COMPMID-998: Release unused trainable parameters.

Change-Id: I361a520f34080016a25bc86e1e6789777c5152c1
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/124432
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

16 files changed, 113 insertions, 9 deletions

diff --git a/arm_compute/core/ITensor.h b/arm_compute/core/ITensor.h
index 1ef9c6d3f6..255a236861 100644
--- a/arm_compute/core/ITensor.h
+++ b/arm_compute/core/ITensor.h
@@ -83,6 +83,16 @@ public:
      * @param io_fmt Format information
      */
     void print(std::ostream &s, IOFormatInfo io_fmt = IOFormatInfo()) const;
+    /** Flags if the tensor is used or not
+     *
+     * @return True if it is used else false
+     */
+    bool is_used() const;
+    /** Marks a tensor as unused */
+    void mark_as_unused() const;
+
+private:
+    mutable bool _is_used = { true }; /**< Flag that marks if the tensor is used or not */
 };
 
 using IImage = ITensor;
diff --git a/arm_compute/graph2/ITensorHandle.h b/arm_compute/graph2/ITensorHandle.h
index 68f79d8b01..c92a213121 100644
--- a/arm_compute/graph2/ITensorHandle.h
+++ b/arm_compute/graph2/ITensorHandle.h
@@ -30,25 +30,34 @@ namespace arm_compute
 {
 namespace graph2
 {
-/** Tensor handle interface object **/
+/** Tensor handle interface object */
 class ITensorHandle
 {
 public:
-    /** Default virtual destructor **/
+    /** Default virtual destructor */
     virtual ~ITensorHandle() = default;
-    /** Allocates backend memory for the handle **/
+    /** Allocates backend memory for the handle */
     virtual void allocate() = 0;
-    /** Backend tensor object accessor **/
+    /** Backend tensor object accessor */
     virtual arm_compute::ITensor &tensor() = 0;
-    /** Backend tensor object const accessor **/
+    /** Backend tensor object const accessor */
     virtual const arm_compute::ITensor &tensor() const = 0;
     /** Maps backend tensor object
      *
      * @param[in] blocking Flags if the mapping operations should be blocking
      */
     virtual void map(bool blocking) = 0;
-    /** Un-maps a backend tensor object **/
+    /** Un-maps a backend tensor object */
     virtual void unmap() = 0;
+    /** Releases backend tensor if is marked as unused
+     *
+     *
+     * @note This has no effect on sub-tensors
+     * @warning Parent tensors don't keep track of sub-tensors,
+     *          thus if a parent is set as unused then all sub-tensors will be invalidated,
+     *          on the other hand if a sub-tensor is marked as unused then the parent tensor won't be released
+     */
+    virtual void release_if_unused() = 0;
     /** Checks if a backing tensor is a sub-tensor object or not
      *
      * @return True if the backend tensor is a sub-tensor else false
diff --git a/arm_compute/graph2/backends/CL/CLSubTensorHandle.h b/arm_compute/graph2/backends/CL/CLSubTensorHandle.h
index 9910980e59..6f3c00c041 100644
--- a/arm_compute/graph2/backends/CL/CLSubTensorHandle.h
+++ b/arm_compute/graph2/backends/CL/CLSubTensorHandle.h
@@ -59,6 +59,7 @@ public:
     const arm_compute::ITensor &tensor() const override;
     void map(bool blocking) override;
     void unmap() override;
+    void release_if_unused() override;
     bool is_subtensor() const override;
 
 private:
diff --git a/arm_compute/graph2/backends/CL/CLTensorHandle.h b/arm_compute/graph2/backends/CL/CLTensorHandle.h
index 37d7147b6b..0b20d1d8fc 100644
--- a/arm_compute/graph2/backends/CL/CLTensorHandle.h
+++ b/arm_compute/graph2/backends/CL/CLTensorHandle.h
@@ -56,6 +56,7 @@ public:
     const arm_compute::ITensor &tensor() const override;
     void map(bool blocking) override;
     void unmap() override;
+    void release_if_unused() override;
     bool is_subtensor() const override;
 
