[OpenCL] Support cl_ext_float_atomics

[haonanya]

Summary

This aims to support cl_ext_float_atomics on CFE, see https://github.com/KhronosGroup/OpenCL-Docs/pull/552 for details.

[Anastasia]

The extension spec seems to also mention atomic_half. Are planning to add it too?

clang/include/clang/Basic/OpenCLExtensions.def

85 ↗

(On Diff #360018)

If the only purpose of adding this extension here is to define the macros you should just use the internal header for it. See guidelines: https://clang.llvm.org/docs/OpenCLSupport.html#implementation-guidelines

The same applies to all the feature macros below.

You could check how new subgroup extensions are implemented: https://clang.llvm.org/doxygen/opencl-c-base_8h_source.html

// For SPIR all extensions are supported.
#if defined(__SPIR__)
#define cl_khr_subgroup_extended_types 1
#define cl_khr_subgroup_non_uniform_vote 1
#define cl_khr_subgroup_ballot 1
#define cl_khr_subgroup_non_uniform_arithmetic 1
#define cl_khr_subgroup_shuffle 1
#define cl_khr_subgroup_shuffle_relative 1
#define cl_khr_subgroup_clustered_reduce 1
#endif // defined(__SPIR__)

clang/lib/Headers/opencl-c-base.h

40

From the spec it feels like those macros should be defined conditionally?

clang/lib/Headers/opencl-c.h

13419

should this not be conditioned on __opencl_c_ext_fp32_global_atomic? Otherwise, I am missing what is the intent of those macros...

[Anastasia]

Please consider uploading full diff: https://llvm.org/docs/Phabricator.html#requesting-a-review-via-the-web-interface

[Anastasia]

FYI this revision has not been added to cfe-commits, is this intensional?

[haonanya]

Add extension and all the feature macros to internal header.

[haonanya]

In D106343#2892960, @Anastasia wrote:

FYI this revision has not been added to cfe-commits, is this intensional?

Hi, Anastasia. I am not very familiar with the process, could you please help to add to cfe-commits if possible? Thanks very much.

[Anastasia]

Just to make sure you are aware Clang doesn't use this header by default, so the upstream users won't be able to call those functions unless you add them into OpenCLBuiltins.td:
https://clang.llvm.org/docs/OpenCLSupport.html#opencl-builtins

This header is only accessible via the frontend options: https://clang.llvm.org/docs/OpenCLSupport.html#cmdoption-finclude-default-header

clang/lib/Headers/opencl-c.h

13657

Can you annotate the #endifs with a comment describing what they correspond to. i.e. something like:

#endif //defined(__opencl_c_ext_fp32_global_atomic_min_max)

[svenvh]

clang/lib/Headers/opencl-c-base.h

24

Should this be defined as 1?

Should this define be tested in clang/test/Headers/opencl-c-header.cl too?

[Anastasia]

clang/lib/Headers/opencl-c-base.h

24

Actually, now that I think more about this, it seems incorrect to add this here without adding the functions to OpenCLBuiltins.td because then the feature macro will be present without the feature when the default header is used?

So I guess we either need to extend OpenCLBuiltins.td with new functions or move the new macros into opencl-c.h?

[svenvh]

clang/lib/Headers/opencl-c-base.h

24

Good catch! Indeed, adding the define in the shared opencl-c-base.h without also providing the builtins through OpenCLBuiltins.td is incorrect.

The preferred solution would be to add the new builtins to clang/lib/Sema/OpenCLBuiltins.td too, to avoid diverging the header and tablegen-driven code paths.

[haonanya]

Add the new builtins to clang/lib/Sema/OpenCLBuiltins.td.

[haonanya]

Hi, Anastasia and svenvh.
Sorry for late reply. I have updated patch per your comments.
Many thanks for your comments.

[svenvh]

Thanks for the update! I have a few points to improve the patch.

clang/lib/Sema/OpenCLBuiltins.td

1117

Do we really need to guard these additions behind OpenCL 3.0? The spec mentions

The functionality added by this extension uses the OpenCL C 2.0 atomic syntax and hence requires OpenCL 2.0 or newer.

(same applies to the opencl-h.c changes of course)

1122

The feature macros seem to be missing. See FuncExtOpenCLCPipes for an example how to do that.

1142–1145

This can be merged into the preceeding foreach parts I think?

clang/test/SemaOpenCL/fdeclare-opencl-builtins.cl

137

As mentioned in the comment on lines 13-17, this test is not meant to be exhaustive. So you don't have to test every overload, checking one or two builtins should suffice.

[haonanya]

Remove OpenCL3.0 macro guards on opencl-c.h and OpenCLBuiltins.td.
Add missing feature macro.
simplify test.

[svenvh]

clang/lib/Sema/OpenCLBuiltins.td

1118

So now all of those builtins are guarded by cl_ext_float_atomics, which is good, but not by any of the __opencl_c_ext_... macros yet.

To guard by multiple macros, we'd need to do something like:

def FuncExtFloatAtomicsFp32GlobalMinMax  : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp32_global_atomic_min_max">;
def FuncExtFloatAtomicsFp32LocalMinMax   : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp32_local_atomic_min_max">;

And then use let Extension = FuncExtFloatAtomics... around the corresponding builtins. You shouldn't have to change the loop structure much for this, as you can hopefully use # concatenation to construct the appropriate FuncExt name (and then !cast it to a record).

However, I do see some problematic cases: the generic address space builtins are enabled by one of multiple feature macros, which is something that is currently not supported by the OpenCLBuiltins.td handling. If it's not too late, could we ask the extension spec editors to provide a dedicated feature macro for generic perhaps?

[haonanya]

Let builtins are guarded by related macro

[svenvh]

clang/lib/Sema/OpenCLBuiltins.td

1120

Please try to follow the formatting used in the rest of this file:

def : Builtin<...

So a space after def, then no newline after the :.

This applies to all the new defs below too.

