reference_conv_fwd.hpp

#ifndef REFERENCE_CONV_FWD_HPP
#define REFERENCE_CONV_FWD_HPP

#include <iostream>
#include <type_traits>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"

namespace ck {
namespace tensor_operation {
namespace host {

//
// @brief      Reference implementation for forward convolution.
//
// @paragraph Supported tensor layouts. Input tensor supports NCHiWi data layout.
//             Weights tensor supports KCYX data layout. Output tensor supports
//             NKHoWo data layout.
//
// @tparam     InDataType               Input tensor data type.
// @tparam     WeiDataType              Weights tensor data type.
// @tparam     OutDataType              Output tensor data type.
// @tparam     InElementwiseOperation   Functor for input tensor elementwise
//                                      operation.
// @tparam     WeiElementwiseOperation  Functor for weights tensor elementwise
//                                      operation.
// @tparam     NumDimSpatial  Number of spatial dimensions.
//
template <typename InDataType,
          typename WeiDataType,
          typename OutDataType,
          typename InElementwiseOperation,
          typename WeiElementwiseOperation,
          typename OutElementwiseOperation,
          ck::index_t NumDimSpatial                                                     = 2,
          typename std::enable_if<NumDimSpatial >= 1 && NumDimSpatial <= 3, bool>::type = false>
struct ReferenceConvFwd : public device::BaseOperator
{
    // Argument
    struct Argument : public device::BaseArgument
    {
        Argument(const Tensor<InDataType>& input,
                 const Tensor<WeiDataType>& weight,
                 Tensor<OutDataType>& output,
                 std::vector<ck::index_t> conv_filter_strides,
                 std::vector<ck::index_t> conv_filter_dilations,
                 std::vector<ck::index_t> input_left_pads,
                 std::vector<ck::index_t> input_right_pads,
                 InElementwiseOperation in_element_op,
                 WeiElementwiseOperation wei_element_op,
                 OutElementwiseOperation out_element_op)
            : input_{input},
              weight_{weight},
              output_{output},
              conv_strides_{conv_filter_strides},
              conv_dilations_{conv_filter_dilations},
              in_left_pads_{input_left_pads},
              in_right_pads_{input_right_pads},
              in_element_op_{in_element_op},
              wei_element_op_{wei_element_op},
              out_element_op_{out_element_op}
        {
        }

        const Tensor<InDataType>& input_;
        const Tensor<WeiDataType>& weight_;
        Tensor<OutDataType>& output_;

        std::vector<index_t> conv_strides_;
        std::vector<index_t> conv_dilations_;
        std::vector<index_t> in_left_pads_;
        std::vector<index_t> in_right_pads_;

        InElementwiseOperation in_element_op_;
        WeiElementwiseOperation wei_element_op_;
        OutElementwiseOperation out_element_op_;
    };

    struct Invoker : public device::BaseInvoker
    {
        using Argument = ReferenceConvFwd::Argument;

        float Run(const Argument& arg)
        {
            if constexpr(NumDimSpatial == 1)
            {
                auto f_ncw = [&](auto n, auto k, auto wo) {
                    float v_acc = 0;

                    for(int c = 0; c < arg.weight_.mDesc.GetLengths()[1]; ++c)
                    {
                        for(int x = 0; x < arg.weight_.mDesc.GetLengths()[2]; ++x)
                        {
                            int wi = wo * arg.conv_strides_[0] + x * arg.conv_dilations_[0] -
                                     arg.in_left_pads_[0];
                            if(wi >= 0 && wi < arg.input_.mDesc.GetLengths()[2])
                            {
                                float v_in;
                                float v_wei;

                                arg.in_element_op_(v_in,
                                                   static_cast<const float>(arg.input_(n, c, wi)));
                                arg.wei_element_op_(v_wei,
                                                    static_cast<const float>(arg.weight_(k, c, x)));

                                v_acc += v_in * v_wei;
                            }
                        }
                    }

                    float v_out;

                    arg.out_element_op_(v_out, v_acc);
                    arg.output_(n, k, wo) = v_out;
                };

                make_ParallelTensorFunctor(f_ncw,
                                           arg.output_.mDesc.GetLengths()[0],
                                           arg.output_.mDesc.GetLengths()[1],
                                           arg.output_.mDesc.GetLengths()[2])(
                    std::thread::hardware_concurrency());

                return 0;
            }
            else if constexpr(NumDimSpatial == 2)
            {
                auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
                    float v_acc = 0;

                    for(int c = 0; c < arg.weight_.mDesc.GetLengths()[1]; ++c)
                    {
                        for(int y = 0; y < arg.weight_.mDesc.GetLengths()[2]; ++y)
                        {
                            int hi = ho * arg.conv_strides_[0] + y * arg.conv_dilations_[0] -
                                     arg.in_left_pads_[0];
                            for(int x = 0; x < arg.weight_.mDesc.GetLengths()[3]; ++x)
                            {
                                int wi = wo * arg.conv_strides_[1] + x * arg.conv_dilations_[1] -
                                         arg.in_left_pads_[1];
                                if(hi >= 0 && hi < arg.input_.mDesc.GetLengths()[2] && wi >= 0 &&
                                   wi < arg.input_.mDesc.GetLengths()[3])
                                {
                                    float v_in;
                                    float v_wei;

                                    arg.in_element_op_(
                                        v_in, ck::type_convert<float>(arg.input_(n, c, hi, wi)));
                                    arg.wei_element_op_(
                                        v_wei, ck::type_convert<float>(arg.weight_(k, c, y, x)));
                                    v_acc += v_in * v_wei;
                                }
                            }
                        }
                    }

                    float v_out;

                    arg.out_element_op_(v_out, v_acc);
                    arg.output_(n, k, ho, wo) = ck::type_convert<OutDataType>(v_out);
                };

                make_ParallelTensorFunctor(f_nchw,
                                           arg.output_.mDesc.GetLengths()[0],
                                           arg.output_.mDesc.GetLengths()[1],
                                           arg.output_.mDesc.GetLengths()[2],
                                           arg.output_.mDesc.GetLengths()[3])(
                    std::thread::hardware_concurrency());

                return 0;
            }
        }

        float Run(const device::BaseArgument* p_arg, int) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg));
        }
    };

    static constexpr bool IsValidCompilationParameter()
    {
        // TODO: properly implement this check
        return true;
    }

    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }

    static auto MakeArgument(const Tensor<InDataType>& input,
                             const Tensor<WeiDataType>& weight,
                             Tensor<OutDataType>& output,
                             std::vector<ck::index_t> conv_filter_strides,
                             std::vector<ck::index_t> conv_filter_dilations,
                             std::vector<ck::index_t> input_left_pads,
                             std::vector<ck::index_t> input_right_pads,
                             InElementwiseOperation in_element_op,
                             WeiElementwiseOperation wei_element_op,
                             OutElementwiseOperation out_element_op)
    {
        return Argument{input,
                        weight,
                        output,
                        conv_filter_strides,
                        conv_filter_dilations,
                        input_left_pads,
                        input_right_pads,
                        in_element_op,
                        wei_element_op,
                        out_element_op};
    }

    static auto MakeInvoker() { return Invoker{}; }

    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
    {
        return std::make_unique<Invoker>(Invoker{});
    }

    std::string GetTypeString() const override
    {
        auto str = std::stringstream();

        // clang-format off
        str << "ReferenceConvFwd"
            << std::endl;
        // clang-format on

        return str.str();
    }
};

} // namespace host
} // namespace tensor_operation
} // namespace ck
#endif