_ATOMIC_ATOMIC_LOAD_H__ +#include <clc/opencl/opencl-base.h> +#include <clc/opencl/types.h> + #define __CLC_FUNCTION atomic_load #define __CLC_NO_VALUE_ARG diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_store.h b/libclc/opencl/include/clc/opencl/atomic/atomic_store.h index b3cdfc6ffaeae..f754314918f82 100644 --- a/libclc/opencl/include/clc/opencl/atomic/atomic_store.h +++ b/libclc/opencl/include/clc/opencl/atomic/atomic_store.h @@ -9,6 +9,9 @@ #ifndef __CLC_OPENCL_ATOMIC_ATOMIC_STORE_H__ #define __CLC_OPENCL_ATOMIC_ATOMIC_STORE_H__ +#include <clc/opencl/opencl-base.h> +#include <clc/opencl/types.h> + #define __CLC_FUNCTION atomic_store #define __CLC_RETURN_VOID diff --git a/libclc/opencl/include/clc/opencl/types.h b/libclc/opencl/include/clc/opencl/types.h new file mode 100644 index 0000000000000..b1be88f21bdaa --- /dev/null +++ b/libclc/opencl/include/clc/opencl/types.h @@ -0,0 +1,48 @@ +//===----------------------------------------------------------------------===// +// +// 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 __CLC_OPENCL_TYPES_H__ +#define __CLC_OPENCL_TYPES_H__ + +// Copied from clang/lib/Headers/opencl-c-base.h + +typedef enum memory_scope { + memory_scope_work_item = __OPENCL_MEMORY_SCOPE_WORK_ITEM, + memory_scope_work_group = __OPENCL_MEMORY_SCOPE_WORK_GROUP, + memory_scope_device = __OPENCL_MEMORY_SCOPE_DEVICE, +#if defined(__opencl_c_atomic_scope_all_devices) + memory_scope_all_svm_devices = __OPENCL_MEMORY_SCOPE_ALL_SVM_DEVICES, +#if (__OPENCL_C_VERSION__ >= CL_VERSION_3_0 || __OPENCL_CPP_VERSION__ >= 202100) + memory_scope_all_devices = memory_scope_all_svm_devices, +#endif // (__OPENCL_C_VERSION__ >= CL_VERSION_3_0 || __OPENCL_CPP_VERSION__ >= + // 202100) +#endif // defined(__opencl_c_atomic_scope_all_devices) +/** + * Subgroups have different requirements on forward progress, so just test + * all the relevant macros. + * CL 3.0 sub-groups "they are not guaranteed to make independent forward + * progress" KHR subgroups "Subgroups within a workgroup are independent, make + * forward progress with respect to each other" + */ +#if defined(cl_intel_subgroups) || defined(cl_khr_subgroups) || \ + defined(__opencl_c_subgroups) + memory_scope_sub_group = __OPENCL_MEMORY_SCOPE_SUB_GROUP +#endif +} memory_scope; + +typedef enum memory_order { + memory_order_relaxed = __ATOMIC_RELAXED, + memory_order_acquire = __ATOMIC_ACQUIRE, + memory_order_release = __ATOMIC_RELEASE, + memory_order_acq_rel = __ATOMIC_ACQ_REL, +#if defined(__opencl_c_atomic_order_seq_cst) + memory_order_seq_cst = __ATOMIC_SEQ_CST +#endif +} memory_order; + +#endif // __CLC_OPENCL_TYPES_H__ diff --git a/libclc/opencl/include/clc/opencl/utils.h b/libclc/opencl/include/clc/opencl/utils.h new file mode 100644 index 0000000000000..c677f82ebb67d --- /dev/null +++ b/libclc/opencl/include/clc/opencl/utils.h @@ -0,0 +1,33 @@ +//===----------------------------------------------------------------------===// +// +// 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 __CLC_OPENCL_UTILS_H__ +#define __CLC_OPENCL_UTILS_H__ + +#include <clc/internal/clc.h> +#include <clc/opencl/types.h> + +static _CLC_INLINE int __opencl_get_clang_memory_scope(memory_scope scope) { + switch (scope) { + case __OPENCL_MEMORY_SCOPE_WORK_ITEM: + return __MEMORY_SCOPE_SINGLE; +#if defined(cl_intel_subgroups) || defined(cl_khr_subgroups) || \ + defined(__opencl_c_subgroups) + case __OPENCL_MEMORY_SCOPE_SUB_GROUP: + return __MEMORY_SCOPE_WVFRNT; +#endif + case __OPENCL_MEMORY_SCOPE_WORK_GROUP: + return __MEMORY_SCOPE_WRKGRP; + case __OPENCL_MEMORY_SCOPE_DEVICE: + return __MEMORY_SCOPE_DEVICE; + default: + return __MEMORY_SCOPE_SYSTEM; + } +} + +#endif // __CLC_OPENCL_UTILS_H__ diff --git a/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_strong.cl b/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_strong.cl index 2c1f07d8ca485..c0ca3f8f7d332 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_strong.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_strong.cl @@ -6,11 +6,9 @@ // //===----------------------------------------------------------------------===// -#if defined(__opencl_c_atomic_order_seq_cst) && \ - defined(__opencl_c_atomic_scope_device) - #include <clc/atomic/clc_atomic_compare_exchange.h> #include <clc/opencl/atomic/atomic_compare_exchange_strong.h> +#include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_compare_exchange_strong #define __CLC_COMPARE_EXCHANGE @@ -20,6 +18,3 @@ #define __CLC_BODY <atomic_def.inc> #include <clc/math/gentype.inc> - -#endif // defined(__opencl_c_atomic_order_seq_cst) && - // defined(__opencl_c_atomic_scope_device) diff --git a/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_weak.cl b/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_weak.cl index 69bdf37250f70..39768fb345714 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_weak.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_weak.cl @@ -6,11 +6,9 @@ // //===----------------------------------------------------------------------===// -#if defined(__opencl_c_atomic_order_seq_cst) && \ - defined(__opencl_c_atomic_scope_device) - #include <clc/atomic/clc_atomic_compare_exchange.h> #include <clc/opencl/atomic/atomic_compare_exchange_weak.h> +#include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_compare_exchange_weak #define __CLC_COMPARE_EXCHANGE @@ -20,6 +18,3 @@ #define __CLC_BODY <atomic_def.inc> #include <clc/math/gentype.inc> - -#endif // defined(__opencl_c_atomic_order_seq_cst) && - // defined(__opencl_c_atomic_scope_device) diff --git a/libclc/opencl/lib/generic/atomic/atomic_def.inc b/libclc/opencl/lib/generic/atomic/atomic_def.inc index a4ccab5990888..99fb778a8b342 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_def.inc +++ b/libclc/opencl/lib/generic/atomic/atomic_def.inc @@ -12,7 +12,8 @@ defined(cl_khr_int64_extended_atomics)) #define __CLC_HAVE_64_ATOMIC #endif -#if defined(__CLC_FPSIZE) && (__CLC_FPSIZE < 64 || defined(__CLC_HAVE_64_ATOMIC) +#if defined(__CLC_FPSIZE) && \ + (__CLC_FPSIZE < 64 || defined(__CLC_HAVE_64_ATOMIC)) #define __CLC_HAVE_FP_ATOMIC #endif #if defined(__CLC_GENSIZE) && \ @@ -24,41 +25,134 @@ #define __CLC_ATOMIC_GENTYPE __CLC_XCONCAT(atomic_, __CLC_GENTYPE) +#define __CLC_FUNCTION_EXPLICIT __CLC_XCONCAT(__CLC_FUNCTION, _explicit) + #ifdef __CLC_NO_VALUE_ARG #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ - _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ - volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr) { \ - return __CLC_IMPL_FUNCTION((volatile ADDRSPACE __CLC_GENTYPE *)Ptr, \ - __ATOMIC_SEQ_CST, __MEMORY_SCOPE_DEVICE); \ + _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, memory_order Order, \ + memory_scope Scope) { \ + return __CLC_IMPL_FUNCTION((volatile ADDRSPACE __CLC_GENTYPE *)Ptr, Order, \ + __opencl_get_clang_memory_scope(Scope)); \ } #elif defined(__CLC_RETURN_VOID) #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ - _CLC_OVERLOAD _CLC_DEF void __CLC_FUNCTION( \ - volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value) { \ - __CLC_IMPL_FUNCTION((volatile ADDRSPACE __CLC_GENTYPE *)Ptr, Value, \ - __ATOMIC_SEQ_CST, __MEMORY_SCOPE_DEVICE); \ + _CLC_OVERLOAD _CLC_DEF void __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ + memory_order Order, memory_scope Scope) { \ + __CLC_IMPL_FUNCTION((volatile ADDRSPACE __CLC_GENTYPE *)Ptr, Value, Order, \ + __opencl_get_clang_memory_scope(Scope)); \ } #elif defined(__CLC_COMPARE_EXCHANGE) #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ - _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ + _CLC_OVERLOAD _CLC_DEF bool __CLC_FUNCTION_EXPLICIT( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, \ - ADDRSPACE __CLC_GENTYPE *Expected, __CLC_GENTYPE Desired) { \ + ADDRSPACE __CLC_GENTYPE *Expected, __CLC_GENTYPE Desired, \ + memory_order Success, memory_order Failure, memory_scope Scope) { \ __CLC_GENTYPE Comparator = *Expected; \ __CLC_GENTYPE RetValue = __clc_atomic_compare_exchange( \ - (volatile ADDRSPACE __CLC_GENTYPE *)Ptr, Comparator, Desired, \ - __ATOMIC_SEQ_CST, __ATOMIC_RELAXED, __MEMORY_SCOPE_DEVICE); \ + (volatile ADDRSPACE __CLC_GENTYPE *)Ptr, Comparator, Desired, Success, \ + Failure, __opencl_get_clang_memory_scope(Scope)); \ if (Comparator != RetValue) { \ *Expected = RetValue; \ - return true; \ + return false; \ } \ - return false; \ + return true; \ + } +#else +#define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ + memory_order Order, memory_scope Scope) { \ + return __CLC_IMPL_FUNCTION((volatile ADDRSPACE __CLC_GENTYPE *)Ptr, Value, \ + Order, __opencl_get_clang_memory_scope(Scope)); \ + } +#endif + +__CLC_DEFINE_ATOMIC(global) +__CLC_DEFINE_ATOMIC(local) +#if _CLC_GENERIC_AS_SUPPORTED +__CLC_DEFINE_ATOMIC() +#endif + +#undef __CLC_DEFINE_ATOMIC + +#if defined(__opencl_c_atomic_scope_device) + +#ifdef __CLC_NO_VALUE_ARG +#define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, memory_order Order) { \ + return __CLC_FUNCTION_EXPLICIT(Ptr, Order, __OPENCL_MEMORY_SCOPE_DEVICE); \ + } +#elif defined(__CLC_RETURN_VOID) +#define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DEF void __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ + memory_order Order) { \ + __CLC_FUNCTION_EXPLICIT(Ptr, Value, Order, __OPENCL_MEMORY_SCOPE_DEVICE); \ + } +#elif defined(__CLC_COMPARE_EXCHANGE) +#define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DEF bool __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, \ + ADDRSPACE __CLC_GENTYPE *Expected, __CLC_GENTYPE Desired, \ + memory_order Success, memory_order Failure) { \ + return __CLC_FUNCTION_EXPLICIT(Ptr, Expected, Desired, Success, Failure, \ + __OPENCL_MEMORY_SCOPE_DEVICE); \ + } +#else +#define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION_EXPLICIT( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ + memory_order Order) { \ + return __CLC_FUNCTION_EXPLICIT(Ptr, Value, Order, \ + __OPENCL_MEMORY_SCOPE_DEVICE); \ + } +#endif + +__CLC_DEFINE_ATOMIC(global) +__CLC_DEFINE_ATOMIC(local) +#if _CLC_GENERIC_AS_SUPPORTED +__CLC_DEFINE_ATOMIC() +#endif + +#undef __CLC_DEFINE_ATOMIC + +#endif // defined(__opencl_c_atomic_scope_device) + +#if defined(__opencl_c_atomic_order_seq_cst) && \ + defined(__opencl_c_atomic_scope_device) + +#ifdef __CLC_NO_VALUE_ARG +#define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr) { \ + return __CLC_FUNCTION_EXPLICIT(Ptr, __ATOMIC_SEQ_CST, \ + __OPENCL_MEMORY_SCOPE_DEVICE); \ + } +#elif defined(__CLC_RETURN_VOID) +#define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DEF void __CLC_FUNCTION( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value) { \ + __CLC_FUNCTION_EXPLICIT(Ptr, Value, __ATOMIC_SEQ_CST, \ + __OPENCL_MEMORY_SCOPE_DEVICE); \ + } +#elif defined(__CLC_COMPARE_EXCHANGE) +#define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ + _CLC_OVERLOAD _CLC_DEF bool __CLC_FUNCTION( \ + volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, \ + ADDRSPACE __CLC_GENTYPE *Expected, __CLC_GENTYPE Desired) { \ + return __CLC_FUNCTION_EXPLICIT(Ptr, Expected, Desired, __ATOMIC_SEQ_CST, \ + __ATOMIC_SEQ_CST, \ + __OPENCL_MEMORY_SCOPE_DEVICE); \ } #else #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value) { \ - return __CLC_IMPL_FUNCTION((volatile ADDRSPACE __CLC_GENTYPE *)Ptr, Value, \ - __ATOMIC_SEQ_CST, __MEMORY_SCOPE_DEVICE); \ + return __CLC_FUNCTION_EXPLICIT(Ptr, Value, __ATOMIC_SEQ_CST, \ + __OPENCL_MEMORY_SCOPE_DEVICE); \ } #endif @@ -70,6 +164,9 @@ __CLC_DEFINE_ATOMIC() #undef __CLC_DEFINE_ATOMIC +#endif // defined(__opencl_c_atomic_order_seq_cst) && + // defined(__opencl_c_atomic_scope_device) + #endif // __CLC_HAVE_FP_ATOMIC || __CLC_HAVE_INT_ATOMIC #undef __CLC_HAVE_INT_ATOMIC diff --git a/libclc/opencl/lib/generic/atomic/atomic_exchange.cl b/libclc/opencl/lib/generic/atomic/atomic_exchange.cl index 5f7e2fa593e3f..f7568f6ace38c 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_exchange.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_exchange.cl @@ -6,11 +6,9 @@ // //===----------------------------------------------------------------------===// -#if defined(__opencl_c_atomic_order_seq_cst) && \ - defined(__opencl_c_atomic_scope_device) - #include <clc/atomic/clc_atomic_exchange.h> #include <clc/opencl/atomic/atomic_exchange.h> +#include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_exchange #define __CLC_IMPL_FUNCTION __clc_atomic_exchange @@ -20,6 +18,3 @@ #define __CLC_BODY <atomic_def.inc> #include <clc/math/gentype.inc> - -#endif // defined(__opencl_c_atomic_order_seq_cst) && - // defined(__opencl_c_atomic_scope_device) diff --git a/libclc/opencl/lib/generic/atomic/atomic_fetch_add.cl b/libclc/opencl/lib/generic/atomic/atomic_fetch_add.cl index 0362ff89d1d78..d27cc7120ccce 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_fetch_add.