COMPMID-605: Transition buffer memory manager

Change-Id: Ide7c6124eb19f13f15f517e62d705646a0cd1ecd
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/130184
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

8 files changed, 180 insertions, 29 deletions

diff --git a/src/graph/backends/CL/CLDeviceBackend.cpp b/src/graph/backends/CL/CLDeviceBackend.cpp
index 37cbcd72d7..7f2be674f6 100644
--- a/src/graph/backends/CL/CLDeviceBackend.cpp
+++ b/src/graph/backends/CL/CLDeviceBackend.cpp
@@ -37,6 +37,7 @@
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/runtime/BlobLifetimeManager.h"
 #include "arm_compute/runtime/CL/CLBufferAllocator.h"
+#include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 #include "arm_compute/runtime/MemoryManagerOnDemand.h"
 #include "arm_compute/runtime/PoolManager.h"
@@ -107,8 +108,10 @@ void CLDeviceBackend::setup_backend_context(GraphContext &ctx)
     if(ctx.memory_management_ctx(Target::CL) == nullptr)
     {
         MemoryManagerContext mm_ctx;
-        mm_ctx.target = Target::CL;
-        mm_ctx.mm     = create_memory_manager(MemoryManagerAffinity::Buffer);
+        mm_ctx.target      = Target::CL;
+        mm_ctx.intra_mm    = create_memory_manager(MemoryManagerAffinity::Buffer);
+        mm_ctx.cross_mm    = create_memory_manager(MemoryManagerAffinity::Buffer);
+        mm_ctx.cross_group = std::make_shared<CLMemoryGroup>(mm_ctx.cross_mm);
 
         ctx.insert_memory_management_ctx(std::move(mm_ctx));
     }
@@ -119,6 +122,11 @@ bool CLDeviceBackend::is_backend_supported()
     return arm_compute::opencl_is_available();
 }
 
+IAllocator *CLDeviceBackend::backend_allocator()
+{
+    return &_allocator;
+}
+
 std::unique_ptr<ITensorHandle> CLDeviceBackend::create_tensor(const Tensor &tensor)
 {
     // Get tensor descriptor
diff --git a/src/graph/backends/CL/CLSubTensorHandle.cpp b/src/graph/backends/CL/CLSubTensorHandle.cpp
index a1bc8a1dd3..016dca753b 100644
--- a/src/graph/backends/CL/CLSubTensorHandle.cpp
+++ b/src/graph/backends/CL/CLSubTensorHandle.cpp
@@ -32,11 +32,12 @@ namespace graph
 namespace backends
 {
 CLSubTensorHandle::CLSubTensorHandle(ITensorHandle *parent_handle, const TensorShape &shape, const Coordinates &coords, bool extend_parent)
-    : _sub_tensor()
+    : _sub_tensor(), _parent_handle(nullptr)
 {
     ARM_COMPUTE_ERROR_ON(!parent_handle);
     auto parent_tensor = arm_compute::utils::cast::polymorphic_downcast<ICLTensor *>(&parent_handle->tensor());
     _sub_tensor        = arm_compute::CLSubTensor(parent_tensor, shape, coords, extend_parent);
+    _parent_handle     = parent_handle;
 }
 
 void CLSubTensorHandle::allocate()
@@ -44,14 +45,15 @@ void CLSubTensorHandle::allocate()
     // noop
 }
 
-const arm_compute::ITensor &CLSubTensorHandle::tensor() const
+void CLSubTensorHandle::free()
 {
-    return _sub_tensor;
+    // noop
 }
 
-arm_compute::ITensor &CLSubTensorHandle::tensor()
+void CLSubTensorHandle::manage(IMemoryGroup *mg)
 {
-    return _sub_tensor;
+    ARM_COMPUTE_UNUSED(mg);
+    // noop
 }
 
 void CLSubTensorHandle::map(bool blocking)
@@ -69,10 +71,31 @@ void CLSubTensorHandle::release_if_unused()
     // noop
 }
 