1121

The paste operator # is a binary operator, so it makes more sense to put a space on both sides.

1197

Wrong extension guard.

1293

Wrong extension guard.

[haonanya]

Unify formatting and fix some errors on OpenCLBuiltins.td

[svenvh]

Thanks for the update! I have a comment about indentation, other than that this is looking good to me.

clang/lib/Sema/OpenCLBuiltins.td

1128

Please indent the content inside all let blocks.

[haonanya]

Fix formatting issues

[svenvh]

LGTM, thanks!

[haonanya]

In D106343#2892946, @Anastasia wrote:

The extension spec seems to also mention atomic_half. Are planning to add it too?

Hi, Anastasia. I also have some work to translate these atomic builtins on SPIRV currently, so I'd like to have another patch to add atomic_half later.
Do you have any comments?
Thanks very much.

[haonanya]

Hi, svenvh.
Should we use cl_khr_int64_base_atomics and cl_khr_int64_extended_atomics to guard the functions using atomic_double type?
Thanks very much.

#if defined(__opencl_c_ext_fp64_local_atomic_min_max)
double __ovld atomic_fetch_min(volatile __local atomic_double *object,
                               double operand);
#endif

[haonanya]

Hi, svenvh and Anastasia. If you approve the patch, could you please submit it?
I don't have permission to do it.

[Anastasia]

In D106343#2974055, @haonanya wrote:

Hi, svenvh and Anastasia. If you approve the patch, could you please submit it?
I don't have permission to do it.

Sure, I can commit it on your behalf, atomic_half can be added separately. Thanks!

[haonanya]

In D106343#2967089, @haonanya wrote:

Hi, svenvh.
Should we use cl_khr_int64_base_atomics and cl_khr_int64_extended_atomics to guard the functions using atomic_double type?
Thanks very much.

#if defined(__opencl_c_ext_fp64_local_atomic_min_max)
double __ovld atomic_fetch_min(volatile __local atomic_double *object,
                               double operand);
#endif

Hi, svenvh and Anastasia. Do you have any comments for adding cl_khr_int64_base_atomics and cl_khr_int64_extended_atomics to guard atomic_double type? I'd appreciate it if you have time to answer it.
And if there is no any comment, please commit the patch.
Thanks very much.

[haonanya]

Kindly ping

[svenvh]

Apologies for the delayed response.

In D106343#2967089, @haonanya wrote:

Hi, svenvh.
Should we use cl_khr_int64_base_atomics and cl_khr_int64_extended_atomics to guard the functions using atomic_double type?
Thanks very much.

#if defined(__opencl_c_ext_fp64_local_atomic_min_max)
double __ovld atomic_fetch_min(volatile __local atomic_double *object,
                               double operand);
#endif

This is perhaps something to raise at the specification level?

We can adjust the guards after any followup discussion if needed. To progress the support of this extension, I've just committed your patch (with some minor whitespace fixes).

[haonanya]

Hi, svenvh.
I am ok with the patch. Thanks very much.
I'd appreciate it if you help review https://github.com/KhronosGroup/SPIRV-LLVM-Translator/pull/1116 as well.

clang/lib/Headers/opencl-c-base.h

 #if defined(__SPIR__)
 #define cl_khr_subgroup_extended_types 1
 #define cl_khr_subgroup_non_uniform_vote 1
 #define cl_khr_subgroup_ballot 1
 #define cl_khr_subgroup_non_uniform_arithmetic 1
 #define cl_khr_subgroup_shuffle 1
 #define cl_khr_subgroup_shuffle_relative 1
 #define cl_khr_subgroup_clustered_reduce 1
 #define cl_khr_extended_bit_ops 1
 #define cl_khr_integer_dot_product 1
 #define __opencl_c_integer_dot_product_input_4x8bit 1
 #define __opencl_c_integer_dot_product_input_4x8bit_packed 1
+#define cl_ext_float_atomics 1
+#ifdef cl_khr_fp16
+#define __opencl_c_ext_fp16_global_atomic_load_store 1
+#define __opencl_c_ext_fp16_local_atomic_load_store 1
+#define __opencl_c_ext_fp16_global_atomic_add 1
+#define __opencl_c_ext_fp16_local_atomic_add 1
+#define __opencl_c_ext_fp16_global_atomic_min_max 1
+#define __opencl_c_ext_fp16_local_atomic_min_max 1
+#endif
+#ifdef cl_khr_fp64
+#define __opencl_c_ext_fp64_global_atomic_add 1
+#define __opencl_c_ext_fp64_local_atomic_add 1
+#define __opencl_c_ext_fp64_global_atomic_min_max 1
+#define __opencl_c_ext_fp64_local_atomic_min_max 1
+#endif
+#define __opencl_c_ext_fp32_global_atomic_add 1
+#define __opencl_c_ext_fp32_local_atomic_add 1
+#define __opencl_c_ext_fp32_global_atomic_min_max 1
+#define __opencl_c_ext_fp32_local_atomic_min_max 1
 
 #endif // defined(__SPIR__)
 #endif // (defined(__OPENCL_CPP_VERSION__) || __OPENCL_C_VERSION__ >= 200)
 
