COMPMID-2273: Fuse Batch Normalization with Depthwise Convolution layer at graph level (only for CL)

Change-Id: I1d941c6e66722f39583bf68148c980bb28ff89a1
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1423
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 131 insertions, 0 deletions

diff --git a/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h b/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h
new file mode 100644
index 0000000000..6f70d3c3a0
--- /dev/null
+++ b/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h
@@ -0,0 +1,131 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef __ARM_COMPUTE_GRAPH_BACKENDS_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_FUNCTION_H__
+#define __ARM_COMPUTE_GRAPH_BACKENDS_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_FUNCTION_H__
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/IFunction.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+namespace backends
+{
+/** Wrapper function to first apply {NE, CL}BatchNormalizationLayer on the weights and then run {NE, CL}DepthwiseConvolutionLayer with the modified weights */
+template <typename TargetInfo, typename FusedLayerTypes>
+class FusedDepthwiseConvolutionBatchNormalizationFunction : public IFunction
+{
+public:
+    using TensorType         = typename TargetInfo::TensorType;
+    using TensorConcreteType = typename TargetInfo::TensorConcreteType;
+
+    FusedDepthwiseConvolutionBatchNormalizationFunction()
+        : _depth_conv_layer(), _fused_batch_norm_layer(), _fused_bias(), _is_prepared(false)
+    {
+    }
+
+    /** Set the input and output tensors.
+     *
+     * @param[in]  input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
+     *                              while every optional dimension from 4 and above represent a batch of inputs.
+     *                              Data types supported: F16/F32.
+     * @param[in]  weights          Weights tensor.  These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+     * @param[in]  bias             Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [IFM].
+     *                              Data type supported: Should match @p input data type.
+     * @param[out] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+     *                              Data types supported: Same as @p input.
+     * @param[in]  mean             Mean values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+     * @param[in]  var              Variance values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+     * @param[in]  beta             Beta values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for beta is 0. Data types supported: Same as @p input
+     * @param[in]  gamma            Gamma values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for gamma is 1. Data types supported: Same as @p input
+     * @param[in]  epsilon          Small value to avoid division with zero. Default value is 0.001f.
+     * @param[in]  conv_info        Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in]  depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+     * @param[in]  fused_act        Activation layer information in case of a fused activation.
+     *
+     */
+    void configure(TensorType       *input,
+                   TensorType       *weights,
+                   TensorType       *bias,
+                   TensorType       *output,
+                   const TensorType *mean,
+                   const TensorType *var,
+                   const TensorType *beta,
+                   const TensorType *gamma,
+                   float epsilon, const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo const &fused_act)
+    {
+        // We don't run any validate, as we assume that the layers have been already validated
+        const bool        has_bias = (bias != nullptr);
+        const TensorType *bias_to_use;
+
+        // We check if the layer has a bias. If yes, use it in-place. If not, we need to create one
+        // as batch normalization might end up with a bias != 0
+        if(has_bias)
+        {
+            _fused_batch_norm_layer.configure(weights, mean, var, nullptr, nullptr, bias, beta, gamma, epsilon, FuseBatchNormalizationType::DEPTHWISECONVOLUTION);
+            bias_to_use = bias;
+        }
+        else
+        {
+            _fused_batch_norm_layer.configure(weights, mean, var, nullptr, &_fused_bias, nullptr, beta, gamma, epsilon, FuseBatchNormalizationType::DEPTHWISECONVOLUTION);
+            bias_to_use = &_fused_bias;
+        }
+
+        _depth_conv_layer.configure(input, weights, bias_to_use, output, conv_info, depth_multiplier, fused_act.enabled() ? fused_act : ActivationLayerInfo());
+
+        if(!has_bias)
+        {
+            _fused_bias.allocator()->allocate();
+        }
+    }
+
+    // Inherited methods overridden:
+    void run()
+    {
+        prepare();
+        _depth_conv_layer.run();
+    }
+
+    void prepare()
+    {
+        if(!_is_prepared)
+        {
+            _fused_batch_norm_layer.run();
+            _is_prepared = true;
+        }
+    }
+
+private:
+    typename FusedLayerTypes::DepthwiseConvolutionLayer _depth_conv_layer;
+    typename FusedLayerTypes::FuseBatchNormalization    _fused_batch_norm_layer;
+    TensorConcreteType                                  _fused_bias;
+    bool                                                _is_prepared;
+};
+} // namespace backends
+} // namespace graph
+} // namespace arm_compute
+
+#endif /* __ARM_COMPUTE_GRAPH_BACKENDS_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_FUNCTION_H__ */