+const arm_compute::ITensor &CLSubTensorHandle::tensor() const
+{
+    return _sub_tensor;
+}
+
+arm_compute::ITensor &CLSubTensorHandle::tensor()
+{
+    return _sub_tensor;
+}
+
+ITensorHandle *CLSubTensorHandle::parent_handle()
+{
+    ARM_COMPUTE_ERROR_ON(_parent_handle == nullptr);
+    return _parent_handle->parent_handle();
+}
+
 bool CLSubTensorHandle::is_subtensor() const
 {
     return true;
 }
+
+Target CLSubTensorHandle::target() const
+{
+    return Target::CL;
+}
 } // namespace backends
 } // namespace graph
 } // namespace arm_compute
\ No newline at end of file
diff --git a/src/graph/backends/CL/CLTensorHandle.cpp b/src/graph/backends/CL/CLTensorHandle.cpp
index 563c4d9ac6..219d9d0301 100644
--- a/src/graph/backends/CL/CLTensorHandle.cpp
+++ b/src/graph/backends/CL/CLTensorHandle.cpp
@@ -23,6 +23,9 @@
  */
 #include "arm_compute/graph/backends/CL/CLTensorHandle.h"
 
+#include "arm_compute/core/utils/misc/Cast.h"
+#include "arm_compute/runtime/CL/CLMemoryGroup.h"
+
 namespace arm_compute
 {
 namespace graph
@@ -40,14 +43,18 @@ void CLTensorHandle::allocate()
     _tensor.allocator()->allocate();
 }
 
-const arm_compute::ITensor &CLTensorHandle::tensor() const
+void CLTensorHandle::free()
 {
-    return _tensor;
+    _tensor.allocator()->free();
 }
 
-arm_compute::ITensor &CLTensorHandle::tensor()
+void CLTensorHandle::manage(IMemoryGroup *mg)
 {
-    return _tensor;
+    if(mg != nullptr)
+    {
+        auto *cl_mg = arm_compute::utils::cast::polymorphic_downcast<CLMemoryGroup *>(mg);
+        cl_mg->manage(&_tensor);
+    }
 }
 
 void CLTensorHandle::map(bool blocking)
@@ -69,10 +76,30 @@ void CLTensorHandle::release_if_unused()
     }
 }
 
+const arm_compute::ITensor &CLTensorHandle::tensor() const
+{
+    return _tensor;
+}
+
+arm_compute::ITensor &CLTensorHandle::tensor()
+{
+    return _tensor;
+}
+
+ITensorHandle *CLTensorHandle::parent_handle()
+{
+    return this;
+}
+
 bool CLTensorHandle::is_subtensor() const
 {
     return false;
 }
+
+Target CLTensorHandle::target() const
+{
+    return Target::CL;
+}
 } // namespace backends
 } // namespace graph
 } // namespace arm_compute
\ No newline at end of file
diff --git a/src/graph/backends/GLES/GCDeviceBackend.cpp b/src/graph/backends/GLES/GCDeviceBackend.cpp
index 0185598965..770cca5d42 100644
--- a/src/graph/backends/GLES/GCDeviceBackend.cpp
+++ b/src/graph/backends/GLES/GCDeviceBackend.cpp
@@ -36,6 +36,7 @@
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/runtime/BlobLifetimeManager.h"
 #include "arm_compute/runtime/GLES_COMPUTE/GCBufferAllocator.h"
+#include "arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h"
 #include "arm_compute/runtime/GLES_COMPUTE/GCScheduler.h"
 #include "arm_compute/runtime/MemoryManagerOnDemand.h"
 #include "arm_compute/runtime/PoolManager.h"
@@ -68,8 +69,10 @@ void GCDeviceBackend::setup_backend_context(GraphContext &ctx)
     if(ctx.memory_management_ctx(Target::GC) == nullptr)
     {
         MemoryManagerContext mm_ctx;
-        mm_ctx.target = Target::GC;
-        mm_ctx.mm     = create_memory_manager(MemoryManagerAffinity::Buffer);
+        mm_ctx.target      = Target::GC;
+        mm_ctx.intra_mm    = create_memory_manager(MemoryManagerAffinity::Buffer);
+        mm_ctx.cross_mm    = create_memory_manager(MemoryManagerAffinity::Buffer);
+        mm_ctx.cross_group = std::make_shared<GCMemoryGroup>(mm_ctx.cross_mm);
 
         ctx.insert_memory_management_ctx(std::move(mm_ctx));
     }
@@ -80,6 +83,11 @@ bool GCDeviceBackend::is_backend_supported()
     return arm_compute::opengles31_is_available();
 }
 
+IAllocator *GCDeviceBackend::backend_allocator()
+{
+    return &_allocator;
+}
+
 std::unique_ptr<ITensorHandle> GCDeviceBackend::create_tensor(const Tensor &tensor)
 {
     // Get tensor descriptor
diff --git a/src/graph/backends/GLES/GCTensorHandle.cpp b/src/graph/backends/GLES/GCTensorHandle.cpp
index ae7c778130..4e5c652120 100644
--- a/src/graph/backends/GLES/GCTensorHandle.cpp
+++ b/src/graph/backends/GLES/GCTensorHandle.cpp
@@ -23,6 +23,9 @@
  */
 #include "arm_compute/graph/backends/GLES/GCTensorHandle.h"
 
+#include "arm_compute/core/utils/misc/Cast.h"
+#include "arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h"
+
 namespace arm_compute
 {
 namespace graph
@@ -40,14 +43,18 @@ void GCTensorHandle::allocate()
     _tensor.allocator()->allocate();
 }
 
-const arm_compute::ITensor &GCTensorHandle::tensor() const
+void GCTensorHandle::free()
 {
-    return _tensor;
+    _tensor.allocator()->free();
 }
 
-arm_compute::ITensor &GCTensorHandle::tensor()
+void GCTensorHandle::manage(IMemoryGroup *mg)
 {
-    return _tensor;
+    if(mg != nullptr)
+    {
+        auto *gc_mg = arm_compute::utils::cast::polymorphic_downcast<GCMemoryGroup *>(mg);
+        gc_mg->manage(&_tensor);
+    }
 }
 
 void GCTensorHandle::map(bool blocking)
@@ -69,10 +76,30 @@ void GCTensorHandle::release_if_unused()
     }
 }
 
+const arm_compute::ITensor &GCTensorHandle::tensor() const
+{
+    return _tensor;
+}
+
+arm_compute::ITensor &GCTensorHandle::tensor()
+{
+    return _tensor;
+}
+
+ITensorHandle *GCTensorHandle::parent_handle()
+{
+    return this;
+}
+
 bool GCTensorHandle::is_subtensor() const
 {
     return false;
 }
+
+Target GCTensorHandle::target() const
+{
+    return Target::GC;
+}
 } // namespace backends
 } // namespace graph
 } // namespace arm_compute
\ No newline at end of file
diff --git a/src/graph/backends/NEON/NEDeviceBackend.cpp b/src/graph/backends/NEON/NEDeviceBackend.cpp
index def6c39003..7c2db40260 100644
--- a/src/graph/backends/NEON/NEDeviceBackend.cpp
+++ b/src/graph/backends/NEON/NEDeviceBackend.cpp
@@ -37,6 +37,7 @@
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/runtime/Allocator.h"
 #include "arm_compute/runtime/BlobLifetimeManager.h"
+#include "arm_compute/runtime/MemoryGroup.h"
 #include "arm_compute/runtime/MemoryManagerOnDemand.h"
 #include "arm_compute/runtime/OffsetLifetimeManager.h"
 #include "arm_compute/runtime/PoolManager.h"
@@ -74,8 +75,10 @@ void NEDeviceBackend::setup_backend_context(GraphContext &ctx)
     if(ctx.memory_management_ctx(Target::NEON) == nullptr)
     {
         MemoryManagerContext mm_ctx;
-        mm_ctx.target = Target::NEON;
-        mm_ctx.mm     = create_memory_manager(MemoryManagerAffinity::Buffer);
+        mm_ctx.target      = Target::NEON;
+        mm_ctx.intra_mm    = create_memory_manager(MemoryManagerAffinity::Offset);
+        mm_ctx.cross_mm    = create_memory_manager(MemoryManagerAffinity::Offset);
+        mm_ctx.cross_group = std::make_shared<MemoryGroup>(mm_ctx.cross_mm);
 
         ctx.insert_memory_management_ctx(std::move(mm_ctx));
     }
@@ -86,6 +89,11 @@ bool NEDeviceBackend::is_backend_supported()
     return true;
 }
 
+IAllocator *NEDeviceBackend::backend_allocator()
+{
+    return &_allocator;
+}
+
 std::unique_ptr<ITensorHandle> NEDeviceBackend::create_tensor(const Tensor &tensor)
 {
     // Get tensor descriptor
diff --git a/src/graph/backends/NEON/NESubTensorHandle.cpp b/src/graph/backends/NEON/NESubTensorHandle.cpp
index c48ba6b9d6..c0acedd9f2 100644
--- a/src/graph/backends/NEON/NESubTensorHandle.cpp
+++ b/src/graph/backends/NEON/NESubTensorHandle.cpp
@@ -30,10 +30,11 @@ namespace graph
 namespace backends
 {
 NESubTensorHandle::NESubTensorHandle(ITensorHandle *parent_handle, const TensorShape &shape, const Coordinates &coords, bool extend_parent)
-    : _sub_tensor()
+    : _sub_tensor(), _parent_handle(nullptr)
 {
     ARM_COMPUTE_ERROR_ON(!parent_handle);
-    _sub_tensor = arm_compute::SubTensor(&parent_handle->tensor(), shape, coords, extend_parent);
+    _sub_tensor    = arm_compute::SubTensor(&parent_handle->tensor(), shape, coords, extend_parent);
+    _parent_handle = parent_handle;
 }
 
 void NESubTensorHandle::allocate()
@@ -41,14 +42,15 @@ void NESubTensorHandle::allocate()
     // noop
 }
 
-const arm_compute::ITensor &NESubTensorHandle::tensor() const
+void NESubTensorHandle::free()
 {
-    return _sub_tensor;
+    // noop
 }
 
-arm_compute::ITensor &NESubTensorHandle::tensor()
+void NESubTensorHandle::manage(IMemoryGroup *mg)
 {
-    return _sub_tensor;
+    ARM_COMPUTE_UNUSED(mg);
+    // noop
 }
 
 void NESubTensorHandle::map(bool blocking)
@@ -66,10 +68,31 @@ void NESubTensorHandle::release_if_unused()
     // noop
 }
 
+const arm_compute::ITensor &NESubTensorHandle::tensor() const
+{
+    return _sub_tensor;
+}
+
+arm_compute::ITensor &NESubTensorHandle::tensor()
+{
+    return _sub_tensor;
+}
+
+ITensorHandle *NESubTensorHandle::parent_handle()
+{
+    ARM_COMPUTE_ERROR_ON(_parent_handle == nullptr);
+    return _parent_handle->parent_handle();
+}
+
 bool NESubTensorHandle::is_subtensor() const
 {
     return true;
 }
+
+Target NESubTensorHandle::target() const
+{
+    return Target::NEON;
+}
 } // namespace backends
 } // namespace graph
 } // namespace arm_compute
\ No newline at end of file
diff --git a/src/graph/backends/NEON/NETensorHandle.cpp b/src/graph/backends/NEON/NETensorHandle.cpp
index 8508ac9511..5892116caf 100644
--- a/src/graph/backends/NEON/NETensorHandle.cpp
+++ b/src/graph/backends/NEON/NETensorHandle.cpp
@@ -23,6 +23,9 @@
  */
 #include "arm_compute/graph/backends/NEON/NETensorHandle.h"
 
+#include "arm_compute/core/utils/misc/Cast.h"
+#include "arm_compute/runtime/MemoryGroup.h"
+
 namespace arm_compute
 {
 namespace graph
@@ -40,14 +43,18 @@ void NETensorHandle::allocate()
     _tensor.allocator()->allocate();
 }
 
-const arm_compute::ITensor &NETensorHandle::tensor() const
+void NETensorHandle::free()
 {
-    return _tensor;
+    _tensor.allocator()->free();
 }
 
-arm_compute::ITensor &NETensorHandle::tensor()
+void NETensorHandle::manage(IMemoryGroup *mg)
 {
-    return _tensor;
+    if(mg != nullptr)
+    {
+        auto *ne_mg = arm_compute::utils::cast::polymorphic_downcast<MemoryGroup *>(mg);
+        ne_mg->manage(&_tensor);
+    }
 }
 
 void NETensorHandle::map(bool blocking)
@@ -68,10 +75,30 @@ void NETensorHandle::release_if_unused()
     }
 }
 
+const arm_compute::ITensor &NETensorHandle::tensor() const
+{
+    return _tensor;
+}
+
+arm_compute::ITensor &NETensorHandle::tensor()
+{
+    return _tensor;
+}
+
+ITensorHandle *NETensorHandle::parent_handle()
+{
+    return this;
+}
+
 bool NETensorHandle::is_subtensor() const
 {
     return false;
 }
+
+Target NETensorHandle::target() const
+{
+    return Target::NEON;
+}
 } // namespace backends
 } // namespace graph
 } // namespace arm_compute
\ No newline at end of file