 private:
diff --git a/arm_compute/graph2/backends/GLES/GCTensorHandle.h b/arm_compute/graph2/backends/GLES/GCTensorHandle.h
index 8ead236614..281adee428 100644
--- a/arm_compute/graph2/backends/GLES/GCTensorHandle.h
+++ b/arm_compute/graph2/backends/GLES/GCTensorHandle.h
@@ -56,6 +56,7 @@ public:
     const arm_compute::ITensor &tensor() const override;
     void map(bool blocking) override;
     void unmap() override;
+    void release_if_unused() override;
     bool is_subtensor() const override;
 
 private:
diff --git a/arm_compute/graph2/backends/NEON/NESubTensorHandle.h b/arm_compute/graph2/backends/NEON/NESubTensorHandle.h
index eacdfe0fb4..d62b66f343 100644
--- a/arm_compute/graph2/backends/NEON/NESubTensorHandle.h
+++ b/arm_compute/graph2/backends/NEON/NESubTensorHandle.h
@@ -59,6 +59,7 @@ public:
     const arm_compute::ITensor &tensor() const override;
     void map(bool blocking) override;
     void unmap() override;
+    void release_if_unused() override;
     bool is_subtensor() const override;
 
 private:
diff --git a/arm_compute/graph2/backends/NEON/NETensorHandle.h b/arm_compute/graph2/backends/NEON/NETensorHandle.h
index c22fcdf216..23fd7ccc69 100644
--- a/arm_compute/graph2/backends/NEON/NETensorHandle.h
+++ b/arm_compute/graph2/backends/NEON/NETensorHandle.h
@@ -56,6 +56,7 @@ public:
     const arm_compute::ITensor &tensor() const override;
     void map(bool blocking) override;
     void unmap() override;
+    void release_if_unused() override;
     bool is_subtensor() const override;
 
 private:
diff --git a/arm_compute/graph2/detail/ExecutionHelpers.h b/arm_compute/graph2/detail/ExecutionHelpers.h
index e4523ecf47..bc3cfd5d72 100644
--- a/arm_compute/graph2/detail/ExecutionHelpers.h
+++ b/arm_compute/graph2/detail/ExecutionHelpers.h
@@ -63,6 +63,11 @@ void validate_all_nodes(Graph &g);
  * @return The execution workload
  */
 ExecutionWorkload configure_all_nodes(Graph &g, GraphContext &ctx);
+/** Release the memory of all unused const nodes
+ *
+ * @param[in] g Graph to release the memory from
+ */
+void release_unused_tensors(Graph &g);
 /** Calls accessor of a given tensor
  *
  * @param[in] tensor The tensor of which the accessor should be called
diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
index eb12fe4cca..d6fc8f0fcc 100644
--- a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
@@ -83,6 +83,14 @@ class CLDepthwiseConvolutionLayer : public IFunction
 public:
     /** Default constructor */
     CLDepthwiseConvolutionLayer();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLDepthwiseConvolutionLayer(const CLDepthwiseConvolutionLayer &) = delete;
+    /** Default move constructor */
+    CLDepthwiseConvolutionLayer(CLDepthwiseConvolutionLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLDepthwiseConvolutionLayer &operator=(const CLDepthwiseConvolutionLayer &) = delete;
+    /** Default move assignment operator */
+    CLDepthwiseConvolutionLayer &operator=(CLDepthwiseConvolutionLayer &&) = default;
     /** Initialize the function's source, destination, weights and convolution information.
      *
      * @param[in, out] input     Source tensor. Data type supported: QASYMM8/F32. (Written to only for border filling).
@@ -109,7 +117,9 @@ private:
     CLTensor                                  _weights_reshaped;
     CLTensor                                  _v2mm_output;
     CLTensor                                  _output_reshaped;
+    bool                                      _is_first_run;
     bool                                      _is_quantized;
+    const ICLTensor                          *_original_weights;
 };
 }
 #endif /*__ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
index 1e9ee492ad..ad821318e9 100644
--- a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -76,6 +76,14 @@ class CLFullyConnectedLayer : public IFunction
 public:
     /** Constructor */
     CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLFullyConnectedLayer(const CLFullyConnectedLayer &) = delete;
+    /** Default move constructor */
+    CLFullyConnectedLayer(CLFullyConnectedLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLFullyConnectedLayer &operator=(const CLFullyConnectedLayer &) = delete;
+    /** Default move assignment operator */
+    CLFullyConnectedLayer &operator=(CLFullyConnectedLayer &&) = default;
     /** Set the input and output tensors.
      *
      * @param[in]  input                Source tensor. Data type supported: QS8/QASYMM8/QS16/F16/F32.
@@ -121,6 +129,7 @@ private:
     bool                                                _is_fc_after_conv;
     bool                                                _accumulate_biases;
     bool                                                _is_quantized;
+    const ICLTensor                                    *_original_weights;
 };
 }
 #endif /* __ARM_COMPUTE_CLFULLYCONNECTEDLAYER_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
index dc1211a51a..91d35ca0b2 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
@@ -103,6 +103,14 @@ public:
      * @param[in] memory_manager (Optional) Memory manager.
      */
     CLGEMMConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGEMMConvolutionLayer(const CLGEMMConvolutionLayer &) = delete;
+    /** Default move constructor */
+    CLGEMMConvolutionLayer(CLGEMMConvolutionLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGEMMConvolutionLayer &operator=(const CLGEMMConvolutionLayer &) = delete;
+    /** Default move assignment operator */
+    CLGEMMConvolutionLayer &operator=(CLGEMMConvolutionLayer &&) = default;
     /** Set the input and output tensors.
      *
      * @param[in]  input        Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
@@ -172,6 +180,8 @@ private:
     CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
     CLCol2ImKernel                                      _col2im_kernel;
 
