COMPMID-1401 Implement NEFullyConnectedLayer for QASYMM8

Change-Id: I0404df6d369855e2f458f2db8f26e81c80a1ee87
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/140148
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

1 files changed, 56 insertions, 57 deletions

diff --git a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
index ea0762ea79..92ca17a3a4 100644
--- a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
@@ -26,66 +26,47 @@
 
 #include "arm_compute/runtime/IFunction.h"
 
-#include "arm_compute/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.h"
-#include "arm_compute/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h"
-#include "arm_compute/core/NEON/kernels/NEGEMMTranspose1xWKernel.h"
 #include "arm_compute/core/NEON/kernels/NEIm2ColKernel.h"
 #include "arm_compute/core/NEON/kernels/NETransposeKernel.h"
 #include "arm_compute/runtime/MemoryGroup.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMM.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h"
 #include "arm_compute/runtime/Tensor.h"
 
 namespace arm_compute
 {
 /** Basic function to reshape the weights of Fully Connected layer with NEON. This function calls the following kernels:
  *
- *  -# @ref NETransposeKernel        (if @p transpose_weights is set to true)
- *  -# @ref NEGEMMTranspose1xWKernel (if @p is_batched_fc_layer is set to true)
+ *  -# @ref NETransposeKernel
  *
  * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
  */
-class NEFullyConnectedLayerReshapeWeights : public IFunction
+class NEFullyConnectedLayerReshapeWeights : public INESimpleFunction
 {
 public:
-    /** Constructor */
-    NEFullyConnectedLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Set the input and output tensors.
      *
-     * @param[in]  input               Weights tensor. The weights must be 2 dimensional. Data types supported: F32.
-     * @param[out] output              Destination tensor. Data type supported: Same as @p input.
-     * @param[in]  transpose_weights   True if the weights must be transposed. Data types supported: Same as @p weights.
-     * @param[in]  is_batched_fc_layer True if it is a batched fully connected layer
+     * @param[in]  input  Weights tensor. The weights must be 2 dimensional. Data types supported: QASYMM8/F16/F32.
+     * @param[out] output Destination tensor. Data type supported: Same as @p input.
      */
-    void configure(const ITensor *input, ITensor *output, bool transpose_weights, bool is_batched_fc_layer);
+    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: F32.
-     * @param[in] output              Destination tensor info. Data type supported: Same as @p input.
-     * @param[in] transpose_weights   True if the weights must be transposed. Data types supported: Same as @p weights.
-     * @param[in] is_batched_fc_layer True if it is a batched fully connected layer
+     * @param[in] input  Weights tensor info. The weights must be 2 dimensional. Data types supported: QASYMM8/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, bool transpose_weights, bool is_batched_fc_layer);
-
-    // Inherited methods overridden:
-    void run() override;
-
-private:
-    MemoryGroup              _memory_group;
-    NETransposeKernel        _transpose_kernel;
-    NEGEMMTranspose1xWKernel _transpose1xW_kernel;
-    Tensor                   _transpose_output;
-    bool                     _transpose_weights;
-    bool                     _is_batched_fc_layer;
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output);
 };
 
