tf.cpp

#include <google/protobuf/text_format.h>
#include <google/protobuf/io/zero_copy_stream_impl.h>
#include <graph.pb.h>
#include <iostream>
#include <fstream>
#include <unordered_map>
#include <unordered_set>
#include <functional>
#include <array>
#include <utility>
#include <vector>

#include <migraphx/fallthrough.hpp>
#include <migraphx/program.hpp>
#include <migraphx/operators.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/config.hpp>
#include <migraphx/tf.hpp>
#include <migraphx/pad_calc.hpp>

namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {

struct tf_parser
{
    using attribute_map = std::unordered_map<std::string, tensorflow::AttrValue>;
    using node_map      = std::map<std::string, tensorflow::NodeDef>;
    using op_func = std::function<instruction_ref(attribute_map, std::vector<instruction_ref>)>;

    node_map nodes;
    std::vector<tensorflow::NodeDef> input_nodes;
    std::unordered_map<std::string, instruction_ref> instructions;
    program prog = program();
    bool is_nhwc = true;

    std::unordered_map<std::string, op_func> ops;

    bool should_transpose(instruction_ref ins) const
    {
        return is_nhwc and ins->get_shape().lens().size() == 4;
    }

    instruction_ref to_nhwc(instruction_ref ins)
    {
        if(should_transpose(ins))
            return prog.add_instruction(op::transpose{{0, 2, 3, 1}}, ins);
        return ins;
    }

    instruction_ref to_nchw(instruction_ref ins)
    {
        if(should_transpose(ins))
            return prog.add_instruction(op::transpose{{0, 3, 1, 2}}, ins);
        return ins;
    }

    instruction_ref to_kcxy(instruction_ref ins)
    {
        if(should_transpose(ins))
            return prog.add_instruction(op::transpose{{3, 2, 0, 1}}, ins);
        return ins;
    }

    instruction_ref make_contiguous(instruction_ref ins)
    {
        if(ins->get_shape().standard())
            return ins;
        else
            return prog.add_instruction(op::contiguous{}, ins);
    }

    std::vector<instruction_ref> to_nchw(const std::vector<instruction_ref>& args)
    {
        std::vector<instruction_ref> result(args.size());
        std::transform(
            args.begin(), args.end(), result.begin(), [&](auto ins) { return this->to_nchw(ins); });
        return result;
    }

    std::vector<size_t>
    parse_axes(const attribute_map& attributes, const std::string& s, const size_t num_dims) const
    {
        auto attrs = attributes.at(s).list().i();
        std::vector<size_t> axes;
        copy(attrs.begin(), attrs.end(), std::back_inserter(axes));
        if(is_nhwc)
        {
            std::transform(axes.begin(), axes.end(), axes.begin(), [&](size_t axis) {
                return parse_axis(axis, num_dims);
            });
        }
        return axes;
    }

    template <class T>
    std::vector<T> parse_axes(std::vector<T> axes, const size_t num_dims) const
    {
        if(is_nhwc)
        {
            std::vector<T> new_axes;
            std::transform(axes.begin(),
                           axes.end(),
                           std::back_inserter(new_axes),
                           [&](size_t axis) { return parse_axis(axis, num_dims); });
            return new_axes;
        }
        return axes;
    }

    // tf stores certain attributes such as strides, dilations, as a 4D input.
    // The first and last dims are equal to 1, and the relevant data is in dims 2 and 3.
    // This helper function reorders the data to store for the respective operator member variables.
    template <class T>
    void reorder_data(std::vector<T>& prev_data) const
    {
        std::vector<T> new_data(prev_data.size());
        for(size_t i = 0; i < new_data.size(); i++)
        {
            auto new_idx         = parse_axis(i, new_data.size());
            new_data.at(new_idx) = prev_data.at(i);
        }
        prev_data = new_data;
    }

    template <class T>
    T parse_axis(const T& dim, const size_t num_dims) const
    {
        T new_dim = dim;
        if(is_nhwc and num_dims >= 4)
        {
            switch(dim)
            {
            case 0: new_dim = 0; break;
            case 1: new_dim = 2; break;
            case 2: new_dim = 3; break;
            case 3: new_dim = 1; break;
            default: break;
            }
        }
        return new_dim;
    }

    std::vector<int64_t> get_axes(size_t num_axes) const
    {
        std::vector<int64_t> axes(num_axes);
        std::iota(axes.begin(), axes.end(), 0);
        return axes;
    }

    std::vector<int64_t> get_axes_from_mask(const size_t num_axes, const uint32_t mask)
    {
        uint32_t bitwise_compare = 1;
        std::vector<int64_t> axes;
        for(size_t i = 0; i < num_axes; i++)
        {
            // the LSB corresponds to axis 0 when determining which axes to begin
            if(((mask >> i) & bitwise_compare) == 1)
                axes.push_back(1);
            else
                axes.push_back(0);
        }
        return axes;
    }

    tf_parser()
    {
        add_generic_op("All", op::identity{});
        add_generic_op("Identity", op::identity{});
        add_generic_op("LessEqual", op::identity{});
        add_generic_op("Relu", op::relu{});
        add_generic_op("Relu6", op::clip{6.0, 0.0});
        add_generic_op("Rsqrt", op::rsqrt{});
        add_generic_op("Tanh", op::tanh{});
        add_generic_op("StopGradient", op::identity{});

        add_binary_op("Add", op::add{});
        add_binary_op("Mul", op::mul{});
        add_binary_op("Pow", op::pow{});
        add_binary_op("SquaredDifference", op::sqdiff{});
        add_binary_op("Sub", op::sub{});