 // Define feature macros for OpenCL C 2.0
 #if (__OPENCL_CPP_VERSION__ == 100 || __OPENCL_C_VERSION__ == 200)
 #define __opencl_c_pipes 1
 #define __opencl_c_generic_address_space 1

clang/lib/Headers/opencl-c.h

 uint __ovld atomic_fetch_or_explicit(volatile atomic_uint *object, uint operand, memory_order order);
 int __ovld atomic_fetch_xor_explicit(volatile atomic_int *object, int operand, memory_order order);
 uint __ovld atomic_fetch_xor_explicit(volatile atomic_uint *object, uint operand, memory_order order);
 int __ovld atomic_fetch_and_explicit(volatile atomic_int *object, int operand, memory_order order);
 uint __ovld atomic_fetch_and_explicit(volatile atomic_uint *object, uint operand, memory_order order);
 int __ovld atomic_fetch_min_explicit(volatile atomic_int *object, int operand, memory_order order);
 uint __ovld atomic_fetch_min_explicit(volatile atomic_uint *object, uint operand, memory_order order);
 int __ovld atomic_fetch_max_explicit(volatile atomic_int *object, int operand, memory_order order);
 uint __ovld atomic_fetch_max_explicit(volatile atomic_uint *object, uint operand, memory_order order);
 #if defined(cl_khr_int64_base_atomics) && defined(cl_khr_int64_extended_atomics)
 long __ovld atomic_fetch_add_explicit(volatile atomic_long *object, long operand, memory_order order);
 ulong __ovld atomic_fetch_add_explicit(volatile atomic_ulong *object, ulong operand, memory_order order);
 long __ovld atomic_fetch_sub_explicit(volatile atomic_long *object, long operand, memory_order order);
 ulong __ovld atomic_fetch_sub_explicit(volatile atomic_ulong *object, ulong operand, memory_order order);
 long __ovld atomic_fetch_or_explicit(volatile atomic_long *object, long operand, memory_order order);
 ulong __ovld atomic_fetch_or_explicit(volatile atomic_ulong *object, ulong operand, memory_order order);
 long __ovld atomic_fetch_xor_explicit(volatile atomic_long *object, long operand, memory_order order);

 long __ovld atomic_fetch_max_explicit(volatile __local atomic_long *object, long operand, memory_order order, memory_scope scope);
 ulong __ovld atomic_fetch_max_explicit(volatile __global atomic_ulong *object, ulong operand, memory_order order, memory_scope scope);
 ulong __ovld atomic_fetch_max_explicit(volatile __local atomic_ulong *object, ulong operand, memory_order order, memory_scope scope);
 uintptr_t __ovld atomic_fetch_add_explicit(volatile __global atomic_uintptr_t *object, ptrdiff_t operand, memory_order order, memory_scope scope);
 uintptr_t __ovld atomic_fetch_sub_explicit(volatile __local atomic_uintptr_t *object, ptrdiff_t operand, memory_order order, memory_scope scope);
 #endif //defined(cl_khr_int64_base_atomics) && defined(cl_khr_int64_extended_atomics)
 #endif //__OPENCL_C_VERSION__ >= CL_VERSION_3_0
 
