[Clang] Add support for missing OpenCL extensions

[vmustya]

This patch adds handling for the following OpenCL extensions, passed as
-cl-ext arguments to Clang:

• cl_khr_extended_bit_ops
• cl_khr_integer_dot_product
• cl_khr_subgroup_ballot
• cl_khr_subgroup_clustered_reduce
• cl_khr_subgroup_extended_types
• cl_khr_subgroup_named_barrier
• cl_khr_subgroup_non_uniform_arithmetic
• cl_khr_subgroup_non_uniform_vote
• cl_khr_subgroup_rotate
• cl_khr_subgroup_shuffle
• cl_khr_subgroup_shuffle_relative
• cl_khr_work_group_uniform_arithmetic
• cl_intel_subgroups_char
• cl_intel_subgroups_long

[llvmbot]

@llvm/pr-subscribers-clang

Author: Victor Mustya (vmustya)

Changes

This patch adds handling for the following OpenCL extensions, passed as
-cl-ext arguments to Clang:

• cl_khr_extended_bit_ops
• cl_khr_integer_dot_product
• cl_khr_subgroup_ballot
• cl_khr_subgroup_clustered_reduce
• cl_khr_subgroup_extended_types
• cl_khr_subgroup_named_barrier
• cl_khr_subgroup_non_uniform_arithmetic
• cl_khr_subgroup_non_uniform_vote
• cl_khr_subgroup_rotate
• cl_khr_subgroup_shuffle
• cl_khr_subgroup_shuffle_relative
• cl_khr_work_group_uniform_arithmetic
• cl_intel_subgroups_char
• cl_intel_subgroups_long

---

Full diff: https://github.com/llvm/llvm-project/pull/129777.diff

2 Files Affected:

• (modified) clang/include/clang/Basic/OpenCLExtensions.def (+21)
• (modified) clang/lib/Basic/OpenCLOptions.cpp (+8-1)

diff --git a/clang/include/clang/Basic/OpenCLExtensions.def b/clang/include/clang/Basic/OpenCLExtensions.def
index 6f73b26137500..09ab1245a6a9a 100644
--- a/clang/include/clang/Basic/OpenCLExtensions.def
+++ b/clang/include/clang/Basic/OpenCLExtensions.def
@@ -90,13 +90,29 @@ OPENCL_EXTENSION(__cl_clang_variadic_functions, true, 100)
 OPENCL_EXTENSION(__cl_clang_non_portable_kernel_param_types, true, 100)
 OPENCL_EXTENSION(__cl_clang_bitfields, true, 100)
 
+// Khronos OpenCL extensions
+OPENCL_EXTENSION(cl_khr_extended_bit_ops, true, 100)
+OPENCL_EXTENSION(cl_khr_integer_dot_product, true, 100)
+OPENCL_EXTENSION(cl_khr_subgroup_ballot, true, 100)
+OPENCL_EXTENSION(cl_khr_subgroup_clustered_reduce, true, 100)
+OPENCL_EXTENSION(cl_khr_subgroup_extended_types, true, 100)
+OPENCL_EXTENSION(cl_khr_subgroup_named_barrier, true, 100)
+OPENCL_EXTENSION(cl_khr_subgroup_non_uniform_arithmetic, true, 100)
+OPENCL_EXTENSION(cl_khr_subgroup_non_uniform_vote, true, 100)
+OPENCL_EXTENSION(cl_khr_subgroup_rotate, true, 100)
+OPENCL_EXTENSION(cl_khr_subgroup_shuffle, true, 100)
+OPENCL_EXTENSION(cl_khr_subgroup_shuffle_relative, true, 100)
+OPENCL_EXTENSION(cl_khr_work_group_uniform_arithmetic, true, 200)
+
 // AMD OpenCL extensions
 OPENCL_EXTENSION(cl_amd_media_ops, true, 100)
 OPENCL_EXTENSION(cl_amd_media_ops2, true, 100)
 
 // Intel OpenCL extensions
 OPENCL_EXTENSION(cl_intel_subgroups, true, 120)
+OPENCL_EXTENSION(cl_intel_subgroups_char, true, 120)
 OPENCL_EXTENSION(cl_intel_subgroups_short, true, 120)
+OPENCL_EXTENSION(cl_intel_subgroups_long, true, 120)
 OPENCL_EXTENSION(cl_intel_device_side_avc_motion_estimation, true, 120)
 
 // OpenCL C 3.0 features (6.2.1. Features)
@@ -112,6 +128,11 @@ OPENCL_OPTIONALCOREFEATURE(__opencl_c_program_scope_global_variables, false, 300
 OPENCL_OPTIONALCOREFEATURE(__opencl_c_fp64, false, 300, OCL_C_30)
 OPENCL_OPTIONALCOREFEATURE(__opencl_c_images, false, 300, OCL_C_30)
 
+// Features enabled by the Khronos extensions
+OPENCL_GENERIC_EXTENSION(__opencl_c_work_group_collective_functions, false, 300, 0)
+OPENCL_GENERIC_EXTENSION(__opencl_c_integer_dot_product_input_4x8bit, false, 300, 0)
+OPENCL_GENERIC_EXTENSION(__opencl_c_integer_dot_product_input_4x8bit_packed, false, 300, 0)
+
 #undef OPENCL_OPTIONALCOREFEATURE
 #undef OPENCL_COREFEATURE
 #undef OPENCL_GENERIC_EXTENSION
diff --git a/clang/lib/Basic/OpenCLOptions.cpp b/clang/lib/Basic/OpenCLOptions.cpp
index d39686ea688e5..9a58605294873 100644
--- a/clang/lib/Basic/OpenCLOptions.cpp
+++ b/clang/lib/Basic/OpenCLOptions.cpp
@@ -18,7 +18,14 @@ static const std::pair<StringRef, StringRef> DependentFeaturesList[] = {
     {"__opencl_c_3d_image_writes", "__opencl_c_images"},
     {"__opencl_c_pipes", "__opencl_c_generic_address_space"},
     {"__opencl_c_device_enqueue", "__opencl_c_generic_address_space"},
-    {"__opencl_c_device_enqueue", "__opencl_c_program_scope_global_variables"}};
+    {"__opencl_c_device_enqueue", "__opencl_c_program_scope_global_variables"},
+    {"__opencl_c_work_group_collective_functions",
+     "cl_khr_work_group_uniform_arithmetic"},
+    {"__opencl_c_integer_dot_product_input_4x8bit",
+     "cl_khr_integer_dot_product"},
+    {"__opencl_c_integer_dot_product_input_4x8bit_packed",
+     "cl_khr_integer_dot_product"},
+};
 
 // Extensions and equivalent feature pairs.
 static const std::pair<StringRef, StringRef> FeatureExtensionMap[] = {

[svenvh]

What is the motivation for this change? We are trying to move away from adding extensions to this file for extensions that only add builtin functions, see e.g. https://reviews.llvm.org/D92231 . What functionality are you gaining by adding them here?

[vmustya]

What is the motivation for this change? We are trying to move away from adding extensions to this file for extensions that only add builtin functions, see e.g. https://reviews.llvm.org/D92231 . What functionality are you gaining by adding them here?

@svenvh, I'm trying to unify the extensions enablement in the user interface. We can enable or disable some of the extensions with the -Xclang -cl-ext=+extension command line options. For example, a user can pass -Xclang -cl-ext=+cl_khr_subgroups successfully, but the -Xclang -cl-ext=+cl_khr_subgroup_ballot has no effect on the compiler.

[svenvh]

@svenvh, I'm trying to unify the extensions enablement in the user interface. We can enable or disable some of the extensions with the -Xclang -cl-ext=+extension command line options. For example, a user can pass -Xclang -cl-ext=+cl_khr_subgroups successfully, but the -Xclang -cl-ext=+cl_khr_subgroup_ballot has no effect on the compiler.

Got it, thanks for looking into this!

I'm trying to refresh my memory on this, as this is something Anton, Anastasia and myself looked at back in 2021, but we never concluded the work. From the most recent discussion I could find on this topic, I think the consensus was that for header-only extensions given to -cl-ext (such as -cl-ext=+cl_khr_subgroup_ballot) clang should just define the corresponding macro (-Dcl_khr_subgroup_ballot), probably in TargetInfo::setCommandLineOpenCLOpts. That way, header-only extensions don't need to be added to OpenCLExtensions.def

[vmustya]

That way, header-only extensions don't need to be added to OpenCLExtensions.def

@svenvh, in any case, we need a list of all the supported extensions, including the header ones. When someone passes the -Xclang -cl-ext=+all, the compiler should add defines for all the supported extensions.

[svenvh]

When someone passes the -Xclang -cl-ext=+all, the compiler should add defines for all the supported extensions.

Indeed; we already have such a list in opencl-c-base.h to enable all header-only extensions for the spir/spirv targets, so it should be easy to make -cl-ext=+all also enable those defines.