        add_mem_op("AvgPool", &tf_parser::parse_pooling);
        add_mem_op("BatchMatMul", &tf_parser::parse_matmul, false);
        add_mem_op("BatchMatMulV2", &tf_parser::parse_matmul, false);
        add_mem_op("BiasAdd", &tf_parser::parse_biasadd);
        add_mem_op("Cast", &tf_parser::parse_cast, false);
        add_mem_op("ConcatV2", &tf_parser::parse_concat, false);
        add_mem_op("Const", &tf_parser::parse_constant);
        add_mem_op("Conv2D", &tf_parser::parse_conv);
        add_mem_op("DepthwiseConv2dNative", &tf_parser::parse_depthwiseconv);
        add_mem_op("ExpandDims", &tf_parser::parse_expanddims, false);
        add_mem_op("FusedBatchNorm", &tf_parser::parse_batchnorm);
        add_mem_op("GatherV2", &tf_parser::parse_gather, false);
        add_mem_op("MatMul", &tf_parser::parse_matmul, false);
        add_mem_op("MaxPool", &tf_parser::parse_pooling);
        add_mem_op("Mean", &tf_parser::parse_mean, false);
        add_mem_op("OneHot", &tf_parser::parse_onehot, false);
        add_mem_op("Pack", &tf_parser::parse_pack, false);
        add_mem_op("Pad", &tf_parser::parse_pad);
        add_mem_op("Reshape", &tf_parser::parse_reshape, false);
        add_mem_op("Slice", &tf_parser::parse_slice, false);
        add_mem_op("Softmax", &tf_parser::parse_softmax<op::softmax>, false);
        add_mem_op("Squeeze", &tf_parser::parse_squeeze, false);
        add_mem_op("StridedSlice", &tf_parser::parse_stridedslice, false);
        add_mem_op("Transpose", &tf_parser::parse_transpose, false);
    }

    template <class F>
    void add_op(std::string name, F f, bool transpose = true)
    {
        if(transpose)
        {
            ops.emplace(name,
                        op_func{[=](const attribute_map& attributes,
                                    const std::vector<instruction_ref>& args) -> instruction_ref {
                            return to_nhwc(f(attributes, to_nchw(args)));
                        }});
        }
        else
        {
            ops.emplace(name, f);
        }
    }

    template <class F>
    void add_mem_op(std::string name, F f, bool transpose = true)
    {
        add_op(name,
               [=](auto&&... xs) {
                   return std::mem_fn(f)(*this, name, std::forward<decltype(xs)>(xs)...);
               },
               transpose);
    }

    template <class T>
    void add_binary_op(std::string name, T x)
    {
        add_op(name,
               [this, x](const attribute_map&, std::vector<instruction_ref> args) {
                   if(args.size() != 2)
                       MIGRAPHX_THROW("binary operators should have 2 operands");
                   // TODO
                   // if(contains(attributes, "data_format"))
                   // {
                   //     if(is_nhwc)
                   //     {
                   //         l0 = prog.add_instruction(op::transpose{{0, 3, 1, 2}}, args[1]);
                   //     }
                   // }
                   return add_broadcastable_binary_op(args[0], args[1], x);
               },
               false);
    }

    template <class T>
    instruction_ref add_broadcastable_binary_op(instruction_ref arg0, instruction_ref arg1, T x)
    {
        if(arg0->get_shape().lens() != arg1->get_shape().lens())
        {
            // Example:
            // s0 = (3,2,4,5) and s1 = (2,1,1)
            //
            // In this case we need to broadcast (:,1,1) portion of
            // s1 plus broadcast the 1st dimension of s1
            // giving output_lens = (3,2,4,5)
            //
            // Another example:
            // s0 = (3,2,1,5) and s1 = (2,7,5)
            // In this case we need to broadcast the (:,:,1:,:) axis
            // of s0 plus the 1st dimension of s1 giving
            // output_lens = (3,2,7,5)
            //
            // Get lengths for both arguments
            const std::vector<size_t>* s0 = &arg0->get_shape().lens();
            const std::vector<size_t>* s1 = &arg1->get_shape().lens();

            // Make sure s0 is the smaller size
            if(s0->size() > s1->size())
                std::swap(s0, s1);

            std::vector<size_t> output_lens(*s1);
            auto offset = s1->size() - s0->size();
            std::transform(s0->begin(),
                           s0->end(),
                           s1->begin() + offset,
                           output_lens.begin() + offset,
                           [](auto a, auto b) { return std::max(a, b); });

            auto l0 = prog.add_instruction(op::multibroadcast{output_lens}, arg0);
            auto l1 = prog.add_instruction(op::multibroadcast{output_lens}, arg1);
            return to_nhwc(prog.add_instruction(x, to_nchw(l0), to_nchw(l1)));
        }
        else
        {
            return to_nhwc(prog.add_instruction(x, {to_nchw(arg0), to_nchw(arg1)}));
        }
    }

    template <class T>
    void add_generic_op(std::string name, T x, bool transpose = true)
    {
        add_op(name,
               [this, x](const attribute_map&, std::vector<instruction_ref> args) {
                   return prog.add_instruction(x, args);
               },
               transpose);
    }

    instruction_ref
    parse_batchnorm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        float epsilon                                     = 1e-5f;
        float momentum                                    = 0.9f;
        op::batch_norm_inference::bn_infer_mode_t bn_mode = op::batch_norm_inference::spatial;
        if(contains(attributes, "epsilon"))
        {
            epsilon = attributes.at("epsilon").f();
        }
        op::batch_norm_inference op{epsilon, momentum, bn_mode};
        return prog.add_instruction(op, std::move(args));
    }

    instruction_ref
    parse_biasadd(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
    {
        uint64_t axis = 1; // assume output of previous layer is in NCHW (broadcast on channel)
        auto l0 = prog.add_instruction(op::broadcast{axis, args[0]->get_shape().lens()}, args[1]);
        return prog.add_instruction(op::add{}, args[0], l0);
    }

    instruction_ref
    parse_cast(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        shape::type_t type = parse_type(attributes.at("DstT").type());
        return prog.add_instruction(op::convert{type}, std::move(args));
    }

    instruction_ref
    parse_concat(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        // get index for axis within args
        size_t axis_idx = attributes.at("N").i();
        size_t axis     = args[axis_idx]->eval().at<int64_t>();
        op::concat op{axis};
        // return only first N arguments (assuming last index is the axis value)
        return prog.add_instruction(
            op, std::vector<instruction_ref>(args.begin(), args.begin() + args.size() - 1));
    }

