COMPMID-1386: Add support for converting weights for CL.

Change-Id: I62e3ead903366baeeb1488f233a9b8b0c388c9de
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/140403
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

10 files changed, 67 insertions, 54 deletions

diff --git a/arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h b/arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h
index b85f93e992..40c9dc826f 100644
--- a/arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h
+++ b/arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h
@@ -57,7 +57,7 @@ public:
      *
      * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32.
      * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
-     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer).
      * @param[in]  data_layout          The data layout the weights have been trained in.
      */
     void configure(const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
@@ -65,7 +65,7 @@ public:
      *
      * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32.
      * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
-     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer).
      * @param[in] data_layout          The data layout the weights have been trained in.
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
diff --git a/arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h b/arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h
index 1a276c353e..5b8d7fd457 100644
--- a/arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h
+++ b/arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h
@@ -61,7 +61,7 @@ public:
      *
      * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32.
      * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
-     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer).
      * @param[in]  data_layout          The data layout the weights have been trained in.
      */
     void configure(const ITensor *input, ITensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
@@ -69,7 +69,7 @@ public:
      *
      * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32.
      * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
-     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer).
      * @param[in] data_layout          The data layout the weights have been trained in.
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 1363324e3b..343952f0b2 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -682,6 +682,15 @@ private:
     DimensionRoundingType _round_type;
 };
 
+/** Fully connected layer info */
+struct FullyConnectedLayerInfo
+{
+    DataLayout weights_trained_layout{ DataLayout::NCHW }; /**<  Layout that the weights have been trained with. */
+    bool       transpose_weights{ true };                  /**<  Transpose weights if true. */
+    bool       are_weights_reshaped{ false };              /**<  Reshape the weights tensor if false. */
+    bool       retain_internal_weights{ false };           /**<  Retain internal reshaped weights. */
+};
+
 /** Pooling Layer Information class */
 class PoolingLayerInfo
 {
diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h
index 16536bcb65..978d3bc1a8 100644
--- a/arm_compute/graph/backends/FunctionHelpers.h
+++ b/arm_compute/graph/backends/FunctionHelpers.h
@@ -524,10 +524,11 @@ std::unique_ptr<IFunction> create_fully_connected_layer(FullyConnectedLayerNode
     typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1));
     typename TargetInfo::TensorType *biases  = get_backing_tensor<TargetInfo>(node.input(2));
     typename TargetInfo::TensorType *output  = get_backing_tensor<TargetInfo>(node.output(0));
+    const FullyConnectedLayerInfo    fc_info = node.info();
 
     // Create and configure function
     auto func = support::cpp14::make_unique<FullyConnectedLayerFunction>(get_memory_manager(ctx, TargetInfo::TargetType));
-    func->configure(input, weights, biases, output);
+    func->configure(input, weights, biases, output, fc_info);
     ARM_COMPUTE_ERROR_ON(input == nullptr);
     ARM_COMPUTE_ERROR_ON(weights == nullptr);
     ARM_COMPUTE_ERROR_ON(output == nullptr);
diff --git a/arm_compute/graph/nodes/FullyConnectedLayerNode.h b/arm_compute/graph/nodes/FullyConnectedLayerNode.h
index 3d1b68909a..79201c8bdc 100644
--- a/arm_compute/graph/nodes/FullyConnectedLayerNode.h
+++ b/arm_compute/graph/nodes/FullyConnectedLayerNode.h
@@ -37,8 +37,9 @@ public:
     /** Constructor
      *
      * @param[in] num_outputs Number of neurons in the layer
+     * @param[in] fc_info     (Optional) Additional information about the fully connected layer
      */
-    FullyConnectedLayerNode(unsigned int num_outputs);
+    FullyConnectedLayerNode(unsigned int num_outputs, FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
     /** Computes weights descriptor
      *
      * @warning Works for inputs with 1D batch space
@@ -59,6 +60,11 @@ public:
      * @return Output descriptor
      */
     static TensorDescriptor compute_output_descriptor(const TensorDescriptor &input_descriptor, unsigned int num_outputs);
+    /** Fully connected layer addition information
+     *
+     * @return Additional information about the fully connected layer
+     */
+    FullyConnectedLayerInfo info() const;
 
