SWDEV-545485 - Fix memory leaks in memset performance tests (#541)

Co-authored-by: Satyanvesh Dittakavi <Satyanvesh.Dittakavi@amd.com>
Co-authored-by: Satyanvesh Dittakavi <53337087+satyanveshd@users.noreply.github.com>

projects/hip-tests/catch/performance/memset/hipMemset2DAsync.cc

@@ -40,6 +40,7 @@ class Memset2DAsyncBenchmark : public Benchmark<Memset2DAsyncBenchmark> {
       HIP_CHECK(hipMemset2DAsync(dst_.ptr(), dst_.pitch(), 17, dst_.width(), dst_.height(),
                                  stream_.stream()));
     }
+    HIP_CHECK(hipStreamSynchronize(stream_.stream()));
   }
 
  private:

projects/hip-tests/catch/performance/memset/hipMemset3DAsync.cc

@@ -39,6 +39,7 @@ class Memset3DAsyncBenchmark : public Benchmark<Memset3DAsyncBenchmark> {
     TIMED_SECTION_STREAM(kTimerTypeEvent, stream_.stream()) {
       HIP_CHECK(hipMemset3DAsync(dst_.pitched_ptr(), 17, dst_.extent(), stream_.stream()));
     }
+    HIP_CHECK(hipStreamSynchronize(stream_.stream()));
   }
 
  private:

projects/hip-tests/catch/performance/memset/hipMemsetAsync.cc

@@ -39,6 +39,7 @@ class MemsetAsyncBenchmark : public Benchmark<MemsetAsyncBenchmark> {
     TIMED_SECTION_STREAM(kTimerTypeEvent, stream_.stream()) {
       HIP_CHECK(hipMemsetAsync(dst_.ptr(), 17, size_, stream_.stream()));
     }
+    HIP_CHECK(hipStreamSynchronize(stream_.stream()));
   }
 
  private:

projects/hip-tests/catch/performance/memset/hipMemsetD16Async.cc

@@ -40,6 +40,7 @@ class MemsetD16AsyncBenchmark : public Benchmark<MemsetD16AsyncBenchmark> {
       HIP_CHECK(hipMemsetD16Async(reinterpret_cast<hipDeviceptr_t>(dst_.ptr()), 311, size_,
                                   stream_.stream()));
     }
+    HIP_CHECK(hipStreamSynchronize(stream_.stream()));
   }
 
  private:

projects/hip-tests/catch/performance/memset/hipMemsetD32Async.cc

@@ -40,6 +40,7 @@ class MemsetD32AsyncBenchmark : public Benchmark<MemsetD32AsyncBenchmark> {
       HIP_CHECK(hipMemsetD32Async(reinterpret_cast<hipDeviceptr_t>(dst_.ptr()), 123'456, size_,
                                   stream_.stream()));
     }
+    HIP_CHECK(hipStreamSynchronize(stream_.stream()));
   }
 
  private:

projects/hip-tests/catch/performance/memset/hipMemsetD8Async.cc

@@ -40,6 +40,7 @@ class MemsetD8AsyncBenchmark : public Benchmark<MemsetD8AsyncBenchmark> {
       HIP_CHECK(hipMemsetD8Async(reinterpret_cast<hipDeviceptr_t>(dst_.ptr()), 17, size_,
                                  stream_.stream()));
     }
+    HIP_CHECK(hipStreamSynchronize(stream_.stream()));
   }
 
  private: