[SYCL] Fix memory leak for sycl::device::get_devices (#2256)

Plugins create pi_device objects for root devices, but
it does not call piDeviceRelease and causes memory leaks.
This slows the performance of SYCL apps.

Signed-off-by: Byoungro So <byoungro.so@intel.com>

Author: bso-intel

sycl/source/detail/device_impl.cpp

@@ -77,7 +77,7 @@ device_impl::device_impl(pi_native_handle InteropDeviceHandle,
 }
 
 device_impl::~device_impl() {
-  if (!MIsRootDevice && !MIsHostDevice) {
+  if (!MIsHostDevice) {
     // TODO catch an exception and put it to list of asynchronous exceptions
     const detail::plugin &Plugin = getPlugin();
     RT::PiResult Err = Plugin.call_nocheck<PiApiKind::piDeviceRelease>(MDevice);

sycl/test/basic_tests/memory-consumption.cpp

@@ -0,0 +1,62 @@
+// RUN: %clangxx -fsycl %s -o %t.out
+// RUN: %GPU_RUN_PLACEHOLDER %t.out
+//
+// UNSUPPORTED: windows
+//
+//==-----memory-consumption.cpp - SYCL memory consumption basic test ------==//
+//
+// This test specifically tests that memory consumption does not change
+// when the get_devices() is called repeatedly.
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "CL/sycl.hpp"
+#include <iostream>
+#include <thread>
+
+#ifdef __linux__
+#include <strings.h>
+#include <sys/resource.h>
+#include <sys/time.h>
+
+long get_cpu_mem() {
+  struct rusage usage;
+  memset(&usage, 0, sizeof(rusage));
+  getrusage(RUSAGE_SELF, &usage);
+  return usage.ru_maxrss;
+}
+#endif
+
+using namespace cl::sycl;
+
+int main() {
+  constexpr auto dev_type = info::device_type::gpu;
+  auto devices = device::get_devices(dev_type);
+
+  int startSize = get_cpu_mem();
+  std::cout << startSize << " kb" << std::endl;
+
+  for (int i = 0; i < 1000; i++) {
+    devices = cl::sycl::device::get_devices(dev_type);
+  }
+  int endSize = get_cpu_mem();
+  std::cout << endSize << " kb" << std::endl;
+
+  auto plat = devices[0].get_platform();
+  std::string plat_name = plat.get_info<info::platform::name>();
+  std::cout << "Platform: " << plat_name << std::endl;
+
+  devices.erase(devices.begin(), devices.end());
+
+  if (startSize == endSize) {
+    std::cout << "Passed" << std::endl;
+    return 0;
+  } else {
+    std::cout << "Failed" << std::endl;
+    return 1;
+  }
+}