+// The functionality added by cl_ext_float_atomics extension
+#if defined(cl_ext_float_atomics)
+
+#if defined(__opencl_c_ext_fp32_global_atomic_min_max)
+float __ovld atomic_fetch_min(volatile __global atomic_float *object,
+                              float operand);
+float __ovld atomic_fetch_max(volatile __global atomic_float *object,
+                              float operand);
+float __ovld atomic_fetch_min_explicit(volatile __global atomic_float *object,
+                                       float operand, memory_order order);
+float __ovld atomic_fetch_max_explicit(volatile __global atomic_float *object,
+                                       float operand, memory_order order);
+float __ovld atomic_fetch_min_explicit(volatile __global atomic_float *object,
+                                       float operand, memory_order order,
+                                       memory_scope scope);
+float __ovld atomic_fetch_max_explicit(volatile __global atomic_float *object,
+                                       float operand, memory_order order,
+                                       memory_scope scope);
+#endif // defined(__opencl_c_ext_fp32_global_atomic_min_max)
+
+#if defined(__opencl_c_ext_fp32_local_atomic_min_max)
+float __ovld atomic_fetch_min(volatile __local atomic_float *object,
+                              float operand);
+float __ovld atomic_fetch_max(volatile __local atomic_float *object,
+                              float operand);
+float __ovld atomic_fetch_min_explicit(volatile __local atomic_float *object,
+                                       float operand, memory_order order);
+float __ovld atomic_fetch_max_explicit(volatile __local atomic_float *object,
+                                       float operand, memory_order order);
+float __ovld atomic_fetch_min_explicit(volatile __local atomic_float *object,
+                                       float operand, memory_order order,
+                                       memory_scope scope);
+float __ovld atomic_fetch_max_explicit(volatile __local atomic_float *object,
+                                       float operand, memory_order order,
+                                       memory_scope scope);
+#endif // defined(__opencl_c_ext_fp32_local_atomic_min_max)
+
+#if defined(__opencl_c_ext_fp32_global_atomic_min_max) &&                      \
+    defined(__opencl_c_ext_fp32_local_atomic_min_max)
+float __ovld atomic_fetch_min(volatile atomic_float *object, float operand);
+float __ovld atomic_fetch_max(volatile atomic_float *object, float operand);
+float __ovld atomic_fetch_min_explicit(volatile atomic_float *object,
+                                       float operand, memory_order order);
+float __ovld atomic_fetch_max_explicit(volatile atomic_float *object,
+                                       float operand, memory_order order);
+float __ovld atomic_fetch_min_explicit(volatile atomic_float *object,
+                                       float operand, memory_order order,
+                                       memory_scope scope);
+float __ovld atomic_fetch_max_explicit(volatile atomic_float *object,
+                                       float operand, memory_order order,
+                                       memory_scope scope);
+#endif // defined(__opencl_c_ext_fp32_global_atomic_min_max) &&                \
+    defined(__opencl_c_ext_fp32_local_atomic_min_max)
+
+#if defined(__opencl_c_ext_fp64_global_atomic_min_max)
+double __ovld atomic_fetch_min(volatile __global atomic_double *object,
+                               double operand);
+double __ovld atomic_fetch_max(volatile __global atomic_double *object,
+                               double operand);
+double __ovld atomic_fetch_min_explicit(volatile __global atomic_double *object,
+                                        double operand, memory_order order);
+double __ovld atomic_fetch_max_explicit(volatile __global atomic_double *object,
+                                        double operand, memory_order order);
+double __ovld atomic_fetch_min_explicit(volatile __global atomic_double *object,
+                                        double operand, memory_order order,
+                                        memory_scope scope);
+double __ovld atomic_fetch_max_explicit(volatile __global atomic_double *object,
+                                        double operand, memory_order order,
+                                        memory_scope scope);
+#endif // defined(__opencl_c_ext_fp64_global_atomic_min_max)
+
+#if defined(__opencl_c_ext_fp64_local_atomic_min_max)
+double __ovld atomic_fetch_min(volatile __local atomic_double *object,
+                               double operand);
+double __ovld atomic_fetch_max(volatile __local atomic_double *object,
+                               double operand);
+double __ovld atomic_fetch_min_explicit(volatile __local atomic_double *object,
+                                        double operand, memory_order order);
+double __ovld atomic_fetch_max_explicit(volatile __local atomic_double *object,
+                                        double operand, memory_order order);
+double __ovld atomic_fetch_min_explicit(volatile __local atomic_double *object,
+                                        double operand, memory_order order,
+                                        memory_scope scope);
+double __ovld atomic_fetch_max_explicit(volatile __local atomic_double *object,
+                                        double operand, memory_order order,
+                                        memory_scope scope);
+#endif // defined(__opencl_c_ext_fp64_local_atomic_min_max)
+
+#if defined(__opencl_c_ext_fp64_global_atomic_min_max) &&                      \
+    defined(__opencl_c_ext_fp64_local_atomic_min_max)
+double __ovld atomic_fetch_min(volatile atomic_double *object, double operand);
+double __ovld atomic_fetch_max(volatile atomic_double *object, double operand);
+double __ovld atomic_fetch_min_explicit(volatile atomic_double *object,
+                                        double operand, memory_order order);
+double __ovld atomic_fetch_max_explicit(volatile atomic_double *object,
+                                        double operand, memory_order order);
+double __ovld atomic_fetch_min_explicit(volatile atomic_double *object,
+                                        double operand, memory_order order,
+                                        memory_scope scope);
+double __ovld atomic_fetch_max_explicit(volatile atomic_double *object,
+                                        double operand, memory_order order,
+                                        memory_scope scope);
+#endif // defined(__opencl_c_ext_fp64_global_atomic_min_max) &&                \
+    defined(__opencl_c_ext_fp64_local_atomic_min_max)
+
+#if defined(__opencl_c_ext_fp32_global_atomic_add)
+float __ovld atomic_fetch_add(volatile __global atomic_float *object,
+                              float operand);
+float __ovld atomic_fetch_sub(volatile __global atomic_float *object,
+                              float operand);
+float __ovld atomic_fetch_add_explicit(volatile __global atomic_float *object,
+                                       float operand, memory_order order);
+float __ovld atomic_fetch_sub_explicit(volatile __global atomic_float *object,
+                                       float operand, memory_order order);
+float __ovld atomic_fetch_add_explicit(volatile __global atomic_float *object,
+                                       float operand, memory_order order,
+                                       memory_scope scope);
+float __ovld atomic_fetch_sub_explicit(volatile __global atomic_float *object,
+                                       float operand, memory_order order,
+                                       memory_scope scope);
+#endif // defined(__opencl_c_ext_fp32_global_atomic_add)
+
+#if defined(__opencl_c_ext_fp32_local_atomic_add)
+float __ovld atomic_fetch_add(volatile __local atomic_float *object,
+                              float operand);
+float __ovld atomic_fetch_sub(volatile __local atomic_float *object,
+                              float operand);
+float __ovld atomic_fetch_add_explicit(volatile __local atomic_float *object,
+                                       float operand, memory_order order);
+float __ovld atomic_fetch_sub_explicit(volatile __local atomic_float *object,
+                                       float operand, memory_order order);
+float __ovld atomic_fetch_add_explicit(volatile __local atomic_float *object,
+                                       float operand, memory_order order,
+                                       memory_scope scope);
+float __ovld atomic_fetch_sub_explicit(volatile __local atomic_float *object,
+                                       float operand, memory_order order,
+                                       memory_scope scope);
+#endif // defined(__opencl_c_ext_fp32_local_atomic_add)
+
+#if defined(__opencl_c_ext_fp32_global_atomic_add) &&                          \
+    defined(__opencl_c_ext_fp32_local_atomic_add)
+float __ovld atomic_fetch_add(volatile atomic_float *object, float operand);
+float __ovld atomic_fetch_sub(volatile atomic_float *object, float operand);
+float __ovld atomic_fetch_add_explicit(volatile atomic_float *object,
+                                       float operand, memory_order order);
+float __ovld atomic_fetch_sub_explicit(volatile atomic_float *object,
+                                       float operand, memory_order order);
+float __ovld atomic_fetch_add_explicit(volatile atomic_float *object,
+                                       float operand, memory_order order,
+                                       memory_scope scope);
+float __ovld atomic_fetch_sub_explicit(volatile atomic_float *object,
+                                       float operand, memory_order order,
+                                       memory_scope scope);
+#endif // defined(__opencl_c_ext_fp32_global_atomic_add) &&                    \
+    defined(__opencl_c_ext_fp32_local_atomic_add)
+
+#if defined(__opencl_c_ext_fp64_global_atomic_add)
+double __ovld atomic_fetch_add(volatile __global atomic_double *object,
+                               double operand);
+double __ovld atomic_fetch_sub(volatile __global atomic_double *object,
+                               double operand);
+double __ovld atomic_fetch_add_explicit(volatile __global atomic_double *object,
+                                        double operand, memory_order order);
+double __ovld atomic_fetch_sub_explicit(volatile __global atomic_double *object,
+                                        double operand, memory_order order);
+double __ovld atomic_fetch_add_explicit(volatile __global atomic_double *object,
+                                        double operand, memory_order order,
+                                        memory_scope scope);
+double __ovld atomic_fetch_sub_explicit(volatile __global atomic_double *object,
+                                        double operand, memory_order order,
+                                        memory_scope scope);
+#endif // defined(__opencl_c_ext_fp64_global_atomic_add)
+
+#if defined(__opencl_c_ext_fp64_local_atomic_add)
+double __ovld atomic_fetch_add(volatile __local atomic_double *object,
+                               double operand);
+double __ovld atomic_fetch_sub(volatile __local atomic_double *object,
+                               double operand);
+double __ovld atomic_fetch_add_explicit(volatile __local atomic_double *object,
+                                        double operand, memory_order order);
+double __ovld atomic_fetch_sub_explicit(volatile __local atomic_double *object,
+                                        double operand, memory_order order);
+double __ovld atomic_fetch_add_explicit(volatile __local atomic_double *object,
+                                        double operand, memory_order order,
+                                        memory_scope scope);
+double __ovld atomic_fetch_sub_explicit(volatile __local atomic_double *object,
+                                        double operand, memory_order order,
+                                        memory_scope scope);
+#endif // defined(__opencl_c_ext_fp64_local_atomic_add)
+
+#if defined(__opencl_c_ext_fp64_global_atomic_add) &&                          \
+    defined(__opencl_c_ext_fp64_local_atomic_add)
+double __ovld atomic_fetch_add(volatile atomic_double *object, double operand);
+double __ovld atomic_fetch_sub(volatile atomic_double *object, double operand);
+double __ovld atomic_fetch_add_explicit(volatile atomic_double *object,
+                                        double operand, memory_order order);
+double __ovld atomic_fetch_sub_explicit(volatile atomic_double *object,
+                                        double operand, memory_order order);
+double __ovld atomic_fetch_add_explicit(volatile atomic_double *object,
+                                        double operand, memory_order order,
+                                        memory_scope scope);
+double __ovld atomic_fetch_sub_explicit(volatile atomic_double *object,
+                                        double operand, memory_order order,
+                                        memory_scope scope);
+#endif // defined(__opencl_c_ext_fp64_global_atomic_add) &&                    \
+    defined(__opencl_c_ext_fp64_local_atomic_add)
+
+#endif // cl_ext_float_atomics
+
 // atomic_store()
 