    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
    {
        literal v = parse_tensor(attributes.at("value").tensor());
        return prog.add_literal(v);
    }

    instruction_ref
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::convolution op;
        if(contains(attributes, "strides"))
        {
            std::vector<size_t> stride;
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
            reorder_data(stride);
            if(stride.size() != 4)
            {
                MIGRAPHX_THROW("strides should have 4 values");
            }
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
        }
        if(contains(attributes, "dilations"))
        {
            std::vector<size_t> dilation;
            copy(attributes.at("dilations").list().i(), std::back_inserter(dilation));
            reorder_data(dilation);
            if(dilation.size() != 4)
            {
                MIGRAPHX_THROW("dilation should have 4 values");
            }
            op.dilation[0] = dilation[2];
            op.dilation[1] = dilation[3];
        }

        auto weights = to_kcxy(args[1]);
        auto l0      = args[0];
        if(contains(attributes, "padding"))
        {
            const std::string& pad_mode = attributes.at("padding").s();
            if(pad_mode.find("SAME") != std::string::npos)
            {
                op.padding_mode                 = op::padding_mode_t::same;
                std::vector<size_t> weight_dims = weights->get_shape().lens();
                size_t weight_h                 = weight_dims[2];
                size_t weight_w                 = weight_dims[3];

                auto input_dims = l0->get_shape().lens();
                size_t input_h  = input_dims[2];
                size_t input_w  = input_dims[3];
                std::vector<int64_t> pads(input_dims.size());
                calculate_padding(0, pads, input_h, op.stride[0], op.dilation[0], weight_h);
                calculate_padding(1, pads, input_w, op.stride[1], op.dilation[1], weight_w);

                if(pads[0] != pads[2] || pads[1] != pads[3])
                {
                    std::vector<int64_t> padding = {0, 0, pads[0], pads[1], 0, 0, pads[2], pads[3]};
                    l0 = prog.add_instruction(migraphx::op::pad{padding}, l0);
                }
                else
                {
                    op.padding[0] = pads[0];
                    op.padding[1] = pads[1];
                }
            }
            else if(pad_mode.find("VALID") != std::string::npos)
            {
                op.padding_mode = op::padding_mode_t::valid;
            }
            else if(pad_mode.find("EXPLICIT") != std::string::npos)
            {
                std::vector<size_t> padding;
                copy(attributes.at("explicit_paddings").list().i(), std::back_inserter(padding));
                if(padding.size() != 4)
                {
                    MIGRAPHX_THROW("padding should have 4 values");
                }
                if(padding[0] != padding[2] || padding[1] != padding[3])
                {
                    MIGRAPHX_THROW("migraphx does not support asymetric padding");
                }
                op.padding[0] = padding[0];
                op.padding[1] = padding[1];
            }
        }
        return prog.add_instruction(op, {l0, to_kcxy(args[1])});
    }

    instruction_ref parse_depthwiseconv(const std::string&,
                                        attribute_map attributes,
                                        std::vector<instruction_ref> args)
    {
        op::convolution op;
        size_t num_channels = args[0]->get_shape().lens()[1];
        op.group            = num_channels;

        if(contains(attributes, "strides"))
        {
            std::vector<size_t> stride;
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
            reorder_data(stride);
            if(stride.size() != 4)
            {
                MIGRAPHX_THROW("strides should have 4 values");
            }
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
        }

        auto weights = to_kcxy(args[1]);
        if(contains(attributes, "dilations"))
        {
            std::vector<size_t> dilation;
            copy(attributes.at("dilations").list().i(), std::back_inserter(dilation));
            reorder_data(dilation);
            if(dilation.size() != 4)
            {
                MIGRAPHX_THROW("dilation should have 4 values");
            }
            op.dilation[0] = dilation[2];
            op.dilation[1] = dilation[3];
        }

        auto l0 = args[0];
        if(contains(attributes, "padding"))
        {
            const std::string& pad_mode = attributes.at("padding").s();

            if(pad_mode.find("SAME") != std::string::npos)
            {
                op.padding_mode                 = op::padding_mode_t::same;
                std::vector<size_t> weight_dims = weights->get_shape().lens();
                size_t weight_h                 = weight_dims[2];
                size_t weight_w                 = weight_dims[3];

                auto input_dims = l0->get_shape().lens();
                size_t input_h  = input_dims[2];
                size_t input_w  = input_dims[3];
                std::vector<int64_t> pads(input_dims.size());
                calculate_padding(0, pads, input_h, op.stride[0], op.dilation[0], weight_h);
                calculate_padding(1, pads, input_w, op.stride[1], op.dilation[1], weight_w);

                if(pads[0] != pads[2] || pads[1] != pads[3])
                {
                    std::vector<int64_t> padding = {0, 0, pads[0], pads[1], 0, 0, pads[2], pads[3]};
                    l0 = prog.add_instruction(migraphx::op::pad{padding}, l0);
                }
                else
                {
                    op.padding[0] = pads[0];
                    op.padding[1] = pads[1];
                }
            }
            else if(pad_mode.find("VALID") != std::string::npos)
            {
                op.padding_mode = op::padding_mode_t::valid;
            }
        }

        std::vector<int64_t> new_weights_shape;
        copy(weights->get_shape().lens(), std::back_inserter(new_weights_shape));

