clc: disable 64bit atomic for 32bit spir and nvptx targets to fix warning 'large atomic operation may incur significant performance penalty'

wenju-he · wenju-he · commit c88bf9b9f014 · 2025-07-16T19:33:15.000-07:00
Also rename clc/lib/generic/atomic/atomic_def.inc to clc_atomic_def.inc
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_compare_exchange.inc b/libclc/clc/lib/generic/atomic/clc_atomic_compare_exchange.inc
@@ -7,7 +7,19 @@
 //===----------------------------------------------------------------------===//
 
 #ifdef __CLC_SCALAR
+
+#if defined(__SPIR32__) || defined(CLC_NVPTX)
+#if (defined(__CLC_FPSIZE) && __CLC_FPSIZE <= 32) ||                           \
+    (defined(__CLC_GENSIZE) && (__CLC_GENSIZE == 32))
+#define __CLC_HAS_ATOMIC
+#endif
+#else // defined(__SPIR32__) || defined(CLC_NVPTX)
 #if defined(__CLC_FPSIZE) || (__CLC_GENSIZE >= 32)
+#define __CLC_HAS_ATOMIC
+#endif
+#endif // defined(__SPIR32__) || defined(CLC_NVPTX)
+
+#ifdef __CLC_HAS_ATOMIC
 
 #ifdef __CLC_FPSIZE
 
@@ -46,5 +58,7 @@ __CLC_DEFINE_ATOMIC()
 
 #undef __CLC_DEFINE_ATOMIC
 
-#endif // defined(__CLC_FPSIZE) || (__CLC_GENSIZE >= 32)
+#endif // __CLC_HAS_ATOMIC
+#undef __CLC_HAS_ATOMIC
+
 #endif // __CLC_SCALAR
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_dec.cl b/libclc/clc/lib/generic/atomic/clc_atomic_dec.cl
@@ -12,5 +12,5 @@
 #define __IMPL_FUNCTION __scoped_atomic_fetch_add
 #define __CLC_INC_DEC
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/integer/gentype.inc>
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_def.inc b/libclc/clc/lib/generic/atomic/clc_atomic_def.inc
@@ -7,7 +7,19 @@
 //===----------------------------------------------------------------------===//
 
 #ifdef __CLC_SCALAR
+
+#if defined(__SPIR32__) || defined(CLC_NVPTX)
+#if (defined(__CLC_FPSIZE) && __CLC_FPSIZE <= 32) ||                           \
+    (defined(__CLC_GENSIZE) && (__CLC_GENSIZE == 32))
+#define __CLC_HAS_ATOMIC
+#endif
+#else // defined(__SPIR32__) || defined(CLC_NVPTX)
 #if defined(__CLC_FPSIZE) || (__CLC_GENSIZE >= 32)
+#define __CLC_HAS_ATOMIC
+#endif
+#endif // defined(__SPIR32__) || defined(CLC_NVPTX)
+
+#ifdef __CLC_HAS_ATOMIC
 
 #ifndef __CLC_PTR_CASTTYPE
 #define __CLC_PTR_CASTTYPE __CLC_GENTYPE
@@ -61,5 +73,7 @@ __CLC_DEFINE_ATOMIC()
 
 #undef __CLC_DEFINE_ATOMIC
 
-#endif // defined(__CLC_FPSIZE) || (__CLC_GENSIZE >= 32)
+#endif // __CLC_HAS_ATOMIC
+#undef __CLC_HAS_ATOMIC
+
 #endif // __CLC_SCALAR
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_exchange.cl b/libclc/clc/lib/generic/atomic/clc_atomic_exchange.cl
@@ -11,13 +11,13 @@
 #define FUNCTION __clc_atomic_exchange
 #define __IMPL_FUNCTION __scoped_atomic_exchange_n
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/integer/gentype.inc>
 
 #undef __CLC_PTR_CASTTYPE
 #undef __CLC_AS_RETTYPE
 #define __CLC_PTR_CASTTYPE __CLC_BIT_INTN
 #define __CLC_AS_RETTYPE(x) __CLC_AS_GENTYPE(x)
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/math/gentype.inc>
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_fetch_add.cl b/libclc/clc/lib/generic/atomic/clc_atomic_fetch_add.cl
@@ -11,8 +11,8 @@
 #define FUNCTION __clc_atomic_fetch_add
 #define __IMPL_FUNCTION __scoped_atomic_fetch_add
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/integer/gentype.inc>
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/math/gentype.inc>
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_fetch_and.cl b/libclc/clc/lib/generic/atomic/clc_atomic_fetch_and.cl
@@ -11,5 +11,5 @@
 #define FUNCTION __clc_atomic_fetch_and
 #define __IMPL_FUNCTION __scoped_atomic_fetch_and
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/integer/gentype.inc>
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_fetch_max.cl b/libclc/clc/lib/generic/atomic/clc_atomic_fetch_max.cl
@@ -11,8 +11,8 @@
 #define FUNCTION __clc_atomic_fetch_max
 #define __IMPL_FUNCTION __scoped_atomic_fetch_max
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/integer/gentype.inc>
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/math/gentype.inc>
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_fetch_min.cl b/libclc/clc/lib/generic/atomic/clc_atomic_fetch_min.cl
@@ -11,8 +11,8 @@
 #define FUNCTION __clc_atomic_fetch_min
 #define __IMPL_FUNCTION __scoped_atomic_fetch_min
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/integer/gentype.inc>
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/math/gentype.inc>
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_fetch_or.cl b/libclc/clc/lib/generic/atomic/clc_atomic_fetch_or.cl
@@ -11,5 +11,5 @@
 #define FUNCTION __clc_atomic_fetch_or
 #define __IMPL_FUNCTION __scoped_atomic_fetch_or
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/integer/gentype.inc>
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_fetch_sub.cl b/libclc/clc/lib/generic/atomic/clc_atomic_fetch_sub.cl
@@ -11,8 +11,8 @@
 #define FUNCTION __clc_atomic_fetch_sub
 #define __IMPL_FUNCTION __scoped_atomic_fetch_sub
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/integer/gentype.inc>
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/math/gentype.inc>
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_fetch_xor.cl b/libclc/clc/lib/generic/atomic/clc_atomic_fetch_xor.cl
@@ -11,5 +11,5 @@
 #define FUNCTION __clc_atomic_fetch_xor
 #define __IMPL_FUNCTION __scoped_atomic_fetch_xor
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/integer/gentype.inc>
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_inc.cl b/libclc/clc/lib/generic/atomic/clc_atomic_inc.cl
@@ -12,5 +12,5 @@
 #define __IMPL_FUNCTION __scoped_atomic_fetch_sub
 #define __CLC_INC_DEC
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/integer/gentype.inc>
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_load.cl b/libclc/clc/lib/generic/atomic/clc_atomic_load.cl
@@ -12,13 +12,13 @@
 #define __IMPL_FUNCTION __scoped_atomic_load_n
 #define __CLC_NO_VALUE_ARG
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/integer/gentype.inc>
 