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_fetch_add.cl @@ -6,11 +6,9 @@ // //===----------------------------------------------------------------------===// -#if defined(__opencl_c_atomic_order_seq_cst) && \ - defined(__opencl_c_atomic_scope_device) - #include <clc/atomic/clc_atomic_fetch_add.h> #include <clc/opencl/atomic/atomic_fetch_add.h> +#include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_fetch_add #define __CLC_IMPL_FUNCTION __clc_atomic_fetch_add @@ -20,6 +18,3 @@ #define __CLC_BODY <atomic_def.inc> #include <clc/math/gentype.inc> - -#endif // defined(__opencl_c_atomic_order_seq_cst) && - // defined(__opencl_c_atomic_scope_device) diff --git a/libclc/opencl/lib/generic/atomic/atomic_fetch_and.cl b/libclc/opencl/lib/generic/atomic/atomic_fetch_and.cl index a1796f20c6e44..b8531722911cf 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_fetch_and.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_fetch_and.cl @@ -6,17 +6,12 @@ // //===----------------------------------------------------------------------===// -#if defined(__opencl_c_atomic_order_seq_cst) && \ - defined(__opencl_c_atomic_scope_device) - #include <clc/atomic/clc_atomic_fetch_and.h> #include <clc/opencl/atomic/atomic_fetch_and.h> +#include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_fetch_and #define __CLC_IMPL_FUNCTION __clc_atomic_fetch_and #define __CLC_BODY <atomic_def.inc> #include <clc/integer/gentype.inc> - -#endif // defined(__opencl_c_atomic_order_seq_cst) && - // defined(__opencl_c_atomic_scope_device) diff --git a/libclc/opencl/lib/generic/atomic/atomic_fetch_max.cl b/libclc/opencl/lib/generic/atomic/atomic_fetch_max.cl index 03b5d1d8ae7bd..b644ca336437a 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_fetch_max.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_fetch_max.cl @@ -6,11 +6,9 @@ // //===----------------------------------------------------------------------===// -#if defined(__opencl_c_atomic_order_seq_cst) && \ - defined(__opencl_c_atomic_scope_device) - #include <clc/atomic/clc_atomic_fetch_max.h> #include <clc/opencl/atomic/atomic_fetch_max.h> +#include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_fetch_max #define __CLC_IMPL_FUNCTION __clc_atomic_fetch_max @@ -20,6 +18,3 @@ #define __CLC_BODY <atomic_def.inc> #include <clc/math/gentype.inc> - -#endif // defined(__opencl_c_atomic_order_seq_cst) && - // defined(__opencl_c_atomic_scope_device) diff --git a/libclc/opencl/lib/generic/atomic/atomic_fetch_min.cl b/libclc/opencl/lib/generic/atomic/atomic_fetch_min.cl index 60ffeff04cc6a..f24fcf329b6f2 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_fetch_min.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_fetch_min.cl @@ -6,11 +6,9 @@ // //===----------------------------------------------------------------------===// -#if defined(__opencl_c_atomic_order_seq_cst) && \ - defined(__opencl_c_atomic_scope_device) - #include <clc/atomic/clc_atomic_fetch_min.h> #include <clc/opencl/atomic/atomic_fetch_min.h> +#include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_fetch_min #define __CLC_IMPL_FUNCTION __clc_atomic_fetch_min @@ -20,6 +18,3 @@ #define __CLC_BODY <atomic_def.inc> #include <clc/math/gentype.inc> - -#endif // defined(__opencl_c_atomic_order_seq_cst) && - // defined(__opencl_c_atomic_scope_device) diff --git a/libclc/opencl/lib/generic/atomic/atomic_fetch_or.cl b/libclc/opencl/lib/generic/atomic/atomic_fetch_or.cl index 8f4100bb150e3..1f6fe4cac090a 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_fetch_or.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_fetch_or.cl @@ -6,17 +6,12 @@ // //===----------------------------------------------------------------------===// -#if defined(__opencl_c_atomic_order_seq_cst) && \ - defined(__opencl_c_atomic_scope_device) - #include <clc/atomic/clc_atomic_fetch_or.h> #include <clc/opencl/atomic/atomic_fetch_or.h> +#include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_fetch_or #define __CLC_IMPL_FUNCTION __clc_atomic_fetch_or #define __CLC_BODY <atomic_def.inc> #include <clc/integer/gentype.inc> - -#endif // defined(__opencl_c_atomic_order_seq_cst) && - // defined(__opencl_c_atomic_scope_device) diff --git a/libclc/opencl/lib/generic/atomic/atomic_fetch_sub.cl b/libclc/opencl/lib/generic/atomic/atomic_fetch_sub.cl index ecb5b4315ee86..94323a2c0fcb6 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_fetch_sub.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_fetch_sub.cl @@ -6,11 +6,9 @@ // //===----------------------------------------------------------------------===// -#if defined(__opencl_c_atomic_order_seq_cst) && \ - defined(__opencl_c_atomic_scope_device) - #include <clc/atomic/clc_atomic_fetch_sub.h> #include <clc/opencl/atomic/atomic_fetch_sub.h> +#include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_fetch_sub #define __CLC_IMPL_FUNCTION __clc_atomic_fetch_sub @@ -20,6 +18,3 @@ #define __CLC_BODY <atomic_def.inc> #include <clc/math/gentype.inc> - -#endif // defined(__opencl_c_atomic_order_seq_cst) && - // defined(__opencl_c_atomic_scope_device) diff --git a/libclc/opencl/lib/generic/atomic/atomic_fetch_xor.cl b/libclc/opencl/lib/generic/atomic/atomic_fetch_xor.cl index c49a55820c8d4..13e1db1124f9a 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_fetch_xor.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_fetch_xor.cl @@ -6,17 +6,12 @@ // //===----------------------------------------------------------------------===// -#if defined(__opencl_c_atomic_order_seq_cst) && \ - defined(__opencl_c_atomic_scope_device) - #include <clc/atomic/clc_atomic_fetch_xor.h> #include <clc/opencl/atomic/atomic_fetch_xor.h> +#include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_fetch_xor #define __CLC_IMPL_FUNCTION __clc_atomic_fetch_xor #define __CLC_BODY <atomic_def.inc> #include <clc/integer/gentype.inc> - -#endif // defined(__opencl_c_atomic_order_seq_cst) && - // defined(__opencl_c_atomic_scope_device) diff --git a/libclc/opencl/lib/generic/atomic/atomic_load.cl b/libclc/opencl/lib/generic/atomic/atomic_load.cl index e904330be0064..1b93ce84ea863 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_load.