        // weight format is (out_channels, in_channels, h, w), but in depthwise_conv,
        // out_channels is equal to the multiplier. Adjust by inserting a reshape and
        // setting in_channels to 1
        int64_t multiplier   = new_weights_shape[0];
        int64_t out_channels = num_channels * multiplier;
        new_weights_shape[0] = out_channels;
        new_weights_shape[1] = 1;
        // Make sure weights are contiguous before doing reshape
        auto new_weights =
            prog.add_instruction(op::reshape{new_weights_shape}, make_contiguous(weights));

        return prog.add_instruction(op, {l0, new_weights});
    }

    instruction_ref
    parse_expanddims(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
    {
        std::vector<size_t> input_dims = args[0]->get_shape().lens();
        std::vector<int64_t> new_dims(input_dims.begin(), input_dims.end());
        size_t num_dims = input_dims.size();
        int32_t dim     = args[1]->eval().at<int32_t>();

        if(dim < 0)
        {
            new_dims.insert(new_dims.begin() + (num_dims + dim + 1), 1);
        }
        else
        {
            new_dims.insert(new_dims.begin() + dim, 1);
        }
        return prog.add_instruction(op::reshape{new_dims}, args[0]);
    }

    instruction_ref
    parse_gather(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
    {
        int axis = args[2]->eval().at<int32_t>();
        op::gather op{axis};
        return prog.add_instruction(op, {args[0], args[1]});
    }

    instruction_ref
    parse_matmul(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        bool transa = false;
        bool transb = false;

        if(contains(attributes, "transpose_a"))
        {
            transa = attributes.at("transpose_a").b();
        }
        if(contains(attributes, "transpose_b"))
        {
            transb = attributes.at("transpose_b").b();
        }

        if(contains(attributes, "adj_x"))
        {
            transa = attributes.at("adj_x").b();
        }
        if(contains(attributes, "adj_y"))
        {
            transb = attributes.at("adj_y").b();
        }

        std::vector<int64_t> perm(args[0]->get_shape().lens().size());
        std::iota(perm.begin(), perm.end(), int64_t{0});
        // swap the last two elements
        std::iter_swap(perm.end() - 1, perm.end() - 2);

        auto l1 = (transa) ? prog.add_instruction(op::transpose{perm}, args[0]) : args[0];
        auto l2 = (transb) ? prog.add_instruction(op::transpose{perm}, args[1]) : args[1];

        return prog.add_instruction(op::dot{}, l1, l2);
    }

    instruction_ref
    parse_mean(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        bool keep_dims = attributes.at("keep_dims").b();
        auto axes      = args[1]->eval().get<int32_t>().to_vector<int64_t>();

        if(keep_dims)
        {
            return prog.add_instruction(op::reduce_mean{axes}, args[0]);
        }
        else
        {
            auto ins = prog.add_instruction(op::reduce_mean{axes}, args[0]);
            return prog.add_instruction(op::squeeze{axes}, ins);
        }
    }

    instruction_ref
    parse_onehot(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        size_t depth = static_cast<size_t>(args[1]->eval().at<int32_t>());

        int64_t axis    = -1;
        float on_value  = args[2]->eval().at<float>();
        float off_value = args[3]->eval().at<float>();

        std::vector<float> depth_input(depth * depth, off_value);
        for(int i = 0; i < depth; i++)
        {
            depth_input[depth * i + i] = on_value;
        }

        if(contains(attributes, "axis"))
            axis = attributes.at("axis").i();
        if(axis == -1)
        {
            shape s{shape::float_type, {depth, depth}};
            auto l0 = prog.add_literal({s, depth_input});
            return prog.add_instruction(op::gather{0}, {l0, args[0]});
        }
        MIGRAPHX_THROW("MIGraphX does not support axis != -1");
    }

    instruction_ref parse_pack(const std::string&,
                               const attribute_map& attributes,
                               std::vector<instruction_ref> args)
    {
        // reinterpret as unsqueeze with concat
        std::vector<instruction_ref> unsqueezed_args;
        int64_t axis = 0;
        if(contains(attributes, "axis"))
            axis = attributes.at("axis").i();
        size_t input_size = args.front()->get_shape().lens().size();
        if(axis > input_size)
        {
            MIGRAPHX_THROW("TF_PARSER: axis value of " + to_string(axis) +
                           " must be smaller than input size " + to_string(input_size));
        }

        std::transform(
            args.begin(),
            args.end(),
            std::back_inserter(unsqueezed_args),
            [&](instruction_ref arg) { return prog.add_instruction(op::unsqueeze{{axis}}, arg); });
        return to_nhwc(
            prog.add_instruction(op::concat{static_cast<size_t>(axis)}, unsqueezed_args));
    }

    instruction_ref
    parse_pad(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
    {
        size_t ndims = args.front()->get_shape().lens().size();

        // in tf, the paddings are arranged as a 2d shape (ndims, 2),
        // the last dim contains the left padding and right padding respectively
        std::vector<std::pair<int32_t, int32_t>> pad_per_dim(ndims);
        auto tf_padding = args[1]->eval().get<int32_t>().to_vector();
        for(size_t i = 0; i < 2 * ndims; i += 2)
        {
            pad_per_dim[i / 2].first  = tf_padding[i];
            pad_per_dim[i / 2].second = tf_padding[i + 1];
        }
        reorder_data(pad_per_dim);

        op::pad op;
        std::vector<int64_t> pads(ndims * 2);
        for(size_t i = 0; i < ndims; i++)
        {
            pads[i]         = pad_per_dim[i].first;
            pads[i + ndims] = pad_per_dim[i].second;
        }
        op.pads = pads;
        return prog.add_instruction(op, args.front());
    }

