COMPMID-2244: Extend CLFuseBatchNormalization to support DepthwiseConvolution weights

Change-Id: I7d1907f35cc4899379073759be2f7cce24e51e9d
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1327
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 39 insertions, 37 deletions

diff --git a/arm_compute/core/CL/kernels/CLFuseBatchNormalizationKernel.h b/arm_compute/core/CL/kernels/CLFuseBatchNormalizationKernel.h
index a5b98bb27d..d2df0897d4 100644
--- a/arm_compute/core/CL/kernels/CLFuseBatchNormalizationKernel.h
+++ b/arm_compute/core/CL/kernels/CLFuseBatchNormalizationKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -49,57 +49,59 @@ public:
     ~CLFuseBatchNormalizationKernel() = default;
     /** Set the source, destination of the kernel
      *
-     * @param[in]  conv_weights  Convolution layer weights tensor. Data type supported: F16/F32
-     * @param[in]  bn_mean       Batch normalization layer mean tensor. Same as @p conv_weights
-     * @param[in]  bn_var        Batch normalization layer variance tensor. Same as @p conv_weights
-     * @param[out] fused_weights Output fused weights tensor. It can be a nullptr in case of in-place computation. Same as @p conv_weights
-     * @param[out] fused_bias    Output fused bias tensor. It can be a nullptr in case of in-place computation and conv_bias != nullptr. Same as @p conv_weights
-     * @param[in]  conv_bias     (Optional) Convolution layer bias tensor. It can be a nullptr in case the bias tensor is not required. Same as @p conv_weights
-     * @param[in]  bn_beta       (Optional) Batch normalization layer beta tensor. It can be a nullptr in case the beta tensor is not required. Same as @p conv_weights
+     * @param[in]  input_weights Input weights tensor for convolution or depthwise convolution layer. Data type supported: F16/F32. Data layout supported: NCHW, NHWC
+     * @param[in]  bn_mean       Batch normalization layer mean tensor. Same as @p input_weights
+     * @param[in]  bn_var        Batch normalization layer variance tensor. Same as @p input_weights
+     * @param[out] fused_weights Output fused weights tensor. It can be a nullptr in case of in-place computation. Same as @p input_weights
+     * @param[out] fused_bias    Output fused bias tensor. It can be a nullptr in case of in-place computation and input_bias != nullptr. Same as @p input_weights
+     * @param[in]  input_bias    (Optional) Input bias tensor for convolution or depthwise convolution layer. It can be a nullptr in case the bias tensor is not required. Same as @p input_weights
+     * @param[in]  bn_beta       (Optional) Batch normalization layer beta tensor. It can be a nullptr in case the beta tensor is not required. Same as @p input_weights
      *                           @note if nullptr, bn_beta is set to 0.0
-     * @param[in]  bn_gamma      (Optional) Batch normalization layer gamma tensor. It can be a nullptr in case the gamma tensor is not required. Same as @p conv_weights
+     * @param[in]  bn_gamma      (Optional) Batch normalization layer gamma tensor. It can be a nullptr in case the gamma tensor is not required. Same as @p input_weights
      *                           @note if nullptr, bn_gamma is set to 1.0
      * @param[in]  epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
+     * @param[in]  fbn_type      (Optional) Fused batch normalization type. Defaults to CONVOLUTION.
      */
-    void configure(const ICLTensor *conv_weights, const ICLTensor *bn_mean, const ICLTensor *bn_var, ICLTensor *fused_weights, ICLTensor *fused_bias,
-                   const ICLTensor *conv_bias = nullptr, const ICLTensor *bn_beta = nullptr, const ICLTensor *bn_gamma = nullptr,
-                   float epsilon = 0.001f);
+    void configure(const ICLTensor *input_weights, const ICLTensor *bn_mean, const ICLTensor *bn_var, ICLTensor *fused_weights, ICLTensor *fused_bias,
+                   const ICLTensor *input_bias = nullptr, const ICLTensor *bn_beta = nullptr, const ICLTensor *bn_gamma = nullptr,
+                   float epsilon = 0.001f, FuseBatchNormalizationType fbn_type = FuseBatchNormalizationType::CONVOLUTION);
     /** Static function to check if given info will lead to a valid configuration of @ref CLFuseBatchNormalizationKernel
      *
-     * @param[in]  conv_weights  Convolution layer weights tensor info. Data type supported: F16/F32
-     * @param[in]  bn_mean       Batch normalization layer mean tensor info. Same as @p conv_weights
