[SYCL] Cache 'assert supported' flag for device (#4505)

This should eliminate some part of overhead on host when
running kernels in a loop. Especially small ones.
Another part of improvement is in #4516

Signed-off-by: Sergey Kanaev <[email protected]>

Author: sergei

sycl/include/CL/sycl/queue.hpp

@@ -231,6 +231,10 @@ class __SYCL_EXPORT queue {
   template <info::queue param>
   typename info::param_traits<info::queue, param>::return_type get_info() const;
 
+  // A shorthand for `get_device().has()' which is expected to be a bit quicker
+  // than the long version
+  bool device_has(aspect Aspect) const;
+
 public:
   /// Submits a command group function object to the queue, in order to be
   /// scheduled for execution on the device.
@@ -247,7 +251,7 @@ class __SYCL_EXPORT queue {
     if (!is_host()) {
       auto PostProcess = [this, &CodeLoc](bool IsKernel, bool KernelUsesAssert,
                                           event &E) {
-        if (IsKernel && !get_device().has(aspect::ext_oneapi_native_assert) &&
+        if (IsKernel && !device_has(aspect::ext_oneapi_native_assert) &&
             KernelUsesAssert) {
           // __devicelib_assert_fail isn't supported by Device-side Runtime
           // Linking against fallback impl of __devicelib_assert_fail is
@@ -287,7 +291,7 @@ class __SYCL_EXPORT queue {
 #if __SYCL_USE_FALLBACK_ASSERT
     auto PostProcess = [this, &SecondaryQueue, &CodeLoc](
                            bool IsKernel, bool KernelUsesAssert, event &E) {
-      if (IsKernel && !get_device().has(aspect::ext_oneapi_native_assert) &&
+      if (IsKernel && !device_has(aspect::ext_oneapi_native_assert) &&
           KernelUsesAssert) {
         // __devicelib_assert_fail isn't supported by Device-side Runtime
         // Linking against fallback impl of __devicelib_assert_fail is performed

sycl/source/detail/device_impl.cpp

@@ -17,7 +17,9 @@ namespace sycl {
 namespace detail {
 
 device_impl::device_impl()
-    : MIsHostDevice(true), MPlatform(platform_impl::getHostPlatformImpl()) {}
+    : MIsHostDevice(true), MPlatform(platform_impl::getHostPlatformImpl()),
+      // assert is natively supported by host
+      MIsAssertFailSupported(true) {}
 
 device_impl::device_impl(pi_native_handle InteropDeviceHandle,
                          const plugin &Plugin)
@@ -70,6 +72,9 @@ device_impl::device_impl(pi_native_handle InteropDeviceHandle,
     Platform = platform_impl::getPlatformFromPiDevice(MDevice, Plugin);
   }
   MPlatform = Platform;
+
+  MIsAssertFailSupported =
+      has_extension(PI_DEVICE_INFO_EXTENSION_DEVICELIB_ASSERT);
 }
 
 device_impl::~device_impl() {
@@ -334,11 +339,7 @@ std::shared_ptr<device_impl> device_impl::getHostDeviceImpl() {
 }
 
 bool device_impl::isAssertFailSupported() const {
-  // assert is sort of natively supported by host
-  if (MIsHostDevice)
-    return true;
-
-  return has_extension(PI_DEVICE_INFO_EXTENSION_DEVICELIB_ASSERT);
+  return MIsAssertFailSupported;
 }
 
 } // namespace detail

sycl/source/detail/device_impl.hpp

@@ -232,6 +232,7 @@ class device_impl {
   bool MIsRootDevice = false;
   bool MIsHostDevice;
   PlatformImplPtr MPlatform;
+  bool MIsAssertFailSupported = false;
 }; // class device_impl
 
 } // namespace detail

sycl/source/detail/device_info.hpp

@@ -126,7 +126,7 @@ template <info::device param> struct get_device_info<platform, param> {
 // for string return type in other specializations.
 template <info::device param> struct get_device_info_string {
   static std::string get(RT::PiDevice dev, const plugin &Plugin) {
-    size_t resultSize;
+    size_t resultSize = 0;
     Plugin.call<PiApiKind::piDeviceGetInfo>(
         dev, pi::cast<RT::PiDeviceInfo>(param), 0, nullptr, &resultSize);
     if (resultSize == 0) {

sycl/source/queue.cpp

@@ -191,5 +191,10 @@ pi_native_handle queue::getNative() const { return impl->getNative(); }
 buffer<detail::AssertHappened, 1> &queue::getAssertHappenedBuffer() {
   return impl->getAssertHappenedBuffer();
 }
+
+bool queue::device_has(aspect Aspect) const {
+  // avoid creating sycl object from impl
+  return impl->getDeviceImplPtr()->has(Aspect);
+}
 } // namespace sycl
 } // __SYCL_INLINE_NAMESPACE(cl)

sycl/test/abi/sycl_symbols_linux.dump

@@ -3666,6 +3666,7 @@ _ZN2cl4sycl5eventC1Ev
 _ZN2cl4sycl5eventC2EP9_cl_eventRKNS0_7contextE
 _ZN2cl4sycl5eventC2ESt10shared_ptrINS0_6detail10event_implEE
 _ZN2cl4sycl5eventC2Ev
+_ZNK2cl4sycl5queue10device_hasENS0_6aspectE
 _ZN2cl4sycl5queue10mem_adviseEPKvm14_pi_mem_advice
 _ZN2cl4sycl5queue10mem_adviseEPKvmi
 _ZN2cl4sycl5queue10mem_adviseEPKvmiNS0_5eventE

sycl/test/abi/sycl_symbols_windows.dump

@@ -110,6 +110,7 @@
 ??$get_info@$0BAIA@@context@sycl@cl@@QEBAIXZ
 ??$get_info@$0BAIB@@context@sycl@cl@@QEBA?AV?$vector@Vdevice@sycl@cl@@V?$allocator@Vdevice@sycl@cl@@@std@@@std@@XZ
 ??$get_info@$0BAIE@@context@sycl@cl@@QEBA?AVplatform@12@XZ
+?device_has@queue@sycl@cl@@QEBA_NW4aspect@23@@Z
 ??$get_info@$0BAJA@@queue@sycl@cl@@QEBA?AVcontext@12@XZ
 ??$get_info@$0BAJB@@queue@sycl@cl@@QEBA?AVdevice@12@XZ
 ??$get_info@$0BAJC@@queue@sycl@cl@@QEBAIXZ