    instruction_ref parse_pooling(const std::string& name,
                                  attribute_map attributes,
                                  std::vector<instruction_ref> args)
    {
        op::pooling op{starts_with(name, "Max") ? "max" : "average"};

        if(contains(attributes, "strides"))
        {
            std::vector<size_t> stride;
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
            reorder_data(stride);
            if(stride.size() != 4)
            {
                MIGRAPHX_THROW("strides should have 4 values");
            }
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
        }
        if(contains(attributes, "ksize"))
        {
            std::vector<size_t> ksize;
            copy(attributes.at("ksize").list().i(), std::back_inserter(ksize));
            reorder_data(ksize);
            if(ksize.size() != 4)
            {
                MIGRAPHX_THROW("ksize should have 4 values");
            }
            op.lengths[0] = ksize[2];
            op.lengths[1] = ksize[3];
        }

        auto l0 = args[0];
        if(contains(attributes, "padding"))
        {
            const std::string& pad_mode = attributes.at("padding").s();
            if(pad_mode.find("SAME") != std::string::npos)
            {
                op.padding_mode = op::padding_mode_t::same;
                auto input_dims = l0->get_shape().lens();
                size_t input_h  = input_dims[2];
                size_t input_w  = input_dims[3];
                std::vector<int64_t> pads(input_dims.size());
                calculate_padding(0, pads, input_h, op.stride[0], 1, op.lengths[0]);
                calculate_padding(1, pads, input_w, op.stride[1], 1, op.lengths[1]);

                if(pads[0] != pads[2] || pads[1] != pads[3])
                {
                    std::vector<int64_t> padding = {0, 0, pads[0], pads[1], 0, 0, pads[2], pads[3]};
                    l0                           = prog.add_instruction(
                        migraphx::op::pad{padding, std::numeric_limits<float>::lowest()}, l0);
                }
                else
                {
                    op.padding[0] = pads[0];
                    op.padding[1] = pads[1];
                }
            }
            else if(pad_mode.find("VALID") != std::string::npos)
            {
                op.padding_mode = op::padding_mode_t::valid;
            }
        }
        return prog.add_instruction(op, l0);
    }

    instruction_ref
    parse_reshape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
    {
        op::reshape op;
        if(args.size() != 2)
            MIGRAPHX_THROW("reshape needs 2 arguments (input, new_shape)");
        auto s = args[1]->eval();
        s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
        return prog.add_instruction(op, make_contiguous(args[0]));
    }

    void parse_from(std::istream& is)
    {
        tensorflow::GraphDef graph;
        if(graph.ParseFromIstream(&is))
        {
            this->parse_graph(graph);
        }
        else
        {
            throw std::runtime_error("Failed reading tf file");
        }
    }

    instruction_ref
    parse_slice(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
    {
        op::slice op;
        auto starts     = args[1]->eval().get<int32_t>().to_vector();
        auto size       = args[2]->eval().get<int32_t>().to_vector();
        auto axes       = args[0]->get_shape().lens();
        size_t num_axes = axes.size();

        op.starts = std::vector<int64_t>(starts.begin(), starts.end());
        op.ends   = std::vector<int64_t>(num_axes);
        op.axes   = std::vector<int64_t>(num_axes);
        std::iota(op.axes.begin(), op.axes.end(), 0);
        for(size_t i = 0; i < num_axes; i++)
        {
            if(size[i] == -1)
                op.ends[i] = axes[i];
            else
                op.ends[i] = starts[i] + size[i];
        }
        return prog.add_instruction(op, make_contiguous(args[0]));
    }

    // template to facilitate the logsoftmax later
    template <class Op>
    instruction_ref parse_softmax(const std::string&,
                                  const attribute_map& attributes,
                                  std::vector<instruction_ref> args)
    {
        int axis      = -1;
        auto num_dims = args[0]->get_shape().lens().size();
        if(contains(attributes, "axis"))
        {
            axis = static_cast<int>(attributes.at("axis").i());
        }
        if(axis < 0)
        {
            axis += num_dims;
        }

        return prog.add_instruction(Op{axis}, make_contiguous(args[0]));
    }

    instruction_ref parse_squeeze(const std::string&,
                                  const attribute_map& attributes,
                                  std::vector<instruction_ref> args)
    {
        op::squeeze op;
        auto input_dims = args[0]->get_shape().lens();
        auto axes       = attributes.at("squeeze_dims").list().i();
        copy(axes, std::back_inserter(op.axes));

        if(op.axes.empty()) // no squeeze_dims provided, remove any dim that equals 1
        {
            for(size_t i = 0; i < input_dims.size(); i++)
            {
                if(input_dims.at(i) == 1)
                {
                    op.axes.push_back(i);
                }
            }
        }
        return prog.add_instruction(op, make_contiguous(args[0]));
    }

    instruction_ref parse_stridedslice(const std::string&,
                                       const attribute_map& attributes,
                                       std::vector<instruction_ref> args)
    {
        op::slice op;
        auto starts              = args[1]->eval().get<int32_t>().to_vector();
        auto ends                = args[2]->eval().get<int32_t>().to_vector();
        auto l0                  = args[0];
        size_t num_axes          = l0->get_shape().lens().size();
        std::vector<size_t> axes = l0->get_shape().lens();

        op.starts = std::vector<int64_t>(starts.begin(), starts.end());
        op.ends   = std::vector<int64_t>(ends.begin(), ends.end());
        op.axes   = std::vector<int64_t>(num_axes);
        std::iota(op.axes.begin(), op.axes.end(), 0);
        uint32_t begin_mask       = 0;
        uint32_t end_mask         = 0;
        uint32_t shrink_axis_mask = 0;
        uint32_t bitwise_compare  = 1;
        std::vector<int64_t> squeeze_axes;

        if(contains(attributes, "begin_mask"))
            begin_mask = static_cast<uint32_t>(attributes.at("begin_mask").i());

