COMPMID-1342 Add grouping support to CLIm2ColKernel

Change-Id: I4afb19751520a90fee27fb49b775cd10e92a94f5
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/140476
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

4 files changed, 34 insertions, 15 deletions

diff --git a/arm_compute/core/CL/kernels/CLIm2ColKernel.h b/arm_compute/core/CL/kernels/CLIm2ColKernel.h
index ae19319047..c678f277cb 100644
--- a/arm_compute/core/CL/kernels/CLIm2ColKernel.h
+++ b/arm_compute/core/CL/kernels/CLIm2ColKernel.h
@@ -76,8 +76,10 @@ public:
      * @param[in]  conv_info   Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in]  has_bias    In case biases are provided expands the matrix with 1.
      * @param[in]  dilation    (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in]  num_groups  (Optional) Number of groups when performing a grouped convolution
      */
-    void configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U));
+    void configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U),
+                   unsigned int num_groups = 1);
     /** Static function to check if given info will lead to a valid configuration of @ref CLIm2ColKernel
      *
      * @param[in] input       The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
@@ -88,10 +90,12 @@ public:
      * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in] has_bias    In case biases are provided expands the matrix with 1.
      * @param[in] dilation    (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in] num_groups  (Optional) Number of groups when performing a grouped convolution
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U));
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U),
+                           unsigned int num_groups = 1);
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
@@ -103,6 +107,7 @@ public:
     unsigned int  _num_elems_processed_per_iteration;
     Size2D        _kernel_dims;
     PadStrideInfo _conv_info;
+    unsigned int  _num_groups;
 };
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_CLIM2COLKERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEIm2ColKernel.h b/arm_compute/core/NEON/kernels/NEIm2ColKernel.h
index 19da7cfd53..37145a38c1 100644
--- a/arm_compute/core/NEON/kernels/NEIm2ColKernel.h
+++ b/arm_compute/core/NEON/kernels/NEIm2ColKernel.h
@@ -83,12 +83,13 @@ public:
      * @param[in]  kernel_dims        The kernel dimensions (width and height).
      * @param[in]  conv_info          Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in]  has_bias           In case biases are provided expands the matrix with 1.
+     * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in]  num_groups         (Optional) Number of groups when performing a grouped convolution
      * @param[in]  is_fully_connected (Optional) Determines whether this kernel will be called by @ref NEFullyConnectedLayer in order to validate the arguments
      * @param[in]  is_flatten         (Optional) Determines whether this kernel will be called by @ref NEFlattenLayer in order to validate the arguments
-     * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
      */
     void configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info,
-                   bool has_bias, bool is_fully_connected = false, bool is_flatten = false, const Size2D &dilation = Size2D(1U, 1U));
+                   bool has_bias, const Size2D &dilation = Size2D(1U, 1U), unsigned int num_groups = 1, bool is_fully_connected = false, bool is_flatten = false);
     /** Static function to check if given info will lead to a valid configuration of @ref NEIm2ColKernel
      *
      * @param[in] input              The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
@@ -98,14 +99,15 @@ public:
      * @param[in] kernel_dims        The kernel dimensions (width and height).
      * @param[in] conv_info          Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in] has_bias           In case biases are provided expands the matrix with 1.
-     * @param[in] is_fully_connected (Optional) Determines whether this kernel will be called by @ref NEFullyConnectedLayer in order to validate the arguments
-     * @param[in] is_flatten         (Optional) Determines whether this kernel will be called by @ref NEFlattenLayer in order to validate the arguments
      * @param[in] dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in] num_groups         (Optional) Number of groups when performing a grouped convolution
+     * @param[in] is_fully_connected (Optional)Determines whether this kernel will be called by @ref NEFullyConnectedLayer in order to validate the arguments
+     * @param[in] is_flatten         (Optional) Determines whether this kernel will be called by @ref NEFlattenLayer in order to validate the arguments
      *
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info,
-                           bool has_bias, bool is_fully_connected = false, bool is_flatten = false, const Size2D &dilation = Size2D(1U, 1U));
+                           bool has_bias, const Size2D &dilation = Size2D(1U, 1U), unsigned int num_groups = 1, bool is_fully_connected = false, bool is_flatten = false);
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index f726ce9ad3..fc6abf95f3 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -192,9 +192,15 @@ inline TensorShape compute_deconvolution_shape(const ITensorInfo &input, unsigne
 