 #undef __CLC_PTR_CASTTYPE
 #undef __CLC_AS_RETTYPE
 #define __CLC_PTR_CASTTYPE __CLC_BIT_INTN
 #define __CLC_AS_RETTYPE(x) __CLC_AS_GENTYPE(x)
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/math/gentype.inc>
diff --git a/libclc/clc/lib/generic/atomic/clc_atomic_store.cl b/libclc/clc/lib/generic/atomic/clc_atomic_store.cl
@@ -12,11 +12,11 @@
 #define __IMPL_FUNCTION __scoped_atomic_store_n
 #define __CLC_RETURN_VOID
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/integer/gentype.inc>
 
 #undef __CLC_PTR_CASTTYPE
 #define __CLC_PTR_CASTTYPE __CLC_BIT_INTN
 
-#define __CLC_BODY <atomic_def.inc>
+#define __CLC_BODY <clc_atomic_def.inc>
 #include <clc/math/gentype.inc>
diff --git a/libclc/opencl/include/clc/opencl/atomic/atomic_decl.inc b/libclc/opencl/include/clc/opencl/atomic/atomic_decl.inc
@@ -8,18 +8,16 @@
 
 #ifdef __CLC_SCALAR
 
-#if (defined(cl_khr_int64_base_atomics) &&                                     \
-     defined(cl_khr_int64_extended_atomics))
+#if defined(__opencl_c_fp64) && (defined(cl_khr_int64_base_atomics) &&         \
+                                 defined(cl_khr_int64_extended_atomics))
 #define HAVE_64_ATOMIC
 #endif
-#if defined(__CLC_FPSIZE) && (__CLC_FPSIZE < 64 || (defined(HAVE_64_ATOMIC) && \
-                                                    defined(__opencl_c_fp64)))
+#if defined(__CLC_FPSIZE) && (__CLC_FPSIZE < 64 || defined(HAVE_64_ATOMIC))
 #define HAVE_FP_ATOMIC
 #endif
 #if defined(__CLC_GENSIZE) &&                                                  \
     ((__CLC_GENSIZE == 32) ||                                                  \
-     (__CLC_GENSIZE == 64 &&                                                   \
-      (defined(HAVE_64_ATOMIC) && defined(__opencl_c_int64))))
+     (__CLC_GENSIZE == 64 && defined(HAVE_64_ATOMIC)))
 #define HAVE_INT_ATOMIC
 #endif
 #if defined(HAVE_FP_ATOMIC) || defined(HAVE_INT_ATOMIC)
diff --git a/libclc/opencl/lib/generic/atomic/atomic_def.inc b/libclc/opencl/lib/generic/atomic/atomic_def.inc
@@ -8,18 +8,16 @@
 
 #ifdef __CLC_SCALAR
 
-#if (defined(cl_khr_int64_base_atomics) &&                                     \
-     defined(cl_khr_int64_extended_atomics))
+#if defined(__opencl_c_fp64) && (defined(cl_khr_int64_base_atomics) &&         \
+                                 defined(cl_khr_int64_extended_atomics))
 #define HAVE_64_ATOMIC
 #endif
-#if defined(__CLC_FPSIZE) && (__CLC_FPSIZE < 64 || (defined(HAVE_64_ATOMIC) && \
-                                                    defined(__opencl_c_fp64)))
+#if defined(__CLC_FPSIZE) && (__CLC_FPSIZE < 64 || defined(HAVE_64_ATOMIC)
 #define HAVE_FP_ATOMIC
 #endif
 #if defined(__CLC_GENSIZE) &&                                                  \
     ((__CLC_GENSIZE == 32) ||                                                  \
-     (__CLC_GENSIZE == 64 &&                                                   \
-      (defined(HAVE_64_ATOMIC) && defined(__opencl_c_int64))))
+     (__CLC_GENSIZE == 64 && defined(HAVE_64_ATOMIC)))
 #define HAVE_INT_ATOMIC
 #endif
 #if defined(HAVE_FP_ATOMIC) || defined(HAVE_INT_ATOMIC)