        if(contains(attributes, "end_mask"))
            end_mask = static_cast<uint32_t>(attributes.at("end_mask").i());

        if(contains(attributes, "shrink_axis_mask"))
            shrink_axis_mask = static_cast<uint32_t>(attributes.at("shrink_axis_mask").i());

        std::vector<int64_t> begin_axes = get_axes_from_mask(num_axes, begin_mask);
        std::vector<int64_t> end_axes   = get_axes_from_mask(num_axes, end_mask);

        for(size_t i = 0; i < num_axes; i++)
        {
            if(begin_axes.at(i) == 1)
            {
                op.starts.at(i) = 0;
            }
            if(end_axes.at(i) == 1)
            {
                op.ends.at(i) = axes.at(i);
            }
        }

        auto l1 = prog.add_instruction(op, l0);
        if(shrink_axis_mask == 0)
            return l1;

        for(size_t i = 0; i < num_axes; i++)
        {
            // the LSB corresponds to axis 0 when determining which axes to squeeze
            if(((shrink_axis_mask >> i) & bitwise_compare) == 1)
                squeeze_axes.push_back(i);
        }

        return prog.add_instruction(op::squeeze{squeeze_axes}, l1);
    }

    instruction_ref
    parse_transpose(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
    {
        auto perm = args[1]->eval().get<int32_t>().to_vector();
        op::transpose op;
        op.dims = std::vector<int64_t>(perm.begin(), perm.end());

        return prog.add_instruction(op, args.front());
    }

    void parse_graph(const tensorflow::GraphDef& graph)
    {
        nodes = get_nodes(graph, input_nodes);
        for(auto&& input : input_nodes)
        {
            const std::string& name   = input.name();
            attribute_map input_attrs = get_attributes(input);
            shape::type_t shape_type  = parse_type(input_attrs.at("dtype").type());
            std::vector<size_t> dims  = parse_dims(input_attrs.at("shape").shape());
            if(is_nhwc and dims.size() >= 4)
            {
                reorder_data(dims);
            }
            shape s            = shape{shape_type, dims};
            instructions[name] = to_nhwc(prog.add_parameter(name, s));
        }
        for(auto&& p : nodes)
        {
            this->parse_node(p.first);
        }
    }

    void parse_node(const std::string& name)
    {
        if(instructions.count(name) == 0)
        {
            auto&& node = nodes.at(name);
            // assert ops ignored
            if(node.op() == "Assert" or contains(name, "Assert"))
                return;
            std::vector<instruction_ref> args;

            for(auto&& input : node.input())
            {
                // control dependencies (signified by ^ before the name) are ignored
                if(contains(input, "^"))
                    continue;
                if(nodes.count(input) > 0)
                {
                    auto&& iname = get_name(nodes.at(input));
                    assert(name != iname);
                    this->parse_node(iname);
                    args.push_back(instructions.at(iname));
                }
                else
                {
                    args.push_back(instructions.at(input));
                }
            }
            if(ops.count(node.op()) == 0)
            {
                instructions[name] = prog.add_instruction(op::unknown{node.op()}, args);
            }
            else
            {
                instructions[name] = ops[node.op()](get_attributes(node), args);
            }
        }
    }

    static attribute_map get_attributes(const tensorflow::NodeDef& node)
    {
        attribute_map result;
        for(auto&& attr : node.attr())
        {
            result[attr.first] = attr.second;
        }
        return result;
    }

    static std::string get_name(const tensorflow::NodeDef& node) { return node.name(); }

    static node_map get_nodes(const tensorflow::GraphDef& graph,
                              std::vector<tensorflow::NodeDef>& input_nodes)
    {
        node_map result;
        for(auto&& node : graph.node())
        {
            auto node_name = get_name(node);
            // assume each node in graph has an associated name
            if(node_name.empty())
                MIGRAPHX_THROW("tf node with no name found");
            result[node_name] = node;
            if(node.op() == "Placeholder")
            {
                input_nodes.push_back(node);
            }
        }
        return result;
    }

    static shape::type_t parse_type(const tensorflow::DataType t)
    {
        shape::type_t shape_type{};
        switch(t)
        {
        case tensorflow::DataType::DT_FLOAT: shape_type = shape::float_type; break;
        case tensorflow::DataType::DT_DOUBLE: shape_type = shape::double_type; break;
        case tensorflow::DataType::DT_INT32: shape_type = shape::int32_type; break;
        case tensorflow::DataType::DT_INT16: shape_type = shape::int16_type; break;
        case tensorflow::DataType::DT_INT8: shape_type = shape::int8_type; break;
        case tensorflow::DataType::DT_INT64: shape_type = shape::int64_type; break;
        case tensorflow::DataType::DT_UINT16: shape_type = shape::uint16_type; break;
        case tensorflow::DataType::DT_HALF: shape_type = shape::half_type; break;
        case tensorflow::DataType::DT_UINT32: shape_type = shape::uint32_type; break;
        case tensorflow::DataType::DT_UINT64: shape_type = shape::uint64_type; break;

