COMPMID-1344 Add grouping support to CLWeightsReshapeKernel

Change-Id: Idde333308db71087ec234b3fd1eb4e36a44db46c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/143049
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

2 files changed, 28 insertions, 16 deletions

diff --git a/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h b/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h
index 664fc3c216..6c93c23cec 100644
--- a/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h
+++ b/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h
@@ -68,26 +68,30 @@ public:
     ~CLWeightsReshapeKernel() = default;
     /** Set the input and output of the kernel.
      *
-     * @param[in]  input  The input tensor to convert. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] if shared,
-     *                    and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM,  num_patches] if unshared. Data types supported: QASYMM8/F16/F32
-     * @param[in]  biases The shared biases tensor to append.  Bias is 1D tensor with dimensions [OFM] if shared and 2D tensor with
-     *                    dimensions [OFM, num_patches] if unshared. Data types supported: Same as @p input
-     *                    @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
-     * @param[out] output The output tensor. Should be a 2D Tensor. Data types supported: Same as @p input
+     * @param[in]  input      The input tensor to convert. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] if shared,
+     *                        and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM,  num_patches] if unshared. Data types supported: QASYMM8/F16/F32
+     * @param[in]  biases     The shared biases tensor to append.  Bias is 1D tensor with dimensions [OFM] if shared and 2D tensor with
+     *                        dimensions [OFM, num_patches] if unshared. Data types supported: Same as @p input
+     *                        @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
+     * @param[out] output     The output tensor. Should be a 2D Tensor if there are no groups and the weights are not shared; a 3D Tensor otherwise.
+     *                        Data types supported: Same as @p input
+     * @param[in]  num_groups (Optional) Number of groups when performing a grouped convolution
      */
-    void configure(const ICLTensor *input, const ICLTensor *biases, ICLTensor *output);
+    void configure(const ICLTensor *input, const ICLTensor *biases, ICLTensor *output, const unsigned int num_groups = 1);
     /** Static function to check if given info will lead to a valid configuration of @ref CLWeightsReshapeKernel
      *
-     * @param[in] input  The input tensor to convert. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] if shared,
-     *                   and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM,  num_patches] if unshared. Data types supported: QASYMM8/F16/F32
-     * @param[in] biases The shared biases tensor to append.  Bias is 1D tensor with dimensions [OFM] if shared and 2D tensor with
-     *                   dimensions [OFM, num_patches] if unshared. Data types supported: Same as @p input
-     *                   @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
-     * @param[in] output The output tensor. Should be a 2D Tensor. Data types supported: Same as @p input
+     * @param[in] input      The input tensor to convert. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] if shared,
+     *                       and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM,  num_patches] if unshared. Data types supported: QASYMM8/F16/F32
+     * @param[in] biases     The shared biases tensor to append.  Bias is 1D tensor with dimensions [OFM] if shared and 2D tensor with
+     *                       dimensions [OFM, num_patches] if unshared. Data types supported: Same as @p input
+     *                       @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
+     * @param[in] output     The output tensor. Should be a 2D Tensor if there are no groups and the weights are not shared; a 3D Tensor otherwise.
+     *                       Data types supported: Same as @p input
+     * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output, const unsigned int num_groups = 1);
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index ac83dcb2cd..1e5b9afd0e 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -55,14 +55,22 @@ inline TensorShape compute_permutation_output_shape(const ITensorInfo &input, co
     permute(output_shape, perm);
     return output_shape;
 }
-inline TensorShape compute_weights_reshaped_shape(const ITensorInfo &weights, bool has_bias = false)
+inline TensorShape compute_weights_reshaped_shape(const ITensorInfo &weights, bool has_bias = false, const unsigned int num_groups = 1)
 {
+    ARM_COMPUTE_ERROR_ON(num_groups == 0);
+    ARM_COMPUTE_ERROR_ON((weights.dimension(3) % num_groups) != 0);
+    ARM_COMPUTE_ERROR_ON(weights.data_layout() == DataLayout::NHWC && num_groups > 1);
+
     // Calculate output shape
     TensorShape weights_reshaped{ weights.tensor_shape() };
     weights_reshaped.collapse(3);
     const size_t tmp_dim = weights_reshaped[0];
-    weights_reshaped.set(0, weights_reshaped[1]);
+    weights_reshaped.set(0, weights_reshaped[1] / num_groups);
     weights_reshaped.set(1, tmp_dim + (has_bias ? 1 : 0));
+    if(weights.num_dimensions() < 5)
+    {
+        weights_reshaped.set(2, num_groups);
+    }
 
     return weights_reshaped;
 }