[tensorpool] Introducing TensorPool

This patch introduces TensorPool which manages the list of all the
tensors with their execution orders and lifespan.
The conversion of lifespan and execution orders to the validity will be
done by the TensorPool.
TensorPool forms the middle ground between memory pool and manager for
easier maintainance.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

diff --git a/nntrainer/tensor/tensor_pool.h b/nntrainer/tensor/tensor_pool.h
new file mode 100644 (file)
index 0000000..cbc2fb1
--- /dev/null
+++ b/nntrainer/tensor/tensor_pool.h
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Copyright (C) 2020 Parichay Kapoor <pk.kapoor@samsung.com>
+ *
+ * @file   tensor_pool.h
+ * @date   18 Aug 2021
+ * @brief  This is TensorPool for all requested tensors
+ * @see    https://github.com/nnstreamer/nntrainer
+ * @author Parichay Kapoor <pk.kapoor@samsung.com>
+ * @bug           No known bugs except for NYI items
+ *
+ */
+
+#ifndef __TENSOR_POOL_H__
+#define __TENSOR_POOL_H__
+#ifdef __cplusplus
+
+#include <functional>
+#include <memory>
+#include <unordered_map>
+#include <vector>
+
+#include <memory_pool.h>
+#include <tensor.h>
+#include <tensor_wrap_specs.h>
+
+namespace nntrainer {
+
+/**
+ * @class   TensorPool
+ * @brief   tensor pool of nntrainer
+ */
+class TensorPool {
+
+public:
+  /**
+   * @brief     Constructor of TensorPool
+   */
+  TensorPool() : mem_pool() {}
+
+  /**
+   * @brief     Destructor of TensorPool
+   */
+  ~TensorPool() = default;
+
+  /**
+   * @brief     Request tensor with the given spec
+   *
+   * @param dim Tensor dimensions
+   * @param exec_order The execution orders for this tensors
+   * @param lifespan Lifespan of this tensor
+   * @param name Name of this tensor
+   *
+   * @return ptr to the created tensor
+   *
+   * @note returns empty tensor which will be filled when allocate is called.
+   * @note we assume that the caller checks if the exec_order and lifespan are
+   * compatible.
+   */
+  Tensor *requestTensor(const TensorDim dim,
+                        const std::vector<unsigned int> &exec_order,
+                        TensorLifespan lifespan, const std::string &name);
+
+  /**
+   * @brief     Request tensor which has been already requested with the given
+   * spec
+   *
+   * @param dim Tensor dimensions
+   * @param exec_order The execution orders for this tensors
+   * @param lifespan Lifespan of this tensor
+   * @param name Name of this tensor
+   *
+   * @return ptr to the tensor
+   *
+   * @note returns empty tensor which will be filled when allocate is called.
+   * @note we assume that the caller checks if the exec_order and lifespan are
+   * compatible.
+   */
+  Tensor *requestPreallocatedTensor(const TensorDim dim,
+                                    const std::vector<unsigned int> &exec_order,
+
+                                    TensorLifespan lifespan,
+                                    const std::string &name);
+
+  /**
+   * @brief finalize the requested tensors
+   * @param planner planner to layout the tensor memories
+   * @param start_order start value for the order_exec (inclusive)
+   * @param end_order end value for the order_exec (inclusive)
+   *
+   * @details finalize the requested tensors, request memory for them and plan
+   * layout for their allocations.
+   */
+  void finalize(const MemoryPlanner &planner, unsigned int start_order,
+                unsigned int end_order);
+
+  /**
+   * @brief Set the batch size for the inputs/outputs of the layers
+   */
+  void setBatchSize(const std::string &name, unsigned int batch);
+
+  /**
+   * @brief Allocate memory for all the managed tensors
+   */
+  void allocate();
+
+  /**
+   * @brief Deallocate memory for all the managed tensors
+   */
+  void deallocate();
+
+  /**
+   * @brief     Expand the lifespan of the tensor with the given name
+   *
+   * @param name The name of the tensor
+   * @param lifespan The lifespan to be expanded to
+   */
+  void expand_lifespan(const std::string &name, TensorLifespan lifespan);
+
+  /**
+   * @brief     Expand the execution order of the tensor with the given name
+   *
+   * @param name The name of the tensor
+   * @param exec_order The execution orders
+   */
+  void expand_lifespan(const std::string &name,
+                       const std::vector<unsigned int> &exec_order);
+
+private:
+  struct requestSpec {
+    std::unique_ptr<Tensor> tensor;       /**< tensor object itself */
+    std::vector<unsigned int> exec_order; /**< tensor exec order list */
+    TensorLifespan lifespan;              /**< tensor lifespan */
+    unsigned int token;                   /**< tensor memory token */
+  };
+
+  std::unordered_map<std::string, requestSpec>
+    pool;              /**< list of requested tensors */
+  MemoryPool mem_pool; /**< memory pool for the tensors */
+};
+
+} // namespace nntrainer
+
+#endif /* __cplusplus */
+#endif /* __TENSOR_POOL_H__ */