COMPMID-3374: Remove memory state from NEConcatenateLayer kernels

* Allow the following kernels to accept backing memory at run-time:
  * NEBatchConcatenateLayerKernel
  * NEDepthConcatenateLayerKernel
  * NEHeightConcatenateLayerKernel
  * NEWidthConcatenateLayerKernel
* Allow the following functions to accept backing memory at run-time:
  * NEConcatenateLayer

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: Ib0b6714cff7f06a52dc74d294bc3e0d72a1c2419
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3569
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

8 files changed, 87 insertions, 42 deletions

diff --git a/arm_compute/core/NEON/kernels/NEBatchConcatenateLayerKernel.h b/arm_compute/core/NEON/kernels/NEBatchConcatenateLayerKernel.h
index 267211fad6..f397a29b48 100644
--- a/arm_compute/core/NEON/kernels/NEBatchConcatenateLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEBatchConcatenateLayerKernel.h
@@ -56,15 +56,15 @@ public:
     ~NEBatchConcatenateLayerKernel() = default;
     /** Initialise the kernel's inputs and output
      *
-     * @param[in]     input        Input tensor. Data types supported: All.
+     * @param[in]     input        Input tensor info. Data types supported: All.
      * @param[in]     batch_offset The offset on axis # 3.
-     * @param[in,out] output       Output tensor. Data types supported: Same as @p input.
+     * @param[in,out] output       Output tensor info. Data types supported: Same as @p input.
      *
      * @note: The output tensor's low two dimensions can't be smaller than the input one's.
      * @note: The gaps between the two lowest dimensions of input and output need to be divisible by 2.
      *
      */
-    void configure(const ITensor *input, unsigned int batch_offset, ITensor *output);
+    void configure(const ITensorInfo *input, unsigned int batch_offset, ITensorInfo *output);
     /**  Static function to check if given info will lead to a valid configuration of @ref NEBatchConcatenateLayerKernel
      *
      * @param[in] input        Input tensor info. Data types supported: All.
@@ -76,15 +76,14 @@ public:
     static Status validate(const ITensorInfo *input, unsigned int batch_offset, const ITensorInfo *output);
 
     // Inherited methods overridden:
-    void run(const Window &window, const ThreadInfo &info) override;
+    void run_op(const InputTensorMap &inputs, const OutputTensorMap &outputs,
+                const Window &window, const ThreadInfo &info) override;
 
 private:
     using BatchConcatFunction = void(const ITensor *in, ITensor *out, unsigned int batch_offset, const Window &window);
 
 private:
     BatchConcatFunction *_func;
-    const ITensor       *_input;
-    ITensor             *_output;
     unsigned int         _batch_offset;
 };
 } // namespace arm_compute
diff --git a/arm_compute/core/NEON/kernels/NEDepthConcatenateLayerKernel.h b/arm_compute/core/NEON/kernels/NEDepthConcatenateLayerKernel.h
index a465146184..e1aaa59f25 100644
--- a/arm_compute/core/NEON/kernels/NEDepthConcatenateLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEDepthConcatenateLayerKernel.h
@@ -56,15 +56,15 @@ public:
     ~NEDepthConcatenateLayerKernel() = default;
     /** Initialise the kernel's inputs and output
      *
-     * @param[in]     input        Input tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
+     * @param[in]     input        Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[in]     depth_offset The offset on the Z axis.
-     * @param[in,out] output       Output tensor. Data types supported: Same as @p input.
+     * @param[in,out] output       Output tensor info. Data types supported: Same as @p input.
      *
      * @note: The output tensor's low two dimensions can't be smaller than the input one's.
      * @note: The gaps between the two lowest dimensions of input and output need to be divisible by 2.
      *
      */
-    void configure(const ITensor *input, unsigned int depth_offset, ITensor *output);
+    void configure(const ITensorInfo *input, unsigned int depth_offset, ITensorInfo *output);
     /**  Static function to check if given info will lead to a valid configuration of @ref NEDepthConcatenateLayerKernel
      *
      * @param[in] input        Input tensor info. Data types supported:  QASYMM8/QASYMM8_SIGNED/F16/F32.
@@ -76,15 +76,14 @@ public:
     static Status validate(const ITensorInfo *input, unsigned int depth_offset, const ITensorInfo *output);
 
