Add Auxiliary tensors

The asssign_memory_descriptors method could not automatically assign
Auxiliary tensors. Therefore changes are made to allow developers to
explicitly mark auxiliary tensors.

However, to avoid ambiguity between auxiliary and "intermediate"
tensors, we solidify the definitions of both:

Intermediate tensors are a strictly topological term. They are defined
as "inner" tensors within a workload, hidden from the user, as opposed
to input and output tensors exposed to the users.

Auxiliary tensors are a subcategory of Intermediate tensors, and are
also about memory allocation. They are intermediate tensors that need
real memory backing.

For more details please see the documentation of MemoryType enum

Rename MemoryType::NoAlloc to MemoryType::Virtual

Partially resolves: COMPMID-5523

Signed-off-by: SiCong Li <sicong.li@arm.com>
Change-Id: Ibde44c2ec1570be9423e0fb38b53bb136ffc36dd
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8940
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>

1 files changed, 41 insertions, 9 deletions

diff --git a/src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h b/src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h
index 08796b607b..d5075d5c94 100644
--- a/src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h
+++ b/src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h
@@ -24,6 +24,7 @@
 #ifndef SRC_DYNAMIC_FUSION_SKETCH_GPU_GPUWORKLOADSKETCHIMPL
 #define SRC_DYNAMIC_FUSION_SKETCH_GPU_GPUWORKLOADSKETCHIMPL
 
+#include "arm_compute/dynamic_fusion/sketch/MemoryDescriptor.h"
 #include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h"
 #include "src/dynamic_fusion/sketch/gpu/GpuComponentServices.h"
 #include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.h"
@@ -52,7 +53,8 @@ public:
           _comp_services{},
           _component_graph{ &_comp_services },
           _operator_group{},
-          _interm_tensor_info_list{ std::vector<std::unique_ptr<TensorInfo>>() }
+          _managed_tensor_info_list{ std::vector<std::unique_ptr<TensorInfo>>() },
+          _mem_map{}
     {
     }
     /** Prevent instances of this class from being copy constructed */
@@ -99,18 +101,47 @@ public:
      */
     GpuWorkloadSourceCode generate_source_code() const
     {
-        return component_graph().fuse().write_workload_code();
+        return component_graph().fuse(_mem_map).write_workload_code();
     }
-    /** Create an intermediate tensor info and save it
+    /** Create a virtual (see @ref MemoryType) tensor info and save it
      *
-     * @return ITensorInfo  The created intermediate tensor info object pointer
+     * @return ITensorInfo*  The created virtual tensor info object pointer
      */
-    ITensorInfo *create_intermediate_tensor()
+    ITensorInfo *create_virtual_tensor()
     {
         auto uptr = std::make_unique<TensorInfo>();
-        uptr->set_id(-allocate_new_tensor_id()); // intermediate tensors must have negative id
-        _interm_tensor_info_list.emplace_back(std::move(uptr));
-        return _interm_tensor_info_list.back().get();
+        uptr->set_id(-allocate_new_tensor_id()); // virtual tensors must have negative id
+        register_memory_descriptor(*uptr, MemoryDescriptor{ MemoryType::Virtual });
+        _managed_tensor_info_list.emplace_back(std::move(uptr));
+        return _managed_tensor_info_list.back().get();
+    }
+    /** Create an auxiliary (see @ref MemoryType) tensor info and save it
+     *
+     * @return ITensorInfo*  The created auxiliary tensor info object pointer
+     */
+
+    /** Create an auxiliary (see @ref MemoryType) tensor info and save it
+     *
+     * @param[in] tensor_info @ref ITensorInfo to copy from
+     *
+     * @return ITensorInfo*  The created auxiliary tensor info object pointer
+     */
+    ITensorInfo *create_auxiliary_tensor(const ITensorInfo &tensor_info)
+    {
+        auto uptr = std::make_unique<TensorInfo>(tensor_info);
+        uptr->set_id(allocate_new_tensor_id());
+        register_memory_descriptor(*uptr, MemoryDescriptor{ MemoryType::Auxiliary, AuxMemoryInfo{ uptr->total_size() } });
+        _managed_tensor_info_list.emplace_back(std::move(uptr));
+        return _managed_tensor_info_list.back().get();
+    }
+    /** Register memory descriptor of a tensor info
+     *
+     * @param[in] info     @ref ITensorInfo to be registered
+     * @param[in] mem_desc @ref MemoryDescriptor to be registered with @p info
+     */
+    void register_memory_descriptor(const ITensorInfo &info, const MemoryDescriptor &mem_desc)
+    {
+        _mem_map[info.id()] = mem_desc;
     }
 
 private:
@@ -119,7 +150,8 @@ private:
     GpuKernelComponentGraph                  _component_graph;
     GpuOperatorGroup                         _operator_group;
     ITensorInfo::Id                          _next_id{ ITensorInfo::invalid_tensor_id };
-    std::vector<std::unique_ptr<TensorInfo>> _interm_tensor_info_list;
+    std::vector<std::unique_ptr<TensorInfo>> _managed_tensor_info_list;
+    MemoryDescriptorMap                      _mem_map;
 };
 } // namespace dynamic_fusion
 } // namespace experimental