graph/mutators/InPlaceOperationMutator.cpp

/*
 * Copyright (c) 2018-2021 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.
 */
#include "arm_compute/graph/mutators/InPlaceOperationMutator.h"

#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/graph/Graph.h"
#include "arm_compute/graph/Logger.h"
#include "arm_compute/graph/nodes/DepthwiseConvolutionLayerNode.h"
#include "arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h"
#include "support/Cast.h"

using namespace arm_compute::utils::cast;

namespace arm_compute
{
namespace graph
{
namespace
{
// Check if the output edges of the parent node are separate tensors. If not,
// it means the same output is connected to multiple nodes and computations on
// these nodes cannot be done in-place.
bool output_edges_are_separate_tensors(Graph &g, const Edge *input_edge)
{
    const auto parent_node   = input_edge->producer();
    const auto input_tensor  = input_edge->tensor();
    const auto input_edge_id = input_edge->id();

    if(parent_node == nullptr)
    {
        return false;
    }

    const auto output_edges = parent_node->output_edges();

    // If the output is connected to only one edge, then computations can
    // be done in-place.
    if(output_edges.size() == 1)
    {
        return true;
    }

    return std::all_of(output_edges.begin(),
                       output_edges.end(),
                       [&](const EdgeID & edge_id)
    {
        // Skip check on current input edge
        if(edge_id == input_edge_id)
        {
            return true;
        }

        auto edge = g.edge(edge_id);
        return edge->tensor() != input_tensor;
    });
}

// If do in-place calculation, then need to use the new output and inherit original output's accessor
void set_new_output_and_inherit_accessor(std::unique_ptr<INode> &node, Tensor *orig_output, Tensor *new_output)
{
    ARM_COMPUTE_LOG_GRAPH_INFO("Switching to in-place computation for the node with ID : "
                               << node->id() << " and name : " << node->name() << std::endl);
    // Update accessor
    new_output->set_accessor(orig_output->extract_accessor());
    // Update output
    node->set_output_tensor(new_output->id(), 0);
}

// Try to mutate the node to perform the depthwise in-place calculation
void try_in_place_depthwiseconv(std::unique_ptr<INode> &node)
{
    // Get input edge
    Edge *input_edge  = node->input_edge(0);
    Edge *weight_edge = node->input_edge(1);
    ARM_COMPUTE_ERROR_ON(input_edge == nullptr || weight_edge == nullptr);

    auto input_tensor  = input_edge->tensor();
    auto weight_tensor = weight_edge->tensor();
    ARM_COMPUTE_ERROR_ON(input_tensor == nullptr || weight_tensor == nullptr);

    const auto input_shape = input_tensor->desc().shape;
    const auto qinfo_input = input_tensor->desc().quant_info;

    const auto weight_shape  = weight_tensor->desc().shape;
    const auto weight_layout = weight_tensor->desc().layout;

    // Extract PadStrideInfo and depth multiplier
    PadStrideInfo conv_info{};
    unsigned int  depth_multiplier{};
    if(node->type() == NodeType::FusedDepthwiseConvolutionBatchNormalizationLayer)
    {
        conv_info        = polymorphic_downcast<FusedDepthwiseConvolutionBatchNormalizationNode *>(node.get())->convolution_info();
        depth_multiplier = polymorphic_downcast<FusedDepthwiseConvolutionBatchNormalizationNode *>(node.get())->depth_multiplier();
    }
    else if(node->type() == NodeType::DepthwiseConvolutionLayer)
    {
        conv_info        = polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node.get())->convolution_info();
        depth_multiplier = polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node.get())->depth_multiplier();
    }

    // Get current output tensor
    auto current_output_tensor = node->output(0);
    ARM_COMPUTE_ERROR_ON(current_output_tensor == nullptr);
    const auto out_shape = current_output_tensor->desc().shape;
    const auto qinfo_out = current_output_tensor->desc().quant_info;

    bool input_can_in_place = !arm_compute::detail::have_different_dimensions(out_shape, input_shape, 0) && (qinfo_input == qinfo_out) && (input_tensor->accessor() == nullptr);

    // Specify conditions with which input can be in-placed
    input_can_in_place &= weight_layout == input_tensor->desc().layout && weight_layout == DataLayout::NHWC;

    const int  weights_width_idx  = get_data_layout_dimension_index(weight_layout, DataLayoutDimension::WIDTH);
    const int  weights_height_idx = get_data_layout_dimension_index(weight_layout, DataLayoutDimension::HEIGHT);
    const bool is_1x1             = weight_shape[weights_width_idx] == 1U && weight_shape[weights_height_idx] == 1U;
    input_can_in_place &= is_1x1;

    input_can_in_place &= depth_multiplier == 1;
    input_can_in_place &= conv_info.stride() == std::make_pair(1U, 1U);
    input_can_in_place &= !conv_info.has_padding();
    // NOTE: Dilation should also be (1, 1). However currently dilation is not supported in the depthwise conv node

    if(input_can_in_place)
    {
        set_new_output_and_inherit_accessor(node, current_output_tensor, input_tensor);
    }
    else
    {
        ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented in-place operation as there is an accessor bound to the input tensor or the quantization info are different.\n");
    }
}