     // Inherited overridden methods:
     NodeType         type() const override;
@@ -67,7 +73,8 @@ public:
     void accept(INodeVisitor &v) override;
 
 private:
-    unsigned int _num_outputs;
+    unsigned int            _num_outputs;
+    FullyConnectedLayerInfo _info;
 };
 } // namespace graph
 } // namespace arm_compute
diff --git a/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h b/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h
index 77e9d15e7f..9bfade4894 100644
--- a/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h
+++ b/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h
@@ -39,7 +39,7 @@ public:
      *
      * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32.
      * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
-     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer).
      * @param[in]  data_layout          The data layout the weights have been trained in.
      */
     void configure(const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
@@ -47,7 +47,7 @@ public:
      *
      * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32.
      * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
-     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer).
      * @param[in] data_layout          The data layout the weights have been trained in.
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
diff --git a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
index 3357868968..6b8d7a97ec 100644
--- a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
@@ -31,6 +31,7 @@
 #include "arm_compute/core/CL/kernels/CLTransposeKernel.h"
 #include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h"
 #include "arm_compute/runtime/CL/functions/CLGEMM.h"
 #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
 #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
@@ -86,32 +87,26 @@ public:
     CLFullyConnectedLayer &operator=(CLFullyConnectedLayer &&) = default;
     /** Set the input and output tensors.
      *
-     * @param[in]  input                   Source tensor. Data type supported: QASYMM8/F16/F32.
-     * @param[in]  weights                 Weights tensor. The weights must be 2 dimensional. Data type supported: Same as @p input
-     * @param[in]  biases                  Bias tensor. It can be nullptr. Data type supported:Same as @p input.
-     * @param[out] output                  Destination tensor. Data type supported: Same as @p input.
-     * @param[in]  transpose_weights       (Optional) Transpose weights if true. Defaults to true.
-     * @param[in]  are_weights_reshaped    (Optional) Reshape the weights tensor if false. Defaults to false.
-     * @param[in]  retain_internal_weights (Optional) Retain internal reshaped weights. Defaults to false.
-     *                                     Used for reconfiguration purposes.
+     * @param[in]  input   Source tensor. Data type supported: QASYMM8/F16/F32.
+     * @param[in]  weights Weights tensor. The weights must be 2 dimensional. Data type supported: Same as @p input
+     * @param[in]  biases  Bias tensor. It can be nullptr. Data type supported:Same as @p input.
+     * @param[out] output  Destination tensor. Data type supported: Same as @p input.
+     * @param[in]  fc_info (Optional) Fully connected layer additional info
      */
-    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights = true, bool are_weights_reshaped = false,
-                   bool retain_internal_weights = false);
+    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
+                   FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
     /** Static function to check if given info will lead to a valid configuration of @ref CLFullyConnectedLayer
      *
-     * @param[in] input                   Source tensor. Data type supported: QASYMM8/F16/F32.
-     * @param[in] weights                 Weights tensor. The weights must be 2 dimensional. Data type supported: Same as @p input
-     * @param[in] biases                  Bias tensor. It can be nullptr. Data type supported:Same as @p input.
-     * @param[in] output                  Destination tensor. Data type supported: Same as @p input.
-     * @param[in] transpose_weights       (Optional) Transpose weights if true. Defaults to true.
-     * @param[in] are_weights_reshaped    (Optional) Reshape the weights tensor if false. Defaults to false.
-     * @param[in] retain_internal_weights (Optional) Retain internal reshaped weights. Defaults to false.
-     *                                    Used for reconfiguration purposes.
+     * @param[in] input   Source tensor. Data type supported: QASYMM8/F16/F32.
+     * @param[in] weights Weights tensor. The weights must be 2 dimensional. Data type supported: Same as @p input
+     * @param[in] biases  Bias tensor. It can be nullptr. Data type supported:Same as @p input.
+     * @param[in] output  Destination tensor. Data type supported: Same as @p input.
+     * @param[in] fc_info (Optional) Fully connected layer additional info
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose_weights = true, bool are_weights_reshaped = false,
-                           bool retain_internal_weights = false);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                           FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
 