     return scale_out_shape;
 }
-inline TensorShape compute_im2col_conv_shape(const ITensorInfo *input, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation, bool batch_size_on_z)
+inline TensorShape compute_im2col_conv_shape(const ITensorInfo *input, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation, bool batch_size_on_z,
+                                             unsigned int num_groups = 1)
 {
-    // The output shape will be the 2D shape used as input for GEMM [ out_channels * kernel_area, num_elems_per_out_channel ]
+    // The output shape will be the 3D shape [ out_channels * kernel_area, num_elems_per_out_channel, batches ]                           if batch_size_on_z == true
+    //                       or the 4D shape [ out_channels * kernel_area / num_groups, num_elems_per_out_channel, num_groups, batches ]  if batch_size_on_z == false
+
+    ARM_COMPUTE_ERROR_ON(num_groups == 0);
+    ARM_COMPUTE_ERROR_ON(num_groups > 1 && input->data_layout() != DataLayout::NCHW);
+    ARM_COMPUTE_ERROR_ON(num_groups > 1 && batch_size_on_z);
 
     TensorShape output_shape{ input->tensor_shape() };
 
@@ -204,7 +210,7 @@ inline TensorShape compute_im2col_conv_shape(const ITensorInfo *input, const Siz
     const int        channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
 
     std::pair<unsigned int, unsigned int> out_dims = scaled_dimensions(output_shape[width_idx], output_shape[height_idx], kernel_dims.width, kernel_dims.height, conv_info, dilation);
-    output_shape.set(0, (output_shape[channel_idx] * kernel_dims.area() + (has_bias ? 1 : 0)));
+    output_shape.set(0, (output_shape[channel_idx] / num_groups * kernel_dims.area() + (has_bias ? 1 : 0))); // NOLINT
     output_shape.set(1, (out_dims.first * out_dims.second));
     if(batch_size_on_z && output_shape.num_dimensions() >= 3)
     {
@@ -212,7 +218,7 @@ inline TensorShape compute_im2col_conv_shape(const ITensorInfo *input, const Siz
     }
     else
     {
-        output_shape.set(2, 1);
+        output_shape.set(2, num_groups);
     }
 
     return output_shape;
diff --git a/arm_compute/runtime/NEON/functions/NEIm2Col.h b/arm_compute/runtime/NEON/functions/NEIm2Col.h
index d888b7e8f5..9df4f070d8 100644
--- a/arm_compute/runtime/NEON/functions/NEIm2Col.h
+++ b/arm_compute/runtime/NEON/functions/NEIm2Col.h
@@ -49,10 +49,13 @@ public:
      * @param[in]  kernel_dims        The kernel dimensions (width and height).
      * @param[in]  conv_info          Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in]  has_bias           In case biases are provided expands the matrix with 1.
+     * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in]  num_groups         (Optional) Number of groups when performing a grouped convolution
      * @param[in]  is_fully_connected (Optional) Determines whether this function will be called by @ref NEFullyConnectedLayer in order to validate the arguments
      * @param[in]  is_flatten         (Optional) Determines whether this function will be called by @ref NEFlattenLayer in order to validate the arguments
      */
-    void configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, bool is_fully_connected = false, bool is_flatten = false);
+    void configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U),
+                   unsigned int num_groups = 1, bool is_fully_connected = false, bool is_flatten = false);
     /** Static function to check if given info will lead to a valid configuration of @ref NEIm2Col
      *
      * @param[in] input              The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
@@ -62,12 +65,15 @@ public:
      * @param[in] kernel_dims        The kernel dimensions (width and height).
      * @param[in] conv_info          Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in] has_bias           In case biases are provided expands the matrix with 1.
-     * @param[in] is_fully_connected Determines whether this function will be called by @ref NEFullyConnectedLayer in order to validate the arguments
-     * @param[in] is_flatten         Determines whether this function will be called by @ref NEFlattenLayer in order to validate the arguments
+     * @param[in] dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in] num_groups         (Optional) Number of groups when performing a grouped convolution
+     * @param[in] is_fully_connected (Optional) Determines whether this function will be called by @ref NEFullyConnectedLayer in order to validate the arguments
+     * @param[in] is_flatten         (Optional) Determines whether this function will be called by @ref NEFlattenLayer in order to validate the arguments
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, bool is_fully_connected, bool is_flatten);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U),
+                           unsigned int num_groups = 1, bool is_fully_connected = false, bool is_flatten = false);
 
     // Inherited methods overridden:
     void run() override;