        case tensorflow::DataType::DT_INVALID:
        case tensorflow::DataType::DT_UINT8:
        case tensorflow::DataType::DT_STRING:
        case tensorflow::DataType::DT_COMPLEX64:
        case tensorflow::DataType::DT_BOOL:
        case tensorflow::DataType::DT_QINT8:
        case tensorflow::DataType::DT_QUINT8:
        case tensorflow::DataType::DT_QINT32:
        case tensorflow::DataType::DT_BFLOAT16:
        case tensorflow::DataType::DT_QINT16:
        case tensorflow::DataType::DT_QUINT16:
        case tensorflow::DataType::DT_COMPLEX128:
        case tensorflow::DataType::DT_RESOURCE:
        case tensorflow::DataType::DT_VARIANT:
        // tf pb should not use these types
        case tensorflow::DataType::DT_FLOAT_REF:
        case tensorflow::DataType::DT_DOUBLE_REF:
        case tensorflow::DataType::DT_INT32_REF:
        case tensorflow::DataType::DT_UINT8_REF:
        case tensorflow::DataType::DT_INT16_REF:
        case tensorflow::DataType::DT_INT8_REF:
        case tensorflow::DataType::DT_STRING_REF:
        case tensorflow::DataType::DT_COMPLEX64_REF:
        case tensorflow::DataType::DT_INT64_REF:
        case tensorflow::DataType::DT_BOOL_REF:
        case tensorflow::DataType::DT_QINT8_REF:
        case tensorflow::DataType::DT_QUINT8_REF:
        case tensorflow::DataType::DT_QINT32_REF:
        case tensorflow::DataType::DT_BFLOAT16_REF:
        case tensorflow::DataType::DT_QINT16_REF:
        case tensorflow::DataType::DT_QUINT16_REF:
        case tensorflow::DataType::DT_UINT16_REF:
        case tensorflow::DataType::DT_COMPLEX128_REF:
        case tensorflow::DataType::DT_HALF_REF:
        case tensorflow::DataType::DT_RESOURCE_REF:
        case tensorflow::DataType::DT_VARIANT_REF:
        case tensorflow::DataType::DT_UINT32_REF:
        case tensorflow::DataType::DT_UINT64_REF:
        case tensorflow::DataType::DataType_INT_MAX_SENTINEL_DO_NOT_USE_:
        case tensorflow::DataType::DataType_INT_MIN_SENTINEL_DO_NOT_USE_: break;
        }
        return shape_type;
    }

    static literal parse_tensor(const tensorflow::TensorProto& t)
    {
        std::vector<size_t> dims = parse_dims(t.tensor_shape());
        size_t shape_size = std::accumulate(dims.begin(), dims.end(), 1, std::multiplies<size_t>());
        if(!t.tensor_content().empty()) // has raw data
        {
            const std::string& s = t.tensor_content();
            switch(t.dtype())
            {
            case tensorflow::DataType::DT_FLOAT:
                return literal{{shape::float_type, dims}, s.data()};
            case tensorflow::DataType::DT_BOOL:
            case tensorflow::DataType::DT_INT8: return literal{{shape::int8_type, dims}, s.data()};
            case tensorflow::DataType::DT_UINT16:
            case tensorflow::DataType::DT_INT16:
                return literal{{shape::int16_type, dims}, s.data()};
            case tensorflow::DataType::DT_INT32:
                return literal{{shape::int32_type, dims}, s.data()};
            case tensorflow::DataType::DT_INT64:
                return literal{{shape::int64_type, dims}, s.data()};
            case tensorflow::DataType::DT_HALF: return literal{{shape::half_type, dims}, s.data()};
            case tensorflow::DataType::DT_DOUBLE:
                return literal{{shape::double_type, dims}, s.data()};
            case tensorflow::DataType::DT_INVALID:
            case tensorflow::DataType::DT_UINT8:
            case tensorflow::DataType::DT_STRING:
            case tensorflow::DataType::DT_UINT32:
            case tensorflow::DataType::DT_UINT64:
            case tensorflow::DataType::DT_COMPLEX64:
            case tensorflow::DataType::DT_COMPLEX128:
            case tensorflow::DataType::DT_QINT8:
            case tensorflow::DataType::DT_QUINT8:
            case tensorflow::DataType::DT_QINT32:
            case tensorflow::DataType::DT_BFLOAT16:
            case tensorflow::DataType::DT_QINT16:
            case tensorflow::DataType::DT_QUINT16:
            case tensorflow::DataType::DT_RESOURCE:
            case tensorflow::DataType::DT_VARIANT:
            case tensorflow::DataType::DT_FLOAT_REF:
            case tensorflow::DataType::DT_DOUBLE_REF:
            case tensorflow::DataType::DT_INT32_REF:
            case tensorflow::DataType::DT_UINT8_REF:
            case tensorflow::DataType::DT_INT16_REF:
            case tensorflow::DataType::DT_INT8_REF:
            case tensorflow::DataType::DT_STRING_REF:
            case tensorflow::DataType::DT_COMPLEX64_REF:
            case tensorflow::DataType::DT_INT64_REF:
            case tensorflow::DataType::DT_BOOL_REF:
            case tensorflow::DataType::DT_QINT8_REF:
            case tensorflow::DataType::DT_QUINT8_REF:
            case tensorflow::DataType::DT_QINT32_REF:
            case tensorflow::DataType::DT_BFLOAT16_REF:
            case tensorflow::DataType::DT_QINT16_REF:
            case tensorflow::DataType::DT_QUINT16_REF:
            case tensorflow::DataType::DT_UINT16_REF:
            case tensorflow::DataType::DT_COMPLEX128_REF:
            case tensorflow::DataType::DT_HALF_REF:
            case tensorflow::DataType::DT_RESOURCE_REF:
            case tensorflow::DataType::DT_VARIANT_REF:
            case tensorflow::DataType::DT_UINT32_REF:
            case tensorflow::DataType::DT_UINT64_REF:
            case tensorflow::DataType::DataType_INT_MAX_SENTINEL_DO_NOT_USE_:
            case tensorflow::DataType::DataType_INT_MIN_SENTINEL_DO_NOT_USE_:
                throw std::runtime_error("");
            }
            MIGRAPHX_THROW("Invalid tensor type");
        }
        switch(t.dtype())
        {
        case tensorflow::DataType::DT_FLOAT:
            return create_literal(
                shape::float_type, dims, get_data_vals(t.float_val(), shape_size));
        case tensorflow::DataType::DT_INT8:
            return create_literal(shape::int8_type, dims, get_data_vals(t.int_val(), shape_size));
        case tensorflow::DataType::DT_UINT16:
            return create_literal(shape::uint16_type, dims, get_data_vals(t.int_val(), shape_size));
        case tensorflow::DataType::DT_INT16:
            return create_literal(shape::int16_type, dims, get_data_vals(t.int_val(), shape_size));
        case tensorflow::DataType::DT_INT32:
            return create_literal(shape::int32_type, dims, get_data_vals(t.int_val(), shape_size));
        case tensorflow::DataType::DT_INT64:
            return create_literal(
                shape::int64_type, dims, get_data_vals(t.int64_val(), shape_size));
        case tensorflow::DataType::DT_BOOL:
            return create_literal(shape::int32_type, dims, get_data_vals(t.bool_val(), shape_size));
        case tensorflow::DataType::DT_HALF:
        {
            std::vector<int> data_int32 = get_data_vals(t.half_val(), shape_size);
            std::vector<uint16_t> data_uint16(data_int32.begin(), data_int32.end());
            std::vector<half> data_half;
            std::transform(data_uint16.begin(),
                           data_uint16.end(),
                           std::back_inserter(data_half),
                           [](uint16_t raw_val) { return *reinterpret_cast<half*>(&raw_val); });
            return create_literal(shape::half_type, dims, data_half);
        }
        case tensorflow::DataType::DT_DOUBLE:
            return literal{{shape::double_type, dims}, get_data_vals(t.double_val(), shape_size)};
        case tensorflow::DataType::DT_INVALID:
        case tensorflow::DataType::DT_UINT8:
        case tensorflow::DataType::DT_STRING:
        case tensorflow::DataType::DT_UINT32:
        case tensorflow::DataType::DT_UINT64:
        case tensorflow::DataType::DT_COMPLEX64:
        case tensorflow::DataType::DT_COMPLEX128:
        case tensorflow::DataType::DT_QINT8:
        case tensorflow::DataType::DT_QUINT8:
        case tensorflow::DataType::DT_QINT32:
        case tensorflow::DataType::DT_BFLOAT16:
        case tensorflow::DataType::DT_QINT16:
        case tensorflow::DataType::DT_QUINT16:
        case tensorflow::DataType::DT_RESOURCE:
        case tensorflow::DataType::DT_VARIANT:
        case tensorflow::DataType::DT_FLOAT_REF:
        case tensorflow::DataType::DT_DOUBLE_REF:
        case tensorflow::DataType::DT_INT32_REF:
        case tensorflow::DataType::DT_UINT8_REF:
        case tensorflow::DataType::DT_INT16_REF:
        case tensorflow::DataType::DT_INT8_REF:
        case tensorflow::DataType::DT_STRING_REF:
        case tensorflow::DataType::DT_COMPLEX64_REF:
        case tensorflow::DataType::DT_INT64_REF:
        case tensorflow::DataType::DT_BOOL_REF:
        case tensorflow::DataType::DT_QINT8_REF:
        case tensorflow::DataType::DT_QUINT8_REF:
        case tensorflow::DataType::DT_QINT32_REF:
        case tensorflow::DataType::DT_BFLOAT16_REF:
        case tensorflow::DataType::DT_QINT16_REF:
        case tensorflow::DataType::DT_QUINT16_REF:
        case tensorflow::DataType::DT_UINT16_REF:
        case tensorflow::DataType::DT_COMPLEX128_REF:
        case tensorflow::DataType::DT_HALF_REF:
        case tensorflow::DataType::DT_RESOURCE_REF:
        case tensorflow::DataType::DT_VARIANT_REF:
        case tensorflow::DataType::DT_UINT32_REF:
        case tensorflow::DataType::DT_UINT64_REF:
        case tensorflow::DataType::DataType_INT_MAX_SENTINEL_DO_NOT_USE_:
        case tensorflow::DataType::DataType_INT_MIN_SENTINEL_DO_NOT_USE_:
            throw std::runtime_error("");
        }
        MIGRAPHX_THROW("Invalid tensor type");
    }