     // Inherited methods overridden:
-    void run(const Window &window, const ThreadInfo &info) override;
+    void run_op(const InputTensorMap &inputs, const OutputTensorMap &outputs,
+                const Window &window, const ThreadInfo &info) override;
 
 private:
     using DepthConcatFunction = void(const ITensor *in, ITensor *out, unsigned int depth_offset, const Window &window);
 
 private:
     DepthConcatFunction *_func;
-    const ITensor       *_input;
-    ITensor             *_output;
     unsigned int         _depth_offset;
 };
 } // namespace arm_compute
diff --git a/arm_compute/core/NEON/kernels/NEHeightConcatenateLayerKernel.h b/arm_compute/core/NEON/kernels/NEHeightConcatenateLayerKernel.h
index e8a9d9ad94..d463b53e2c 100644
--- a/arm_compute/core/NEON/kernels/NEHeightConcatenateLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEHeightConcatenateLayerKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 Arm Limited.
+ * Copyright (c) 2019-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -57,12 +57,12 @@ public:
     ~NEHeightConcatenateLayerKernel() = default;
     /** Initialise the kernel's inputs and output
      *
-     * @param[in]     input         Input tensor. Data types supported: All
+     * @param[in]     input         Input tensor info. Data types supported: All
      * @param[in]     height_offset The starting offset on the Y axis for the output tensor.
-     * @param[in,out] output        Output tensor. Data types supported: Same as @p input.
+     * @param[in,out] output        Output tensor info. Data types supported: Same as @p input.
      *
      */
-    void configure(const ITensor *input, unsigned int height_offset, ITensor *output);
+    void configure(const ITensorInfo *input, unsigned int height_offset, ITensorInfo *output);
     /**  Static function to check if given info will lead to a valid configuration of @ref NEHeightConcatenateLayerKernel
      *
      * @param[in] input         Input tensor info. Data types supported: All
@@ -74,12 +74,11 @@ public:
     static Status validate(const ITensorInfo *input, unsigned int height_offset, const ITensorInfo *output);
 
     // Inherited methods overridden:
-    void run(const Window &window, const ThreadInfo &info) override;
+    void run_op(const InputTensorMap &inputs, const OutputTensorMap &outputs,
+                const Window &window, const ThreadInfo &info) override;
 
 private:
-    const ITensor *_input;
-    ITensor       *_output;
-    unsigned int   _height_offset;
+    unsigned int _height_offset;
 };
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_NEHEIGHTCONCATENATELAYERKERNEL_H */
diff --git a/arm_compute/core/NEON/kernels/NEWidthConcatenateLayerKernel.h b/arm_compute/core/NEON/kernels/NEWidthConcatenateLayerKernel.h
index 442d35c656..b5336ad026 100644
--- a/arm_compute/core/NEON/kernels/NEWidthConcatenateLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEWidthConcatenateLayerKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 Arm Limited.
+ * Copyright (c) 2018-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -57,12 +57,11 @@ public:
     ~NEWidthConcatenateLayerKernel() = default;
     /** Initialise the kernel's inputs and output
      *
-     * @param[in]     input        Input tensor. Data types supported: All
+     * @param[in]     input        Input tensor info. Data types supported: All
      * @param[in]     width_offset The offset on the X axis.
-     * @param[in,out] output       Output tensor. Data types supported: Same as @p input.
-     *
+     * @param[in,out] output       Output tensor info. Data types supported: Same as @p input.
      */
-    void configure(const ITensor *input, unsigned int width_offset, ITensor *output);
+    void configure(const ITensorInfo *input, unsigned int width_offset, ITensorInfo *output);
     /**  Static function to check if given info will lead to a valid configuration of @ref NEWidthConcatenateLayerKernel
      *
      * @param[in] input        Input tensor info. Data types supported: All
@@ -74,12 +73,11 @@ public:
     static Status validate(const ITensorInfo *input, unsigned int width_offset, const ITensorInfo *output);
 