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_load.cl @@ -6,11 +6,9 @@ // //===----------------------------------------------------------------------===// -#if defined(__opencl_c_atomic_order_seq_cst) && \ - defined(__opencl_c_atomic_scope_device) - #include <clc/atomic/clc_atomic_load.h> #include <clc/opencl/atomic/atomic_load.h> +#include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_load #define __CLC_IMPL_FUNCTION __clc_atomic_load @@ -21,6 +19,3 @@ #define __CLC_BODY <atomic_def.inc> #include <clc/math/gentype.inc> - -#endif // defined(__opencl_c_atomic_order_seq_cst) && - // defined(__opencl_c_atomic_scope_device) diff --git a/libclc/opencl/lib/generic/atomic/atomic_store.cl b/libclc/opencl/lib/generic/atomic/atomic_store.cl index 584e29ef99a5f..fcaa4d3128f7d 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_store.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_store.cl @@ -6,11 +6,9 @@ // //===----------------------------------------------------------------------===// -#if defined(__opencl_c_atomic_order_seq_cst) && \ - defined(__opencl_c_atomic_scope_device) - #include <clc/atomic/clc_atomic_store.h> #include <clc/opencl/atomic/atomic_store.h> +#include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_store #define __CLC_IMPL_FUNCTION __clc_atomic_store @@ -21,6 +19,3 @@ #define __CLC_BODY <atomic_def.inc> #include <clc/math/gentype.inc> - -#endif // defined(__opencl_c_atomic_order_seq_cst) && - // defined(__opencl_c_atomic_scope_device) >From 857b7951eca428ea5e1a5f3fa60109b53a45fd5d Mon Sep 17 00:00:00 2001 From: Wenju He <[email protected]> Date: Tue, 18 Nov 2025 02:32:52 +0100 Subject: [PATCH 2/4] change default order to __ATOMIC_RELAXED --- libclc/opencl/lib/generic/atomic/atomic_def.inc | 10 +++++----- libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc | 2 +- 2 files changed, 6 insertions(+), 6 deletions(-) diff --git a/libclc/opencl/lib/generic/atomic/atomic_def.inc b/libclc/opencl/lib/generic/atomic/atomic_def.inc index 99fb778a8b342..059f00f528425 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_def.inc +++ b/libclc/opencl/lib/generic/atomic/atomic_def.inc @@ -128,14 +128,14 @@ __CLC_DEFINE_ATOMIC() #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr) { \ - return __CLC_FUNCTION_EXPLICIT(Ptr, __ATOMIC_SEQ_CST, \ + return __CLC_FUNCTION_EXPLICIT(Ptr, __ATOMIC_RELAXED, \ __OPENCL_MEMORY_SCOPE_DEVICE); \ } #elif defined(__CLC_RETURN_VOID) #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF void __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value) { \ - __CLC_FUNCTION_EXPLICIT(Ptr, Value, __ATOMIC_SEQ_CST, \ + __CLC_FUNCTION_EXPLICIT(Ptr, Value, __ATOMIC_RELAXED, \ __OPENCL_MEMORY_SCOPE_DEVICE); \ } #elif defined(__CLC_COMPARE_EXCHANGE) @@ -143,15 +143,15 @@ __CLC_DEFINE_ATOMIC() _CLC_OVERLOAD _CLC_DEF bool __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, \ ADDRSPACE __CLC_GENTYPE *Expected, __CLC_GENTYPE Desired) { \ - return __CLC_FUNCTION_EXPLICIT(Ptr, Expected, Desired, __ATOMIC_SEQ_CST, \ - __ATOMIC_SEQ_CST, \ + return __CLC_FUNCTION_EXPLICIT(Ptr, Expected, Desired, __ATOMIC_RELAXED, \ + __ATOMIC_RELAXED, \ __OPENCL_MEMORY_SCOPE_DEVICE); \ } #else #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value) { \ - return __CLC_FUNCTION_EXPLICIT(Ptr, Value, __ATOMIC_SEQ_CST, \ + return __CLC_FUNCTION_EXPLICIT(Ptr, Value, __ATOMIC_RELAXED, \ __OPENCL_MEMORY_SCOPE_DEVICE); \ } #endif diff --git a/libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc b/libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc index 03eb5d1b33057..dd8dafb38c883 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc +++ b/libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc @@ -13,7 +13,7 @@ #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_GENTYPE *Ptr) { \ - return __CLC_IMPL_FUNCTION(Ptr, __ATOMIC_SEQ_CST, __MEMORY_SCOPE_DEVICE); \ + return __CLC_IMPL_FUNCTION(Ptr, __ATOMIC_RELAXED, __MEMORY_SCOPE_DEVICE); \ } __CLC_DEFINE_ATOMIC(global) >From 5e2939eb1e6ababd7e45ef359dc62893a8f9f555 Mon Sep 17 00:00:00 2001 From: Wenju He <[email protected]> Date: Tue, 18 Nov 2025 02:38:49 +0100 Subject: [PATCH 3/4] Revert "change default order to __ATOMIC_RELAXED" This reverts commit 857b7951eca428ea5e1a5f3fa60109b53a45fd5d. --- libclc/opencl/lib/generic/atomic/atomic_def.inc | 10 +++++----- libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc | 2 +- 2 files changed, 6 insertions(+), 6 deletions(-) diff --git a/libclc/opencl/lib/generic/atomic/atomic_def.inc b/libclc/opencl/lib/generic/atomic/atomic_def.inc index 059f00f528425..99fb778a8b342 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_def.inc +++ b/libclc/opencl/lib/generic/atomic/atomic_def.inc @@ -128,14 +128,14 @@ __CLC_DEFINE_ATOMIC() #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr) { \ - return __CLC_FUNCTION_EXPLICIT(Ptr, __ATOMIC_RELAXED, \ + return __CLC_FUNCTION_EXPLICIT(Ptr, __ATOMIC_SEQ_CST, \ __OPENCL_MEMORY_SCOPE_DEVICE); \ } #elif defined(__CLC_RETURN_VOID) #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF void __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value) { \ - __CLC_FUNCTION_EXPLICIT(Ptr, Value, __ATOMIC_RELAXED, \ + __CLC_FUNCTION_EXPLICIT(Ptr, Value, __ATOMIC_SEQ_CST, \ __OPENCL_MEMORY_SCOPE_DEVICE); \ } #elif defined(__CLC_COMPARE_EXCHANGE) @@ -143,15 +143,15 @@ __CLC_DEFINE_ATOMIC() _CLC_OVERLOAD _CLC_DEF bool __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, \ ADDRSPACE __CLC_GENTYPE *Expected, __CLC_GENTYPE Desired) { \ - return __CLC_FUNCTION_EXPLICIT(Ptr, Expected, Desired, __ATOMIC_RELAXED, \ - __ATOMIC_RELAXED, \ + return __CLC_FUNCTION_EXPLICIT(Ptr, Expected, Desired, __ATOMIC_SEQ_CST, \ + __ATOMIC_SEQ_CST, \ __OPENCL_MEMORY_SCOPE_DEVICE); \ } #else #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value) { \ - return __CLC_FUNCTION_EXPLICIT(Ptr, Value, __ATOMIC_RELAXED, \ + return __CLC_FUNCTION_EXPLICIT(Ptr, Value, __ATOMIC_SEQ_CST, \ __OPENCL_MEMORY_SCOPE_DEVICE); \ } #endif diff --git a/libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc b/libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc index dd8dafb38c883..03eb5d1b33057 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc +++ b/libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc @@ -13,7 +13,7 @@ #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_GENTYPE *Ptr) { \ - return __CLC_IMPL_FUNCTION(Ptr, __ATOMIC_RELAXED, __MEMORY_SCOPE_DEVICE); \ + return __CLC_IMPL_FUNCTION(Ptr, __ATOMIC_SEQ_CST, __MEMORY_SCOPE_DEVICE); \ } __CLC_DEFINE_ATOMIC(global) >From 808f3d8c358ed97706054c878969d71a801edba2 Mon Sep 17 00:00:00 2001 From: Wenju He <[email protected]> Date: Mon, 24 Nov 2025 08:12:43 +0100 Subject: [PATCH 4/4] remove volatile from CLC functions --- libclc/clc/include/clc/atomic/atomic_decl.inc | 13 ++++++------- .../generic/atomic/clc_atomic_compare_exchange.inc | 4 ++-- libclc/clc/lib/generic/atomic/clc_atomic_def.inc | 14 ++++++-------- libclc/opencl/lib/generic/atomic/atom_add.cl | 6 +++--- libclc/opencl/lib/generic/atomic/atom_and.cl | 6 +++--- libclc/opencl/lib/generic/atomic/atom_cmpxchg.cl | 8 ++++---- libclc/opencl/lib/generic/atomic/atom_dec.cl | 6 +++--- libclc/opencl/lib/generic/atomic/atom_inc.cl | 6 +++--- libclc/opencl/lib/generic/atomic/atom_max.cl | 6 +++--- libclc/opencl/lib/generic/atomic/atom_min.cl | 6 +++--- libclc/opencl/lib/generic/atomic/atom_or.cl | 6 +++--- libclc/opencl/lib/generic/atomic/atom_sub.cl | 6 +++--- libclc/opencl/lib/generic/atomic/atom_xchg.cl | 6 +++--- libclc/opencl/lib/generic/atomic/atom_xor.cl | 6 +++--- libclc/opencl/lib/generic/atomic/atomic_add.cl | 2 +- libclc/opencl/lib/generic/atomic/atomic_and.cl | 2 +- libclc/opencl/lib/generic/atomic/atomic_cmpxchg.cl | 4 ++-- .../atomic/atomic_compare_exchange_strong.cl | 1 + .../generic/atomic/atomic_compare_exchange_weak.cl | 1 + libclc/opencl/lib/generic/atomic/atomic_def.inc | 14 +++++++------- .../opencl/lib/generic/atomic/atomic_inc_dec.inc | 3 ++- libclc/opencl/lib/generic/atomic/atomic_max.cl | 12 ++++++------ libclc/opencl/lib/generic/atomic/atomic_min.cl | 12 ++++++------ libclc/opencl/lib/generic/atomic/atomic_or.cl | 2 +- libclc/opencl/lib/generic/atomic/atomic_sub.cl | 2 +- libclc/opencl/lib/generic/atomic/atomic_xchg.cl | 2 +- libclc/opencl/lib/generic/atomic/atomic_xor.cl | 2 +- 27 files changed, 79 insertions(+), 79 deletions(-) diff --git a/libclc/clc/include/clc/atomic/atomic_decl.inc b/libclc/clc/include/clc/atomic/atomic_decl.inc index 5e0f456e34009..87ebe2b355e69 100644 --- a/libclc/clc/include/clc/atomic/atomic_decl.inc +++ b/libclc/clc/include/clc/atomic/atomic_decl.inc @@ -15,24 +15,23 @@ #ifdef __CLC_NO_VALUE_ARG #define __CLC_DECLARE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE __CLC_FUNCTION( \ - volatile ADDRSPACE __CLC_GENTYPE *Ptr, int MemoryOrder, \ - int MemoryScope); + ADDRSPACE __CLC_GENTYPE *Ptr, int MemoryOrder, int MemoryScope); #elif defined(__CLC_RETURN_VOID) #define __CLC_DECLARE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DECL void __CLC_FUNCTION( \ - volatile ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ - int MemoryOrder, int MemoryScope); + ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Value, int MemoryOrder, \ + int MemoryScope); #elif defined(__CLC_COMPARE_EXCHANGE) #define __CLC_DECLARE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE __CLC_FUNCTION( \ - volatile ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Comparator, \ + ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Comparator, \ __CLC_GENTYPE Value, int MemoryOrderEqual, int MemoryOrderUnequal, \ int MemoryScope); #else #define __CLC_DECLARE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE __CLC_FUNCTION( \ - volatile ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ - int MemoryOrder, int MemoryScope); + ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Value, int MemoryOrder, \ + int MemoryScope); #endif __CLC_DECLARE_ATOMIC(global) diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_compare_exchange.inc b/libclc/clc/lib/generic/atomic/clc_atomic_compare_exchange.inc index 74284fd61024c..07ef69d426768 100644 --- a/libclc/clc/lib/generic/atomic/clc_atomic_compare_exchange.inc +++ b/libclc/clc/lib/generic/atomic/clc_atomic_compare_exchange.inc @@ -25,7 +25,7 @@ #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_atomic_compare_exchange( \ - volatile ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Comparator, \ + ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Comparator, \ __CLC_GENTYPE Value, int MemoryOrderEqual, int MemoryOrderUnequal, \ int MemoryScope) { \ __CLC_U_GENTYPE Comp = __CLC_AS_U_GENTYPE(Comparator); \ @@ -39,7 +39,7 @@ #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_atomic_compare_exchange( \ - volatile ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Comparator, \ + ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Comparator, \ __CLC_GENTYPE Value, int MemoryOrderEqual, int MemoryOrderUnequal, \ int MemoryScope) { \ __scoped_atomic_compare_exchange_n(Ptr, &Comparator, Value, false, \ diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_def.inc b/libclc/clc/lib/generic/atomic/clc_atomic_def.inc index 75561430b33ad..ab56cc783da00 100644 --- a/libclc/clc/lib/generic/atomic/clc_atomic_def.inc +++ b/libclc/clc/lib/generic/atomic/clc_atomic_def.inc @@ -36,32 +36,30 @@ #ifdef __CLC_NO_VALUE_ARG #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ - volatile ADDRSPACE __CLC_GENTYPE *Ptr, int MemoryOrder, \ - int MemoryScope) { \ + ADDRSPACE __CLC_GENTYPE *Ptr, int MemoryOrder, int MemoryScope) { \ return __CLC_AS_RETTYPE(__CLC_IMPL_FUNCTION( \ (ADDRSPACE __CLC_CASTTYPE *)Ptr, MemoryOrder, MemoryScope)); \ } #elif defined(__CLC_INC_DEC) #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ - volatile ADDRSPACE __CLC_GENTYPE *Ptr, int MemoryOrder, \ - int MemoryScope) { \ + ADDRSPACE __CLC_GENTYPE *Ptr, int MemoryOrder, int MemoryScope) { \ return __CLC_IMPL_FUNCTION(Ptr, (__CLC_GENTYPE)1, MemoryOrder, \ MemoryScope); \ } #elif defined(__CLC_RETURN_VOID) #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF void __CLC_FUNCTION( \ - volatile ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ - int MemoryOrder, int MemoryScope) { \ + ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Value, int