 #if defined(__opencl_c_atomic_order_seq_cst) && defined(__opencl_c_atomic_scope_device)
 #if defined(__opencl_c_generic_address_space)
 void __ovld atomic_store(volatile atomic_int *object, int desired);
 void __ovld atomic_store(volatile atomic_uint *object, uint desired);
 void __ovld atomic_store(volatile atomic_float *object, float desired);
 

clang/lib/Sema/OpenCLBuiltins.td

 def FuncExtKhrInt64BaseAtomics           : FunctionExtension<"cl_khr_int64_base_atomics">;
 def FuncExtKhrInt64ExtendedAtomics       : FunctionExtension<"cl_khr_int64_extended_atomics">;
 def FuncExtKhrMipmapImage                : FunctionExtension<"cl_khr_mipmap_image">;
 def FuncExtKhrMipmapImageWrites          : FunctionExtension<"cl_khr_mipmap_image_writes">;
 def FuncExtKhrGlMsaaSharing              : FunctionExtension<"cl_khr_gl_msaa_sharing">;
 
 def FuncExtOpenCLCPipes                  : FunctionExtension<"__opencl_c_pipes">;
 def FuncExtOpenCLCWGCollectiveFunctions  : FunctionExtension<"__opencl_c_work_group_collective_functions">;
+def FuncExtFloatAtomicsFp32GlobalAdd     : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp32_global_atomic_add">;
+def FuncExtFloatAtomicsFp64GlobalAdd     : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp64_global_atomic_add">;
+def FuncExtFloatAtomicsFp32LocalAdd      : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp32_local_atomic_add">;
+def FuncExtFloatAtomicsFp64LocalAdd      : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp64_local_atomic_add">;
+def FuncExtFloatAtomicsFp32GenericAdd    : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp32_local_atomic_add __opencl_c_ext_fp32_global_atomic_add">;
+def FuncExtFloatAtomicsFp64GenericAdd    : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp64_local_atomic_add __opencl_c_ext_fp64_global_atomic_add">;
+def FuncExtFloatAtomicsFp32GlobalMinMax  : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp32_global_atomic_min_max">;
+def FuncExtFloatAtomicsFp64GlobalMinMax  : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp64_global_atomic_min_max">;
+def FuncExtFloatAtomicsFp32LocalMinMax   : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp32_local_atomic_min_max">;
+def FuncExtFloatAtomicsFp64LocalMinMax   : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp64_local_atomic_min_max">;
+def FuncExtFloatAtomicsFp32GenericMinMax : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp32_local_atomic_min_max __opencl_c_ext_fp32_global_atomic_min_max">;
+def FuncExtFloatAtomicsFp64GenericMinMax : FunctionExtension<"cl_ext_float_atomics __opencl_c_ext_fp64_local_atomic_min_max __opencl_c_ext_fp64_global_atomic_min_max">;
 
 // Not a real extension, but a workaround to add C++ for OpenCL specific builtins.
 def FuncExtOpenCLCxx                     : FunctionExtension<"__cplusplus">;
 
 // Multiple extensions
 def FuncExtKhrMipmapWritesAndWrite3d     : FunctionExtension<"cl_khr_mipmap_image_writes cl_khr_3d_image_writes">;
 
 // Arm extensions.
let MinVersion = CL20 in {
   def : Builtin<"atomic_flag_test_and_set",
       [Bool, PointerType<VolatileType<AtomicFlag>, GenericAS>]>;
   def : Builtin<"atomic_flag_test_and_set_explicit",
       [Bool, PointerType<VolatileType<AtomicFlag>, GenericAS>, MemoryOrder]>;
   def : Builtin<"atomic_flag_test_and_set_explicit",
       [Bool, PointerType<VolatileType<AtomicFlag>, GenericAS>, MemoryOrder, MemoryScope]>;
 }
 