+    const ICLTensor *_original_weights;
+
     CLTensor _im2col_output;
     CLTensor _weights_reshaped;
     CLTensor _gemm_output;
diff --git a/arm_compute/runtime/CL/functions/CLLocallyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLLocallyConnectedLayer.h
index e5aafd8547..b7b2587454 100644
--- a/arm_compute/runtime/CL/functions/CLLocallyConnectedLayer.h
+++ b/arm_compute/runtime/CL/functions/CLLocallyConnectedLayer.h
@@ -53,6 +53,14 @@ class CLLocallyConnectedLayer : public IFunction
 public:
     /** Default constructor */
     CLLocallyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLLocallyConnectedLayer(const CLLocallyConnectedLayer &) = delete;
+    /** Default move constructor */
+    CLLocallyConnectedLayer(CLLocallyConnectedLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLLocallyConnectedLayer &operator=(const CLLocallyConnectedLayer &) = delete;
+    /** Default move assignment operator */
+    CLLocallyConnectedLayer &operator=(CLLocallyConnectedLayer &&) = default;
     /** Set the input and output tensors.
      *
      * @param[in]  input     Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
@@ -93,6 +101,7 @@ private:
     CLTensor                               _weights_reshaped;
     CLTensor                               _gemm_output;
     bool                                   _is_first_run;
+    const ICLTensor                       *_original_weights;
 };
 }
 #endif /* __ARM_COMPUTE_CLLOCALLYCONNECTEDLAYER_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
index f5805f5ad1..fe65ac1a43 100644
--- a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
@@ -97,6 +97,14 @@ class NEDepthwiseConvolutionLayer : public IFunction
 public:
     /** Default constructor */
     NEDepthwiseConvolutionLayer();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDepthwiseConvolutionLayer(const NEDepthwiseConvolutionLayer &) = delete;
+    /** Default move constructor */
+    NEDepthwiseConvolutionLayer(NEDepthwiseConvolutionLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDepthwiseConvolutionLayer &operator=(const NEDepthwiseConvolutionLayer &) = delete;
+    /** Default move assignment operator */
+    NEDepthwiseConvolutionLayer &operator=(NEDepthwiseConvolutionLayer &&) = default;
     /** Initialize the function's source, destination, weights and convolution information.
      *
      * @param[in, out] input     Source tensor. Data type supported: QASYMM8/F32. (Written to only for border filling).
@@ -123,7 +131,9 @@ private:
     Tensor                                    _weights_reshaped;
     Tensor                                    _v2mm_output;
     Tensor                                    _output_reshaped;
+    bool                                      _is_first_run;
     bool                                      _is_quantized;
+    const ITensor                            *_original_weights;
 };
 }
 #endif /* __ARM_COMPUTE_NEDEPTHWISECONVOLUTION_H__ */
\ No newline at end of file
diff --git a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
index 9bc8d21fc4..071eecc3f7 100644
--- a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
@@ -94,6 +94,14 @@ class NEFullyConnectedLayer : public IFunction
 public:
     /** Constructor */
     NEFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEFullyConnectedLayer(const NEFullyConnectedLayer &) = delete;
+    /** Default move constructor */
+    NEFullyConnectedLayer(NEFullyConnectedLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEFullyConnectedLayer &operator=(const NEFullyConnectedLayer &) = delete;
+    /** Default move assignment operator */
+    NEFullyConnectedLayer &operator=(NEFullyConnectedLayer &&) = default;
     /** Set the input and output tensors.
      *
      * @param[in]  input                Source tensor. Data type supported: QS8/QS16/F32.
@@ -134,6 +142,7 @@ private:
     bool                                _is_batched_fc_layer;
     bool                                _linearize_input;
     bool                                _accumulate_biases;
+    const ITensor                      *_original_weights;
 };
-}
+} // namespace arm_compute
 #endif /* __ARM_COMPUTE_NEFULLYCONNECTEDLAYER_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
index 43e9304414..e733fec4b6 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
@@ -101,7 +101,14 @@ class NEGEMMConvolutionLayer : public IFunction
 public:
     /** Constructor */
     NEGEMMConvolutionLayer(const std::shared_ptr<IMemoryManager> &memory_manager = nullptr);
-
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEGEMMConvolutionLayer(const NEGEMMConvolutionLayer &) = delete;
+    /** Default move constructor */
+    NEGEMMConvolutionLayer(NEGEMMConvolutionLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEGEMMConvolutionLayer &operator=(const NEGEMMConvolutionLayer &) = delete;
+    /** Default move assignment operator */
+    NEGEMMConvolutionLayer &operator=(NEGEMMConvolutionLayer &&) = default;
     /** Set the input and output tensors.
      *
      * @param[in]  input        Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
@@ -165,6 +172,8 @@ private:
     NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
     NECol2ImKernel                                      _output_col2im_kernel;
 