 /** Basic function to compute a Fully Connected layer on NEON. This function calls the following NEON kernels:
- *  -# @ref NEIm2ColKernel                      (called when the input comes from a convolutional layer)
- *  -# @ref NEFullyConnectedLayerReshapeWeights (if @p are_weights_reshaped flag is set to false) (called once)
- *  -# @ref NEGEMMInterleave4x4Kernel (called if we have a multi-batch input)
- *  -# @ref NEGEMMMatrixMultiplyKernel
- *  -# @ref NEGEMMMatrixAccumulateBiasesKernel (if @p biases is not equal to nullptr)
+ *  -# @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 NEGEMMMatrixMultiplyKernel or @ref NEGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
+ *  -# @ref NEGEMMMatrixAccumulateBiasesKernel or @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if quantized asymmetric) (if @p biases is not equal to nullptr)
  *
  * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
  */
@@ -104,21 +85,33 @@ public:
     NEFullyConnectedLayer &operator=(NEFullyConnectedLayer &&) = default;
     /** Set the input and output tensors.
      *
-     * @param[in]  input   Source tensor. Data type supported: F16/F32.
-     * @param[in]  weights Weights tensor. The weights must be 2 dimensional. Data type supported: Same as @p input.
+     * @param[in]  input   Source tensor. Data type supported: QASYMM8/F16/F32.
+     * @param[in]  weights Weights tensor. The weights must be 2 dimensional.
+     *                     If this function is called after a Convolution Layer, the (transposed) weights will have as many rows as the product of the first 3 input's dimensions.
+     *                     If it is called after another FullyConnected Layer, the (transposed) weights will have as many rows as the input's first dimension.
+     *                     Data type supported: Same as @p input.
      * @param[in]  biases  Bias tensor. Can be nullptr. Data type supported:Same as @p input.
-     * @param[out] output  Destination tensor. Data type supported: Same as @p input.
+     * @param[out] output  Destination tensor. Its shape should be equal to the output of a matrix multiplication between:
+     *                     - The output of im2col on the input and the (transposed) 2D weights, if the function is called after a Convolution Layer
+     *                     - The input tensor and the (transposed) 2D weights, if the function is called after another FullyConnected Layer.
+     *                     Data type supported: Same as @p input.
      * @param[in]  fc_info (Optional) Fully connected layer additional info
      */
     void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output,
                    FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
-    /** Static function to check if given info will lead to a valid configuration of @ref CLFullyConnectedLayer
+    /** Static function to check if given info will lead to a valid configuration of @ref NEFullyConnectedLayer
      *
-     * @param[in] input   Source tensor info. Data type supported: F16/F32.
-     * @param[in] weights Weights tensor info. The weights must be 2 dimensional. Data type supported: Same as @p input
-     * @param[in] biases  Bias tensor info. It can be nullptr. Data type supported:Same as @p input.
-     * @param[in] output  Destination tensor info. Data type supported: Same as @p input.
-     * @param[in] fc_info (Optional) Fully connected layer additional info
+     * @param[in]  input   Source tensor info. Data type supported: QASYMM8/F16/F32.
+     * @param[in]  weights Weights tensor info. The weights must be 2 dimensional.
+     *                     If this function is called after a Convolution Layer, the (transposed) weights will have as many rows as the product of the first 3 input's dimensions.
+     *                     If it is called after another FullyConnected Layer, the (transposed) weights will have as many rows as the input's first dimension.
+     *                     Data type supported: Same as @p input.
+     * @param[in]  biases  Bias tensor info. Can be nullptr. Data type supported:Same as @p input.
+     * @param[out] output  Destination tensor info. Its shape should be equal to the output of a matrix multiplication between:
+     *                     - The output of im2col on the input and the (transposed) 2D weights, if the function is called after a Convolution Layer
+     *                     - The input tensor and the (transposed) 2D weights, if the function is called after another FullyConnected Layer.
+     *                     Data type supported: Same as @p input.
+     * @param[in]  fc_info (Optional) Fully connected layer additional info
      *
      * @return a status
      */
@@ -130,20 +123,26 @@ public:
     void prepare() override;
 
 private:
-    MemoryGroup                         _memory_group;
-    NEIm2ColKernel                      _im2col_kernel;
-    NEFullyConnectedLayerReshapeWeights _reshape_weights_function;
-    NEGEMMInterleave4x4Kernel           _interleave4x4_kernel;
-    NEGEMMMatrixMultiplyKernel          _mm_kernel;
-    NEGEMMMatrixAccumulateBiasesKernel  _accumulate_biases_kernel;
-    Tensor                              _im2col_output;
-    Tensor                              _interleave4x4_output;
-    Tensor                              _reshape_weights_output;
-    const ITensor                      *_original_weights;
-    bool                                _is_batched_fc_layer;
-    bool                                _linearize_input;
-    bool                                _accumulate_biases;
-    bool                                _is_prepared;
+    void configure_fc_fc(const ITensor *input, const ITensor *weights, ITensor *output);
+    void configure_conv_fc(const ITensor *input, const ITensor *weights, ITensor *output);
+    void configure_mm(const ITensor *input, const ITensor *weights, ITensor *output);
+
+    MemoryGroup                                         _memory_group;
+    NEIm2ColKernel                                      _im2col_kernel;
+    NEFullyConnectedLayerReshapeWeights                 _reshape_weights_function;
+    NEGEMM                                              _mm_gemm;
+    NEGEMMLowpMatrixMultiplyCore                        _mm_gemmlowp;
+    NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
+    NEGEMMMatrixAccumulateBiasesKernel                  _accumulate_biases_kernel;
+    Tensor                                              _im2col_output;
+    Tensor                                              _gemmlowp_output;
+    Tensor                                              _reshape_weights_output;
+    const ITensor                                      *_original_weights;
+    bool                                                _are_weights_reshaped;
+    bool                                                _is_fc_after_conv;
+    bool                                                _accumulate_biases;
+    bool                                                _is_quantized;
+    bool                                                _is_prepared;
 };
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_NEFULLYCONNECTEDLAYER_H__ */