     // Inherited methods overridden:
-    void run(const Window &window, const ThreadInfo &info) override;
+    void run_op(const InputTensorMap &inputs, const OutputTensorMap &outputs,
+                const Window &window, const ThreadInfo &info) override;
 
 private:
-    const ITensor *_input;
-    ITensor       *_output;
-    unsigned int   _width_offset;
+    unsigned int _width_offset;
 };
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_NEWIDTHCONCATENATELAYERKERNEL_H */
diff --git a/arm_compute/core/experimental/Types.h b/arm_compute/core/experimental/Types.h
index a009c0dab0..eca833e708 100644
--- a/arm_compute/core/experimental/Types.h
+++ b/arm_compute/core/experimental/Types.h
@@ -34,7 +34,7 @@ namespace arm_compute
 class ITensor;
 
 /** Memory type */
-enum class TensorType
+enum TensorType : int32_t
 {
     ACL_UNKNOWN = -1,
     ACL_SRC     = 0,
@@ -47,11 +47,12 @@ enum class TensorType
     ACL_INT     = 50,
     ACL_INT_0   = 50,
     ACL_INT_1   = 51,
-    ACL_INT_2   = 52
+    ACL_INT_2   = 52,
+    ACL_SRC_VEC = 256,
 };
 
-using InputTensorMap    = std::map<TensorType, const ITensor *>;
-using OutputTensorMap   = std::map<TensorType, ITensor *>;
+using InputTensorMap    = std::map<int32_t, const ITensor *>;
+using OutputTensorMap   = std::map<int32_t, ITensor *>;
 using OperatorTensorMap = OutputTensorMap;
 
 namespace experimental
diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h
index fecaa9d79c..af748341a5 100644
--- a/arm_compute/graph/backends/FunctionHelpers.h
+++ b/arm_compute/graph/backends/FunctionHelpers.h
@@ -383,7 +383,7 @@ std::unique_ptr<arm_compute::IFunction> create_concatenate_layer(ConcatenateLaye
     }
 
     // Extract IO and info
-    std::vector<typename TargetInfo::TensorType *> inputs;
+    std::vector<typename TargetInfo::SrcTensorType *> inputs;
     for(unsigned int i = 0; i < node.num_inputs(); ++i)
     {
         inputs.push_back(get_backing_tensor<TargetInfo>(node.input(i)));
diff --git a/arm_compute/runtime/NEON/INEOperator.h b/arm_compute/runtime/NEON/INEOperator.h
index 4f42efde7f..004abb245f 100644
--- a/arm_compute/runtime/NEON/INEOperator.h
+++ b/arm_compute/runtime/NEON/INEOperator.h
@@ -54,7 +54,7 @@ public:
     INEOperator &operator=(INEOperator &&) = default;
 
     // Inherited methods overridden:
-    void run(InputTensorMap inputs, OutputTensorMap outputs, OperatorTensorMap workspace) override final;
+    void run(InputTensorMap inputs, OutputTensorMap outputs, OperatorTensorMap workspace) override;
     void prepare(OperatorTensorMap constants) override final;
 
 protected:
diff --git a/arm_compute/runtime/NEON/functions/NEConcatenateLayer.h b/arm_compute/runtime/NEON/functions/NEConcatenateLayer.h
index b5c200b467..02c27e20e6 100644
--- a/arm_compute/runtime/NEON/functions/NEConcatenateLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEConcatenateLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 Arm Limited.
+ * Copyright (c) 2018-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -29,6 +29,7 @@
 #include "arm_compute/core/NEON/INEKernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/utils/misc/Requires.h"
+#include "arm_compute/runtime/NEON/INEOperator.h"
 