    template <class T>
    static std::vector<T> get_data_vals(const google::protobuf::RepeatedField<T>& data,
                                        const size_t& shape_size)
    {
        std::vector<T> data_vals(shape_size);
        // check if shape has enough data values given existing fields
        if(data.size() == 1)
        {
            std::fill(data_vals.begin(), data_vals.end(), data[0]);
        }
        else
            copy(data.begin(), data.end(), std::back_inserter(data_vals));
        return data_vals;
    }

    static std::vector<size_t> parse_dims(const tensorflow::TensorShapeProto& s)
    {
        std::vector<size_t> dims;
        auto input_dims = s.dim();
        std::transform(input_dims.begin(),
                       input_dims.end(),
                       std::back_inserter(dims),
                       [](const tensorflow::TensorShapeProto_Dim& dim) { return dim.size(); });
        return dims;
    }

    template <class T>
    static literal
    create_literal(shape::type_t shape_type, const std::vector<size_t>& dims, std::vector<T> data)
    {
        // assume if explicit value is mentioned in protobuf and dim size <= 1, treat as scalar
        if(dims.empty() or (dims.size() == 1 and dims.front() == 1))
            return literal{{shape_type}, data};
        return literal{{shape_type, dims}, data};
    }
};

program parse_tf(const std::string& name, bool is_nhwc)
{
    std::fstream input(name.c_str(), std::ios::in | std::ios::binary);
    tf_parser parser;
    parser.is_nhwc = is_nhwc;

#ifndef NDEBUG
    // Log the program when it can't be parsed
    try
    {
        parser.parse_from(input);
    }
    catch(...)
    {
        std::cerr << parser.prog << std::endl;
        throw;
    }
#else
    parser.parse_from(input);
#endif
    parser.to_nchw(std::prev(parser.prog.end()));
    return std::move(parser.prog);
}

} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx