[SYCL][COMPAT] Enabled device_info caching within device_ext (#14630)

This patch introduces a simple memoization mechanism within `device_ext`
by caching the `device_info`, which helps avoid redundant/unnecessary
queries of the full device infos whenever a single device info
_(`max_wg_size` for e.g.)_ is requested.

API Tests in `device_ext_api` have been slightly modified to verify some
of the expected new changes to `get_device_info` methods.

Author: OuadiElfarouki

sycl/include/syclcompat/device.hpp

@@ -425,6 +425,12 @@ class device_ext : public sycl::device {
   }
 
   void get_device_info(device_info &out) const {
+    if (_dev_info) {
+      out = *_dev_info;
+      return;
+    }
+
+    std::lock_guard<std::mutex> lock(m_mutex);
     device_info prop;
     prop.set_name(get_info<sycl::info::device::name>().c_str());
 
@@ -439,8 +445,8 @@ class device_ext : public sycl::device {
         // max_work_item_sizes is an enum class element
         get_info<sycl::info::device::max_work_item_sizes>());
 #else
-        // SYCL 2020-conformant code, max_work_item_sizes is a struct templated
-        // by an int
+        // SYCL 2020-conformant code, max_work_item_sizes is a struct
+        // templated by an int
         get_info<sycl::info::device::max_work_item_sizes<3>>());
 #endif
     prop.set_host_unified_memory(has(sycl::aspect::usm_host_allocations));
@@ -512,8 +518,8 @@ Use 64 bits as memory_bus_width default value."
     prop.set_max_nd_range_size(max_nd_range_size);
 #endif
 
-    // Estimates max register size per work group, feel free to update the value
-    // according to device properties.
+    // Estimates max register size per work group, feel free to update the
+    // value according to device properties.
     prop.set_max_register_size_per_work_group(65536);
 
     prop.set_global_mem_cache_size(
@@ -526,13 +532,16 @@ Use 64 bits as memory_bus_width default value."
     prop.set_image3d_max(get_info<sycl::info::device::image3d_max_width>(),
                          get_info<sycl::info::device::image3d_max_height>(),
                          get_info<sycl::info::device::image3d_max_height>());
+
+    _dev_info = prop;
     out = prop;
   }
 
   device_info get_device_info() const {
-    device_info prop;
-    get_device_info(prop);
-    return prop;
+    if (!_dev_info) {
+      this->get_device_info(*_dev_info);
+    }
+    return _dev_info.value();
   }
 
   void reset(bool print_on_async_exceptions = false, bool in_order = true) {
@@ -683,6 +692,7 @@ Use 64 bits as memory_bus_width default value."
   std::vector<std::shared_ptr<sycl::queue>> _queues;
   mutable std::mutex m_mutex;
   std::vector<sycl::event> _events;
+  mutable std::optional<device_info> _dev_info;
 };
 
 namespace detail {

sycl/test-e2e/syclcompat/device/device.cpp

@@ -159,12 +159,19 @@ void test_device_ext_api() {
   auto major = dev_.get_major_version();
   test_major_version(dev_, major);
   dev_.get_minor_version();
-  dev_.get_max_compute_units();
   dev_.get_max_clock_frequency();
   dev_.get_integrated();
+
+  int max_cu = dev_.get_max_compute_units();
+  int max_wg_size = dev_.get_max_work_group_size();
+  size_t global_mem_size = dev_.get_global_mem_size();
+
   syclcompat::device_info Info;
   dev_.get_device_info(Info);
-  Info = dev_.get_device_info();
+  assert(Info.get_max_compute_units() == max_cu);
+  assert(Info.get_max_work_group_size() == max_wg_size);
+  assert(Info.get_global_mem_size() == global_mem_size);
+
   dev_.reset();
   auto QueuePtr = dev_.default_queue();
   dev_.queues_wait_and_throw();