-     * @param[in]  bn_var        Batch normalization layer variance tensor info. Same as @p conv_weights
-     * @param[out] fused_weights Output fused weights tensor info. It can be a nullptr in case of in-place computation. Same as @p conv_weights
-     * @param[out] fused_bias    Output fused bias tensor info. It can be a nullptr in case of in-place computation and conv_bias != nullptr. Same as @p conv_weights
-     * @param[in]  conv_bias     (Optional) Convolution layer bias tensor info. It can be a nullptr in case the bias tensor is not required. Same as @p conv_weights
-     * @param[in]  bn_beta       (Optional) Batch normalization layer beta tensor info. It can be a nullptr in case the beta tensor is not required. Same as @p conv_weights
-     *                           @note if nullptr, bn_beta is set to 0.0
-     * @param[in]  bn_gamma      (Optional) Batch normalization layer gamma tensor info. It can be a nullptr in case the gamma tensor is not required. Same as @p conv_weights
-     *                           @note if nullptr, bn_gamma is set to 1.0
-     * @param[in]  epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
+     * @param[in] input_weights Input weights tensor info for convolution or depthwise convolution layer. Data type supported: F16/F32. Data layout supported: NCHW, NHWC
+     * @param[in] bn_mean       Batch normalization layer mean tensor info. Same as @p input_weights
+     * @param[in] bn_var        Batch normalization layer variance tensor info. Same as @p input_weights
+     * @param[in] fused_weights Output fused weights tensor info. It can be a nullptr in case of in-place computation. Same as @p input_weights
+     * @param[in] fused_bias    Output fused bias tensor info. It can be a nullptr in case of in-place computation and input_bias != nullptr. Same as @p input_weights
+     * @param[in] input_bias    (Optional) Input bias tensor info for convolution or depthwise convolution layer. It can be a nullptr in case the bias tensor is not required. Same as @p input_weights
+     * @param[in] bn_beta       (Optional) Batch normalization layer beta tensor info. It can be a nullptr in case the beta tensor is not required. Same as @p input_weights
+     *                          @note if nullptr, bn_beta is set to 0.0
+     * @param[in] bn_gamma      (Optional) Batch normalization layer gamma tensor info. It can be a nullptr in case the gamma tensor is not required. Same as @p input_weights
+     *                          @note if nullptr, bn_gamma is set to 1.0
+     * @param[in] epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
+     * @param[in] fbn_type      (Optional) Fused batch normalization type. Defaults to CONVOLUTION.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *conv_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
+    static Status validate(const ITensorInfo *input_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
                            const ITensorInfo *fused_weights, const ITensorInfo *fused_bias,
-                           const ITensorInfo *conv_bias = nullptr, const ITensorInfo *bn_beta = nullptr, const ITensorInfo *bn_gamma = nullptr,
-                           float epsilon = 0.001f);
+                           const ITensorInfo *input_bias = nullptr, const ITensorInfo *bn_beta = nullptr, const ITensorInfo *bn_gamma = nullptr,
+                           float epsilon = 0.001f, FuseBatchNormalizationType fbn_type = FuseBatchNormalizationType::CONVOLUTION);
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
 
 private:
-    const ICLTensor *_conv_weights;
-    const ICLTensor *_conv_bias;
-    const ICLTensor *_bn_mean;
-    const ICLTensor *_bn_var;
-    const ICLTensor *_bn_gamma;
-    const ICLTensor *_bn_beta;
-    ICLTensor       *_fused_weights;
-    ICLTensor       *_fused_bias;
-    float            _epsilon;
-    bool             _run_in_place_weights;
-    bool             _run_in_place_bias;
+    const ICLTensor           *_input_weights;
+    const ICLTensor           *_input_bias;
+    const ICLTensor           *_bn_mean;
+    const ICLTensor           *_bn_var;
+    const ICLTensor           *_bn_gamma;
+    const ICLTensor           *_bn_beta;
+    ICLTensor                 *_fused_weights;
+    ICLTensor                 *_fused_bias;
+    float                      _epsilon;
+    bool                       _run_in_place_weights;
+    bool                       _run_in_place_bias;
 };
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_CLFUSEBATCHNORMALIZATIONKERNEL_H__ */