Add experimental check cl_bf16_conversion extention query

intel · Jun 27, 2024 · 9060485 · 9060485
1 parent fd4fce2
commit 9060485
Show file tree

Hide file tree

Showing 4 changed files with 8 additions and 1 deletion.
diff --git a/csrc/gpu/aten/core/DeviceInfo.h b/csrc/gpu/aten/core/DeviceInfo.h
@@ -34,6 +34,7 @@ struct DeviceInfo {
   uint32_t max_num_sub_groups;
   std::vector<size_t> sub_group_sizes;
   bool support_fp64;
+  bool support_cl_bf16_conversion;
 };
 
 } // namespace dpcpp

diff --git a/csrc/gpu/runtime/Device.cpp b/csrc/gpu/runtime/Device.cpp
@@ -322,6 +322,8 @@ static void initDeviceProperty(DeviceId device_id) {
       : 8;
   device_prop.support_atomic64 = device.has(dpcpp_dev_aspect_atomic64);
   device_prop.support_fp64 = device.has(dpcpp_dev_aspect_fp64);
+  sycl::ext::oneapi::experimental::cl_version version{20, 20, 20};
+  device_prop.support_cl_bf16_conversion = device.ext_oneapi_supports_cl_extension("cl_intel_bfloat16_conversions", &version);
 
   device_properties[device_id] = device_prop;
 
@@ -358,6 +360,7 @@ static void initDeviceProperty(DeviceId device_id) {
   dev_info.support_fp64 = device_prop.support_fp64;
 #if (defined(__INTEL_LLVM_COMPILER) && __INTEL_LLVM_COMPILER >= 20240100)
   dev_info.device_arch = static_cast<uint64_t>(device_prop.device_arch);
+  dev_info.support_cl_bf16_conversion = device_prop.support_cl_bf16_conversion;
 #else
   dev_info.device_arch = (uint64_t)0;
 #endif

diff --git a/csrc/gpu/runtime/DeviceProp.h b/csrc/gpu/runtime/DeviceProp.h
@@ -143,6 +143,7 @@ struct DeviceProp {
 
   bool support_fp64;
   bool support_atomic64;
+  bool support_cl_bf16_conversion;
 };
 
 } // namespace dpcpp

diff --git a/intel_extension_for_pytorch/csrc/xpu/Module.cpp b/intel_extension_for_pytorch/csrc/xpu/Module.cpp
@@ -577,6 +577,7 @@ static void register_xpu_device_info(PyObject* module) {
       .def_readonly("max_num_sub_groups", &DeviceInfo::max_num_sub_groups)
       .def_readonly("sub_group_sizes", &DeviceInfo::sub_group_sizes)
       .def_readonly("has_fp64", &DeviceInfo::support_fp64)
+      .def_readonly("support_cl_bf16_conversion", &DeviceInfo::support_cl_bf16_conversion)
       .def_readonly("device_arch", &DeviceInfo::device_arch)
       .def_property_readonly(
           "dev_type", [](const DeviceInfo& info) { return get_dev_type(info); })
@@ -589,7 +590,8 @@ static void register_xpu_device_info(PyObject* module) {
                << ", total_memory=" << info.global_mem_size / (1024 * 1024)
                << "MB, max_compute_units=" << info.max_compute_units
                << ", gpu_eu_count=" << info.gpu_eu_count
-               << ", device_arch=" << info.device_arch << ")";
+               << ", device_arch=" << info.device_arch 
+               << ", support_cl_bf16_conversion=" << info.support_cl_bf16_conversion << ")";
         return stream.str();
       });
 }