 #include <memory>
 #include <vector>
@@ -52,6 +53,16 @@ class NEConcatenateLayer : public IFunction
 public:
     /** Default constructor */
     NEConcatenateLayer();
+    /** Destructor */
+    ~NEConcatenateLayer();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEConcatenateLayer(const NEConcatenateLayer &) = delete;
+    /** Default move constructor */
+    NEConcatenateLayer(NEConcatenateLayer &&);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEConcatenateLayer &operator=(const NEConcatenateLayer &) = delete;
+    /** Default move assignment operator */
+    NEConcatenateLayer &operator=(NEConcatenateLayer &&);
     /** Initialise the kernel's inputs vector and output.
      *
      * @note Input and output tensor dimensions preconditions defer depending on the concatenation axis.
@@ -61,7 +72,6 @@ public:
      * @param[out]    output        Output tensor. Data types supported: Same as @p input.
      * @param[in]     axis          Concatenation axis. Supported underlying concatenation axis are 0, 1, 2 and 3.
      */
-    void configure(std::vector<ITensor *> inputs_vector, ITensor *output, size_t axis);
     void configure(std::vector<const ITensor *> inputs_vector, ITensor *output, size_t axis);
     /** Static function to check if given info will lead to a valid configuration of @ref NEConcatenateLayer
      *
@@ -74,23 +84,62 @@ public:
      *
      * @return a status
      */
-    static Status validate(const std::vector<ITensorInfo *> &inputs_vector, const ITensorInfo *output, size_t axis);
     static Status validate(const std::vector<const ITensorInfo *> &inputs_vector, const ITensorInfo *output, size_t axis);
 
     // Inherited methods overridden:
     void run() override;
 
 private:
-    template <typename TensorType, REQUIRES_TA(std::is_same<typename std::remove_cv<TensorType>::type, ITensor>::value)>
-    void configure_internal(std::vector<TensorType *> &&inputs_vector, ITensor *output, size_t axis);
+    struct Impl;
+    std::unique_ptr<Impl> _impl;
+};
+
+namespace experimental
+{
+/** Basic function to execute concatenate tensors along a given axis. This function calls the following kernels:
+ *
+ * -# @ref NEWidthConcatenateLayerKernel (if underlying concatenation axis is 0).
+ * -# @ref NEHeightConcatenateLayerKernel (if underlying concatenation axis is 1).
+ * -# @ref NEDepthConcatenateLayerKernel (if underlying concatenation axis is 2).
+ * -# @ref NEBatchConcatenateLayerKernel (if underlying concatenation axis is 3).
+ */
+class NEConcatenateLayer : public INEOperator
+{
+public:
+    /** Default constructor */
+    NEConcatenateLayer();
+    /** Initialise the kernel's inputs vector and output.
+     *
+     * @note Input and output tensor dimensions preconditions defer depending on the concatenation axis.
+     * @note Preconditions can be found respectively at @ref NEWidthConcatenateLayerKernel, @ref NEHeightConcatenateLayerKernel and @ref NEDepthConcatenateLayerKernel.
+     *
+     * @param[in,out] inputs_vector The vectors containing all the tensors to concatenate. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
+     * @param[out]    output        Output tensor. Data types supported: Same as @p input.
+     * @param[in]     axis          Concatenation axis. Supported underlying concatenation axis are 0, 1, 2 and 3.
+     */
+    void configure(const std::vector<const ITensorInfo *> &inputs_vector, ITensorInfo *output, size_t axis);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEConcatenateLayer
+     *
+     * @note Input and output tensor dimensions preconditions defer depending on the concatenation axis.
+     * @note Preconditions can be found respectively at @ref NEWidthConcatenateLayerKernel, @ref NEHeightConcatenateLayerKernel and @ref NEDepthConcatenateLayerKernel.
+     *
+     * @param[in] inputs_vector The vectors containing all the tensors info to concatenate. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
+     * @param[in] output        Output tensor info. Data types supported: Same as @p input.
+     * @param[in] axis          Concatenation axis. Supported underlying concatenation axis are 0, 1, 2 and 3.
+     *
+     * @return a status
+     */
+    static Status validate(const std::vector<const ITensorInfo *> &inputs_vector, const ITensorInfo *output, size_t axis);
 
-    template <typename TensorInfoType, REQUIRES_TA(std::is_same<typename std::remove_cv<TensorInfoType>::type, ITensorInfo>::value)>
-    static Status validate_internal(const std::vector<TensorInfoType *> &inputs_vector, const ITensorInfo *output, size_t axis);
+    // Inherited methods overridden:
+    MemoryRequirements workspace() const override;
+    void run(InputTensorMap inputs, OutputTensorMap outputs, OperatorTensorMap workspace) override;
 
 private:
     std::vector<std::unique_ptr<INEKernel>> _concat_kernels;
     unsigned int                            _num_inputs;
     unsigned int                            _axis;
 };
+} // namespace experimental
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_NECONCATENATELAYER_H */