COMPMID-1740: Fuse batch normalization with Convolution Layer at graph level

Change-Id: I77ca51c2c72783cc26a099a6a9c3210cdbbe822d
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/797
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

2 files changed, 195 insertions, 5 deletions

diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h
index 7242bc6ede..d0035d9a84 100644
--- a/arm_compute/graph/backends/FunctionHelpers.h
+++ b/arm_compute/graph/backends/FunctionHelpers.h
@@ -28,6 +28,7 @@
 #include "arm_compute/graph/Tensor.h"
 #include "arm_compute/graph/TypePrinter.h"
 #include "arm_compute/graph/Types.h"
+#include "arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h"
 #include "arm_compute/graph/backends/Utils.h"
 #include "arm_compute/graph/nodes/Nodes.h"
 
@@ -135,11 +136,12 @@ std::unique_ptr<IFunction> create_batch_normalization_layer(BatchNormalizationLa
     validate_node<TargetInfo>(node, 5 /* expected inputs */, 1 /* expected outputs */);
 
     // Extract IO and info
-    typename TargetInfo::TensorType *input     = get_backing_tensor<TargetInfo>(node.input(0));
-    typename TargetInfo::TensorType *mean      = get_backing_tensor<TargetInfo>(node.input(1));
-    typename TargetInfo::TensorType *var       = get_backing_tensor<TargetInfo>(node.input(2));
-    typename TargetInfo::TensorType *beta      = get_backing_tensor<TargetInfo>(node.input(3));
-    typename TargetInfo::TensorType *gamma     = get_backing_tensor<TargetInfo>(node.input(4));
+    typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
+    typename TargetInfo::TensorType *mean  = get_backing_tensor<TargetInfo>(node.input(1));
+    typename TargetInfo::TensorType *var   = get_backing_tensor<TargetInfo>(node.input(2));
+    typename TargetInfo::TensorType *beta  = get_backing_tensor<TargetInfo>(node.input(3));
+    typename TargetInfo::TensorType *gamma = get_backing_tensor<TargetInfo>(node.input(4));
+
     typename TargetInfo::TensorType *output    = get_backing_tensor<TargetInfo>(node.output(0));
     const float                      epsilon   = node.epsilon();
     const ActivationLayerInfo        fused_act = node.fused_activation();
@@ -163,6 +165,61 @@ std::unique_ptr<IFunction> create_batch_normalization_layer(BatchNormalizationLa
     return std::move(func);
 }
 
+/** Create a backend batch normalization layer function
+ *
+ * @tparam BatchNormalizationLayerFunction Backend batch normalization function
+ * @tparam TargetInfo                      Target-specific information
+ *
+ * @param[in] node Node to create the backend function for
+ *
+ * @return Backend batch normalization layer function
+ */
+template <typename FusedLayerTypes, typename TargetInfo>
+std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_layer(FusedConvolutionBatchNormalizationNode &node)
+{
+    validate_node<TargetInfo>(node, 7 /* expected inputs */, 1 /* expected outputs */);
+
+    // Extract IO and info
+    typename TargetInfo::TensorType *input   = get_backing_tensor<TargetInfo>(node.input(0));
+    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 *mean    = get_backing_tensor<TargetInfo>(node.input(3));
+    typename TargetInfo::TensorType *var     = get_backing_tensor<TargetInfo>(node.input(4));
+    typename TargetInfo::TensorType *beta    = get_backing_tensor<TargetInfo>(node.input(5));
+    typename TargetInfo::TensorType *gamma   = get_backing_tensor<TargetInfo>(node.input(6));
+
+    typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
+
+    const PadStrideInfo       conv_info  = node.convolution_info();
+    const unsigned int        num_groups = node.num_groups();
+    const bool                fast_math  = node.fast_math_hint() == FastMathHint::Enabled;
+    const ActivationLayerInfo fused_act  = node.fused_activation();
+    const float               epsilon    = node.epsilon();
+
+    const bool is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
+    if(is_quantized && biases != nullptr)
+    {
+        biases->info()->set_data_type(DataType::S32);
+    }
+
+    // Create and configure function
+    auto func = support::cpp14::make_unique<FusedConvolutionBatchNormalizationFunction<TargetInfo, FusedLayerTypes>>();
+    func->configure(input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, num_groups, fast_math, fused_act);
+
+    // Log info
+    ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
+                               << node.name()
+                               << " Type: " << node.name()
+                               << " Target: " << TargetInfo::TargetType
+                               << " Data Type: " << input->info()->data_type()
+                               << " Input shape: " << input->info()->tensor_shape()
+                               << " Weights shape: " << weights->info()->tensor_shape()
+                               << " Output shape: " << output->info()->tensor_shape()
+                               << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "")
+                               << std::endl);
+    return std::move(func);
+}
+
 /** Create a backend bounding box transform layer function
  *
  * @tparam BoundingBoxTransformLayerFunction    Backend bounding box transform function
diff --git a/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h
new file mode 100644
index 0000000000..92af17b227
--- /dev/null
+++ b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h
@@ -0,0 +1,133 @@
+/*
+ * 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_CONVOLUTION_BATCH_NORMAZLIZATION_FUNCTION_H__
+#define __ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_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}ConvolutionLayer with the modified weights */
+template <typename TargetInfo, typename FusedLayerTypes>
+class FusedConvolutionBatchNormalizationFunction : public IFunction
+{
+public:
+    using TensorType         = typename TargetInfo::TensorType;
+    using TensorConcreteType = typename TargetInfo::TensorConcreteType;
+
+    FusedConvolutionBatchNormalizationFunction()
+        : _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: QASYMM8/F16/F32.
+     * @param[in]  weights    Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
+     * @param[in]  bias       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
+     *                        Data type supported: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 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]  num_groups Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout
+     * @param[in]  fast_math  Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
+     *                        available which may introduce a drop of accuracy as well. Default is false
+     * @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 num_groups, bool fast_math, 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);
+            bias_to_use = bias;
+        }
+        else
+        {
+            _fused_batch_norm_layer.configure(weights, mean, var, nullptr, &_fused_bias, nullptr, beta, gamma, epsilon);
+            bias_to_use = &_fused_bias;
+        }
+
+        _conv_layer.configure(input, weights, bias_to_use, output, conv_info, WeightsInfo(), Size2D(1U, 1U), fused_act, fast_math, num_groups);
+
+        if(!has_bias)
+        {
+            _fused_bias.allocator()->allocate();
+        }
+    }
+
+    // Inherited methods overridden:
+    void run()
+    {
+        prepare();
+        _conv_layer.run();
+    }
+
+    void prepare()
+    {
+        if(!_is_prepared)
+        {
+            _fused_batch_norm_layer.run();
+            _is_prepared = true;
+        }
+    }
+
+private:
+    typename FusedLayerTypes::ConvolutionLayer       _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_CONVOLUTION_BATCH_NORMAZLIZATION_FUNCTION_H__ */