Substitute NEFullyConnectedLayerReshapeWeights by NETranspose

Resolves partially: COMPMID-4359 (2/2)

Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: Id65ef04268575cc9d74be6114e82e116b8ed106d
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5378
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 8 insertions, 84 deletions

diff --git a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
index 58b11744a8..bc45e58b4b 100644
--- a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
@@ -31,55 +31,15 @@
 #include "arm_compute/runtime/NEON/functions/NEFlattenLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEGEMM.h"
 #include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
+#include "arm_compute/runtime/NEON/functions/NETranspose.h"
 #include "arm_compute/runtime/Tensor.h"
 
 namespace arm_compute
 {
-/** Basic function to reshape the weights of Fully Connected layer. This function calls the following kernels:
- *
- * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
- */
-class NEFullyConnectedLayerReshapeWeights : public IFunction
-{
-public:
-    /** Constructor */
-    NEFullyConnectedLayerReshapeWeights();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEFullyConnectedLayerReshapeWeights(const NEFullyConnectedLayerReshapeWeights &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NEFullyConnectedLayerReshapeWeights &operator=(const NEFullyConnectedLayerReshapeWeights &) = delete;
-    /** Prevent instances of this class from being moved (As this class contains non movable objects) */
-    NEFullyConnectedLayerReshapeWeights(NEFullyConnectedLayerReshapeWeights &&) = delete;
-    /** Prevent instances of this class from being moved (As this class contains non movable objects) */
-    NEFullyConnectedLayerReshapeWeights &operator=(NEFullyConnectedLayerReshapeWeights &&) = delete;
-    /** Default destructor */
-    ~NEFullyConnectedLayerReshapeWeights();
-    /** Set the input and output tensors.
-     *
-     * @param[in]  input  Weights tensor. The weights must be 2 dimensional. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[out] output Destination tensor. Data type supported: Same as @p input.
-     */
-    void configure(const ITensor *input, ITensor *output);
-    /** Static function to check if given info will lead to a valid configuration of @ref NEFullyConnectedLayerReshapeWeights
-     *
-     * @param[in] input  Weights tensor info. The weights must be 2 dimensional. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in] output Destination tensor info. Data type supported: Same as @p input.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output);
-
-    // Inherited methods overridden
-    void run() override;
-
-private:
-    struct Impl;
-    std::unique_ptr<Impl> _impl;
-};
 
 namespace weights_transformations
 {
-/** Basic function to manage the reshape weights generated from @ref NEFullyConnectedLayerReshapeWeights */
+/** Basic function to manage the reshape weights generated from @ref NETranspose */
 class NEFullyConnectedLayerReshapeWeightsManaged : public ITransformWeights
 {
 public:
@@ -111,15 +71,15 @@ public:
     }
 
 private:
-    static constexpr uint32_t           _uid = 0x0;
-    Tensor                              _output{};
-    NEFullyConnectedLayerReshapeWeights _func{};
+    static constexpr uint32_t _uid = 0x0;
+    Tensor                    _output{};
+    NETranspose               _func{};
 };
 } // namespace weights_transformations
 
 /** Basic function to compute a Fully Connected layer. This function calls the following kernels:
  *  -# @ref NEIm2ColKernel (called when the input comes from a convolutional layer)
- *  -# @ref NEFullyConnectedLayerReshapeWeights (if @p are_weights_reshaped is set to false and transpose_weights is set to true ) (called once)
+ *  -# @ref NETranspose (if @p are_weights_reshaped is set to false and transpose_weights is set to true ) (called once)
  *  -# @ref NEGEMMMatrixMultiplyKernel or @ref NEGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
  *  -# @ref NEGEMMMatrixAdditionKernel or @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if quantized asymmetric) (if @p biases is not equal to nullptr)
  *
@@ -189,7 +149,7 @@ private:
     NEFlattenLayer                                                      _flatten;
     NEConvertFullyConnectedWeights                                      _convert_weights;
     weights_transformations::NEConvertFullyConnectedWeightsManaged      _convert_weights_managed;
-    NEFullyConnectedLayerReshapeWeights                                 _reshape_weights_function;
+    NETranspose                                                         _reshape_weights_function;
     weights_transformations::NEFullyConnectedLayerReshapeWeightsManaged _reshape_weights_managed_function;
     NEGEMM                                                              _mm_gemm;
     NEGEMMLowpMatrixMultiplyCore                                        _mm_gemmlowp;
diff --git a/docs/06_functions_list.dox b/docs/06_functions_list.dox
index 2cd16d0603..0b2263555d 100644
--- a/docs/06_functions_list.dox
+++ b/docs/06_functions_list.dox
@@ -67,7 +67,6 @@ namespace arm_compute
         - @ref NEDepthConvertLayer
         - @ref NEFlattenLayer
         - @ref NEFloor
-        - @ref NEFullyConnectedLayerReshapeWeights
         - @ref NEGather
         - @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
         - @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index 0a5318ac30..6ff102cd4a 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -142,41 +142,6 @@ Status validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const I
 }
 } // namespace
 
-struct NEFullyConnectedLayerReshapeWeights::Impl
-{
-    const ITensor                                    *src{ nullptr };
-    ITensor                                          *dst{ nullptr };
-    std::unique_ptr<cpu::kernels::CpuTransposeKernel> op{ nullptr };
-};
-
-NEFullyConnectedLayerReshapeWeights::NEFullyConnectedLayerReshapeWeights()
-    : _impl(std::make_unique<Impl>())
-{
-}
-
-NEFullyConnectedLayerReshapeWeights::~NEFullyConnectedLayerReshapeWeights() = default;
-
-void NEFullyConnectedLayerReshapeWeights::configure(const ITensor *input, ITensor *output)
-{
-    _impl->op = std::make_unique<cpu::kernels::CpuTransposeKernel>();
-    _impl->op->configure(input->info(), output->info());
-    _impl->src = input;
-    _impl->dst = output;
-}
-
-Status NEFullyConnectedLayerReshapeWeights::validate(const ITensorInfo *input, const ITensorInfo *output)
-{
-    return cpu::kernels::CpuTransposeKernel::validate(input, output);
-}
-
-void NEFullyConnectedLayerReshapeWeights::run()
-{
-    ITensorPack pack{};
-    pack.add_tensor(TensorType::ACL_SRC, _impl->src);
-    pack.add_tensor(TensorType::ACL_DST, _impl->dst);
-    NEScheduler::get().schedule_op(_impl->op.get(), Window::DimY, _impl->op->window(), pack);
-}
-
 NEFullyConnectedLayer::~NEFullyConnectedLayer() = default;
 
 NEFullyConnectedLayer::NEFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager)
@@ -392,7 +357,7 @@ Status NEFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorIn
     if(!weights_reshaped)
     {
         // Validate reshape weights kernel
-        ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayerReshapeWeights::validate(weights, &reshaped_weights));
+        ARM_COMPUTE_RETURN_ON_ERROR(NETranspose::validate(weights, &reshaped_weights));
         weights_to_use = &reshaped_weights;
     }