MemoryOrder, \ + int MemoryScope) { \ __CLC_IMPL_FUNCTION((ADDRSPACE __CLC_CASTTYPE *)Ptr, \ __CLC_AS_CASTTYPE(Value), MemoryOrder, MemoryScope); \ } #else #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ - volatile ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ - int MemoryOrder, int MemoryScope) { \ + ADDRSPACE __CLC_GENTYPE *Ptr, __CLC_GENTYPE Value, int MemoryOrder, \ + int MemoryScope) { \ return __CLC_AS_RETTYPE(__CLC_IMPL_FUNCTION( \ (ADDRSPACE __CLC_CASTTYPE *)Ptr, __CLC_AS_CASTTYPE(Value), \ MemoryOrder, MemoryScope)); \ diff --git a/libclc/opencl/lib/generic/atomic/atom_add.cl b/libclc/opencl/lib/generic/atomic/atom_add.cl index 368bbb790fd88..1628ea39de434 100644 --- a/libclc/opencl/lib/generic/atomic/atom_add.cl +++ b/libclc/opencl/lib/generic/atomic/atom_add.cl @@ -12,12 +12,12 @@ // Non-volatile overloads are for backward compatibility with OpenCL 1.0. #define __CLC_IMPL(AS, TYPE) \ - _CLC_OVERLOAD _CLC_DEF TYPE atom_add(volatile AS TYPE *p, TYPE val) { \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_add(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); \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_add(volatile AS TYPE *p, TYPE val) { \ + return atom_add((AS TYPE *)p, val); \ } #ifdef cl_khr_global_int32_base_atomics diff --git a/libclc/opencl/lib/generic/atomic/atom_and.cl b/libclc/opencl/lib/generic/atomic/atom_and.cl index ffcc5bffaafac..bad85046c12e5 100644 --- a/libclc/opencl/lib/generic/atomic/atom_and.cl +++ b/libclc/opencl/lib/generic/atomic/atom_and.cl @@ -12,12 +12,12 @@ // 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) { \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_and(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); \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_and(volatile AS TYPE *p, TYPE val) { \ + return atom_and((AS TYPE *)p, val); \ } #ifdef cl_khr_global_int32_extended_atomics diff --git a/libclc/opencl/lib/generic/atomic/atom_cmpxchg.cl b/libclc/opencl/lib/generic/atomic/atom_cmpxchg.cl index 2e72ec529c45e..41ae9328e9480 100644 --- a/libclc/opencl/lib/generic/atomic/atom_cmpxchg.cl +++ b/libclc/opencl/lib/generic/atomic/atom_cmpxchg.cl @@ -12,14 +12,14 @@ // 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) { \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_cmpxchg(AS TYPE *p, TYPE cmp, TYPE 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); \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_cmpxchg(volatile AS TYPE *p, TYPE cmp, \ + TYPE val) { \ + return atom_cmpxchg((AS TYPE *)p, cmp, val); \ } #ifdef cl_khr_global_int32_base_atomics diff --git a/libclc/opencl/lib/generic/atomic/atom_dec.cl b/libclc/opencl/lib/generic/atomic/atom_dec.cl index a1c7e58ef9e03..1f81f1d37b510 100644 --- a/libclc/opencl/lib/generic/atomic/atom_dec.cl +++ b/libclc/opencl/lib/generic/atomic/atom_dec.cl @@ -12,11 +12,11 @@ // 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) { \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_dec(AS TYPE *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); \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_dec(volatile AS TYPE *p) { \ + return atom_dec((AS TYPE *)p); \ } #ifdef cl_khr_global_int32_base_atomics diff --git a/libclc/opencl/lib/generic/atomic/atom_inc.cl b/libclc/opencl/lib/generic/atomic/atom_inc.cl index f3636d85693b8..5ae5bbb67e791 100644 --- a/libclc/opencl/lib/generic/atomic/atom_inc.cl +++ b/libclc/opencl/lib/generic/atomic/atom_inc.cl @@ -12,11 +12,11 @@ // 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) { \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_inc(AS TYPE *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); \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_inc(volatile AS TYPE *p) { \ + return atom_inc((AS TYPE *)p); \ } #ifdef cl_khr_global_int32_base_atomics diff --git a/libclc/opencl/lib/generic/atomic/atom_max.cl b/libclc/opencl/lib/generic/atomic/atom_max.cl index c2095ec36ba1e..249f5efbd4491 100644 --- a/libclc/opencl/lib/generic/atomic/atom_max.cl +++ b/libclc/opencl/lib/generic/atomic/atom_max.cl @@ -12,12 +12,12 @@ // 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) { \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_max(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); \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_max(volatile AS TYPE *p, TYPE val) { \ + return atom_max((AS TYPE *)p, val); \ } #ifdef cl_khr_global_int32_extended_atomics diff --git a/libclc/opencl/lib/generic/atomic/atom_min.cl b/libclc/opencl/lib/generic/atomic/atom_min.cl index 6360d018d1e90..029c601dd53af 100644 --- a/libclc/opencl/lib/generic/atomic/atom_min.cl +++ b/libclc/opencl/lib/generic/atomic/atom_min.cl @@ -12,12 +12,12 @@ // Non-volatile overloads are for backward compatibility with OpenCL 1.0. #define __CLC_IMPL(AS, TYPE) \ - _CLC_OVERLOAD _CLC_DEF TYPE atom_min(volatile AS TYPE *p, TYPE val) { \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_min(AS TYPE *p, TYPE val) { \ return __clc_atomic_fetch_min(p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } \ - _CLC_OVERLOAD _CLC_DEF TYPE atom_min(AS TYPE *p, TYPE val) { \ - return atom_min((volatile AS TYPE *)p, val); \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_min(volatile AS TYPE *p, TYPE val) { \ + return atom_min((AS TYPE *)p, val); \ } #ifdef cl_khr_global_int32_extended_atomics diff --git a/libclc/opencl/lib/generic/atomic/atom_or.cl b/libclc/opencl/lib/generic/atomic/atom_or.cl index ad28aa436de8c..91f745bdda61a 100644 --- a/libclc/opencl/lib/generic/atomic/atom_or.cl +++ b/libclc/opencl/lib/generic/atomic/atom_or.cl @@ -12,12 +12,12 @@ // Non-volatile overloads are for backward compatibility with OpenCL 1.0. #define __CLC_IMPL(AS, TYPE) \ - _CLC_OVERLOAD _CLC_DEF TYPE atom_or(volatile AS TYPE *p, TYPE val) { \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_or(AS TYPE *p, TYPE val) { \ return __clc_atomic_fetch_or(p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } \ - _CLC_OVERLOAD _CLC_DEF TYPE atom_or(AS TYPE *p, TYPE val) { \ - return atom_or((volatile AS TYPE *)p, val); \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_or(volatile AS TYPE *p, TYPE val) { \ + return atom_or((AS TYPE *)p, val); \ } #ifdef cl_khr_global_int32_extended_atomics diff --git a/libclc/opencl/lib/generic/atomic/atom_sub.cl b/libclc/opencl/lib/generic/atomic/atom_sub.cl index 9daaa1b3ce154..7eeabd51dad48 100644 --- a/libclc/opencl/lib/generic/atomic/atom_sub.cl +++ b/libclc/opencl/lib/generic/atomic/atom_sub.cl @@ -12,12 +12,12 @@ // Non-volatile overloads are for backward compatibility with OpenCL 1.0. #define __CLC_IMPL(AS, TYPE) \ - _CLC_OVERLOAD _CLC_DEF TYPE atom_sub(volatile AS TYPE *p, TYPE val) { \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_sub(AS TYPE *p, TYPE val) { \ return __clc_atomic_fetch_sub(p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } \ - _CLC_OVERLOAD _CLC_DEF TYPE atom_sub(AS TYPE *p, TYPE val) { \ - return atom_sub((volatile AS TYPE *)p, val); \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_sub(volatile AS TYPE *p, TYPE val) { \ + return atom_sub((AS TYPE *)p, val); \ } #ifdef cl_khr_global_int32_base_atomics diff --git a/libclc/opencl/lib/generic/atomic/atom_xchg.cl b/libclc/opencl/lib/generic/atomic/atom_xchg.cl index 5b75873f29760..770e4366de834 100644 --- a/libclc/opencl/lib/generic/atomic/atom_xchg.cl +++ b/libclc/opencl/lib/generic/atomic/atom_xchg.cl @@ -12,12 +12,12 @@ // Non-volatile overloads are for backward compatibility with OpenCL 1.0. #define __CLC_IMPL(AS, TYPE) \ - _CLC_OVERLOAD _CLC_DEF TYPE atom_xchg(volatile AS TYPE *p, TYPE val) { \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_xchg(AS TYPE *p, TYPE val) { \ return __clc_atomic_exchange(p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } \ - _CLC_OVERLOAD _CLC_DEF TYPE atom_xchg(AS TYPE *p, TYPE val) { \ - return atom_xchg((volatile AS TYPE *)p, val); \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_xchg(volatile AS TYPE *p, TYPE val) { \ + return atom_xchg((AS TYPE *)p, val); \ } #ifdef cl_khr_global_int32_base_atomics diff --git a/libclc/opencl/lib/generic/atomic/atom_xor.cl b/libclc/opencl/lib/generic/atomic/atom_xor.cl index 21aba01267e18..3fcf54e1c8832 100644 --- a/libclc/opencl/lib/generic/atomic/atom_xor.cl +++ b/libclc/opencl/lib/generic/atomic/atom_xor.cl @@ -12,12 +12,12 @@ // Non-volatile overloads are for backward compatibility with OpenCL 1.0. #define __CLC_IMPL(AS, TYPE) \ - _CLC_OVERLOAD _CLC_DEF TYPE atom_xor(volatile AS TYPE *p, TYPE val) { \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_xor(AS TYPE *p, TYPE val) { \ return __clc_atomic_fetch_xor(p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } \ - _CLC_OVERLOAD _CLC_DEF TYPE atom_xor(AS TYPE *p, TYPE val) { \ - return atom_xor((volatile AS TYPE *)p, val); \ + _CLC_OVERLOAD _CLC_DEF TYPE atom_xor(volatile AS TYPE *p, TYPE val) { \ + return atom_xor((AS TYPE *)p, val); \ } #ifdef cl_khr_global_int32_extended_atomics diff --git a/libclc/opencl/lib/generic/atomic/atomic_add.cl b/libclc/opencl/lib/generic/atomic/atomic_add.cl index 5501d30544e7c..2f6606eb33338 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_add.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_add.cl @@ -11,7 +11,7 @@ #define __CLC_IMPL(TYPE, AS) \ _CLC_OVERLOAD _CLC_DEF TYPE atomic_add(volatile AS TYPE *p, TYPE val) { \ - return __clc_atomic_fetch_add(p, val, __ATOMIC_RELAXED, \ + return __clc_atomic_fetch_add((AS TYPE *)p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } diff --git a/libclc/opencl/lib/generic/atomic/atomic_and.cl b/libclc/opencl/lib/generic/atomic/atomic_and.cl index ce1adbb6f8235..3bb1f7a6af273 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_and.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_and.cl @@ -11,7 +11,7 @@ #define __CLC_IMPL(TYPE, AS) \ _CLC_OVERLOAD _CLC_DEF TYPE atomic_and(volatile AS TYPE *p, TYPE val) { \ - return __clc_atomic_fetch_and(p, val, __ATOMIC_RELAXED, \ + return __clc_atomic_fetch_and((AS TYPE *)p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } diff --git a/libclc/opencl/lib/generic/atomic/atomic_cmpxchg.cl b/libclc/opencl/lib/generic/atomic/atomic_cmpxchg.cl index 16a8db43e9374..135813b8b879c 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_cmpxchg.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_cmpxchg.cl @@ -12,8 +12,8 @@ #define __CLC_IMPL(TYPE, AS) \ _CLC_OVERLOAD _CLC_DEF TYPE atomic_cmpxchg(volatile AS TYPE *p, TYPE cmp, \ TYPE val) { \ - return __clc_atomic_compare_exchange(p, cmp, val, __ATOMIC_RELAXED, \ - __ATOMIC_RELAXED, \ + return __clc_atomic_compare_exchange((AS TYPE *)p, cmp, val, \ + __ATOMIC_RELAXED, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } diff --git a/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_strong.cl b/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_strong.cl index c0ca3f8f7d332..f113e82285a02 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_strong.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_strong.cl @@ -11,6 +11,7 @@ #include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_compare_exchange_strong +#define __CLC_IMPL_FUNCTION __clc_atomic_compare_exchange #define __CLC_COMPARE_EXCHANGE #define __CLC_BODY <atomic_def.inc> diff --git a/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_weak.cl b/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_weak.cl index 39768fb345714..f5c2899bb18ed 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_weak.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_compare_exchange_weak.cl @@ -11,6 +11,7 @@ #include <clc/opencl/utils.h> #define __CLC_FUNCTION atomic_compare_exchange_weak +#define __CLC_IMPL_FUNCTION __clc_atomic_compare_exchange #define __CLC_COMPARE_EXCHANGE #define __CLC_BODY <atomic_def.inc> diff --git a/libclc/opencl/lib/generic/atomic/atomic_def.inc b/libclc/opencl/lib/generic/atomic/atomic_def.inc index 99fb778a8b342..d13ec24db087d 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_def.inc +++ b/libclc/opencl/lib/generic/atomic/atomic_def.inc @@ -32,7 +32,7 @@ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION_EXPLICIT( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, memory_order Order, \ memory_scope Scope) { \ - return __CLC_IMPL_FUNCTION((volatile ADDRSPACE __CLC_GENTYPE *)Ptr, Order, \ + return __CLC_IMPL_FUNCTION((ADDRSPACE __CLC_GENTYPE *)Ptr, Order, \ __opencl_get_clang_memory_scope(Scope)); \ } #elif defined(__CLC_RETURN_VOID) @@ -40,7 +40,7 @@ _CLC_OVERLOAD _CLC_DEF void __CLC_FUNCTION_EXPLICIT( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ memory_order Order, memory_scope Scope) { \ - __CLC_IMPL_FUNCTION((volatile ADDRSPACE __CLC_GENTYPE *)Ptr, Value, Order, \ + __CLC_IMPL_FUNCTION((ADDRSPACE __CLC_GENTYPE *)Ptr, Value, Order, \ __opencl_get_clang_memory_scope(Scope)); \ } #elif defined(__CLC_COMPARE_EXCHANGE) @@ -50,9 +50,9 @@ ADDRSPACE __CLC_GENTYPE *Expected, __CLC_GENTYPE Desired, \ memory_order Success, memory_order Failure, memory_scope Scope) { \ __CLC_GENTYPE Comparator = *Expected; \ - __CLC_GENTYPE RetValue = __clc_atomic_compare_exchange( \ - (volatile ADDRSPACE __CLC_GENTYPE *)Ptr, Comparator, Desired, Success, \ - Failure, __opencl_get_clang_memory_scope(Scope)); \ + __CLC_GENTYPE RetValue = __CLC_IMPL_FUNCTION( \ + (ADDRSPACE __CLC_GENTYPE *)Ptr, Comparator, Desired, Success, Failure, \ + __opencl_get_clang_memory_scope(Scope)); \ if (Comparator != RetValue) { \ *Expected = RetValue; \ return false; \ @@ -64,8 +64,8 @@ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION_EXPLICIT( \ volatile ADDRSPACE __CLC_ATOMIC_GENTYPE *Ptr, __CLC_GENTYPE Value, \ memory_order Order, memory_scope Scope) { \ - return __CLC_IMPL_FUNCTION((volatile ADDRSPACE __CLC_GENTYPE *)Ptr, Value, \ - Order, __opencl_get_clang_memory_scope(Scope)); \ + return __CLC_IMPL_FUNCTION((ADDRSPACE __CLC_GENTYPE *)Ptr, Value, Order, \ + __opencl_get_clang_memory_scope(Scope)); \ } #endif diff --git a/libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc b/libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc index dd8dafb38c883..fe517ef7a7de7 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc +++ b/libclc/opencl/lib/generic/atomic/atomic_inc_dec.inc @@ -13,7 +13,8 @@ #define __CLC_DEFINE_ATOMIC(ADDRSPACE) \ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNCTION( \ volatile ADDRSPACE __CLC_GENTYPE *Ptr) { \ - return __CLC_IMPL_FUNCTION(Ptr, __ATOMIC_RELAXED, __MEMORY_SCOPE_DEVICE); \ + return __CLC_IMPL_FUNCTION((ADDRSPACE __CLC_GENTYPE *)Ptr, \ + __ATOMIC_RELAXED, __MEMORY_SCOPE_DEVICE); \ } __CLC_DEFINE_ATOMIC(global) diff --git a/libclc/opencl/lib/generic/atomic/atomic_max.cl b/libclc/opencl/lib/generic/atomic/atomic_max.cl index 362a0ed90ca0e..fe0efdbb55c82 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_max.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_max.cl @@ -9,14 +9,14 @@ #include <clc/atomic/clc_atomic_fetch_max.h> #include <clc/opencl/atomic/atomic_max.h> -#define __CLC_IMPL(TYPE, AS, OP) \ +#define __CLC_IMPL(TYPE, AS) \ _CLC_OVERLOAD _CLC_DEF TYPE atomic_max(volatile AS TYPE *p, TYPE val) { \ - return __clc_atomic_fetch_max(p, val, __ATOMIC_RELAXED, \ + return __clc_atomic_fetch_max((AS TYPE *)p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } -__CLC_IMPL(int, global, max) -__CLC_IMPL(unsigned int, global, umax) -__CLC_IMPL(int, local, max) -__CLC_IMPL(unsigned int, local, umax) +__CLC_IMPL(int, global) +__CLC_IMPL(unsigned int, global) +__CLC_IMPL(int, local) +__CLC_IMPL(unsigned int, local) #undef __CLC_IMPL diff --git a/libclc/opencl/lib/generic/atomic/atomic_min.cl b/libclc/opencl/lib/generic/atomic/atomic_min.cl index 1976be0014d70..6a4586b5d26b6 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_min.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_min.cl @@ -9,14 +9,14 @@ #include <clc/atomic/clc_atomic_fetch_min.h> #include <clc/opencl/atomic/atomic_min.h> -#define __CLC_IMPL(TYPE, AS, OP) \ +#define __CLC_IMPL(TYPE, AS) \ _CLC_OVERLOAD _CLC_DEF TYPE atomic_min(volatile AS TYPE *p, TYPE val) { \ - return __clc_atomic_fetch_min(p, val, __ATOMIC_RELAXED, \ + return __clc_atomic_fetch_min((AS TYPE *)p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } -__CLC_IMPL(int, global, min) -__CLC_IMPL(unsigned int, global, umin) -__CLC_IMPL(int, local, min) -__CLC_IMPL(unsigned int, local, umin) +__CLC_IMPL(int, global) +__CLC_IMPL(unsigned int, global) +__CLC_IMPL(int, local) +__CLC_IMPL(unsigned int, local) #undef __CLC_IMPL diff --git a/libclc/opencl/lib/generic/atomic/atomic_or.cl b/libclc/opencl/lib/generic/atomic/atomic_or.cl index ef8bc00f45593..1720b66b87fff 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_or.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_or.cl @@ -11,7 +11,7 @@ #define __CLC_IMPL(TYPE, AS) \ _CLC_OVERLOAD _CLC_DEF TYPE atomic_or(volatile AS TYPE *p, TYPE val) { \ - return __clc_atomic_fetch_or(p, val, __ATOMIC_RELAXED, \ + return __clc_atomic_fetch_or((AS TYPE *)p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } diff --git a/libclc/opencl/lib/generic/atomic/atomic_sub.cl b/libclc/opencl/lib/generic/atomic/atomic_sub.cl index 397737d113c0d..ddb7cc29d3bb0 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_sub.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_sub.cl @@ -11,7 +11,7 @@ #define __CLC_IMPL(TYPE, AS) \ _CLC_OVERLOAD _CLC_DEF TYPE atomic_sub(volatile AS TYPE *p, TYPE val) { \ - return __clc_atomic_fetch_sub(p, val, __ATOMIC_RELAXED, \ + return __clc_atomic_fetch_sub((AS TYPE *)p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } diff --git a/libclc/opencl/lib/generic/atomic/atomic_xchg.cl b/libclc/opencl/lib/generic/atomic/atomic_xchg.cl index 2b4bbf06d9400..8e0b87cc9343c 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_xchg.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_xchg.cl @@ -11,7 +11,7 @@ #define __CLC_IMPL(TYPE, AS) \ _CLC_OVERLOAD _CLC_DEF TYPE atomic_xchg(volatile AS TYPE *p, TYPE val) { \ - return __clc_atomic_exchange(p, val, __ATOMIC_RELAXED, \ + return __clc_atomic_exchange((AS TYPE *)p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } diff --git a/libclc/opencl/lib/generic/atomic/atomic_xor.cl b/libclc/opencl/lib/generic/atomic/atomic_xor.cl index 1f200c58edbff..46dc6c3e9111f 100644 --- a/libclc/opencl/lib/generic/atomic/atomic_xor.cl +++ b/libclc/opencl/lib/generic/atomic/atomic_xor.cl @@ -11,7 +11,7 @@ #define __CLC_IMPL(TYPE, AS) \ _CLC_OVERLOAD _CLC_DEF TYPE atomic_xor(volatile AS TYPE *p, TYPE val) { \ - return __clc_atomic_fetch_xor(p, val, __ATOMIC_RELAXED, \ + return __clc_atomic_fetch_xor((AS TYPE *)p, val, __ATOMIC_RELAXED, \ __MEMORY_SCOPE_DEVICE); \ } _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
