1 files changed, 33 insertions, 18 deletions

diff --git a/src/runtime/CL/CLMemoryRegion.cpp b/src/runtime/CL/CLMemoryRegion.cpp
index 780a563d63..c9ddf9b85c 100644
--- a/src/runtime/CL/CLMemoryRegion.cpp
+++ b/src/runtime/CL/CLMemoryRegion.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021 Arm Limited.
+ * Copyright (c) 2018-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,14 +26,12 @@
 #include "arm_compute/core/Error.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 
+#include "src/common/utils/Log.h"
+
 namespace arm_compute
 {
 ICLMemoryRegion::ICLMemoryRegion(size_t size)
-    : IMemoryRegion(size),
-      _queue(CLScheduler::get().queue()),
-      _ctx(CLScheduler::get().context()),
-      _mapping(nullptr),
-      _mem()
+    : IMemoryRegion(size), _ctx(CLScheduler::get().context()), _mapping(nullptr), _mem()
 {
 }
 
@@ -58,21 +56,34 @@ std::unique_ptr<IMemoryRegion> ICLMemoryRegion::extract_subregion(size_t offset,
     return nullptr;
 }
 
-CLBufferMemoryRegion::CLBufferMemoryRegion(cl_mem_flags flags, size_t size)
-    : ICLMemoryRegion(size)
+CLBufferMemoryRegion::CLBufferMemoryRegion(cl_mem_flags flags, size_t size) : ICLMemoryRegion(size)
 {
-    if(_size != 0)
+    if (_size != 0)
     {
         _mem = cl::Buffer(CLScheduler::get().context(), flags, _size);
     }
 }
 
-CLBufferMemoryRegion::CLBufferMemoryRegion(const cl::Buffer &buffer)
-    : ICLMemoryRegion(buffer.getInfo<CL_MEM_SIZE>())
+CLBufferMemoryRegion::CLBufferMemoryRegion(const cl::Buffer &buffer) : ICLMemoryRegion(buffer.getInfo<CL_MEM_SIZE>())
 {
     _mem = buffer;
 }
 
+CLBufferMemoryRegion::~CLBufferMemoryRegion()
+{
+    // Flush the command queue to ensure all commands that may use this memory buffer are scheduled to be finished before
+    // this buffer is freed
+    // Do not call finish as it is a blocking call which affects the performance
+    try
+    {
+        CLScheduler::get().queue().flush();
+    }
+    catch (const std::exception &e)
+    {
+        ARM_COMPUTE_LOG_ERROR_ACL(e.what());
+    }
+}
+
 void *CLBufferMemoryRegion::ptr()
 {
     return nullptr;
@@ -95,10 +106,10 @@ void CLBufferMemoryRegion::unmap(cl::CommandQueue &q)
 ICLSVMMemoryRegion::ICLSVMMemoryRegion(cl_mem_flags flags, size_t size, size_t alignment)
     : ICLMemoryRegion(size), _ptr(nullptr)
 {
-    if(size != 0)
+    if (size != 0)
     {
         _ptr = clSVMAlloc(CLScheduler::get().context().get(), flags, size, alignment);
-        if(_ptr != nullptr)
+        if (_ptr != nullptr)
         {
             _mem = cl::Buffer(CLScheduler::get().context(), CL_MEM_READ_WRITE | CL_MEM_USE_HOST_PTR, _size, _ptr);
         }
@@ -107,15 +118,18 @@ ICLSVMMemoryRegion::ICLSVMMemoryRegion(cl_mem_flags flags, size_t size, size_t a
 
 ICLSVMMemoryRegion::~ICLSVMMemoryRegion()
 {
-    if(_ptr != nullptr)
+    if (_ptr != nullptr)
     {
         try
         {
-            clFinish(_queue.get());
+            // Can only use the blocking finish instead of the non-blocking flush here, because clSVMFree requires all
+            // commands that may use the svm pointer to finish beforehand
+            // https://registry.khronos.org/OpenCL/sdk/3.0/docs/man/html/clSVMFree.html
+            clFinish(CLScheduler::get().queue().get());
             _mem = cl::Buffer();
             clSVMFree(_ctx.get(), _ptr);
         }
-        catch(...)
+        catch (...)
         {
         }
     }
@@ -134,7 +148,8 @@ CLCoarseSVMMemoryRegion::CLCoarseSVMMemoryRegion(cl_mem_flags flags, size_t size
 void *CLCoarseSVMMemoryRegion::map(cl::CommandQueue &q, bool blocking)
 {
     ARM_COMPUTE_ERROR_ON(_ptr == nullptr);
-    clEnqueueSVMMap(q.get(), blocking ? CL_TRUE : CL_FALSE, CL_MAP_READ | CL_MAP_WRITE, _ptr, _size, 0, nullptr, nullptr);
+    clEnqueueSVMMap(q.get(), blocking ? CL_TRUE : CL_FALSE, CL_MAP_READ | CL_MAP_WRITE, _ptr, _size, 0, nullptr,
+                    nullptr);
     _mapping = _ptr;
     return _mapping;
 }
@@ -153,7 +168,7 @@ CLFineSVMMemoryRegion::CLFineSVMMemoryRegion(cl_mem_flags flags, size_t size, si
 
 void *CLFineSVMMemoryRegion::map(cl::CommandQueue &q, bool blocking)
 {
-    if(blocking)
+    if (blocking)
     {
         clFinish(q.get());
     }