+// The functionality added by cl_ext_float_atomics extension
+let MinVersion = CL20 in {
+  foreach ModOp = ["add", "sub"] in {
+    let Extension = FuncExtFloatAtomicsFp32GlobalAdd in {
+      def : Builtin<"atomic_fetch_" # ModOp,
+          [Float, PointerType<VolatileType<AtomicFloat>, GlobalAS>, Float]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Float, PointerType<VolatileType<AtomicFloat>, GlobalAS>, Float, MemoryOrder]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Float, PointerType<VolatileType<AtomicFloat>, GlobalAS>, Float, MemoryOrder, MemoryScope]>;
+    }
+    let Extension = FuncExtFloatAtomicsFp64GlobalAdd in {
+      def : Builtin<"atomic_fetch_" # ModOp,
+          [Double, PointerType<VolatileType<AtomicDouble>, GlobalAS>, Double]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Double, PointerType<VolatileType<AtomicDouble>, GlobalAS>, Double, MemoryOrder]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Double, PointerType<VolatileType<AtomicDouble>, GlobalAS>, Double, MemoryOrder, MemoryScope]>;
+    }
+    let Extension = FuncExtFloatAtomicsFp32LocalAdd in {
+      def : Builtin<"atomic_fetch_" # ModOp,
+          [Float, PointerType<VolatileType<AtomicFloat>, LocalAS>, Float]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Float, PointerType<VolatileType<AtomicFloat>, LocalAS>, Float, MemoryOrder]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Float, PointerType<VolatileType<AtomicFloat>, LocalAS>, Float, MemoryOrder, MemoryScope]>;
+    }
+    let Extension = FuncExtFloatAtomicsFp64LocalAdd in {
+      def : Builtin<"atomic_fetch_" # ModOp,
+          [Double, PointerType<VolatileType<AtomicDouble>, LocalAS>, Double]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Double, PointerType<VolatileType<AtomicDouble>, LocalAS>, Double, MemoryOrder]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Double, PointerType<VolatileType<AtomicDouble>, LocalAS>, Double, MemoryOrder, MemoryScope]>;
+    }
+    let Extension = FuncExtFloatAtomicsFp32GenericAdd in {
+      def : Builtin<"atomic_fetch_" # ModOp,
+          [Float, PointerType<VolatileType<AtomicFloat>, GenericAS>, Float]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Float, PointerType<VolatileType<AtomicFloat>, GenericAS>, Float, MemoryOrder]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Float, PointerType<VolatileType<AtomicFloat>, GenericAS>, Float, MemoryOrder, MemoryScope]>;
+    }
+    let Extension = FuncExtFloatAtomicsFp64GenericAdd in {
+      def : Builtin<"atomic_fetch_" # ModOp,
+          [Double, PointerType<VolatileType<AtomicDouble>, GenericAS>, Double]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Double, PointerType<VolatileType<AtomicDouble>, GenericAS>, Double, MemoryOrder]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Double, PointerType<VolatileType<AtomicDouble>, GenericAS>, Double, MemoryOrder, MemoryScope]>;
+    }
+  }
+  foreach ModOp = ["min", "max"] in {
+    let Extension = FuncExtFloatAtomicsFp32GlobalMinMax in {
+      def : Builtin<"atomic_fetch_" # ModOp,
+          [Float, PointerType<VolatileType<AtomicFloat>, GlobalAS>, Float]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Float, PointerType<VolatileType<AtomicFloat>, GlobalAS>, Float, MemoryOrder]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Float, PointerType<VolatileType<AtomicFloat>, GlobalAS>, Float, MemoryOrder, MemoryScope]>;
+    }
+    let Extension = FuncExtFloatAtomicsFp64GlobalMinMax in {
+      def : Builtin<"atomic_fetch_" # ModOp,
+          [Double, PointerType<VolatileType<AtomicDouble>, GlobalAS>, Double]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Double, PointerType<VolatileType<AtomicDouble>, GlobalAS>, Double, MemoryOrder]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Double, PointerType<VolatileType<AtomicDouble>, GlobalAS>, Double, MemoryOrder, MemoryScope]>;
+    }
+    let Extension = FuncExtFloatAtomicsFp32LocalMinMax in {
+      def : Builtin<"atomic_fetch_" # ModOp,
+          [Float, PointerType<VolatileType<AtomicFloat>, LocalAS>, Float]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Float, PointerType<VolatileType<AtomicFloat>, LocalAS>, Float, MemoryOrder]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Float, PointerType<VolatileType<AtomicFloat>, LocalAS>, Float, MemoryOrder, MemoryScope]>;
+    }
+    let Extension = FuncExtFloatAtomicsFp64LocalMinMax in {
+      def : Builtin<"atomic_fetch_" # ModOp,
+          [Double, PointerType<VolatileType<AtomicDouble>, LocalAS>, Double]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Double, PointerType<VolatileType<AtomicDouble>, LocalAS>, Double, MemoryOrder]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Double, PointerType<VolatileType<AtomicDouble>, LocalAS>, Double, MemoryOrder, MemoryScope]>;
+    }
+    let Extension = FuncExtFloatAtomicsFp32GenericMinMax in {
+      def : Builtin<"atomic_fetch_" # ModOp,
+          [Float, PointerType<VolatileType<AtomicFloat>, GenericAS>, Float]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Float, PointerType<VolatileType<AtomicFloat>, GenericAS>, Float, MemoryOrder]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Float, PointerType<VolatileType<AtomicFloat>, GenericAS>, Float, MemoryOrder, MemoryScope]>;
+    }
+    let Extension = FuncExtFloatAtomicsFp64GenericMinMax in {
+      def : Builtin<"atomic_fetch_" # ModOp,
+          [Double, PointerType<VolatileType<AtomicDouble>, GenericAS>, Double]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Double, PointerType<VolatileType<AtomicDouble>, GenericAS>, Double, MemoryOrder]>;
+      def : Builtin<"atomic_fetch_" # ModOp # "_explicit",
+          [Double, PointerType<VolatileType<AtomicDouble>, GenericAS>, Double, MemoryOrder, MemoryScope]>;
+    }
+  }
+}
+
 //--------------------------------------------------------------------
 // OpenCL v1.1 s6.11.12, v1.2 s6.12.12, v2.0 s6.13.12 - Miscellaneous Vector Functions
 // --- Table 19 ---
 foreach VSize1 = [2, 4, 8, 16] in {
   foreach VSize2 = [2, 4, 8, 16] in {
     foreach VecAndMaskType = [[Char, UChar], [UChar, UChar],
                               [Short, UShort], [UShort, UShort],
                               [Int, UInt], [UInt, UInt],
let MinVersion = CL12 in {
   foreach aQual = ["RO", "RW"] in {
     foreach imgTy = [Image2d, Image1dArray] in {
       def : Builtin<"read_imagef", [VectorType<Float, 4>, ImageType<imgTy, aQual>, VectorType<Int, 2>], Attr.Pure>;
       def : Builtin<"read_imagei", [VectorType<Int, 4>, ImageType<imgTy, aQual>, VectorType<Int, 2>], Attr.Pure>;
       def : Builtin<"read_imageui", [VectorType<UInt, 4>, ImageType<imgTy, aQual>, VectorType<Int, 2>], Attr.Pure>;
     }
     foreach imgTy = [Image3d, Image2dArray] in {
       def : Builtin<"read_imagef", [VectorType<Float, 4>, ImageType<imgTy, aQual>, VectorType<Int, 4>], Attr.Pure>;
       def : Builtin<"read_imagei", [VectorType<Int, 4>, ImageType<imgTy, aQual>, VectorType<Int, 4>], Attr.Pure>;
       def : Builtin<"read_imageui", [VectorType<UInt, 4>, ImageType<imgTy, aQual>, VectorType<Int, 4>], Attr.Pure>;
     }
     foreach imgTy = [Image1d, Image1dBuffer] in {
       def : Builtin<"read_imagef", [VectorType<Float, 4>, ImageType<imgTy, aQual>, Int], Attr.Pure>;
       def : Builtin<"read_imagei", [VectorType<Int, 4>, ImageType<imgTy, aQual>, Int], Attr.Pure>;
       def : Builtin<"read_imageui", [VectorType<UInt, 4>, ImageType<imgTy, aQual>, Int], Attr.Pure>;
     }
     def : Builtin<"read_imagef", [Float, ImageType<Image2dDepth, aQual>, VectorType<Int, 2>], Attr.Pure>;

clang/test/Headers/opencl-c-header.cl

 #error "Incorrectly defined cl_khr_integer_dot_product"
 #endif
 #if __opencl_c_integer_dot_product_input_4x8bit != 1
 #error "Incorrectly defined __opencl_c_integer_dot_product_input_4x8bit"
 #endif
 #if __opencl_c_integer_dot_product_input_4x8bit_packed != 1
 #error "Incorrectly defined __opencl_c_integer_dot_product_input_4x8bit_packed"
 #endif
+#if cl_ext_float_atomics != 1
+#error "Incorrectly defined cl_ext_float_atomics"
+#endif
+#if __opencl_c_ext_fp16_global_atomic_load_store != 1
+#error "Incorrectly defined __opencl_c_ext_fp16_global_atomic_load_store"
+#endif
+#if __opencl_c_ext_fp16_local_atomic_load_store != 1
+#error "Incorrectly defined __opencl_c_ext_fp16_local_atomic_load_store"
+#endif
+#if __opencl_c_ext_fp16_global_atomic_add != 1
+#error "Incorrectly defined __opencl_c_ext_fp16_global_atomic_add"
+#endif
+#if __opencl_c_ext_fp32_global_atomic_add != 1
+#error "Incorrectly defined __opencl_c_ext_fp32_global_atomic_add"
+#endif
+#if __opencl_c_ext_fp64_global_atomic_add != 1
+#error "Incorrectly defined __opencl_c_ext_fp64_global_atomic_add"
+#endif
+#if __opencl_c_ext_fp16_local_atomic_add != 1
+#error "Incorrectly defined __opencl_c_ext_fp16_local_atomic_add"
+#endif
+#if __opencl_c_ext_fp32_local_atomic_add != 1
+#error "Incorrectly defined __opencl_c_ext_fp32_local_atomic_add"
+#endif
+#if __opencl_c_ext_fp64_local_atomic_add != 1
+#error "Incorrectly defined __opencl_c_ext_fp64_local_atomic_add"
+#endif
+#if __opencl_c_ext_fp16_global_atomic_min_max != 1
+#error "Incorrectly defined __opencl_c_ext_fp16_global_atomic_min_max"
+#endif
+#if __opencl_c_ext_fp32_global_atomic_min_max != 1
+#error "Incorrectly defined __opencl_c_ext_fp32_global_atomic_min_max"
+#endif
+#if __opencl_c_ext_fp64_global_atomic_min_max != 1
+#error "Incorrectly defined __opencl_c_ext_fp64_global_atomic_min_max"
+#endif
+#if __opencl_c_ext_fp16_local_atomic_min_max != 1
+#error "Incorrectly defined __opencl_c_ext_fp16_local_atomic_min_max"
+#endif
+#if __opencl_c_ext_fp32_local_atomic_min_max != 1
+#error "Incorrectly defined __opencl_c_ext_fp32_local_atomic_min_max"
+#endif
+#if __opencl_c_ext_fp64_local_atomic_min_max != 1
+#error "Incorrectly defined __opencl_c_ext_fp64_local_atomic_min_max"
+#endif
 
 #else
 
 #ifdef cl_khr_subgroup_extended_types
 #error "Incorrect cl_khr_subgroup_extended_types define"
 #endif
 #ifdef cl_khr_subgroup_non_uniform_vote
 #error "Incorrect cl_khr_subgroup_non_uniform_vote define"

 #error "Incorrect cl_khr_integer_dot_product define"
 #endif
 #ifdef __opencl_c_integer_dot_product_input_4x8bit
 #error "Incorrect __opencl_c_integer_dot_product_input_4x8bit define"
 #endif
 #ifdef __opencl_c_integer_dot_product_input_4x8bit_packed
 #error "Incorrect __opencl_c_integer_dot_product_input_4x8bit_packed define"
 #endif
+#ifdef cl_ext_float_atomics
+#error "Incorrect cl_ext_float_atomics define"
+#endif
+#ifdef __opencl_c_ext_fp16_global_atomic_load_store
+#error "Incorrectly __opencl_c_ext_fp16_global_atomic_load_store defined"
+#endif
+#ifdef __opencl_c_ext_fp16_local_atomic_load_store
+#error "Incorrectly __opencl_c_ext_fp16_local_atomic_load_store defined"
+#endif
+#ifdef __opencl_c_ext_fp16_global_atomic_add
+#error "Incorrectly __opencl_c_ext_fp16_global_atomic_add defined"
+#endif
+#ifdef __opencl_c_ext_fp32_global_atomic_add
+#error "Incorrectly __opencl_c_ext_fp32_global_atomic_add defined"
+#endif
+#ifdef __opencl_c_ext_fp64_global_atomic_add
+#error "Incorrectly __opencl_c_ext_fp64_global_atomic_add defined"
+#endif
+#ifdef __opencl_c_ext_fp16_local_atomic_add
+#error "Incorrectly __opencl_c_ext_fp16_local_atomic_add defined"
+#endif
+#ifdef __opencl_c_ext_fp32_local_atomic_add
+#error "Incorrectly __opencl_c_ext_fp32_local_atomic_add defined"
+#endif
+#ifdef __opencl_c_ext_fp64_local_atomic_add
+#error "Incorrectly __opencl_c_ext_fp64_local_atomic_add defined"
+#endif
+#ifdef __opencl_c_ext_fp16_global_atomic_min_max
+#error "Incorrectly __opencl_c_ext_fp16_global_atomic_min_max defined"
+#endif
+#ifdef __opencl_c_ext_fp32_global_atomic_min_max
+#error "Incorrectly __opencl_c_ext_fp32_global_atomic_min_max defined"
+#endif
+#ifdef __opencl_c_ext_fp64_global_atomic_min_max
+#error "Incorrectly __opencl_c_ext_fp64_global_atomic_min_max defined"
+#endif
+#ifdef __opencl_c_ext_fp16_local_atomic_min_max
+#error "Incorrectly __opencl_c_ext_fp16_local_atomic_min_max defined"
+#endif
+#ifdef __opencl_c_ext_fp32_local_atomic_min_max
+#error "Incorrectly __opencl_c_ext_fp32_local_atomic_min_max defined"
+#endif
+#ifdef __opencl_c_ext_fp64_local_atomic_min_max
+#error "Incorrectly __opencl_c_ext_fp64_local_atomic_min_max defined"
+#endif
 
 #endif //(defined(__OPENCL_CPP_VERSION__) || __OPENCL_C_VERSION__ >= 200)
 
 // OpenCL C features.
 #if (__OPENCL_CPP_VERSION__ == 202100 || __OPENCL_C_VERSION__ == 300)
 
 #if __opencl_c_atomic_scope_all_devices != 1
 #error "Incorrectly defined feature macro __opencl_c_atomic_scope_all_devices"

clang/test/SemaOpenCL/fdeclare-opencl-builtins.cl

void test_atomic_fetch(volatile __generic atomic_int *a_int,
   ip = atomic_fetch_add(a_intptr, ptrdiff);
   uip = atomic_fetch_add(a_uintptr, ptrdiff);
 
   ip = atomic_fetch_or(a_intptr, ip);
   uip = atomic_fetch_or(a_uintptr, uip);
 }
 #endif
 
+#if !defined(NO_HEADER) && !defined(NO_FP64) && __OPENCL_C_VERSION__ >= 200
+// Check added atomic_fetch_ functions by cl_ext_float_atomics
+// extension can be called
+void test_atomic_fetch_with_address_space(volatile __generic atomic_float *a_float,
+                                          volatile __generic atomic_double *a_double,
+                                          volatile __local atomic_float *a_float_local,
+                                          volatile __local atomic_double *a_double_local,
+                                          volatile __global atomic_float *a_float_global,
+                                          volatile __global atomic_double *a_double_global) {
+  float f1, resf1;
+  double d1, resd1;
+  resf1 = atomic_fetch_min(a_float, f1);
+  resf1 = atomic_fetch_max_explicit(a_float_local, f1, memory_order_seq_cst);
+  resf1 = atomic_fetch_add_explicit(a_float_global, f1, memory_order_seq_cst, memory_scope_work_group);
+
+  resd1 = atomic_fetch_min(a_double, d1);
+  resd1 = atomic_fetch_max_explicit(a_double_local, d1, memory_order_seq_cst);
+  resd1 = atomic_fetch_add_explicit(a_double_global, d1, memory_order_seq_cst, memory_scope_work_group);
+}
+#endif // !defined(NO_HEADER) && __OPENCL_C_VERSION__ >= 200
+
 // Test old atomic overloaded with generic address space in C++ for OpenCL.
 #if __OPENCL_C_VERSION__ >= 200
 void test_legacy_atomics_cpp(__generic volatile unsigned int *a) {
   atomic_add(a, 1);
 #if !defined(__cplusplus)
   // expected-error@-2{{no matching function for call to 'atomic_add'}}
   // expected-note@-3 4 {{candidate function not viable}}
 #endif