     //Inherited methods override
     void run() override;
@@ -124,6 +119,7 @@ private:
 
     CLMemoryGroup                                       _memory_group;
     CLIm2ColKernel                                      _im2col_kernel;
+    CLConvertFullyConnectedWeights                      _convert_weights;
     CLFullyConnectedLayerReshapeWeights                 _reshape_weights_kernel;
     CLGEMM                                              _mm_gemm;
     CLGEMMLowpMatrixMultiplyCore                        _mm_gemmlowp;
@@ -131,11 +127,14 @@ private:
     CLGEMMMatrixAccumulateBiasesKernel                  _accumulate_biases_kernel;
     CLTensor                                            _im2col_output;
     CLTensor                                            _gemmlowp_output;
+    CLTensor                                            _converted_weights_output;
     CLTensor                                            _reshape_weights_output;
+    bool                                                _are_weights_converted;
     bool                                                _are_weights_reshaped;
     bool                                                _is_fc_after_conv;
     bool                                                _accumulate_biases;
     bool                                                _is_quantized;
+    bool                                                _is_prepared;
     const ICLTensor                                    *_original_weights;
 };
 }
diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.h
index cd108c3eab..63565df1a7 100644
--- a/arm_compute/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.h
+++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.h
@@ -75,17 +75,14 @@ public:
     GCFullyConnectedLayer &operator=(GCFullyConnectedLayer &&) = 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]  biases                  Bias tensor. It can be nullptr. Data type supported:Same as @p input.
-     * @param[out] output                  Destination tensor. Data type supported: Same as @p input.
-     * @param[in]  transpose_weights       (Optional) Transpose weights if true. Defaults to true.
-     * @param[in]  are_weights_reshaped    (Optional) Reshape the weights tensor if false. Defaults to false.
-     * @param[in]  retain_internal_weights (Optional) Retain internal reshaped weights. Defaults to false.
-     *                                     Used for reconfiguration purposes.
+     * @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]  biases  Bias tensor. It can be nullptr. Data type supported:Same as @p input.
+     * @param[out] output  Destination tensor. Data type supported: Same as @p input.
+     * @param[in]  fc_info (Optional) Fully connected layer additional info
      */
     void configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output,
-                   bool transpose_weights = true, bool are_weights_reshaped = false, bool retain_internal_weights = false);
+                   FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
 
     //Inherited methods override
     void run() override;
diff --git a/arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h b/arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h
index acbba28040..8f261421e6 100644
--- a/arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h
+++ b/arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h
@@ -42,7 +42,7 @@ public:
      *
      * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32.
      * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
-     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer).
      * @param[in]  data_layout          The data layout the weights have been trained in.
      */
     void configure(const ITensor *input, ITensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
@@ -50,7 +50,7 @@ public:
      *
      * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32.
      * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
-     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer).
      * @param[in] data_layout          The data layout the weights have been trained in.
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
diff --git a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
index 33ac8ecb8a..ea0762ea79 100644
--- a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
@@ -104,26 +104,26 @@ 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]  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[in]  transpose_weights    (Optional) Transpose the weights tensor if true. Defaults to true.
-     * @param[in]  are_weights_reshaped (Optional) Reshape the weights tensor if false. Defaults to false.
+     * @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]  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[in]  fc_info (Optional) Fully connected layer additional info
      */
-    void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose_weights = true, bool are_weights_reshaped = false);
+    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
      *
-     * @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] transpose_weights    (Optional) Transpose weights if true. Defaults to true.
-     * @param[in] are_weights_reshaped (Optional) Reshape the weights tensor if false. Defaults to false.
+     * @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
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose_weights = true, bool are_weights_reshaped = false);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                           FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
 
     //Inherited methods override
     void run() override;