+    const ITensor *_original_weights;
+
     Tensor _input_im2col_reshaped;
     Tensor _input_interleaved_reshaped;
     Tensor _weights_reshaped;
diff --git a/arm_compute/runtime/NEON/functions/NELocallyConnectedLayer.h b/arm_compute/runtime/NEON/functions/NELocallyConnectedLayer.h
index e07f09c588..18cd27414e 100644
--- a/arm_compute/runtime/NEON/functions/NELocallyConnectedLayer.h
+++ b/arm_compute/runtime/NEON/functions/NELocallyConnectedLayer.h
@@ -53,6 +53,14 @@ class NELocallyConnectedLayer : public IFunction
 public:
     /** Default constructor */
     NELocallyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NELocallyConnectedLayer(const NELocallyConnectedLayer &) = delete;
+    /** Default move constructor */
+    NELocallyConnectedLayer(NELocallyConnectedLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NELocallyConnectedLayer &operator=(const NELocallyConnectedLayer &) = delete;
+    /** Default move assignment operator */
+    NELocallyConnectedLayer &operator=(NELocallyConnectedLayer &&) = default;
     /** Set the input and output tensors.
      *
      * @param[in]  input     Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
@@ -93,6 +101,7 @@ private:
     Tensor                                 _weights_reshaped;
     Tensor                                 _gemm_output;
     bool                                   _is_first_run;
+    const ITensor                         *_original_weights;
 };
 }
 #endif /* __ARM_COMPUTE_NELOCALLYCONNECTEDLAYER_H__ */