// Try to mutate the node to perform the elementwise in-place calculation
void try_in_place_elementwise(std::unique_ptr<INode> &node)
{
    // Get input edge
    Edge *input0_edge = node->input_edge(0);
    Edge *input1_edge = node->input_edge(1);
    ARM_COMPUTE_ERROR_ON(input0_edge == nullptr || input1_edge == nullptr);

    auto input0_tensor = input0_edge->tensor();
    auto input1_tensor = input1_edge->tensor();
    ARM_COMPUTE_ERROR_ON(input0_tensor == nullptr || input1_tensor == nullptr);

    const auto shape0 = input0_tensor->desc().shape;
    const auto shape1 = input1_tensor->desc().shape;
    const auto qinfo0 = input0_tensor->desc().quant_info;
    const auto qinfo1 = input1_tensor->desc().quant_info;

    const TensorShape out_shape = TensorShape::broadcast_shape(shape0, shape1);
    // Inputs are not broadcast compatible
    if(out_shape.total_size() == 0)
    {
        return;
    }

    // Get current output tensor
    auto current_output_tensor = node->output(0);
    ARM_COMPUTE_ERROR_ON(current_output_tensor == nullptr);
    const auto qinfo_out = current_output_tensor->desc().quant_info;

    // Can do in place, if the input has same shape as output, has same quntisation info as output, has same data type as output and input doesn't have accessor.
    bool input0_can_in_place = !arm_compute::detail::have_different_dimensions(out_shape, shape0, 0) && (qinfo0 == qinfo_out)
                               && (input0_tensor->desc().data_type == current_output_tensor->desc().data_type) && (input0_tensor->accessor() == nullptr);
    bool input1_can_in_place = !arm_compute::detail::have_different_dimensions(out_shape, shape1, 0) && (qinfo1 == qinfo_out)
                               && (input1_tensor->desc().data_type == current_output_tensor->desc().data_type) && (input1_tensor->accessor() == nullptr);

    if(input0_can_in_place)
    {
        set_new_output_and_inherit_accessor(node, current_output_tensor, input0_tensor);
    }
    else if(input1_can_in_place)
    {
        set_new_output_and_inherit_accessor(node, current_output_tensor, input1_tensor);
    }
    else
    {
        ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented in-place operation as there is an accessor bound to the input tensor or the quantization info are different.\n");
    }
}
} // namespace

const char *InPlaceOperationMutator::name()
{
    return "InPlaceOperationMutator";
}

IGraphMutator::MutationType InPlaceOperationMutator::type() const
{
    return IGraphMutator::MutationType::Backend;
}

void InPlaceOperationMutator::mutate(Graph &g)
{
    std::set<NodeType> in_place_nodes =
    {
        NodeType::ActivationLayer,
        NodeType::BatchNormalizationLayer,
        NodeType::EltwiseLayer,
        NodeType::UnaryEltwiseLayer,
        NodeType::DepthwiseConvolutionLayer,
        NodeType::FusedDepthwiseConvolutionBatchNormalizationLayer,
        NodeType::PrintLayer
    };

    // Not interested in the order of nodes
    for(auto &node : g.nodes())
    {
        if(node && in_place_nodes.find(node->type()) != std::end(in_place_nodes))
        {
            // Get input edge
            Edge *input_edge = node->input_edge(0);

            // Check if parent has a single output if yes then force in place calculation else not
            if((input_edge != nullptr) && output_edges_are_separate_tensors(g, input_edge))
            {
                if(node->type() == NodeType::EltwiseLayer)
                {
                    try_in_place_elementwise(node);
                }
                else if(node->type() == NodeType::FusedDepthwiseConvolutionBatchNormalizationLayer || node->type() == NodeType::DepthwiseConvolutionLayer)
                {
                    try_in_place_depthwiseconv(node);
                }
                else
                {
                    // Get current and new output tensors
                    auto current_output_tensor = node->output(0);
                    auto new_output_tensor     = input_edge->tensor();

                    ARM_COMPUTE_ERROR_ON(current_output_tensor == nullptr || new_output_tensor == nullptr);

                    // Prevent in-place operation if there is an accessor bound to the in-place tensor or quantization info are different
                    if(new_output_tensor->accessor() != nullptr || current_output_tensor->desc().quant_info != new_output_tensor->desc().quant_info)
                    {
                        ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented in-place operation as there is an accessor bound to the input tensor or the quantization info are different.\n");
                    }
                    else
                    {
                        set_new_output_and_inherit_accessor(node, current_output_tensor, new_output_tensor);
                    }
                }
            }
        }
    }
}
} // namespace graph
} // namespace arm_compute