PointwiseConvGenstats.cpp

#include <iostream>

#include "utils.hpp"

#include <hipdnn_data_sdk/utilities/Tensor.hpp>
#include <hipdnn_data_sdk/utilities/Workspace.hpp>
#include <hipdnn_frontend.hpp>

int main()
{
    using InputType = float;

    const int64_t n = 4; // Batch size
    // Input
    const int64_t c = 64; // Number of channels
    const int64_t h = 16; // Height
    const int64_t w = 16; // Width

    // Filter
    const int64_t k = 32; // Number of filters
    const int64_t r = 3; // Height
    const int64_t s = 3; // Width

    auto buildConvBiasPreluAddGraph = [=](hipdnnHandle_t handle) {
        auto graph = std::make_shared<hipdnn_frontend::graph::Graph>();
        graph->set_name("pw_conv_genstats_graph")
            .set_io_data_type(hipdnn_frontend::getDataTypeEnumFromType<InputType>())
            .set_intermediate_data_type(hipdnn_frontend::getDataTypeEnumFromType<InputType>())
            .set_compute_data_type(hipdnn_frontend::DataType::FLOAT); //

        // create conv
        auto input = std::make_shared<hipdnn_frontend::graph::TensorAttributes>(
            hipdnn_frontend::graph::Tensor_attributes()
                .set_name("input")
                .set_dim({n, c, h, w})
                .set_stride({c * h * w, 1, c * w, c}));

        // create bias
        auto scale = std::make_shared<hipdnn_frontend::graph::TensorAttributes>(
            hipdnn_frontend::graph::Tensor_attributes()
                .set_name("scale")
                .set_dim({1, c, 1, 1})
                .set_stride({c, 1, c, c}));
        auto mulAttributes = hipdnn_frontend::graph::PointwiseAttributes()
                                 .set_name("mul_node")
                                 .set_mode(hipdnn_frontend::PointwiseMode_t::MUL);
        auto mulOutput = graph->pointwise(input, scale, mulAttributes);

        // create add
        auto bias = std::make_shared<hipdnn_frontend::graph::TensorAttributes>(
            hipdnn_frontend::graph::Tensor_attributes()
                .set_name("bias")
                .set_dim({1, c, 1, 1})
                .set_stride({c, 1, c, c}));
        auto addAttributes = hipdnn_frontend::graph::PointwiseAttributes()
                                 .set_name("add_node")
                                 .set_mode(hipdnn_frontend::PointwiseMode_t::ADD);
        auto addOutput = graph->pointwise(mulOutput, bias, addAttributes);

        // create relu
        auto reluAttributes = hipdnn_frontend::graph::PointwiseAttributes()
                                  .set_name("relu_node")
                                  .set_mode(hipdnn_frontend::PointwiseMode_t::RELU_FWD);
        auto reluOutput = graph->pointwise(addOutput, reluAttributes);

        auto filter = std::make_shared<hipdnn_frontend::graph::TensorAttributes>(
            hipdnn_frontend::graph::Tensor_attributes()
                .set_name("filter")
                .set_dim({k, c, r, s})
                .set_stride({c * r * s, 1, c * s, c}));
        auto convFpropAttributes = hipdnn_frontend::graph::ConvFpropAttributes()
                                       .set_name("conv_fprop_node")
                                       .set_padding({1, 1})
                                       .set_stride({1, 1})
                                       .set_dilation({1, 1});
        auto convOutput = graph->conv_fprop(reluOutput, filter, convFpropAttributes);
        convOutput->set_output(true);

        auto genstatsAttributes = hipdnn_frontend::graph::GenstatsAttributes();
        auto [sum, sqSum] = graph->genstats(convOutput, genstatsAttributes);

        sum->set_output(true);
        sqSum->set_output(true);

        // build graph
        HIPDNN_FE_CHECK(graph->build(handle));

        return std::make_tuple(graph, input, filter, scale, bias, convOutput, sum, sqSum);
    };

    auto backend = hipdnn_frontend::detail::hipdnnBackend();
    if(!backend)
    {
        std::cout << "Creat backend failed. \n";
        return 1;
    }

    hipdnnHandle_t handle;
    HIPDNN_CHECK(backend->create(&handle));

    auto [graph, input, filter, scale, bias, convOutput, sum, sqSum]
        = buildConvBiasPreluAddGraph(handle);

    // Allocate DCU memory
    hipdnn_data_sdk::utilities::Tensor<InputType> inputTensor(input->get_dim(),
                                                              input->get_stride());
    hipdnn_data_sdk::utilities::Tensor<InputType> wTensor(filter->get_dim(), filter->get_stride());
    hipdnn_data_sdk::utilities::Tensor<InputType> biasTensor(bias->get_dim(), bias->get_stride());
    hipdnn_data_sdk::utilities::Tensor<InputType> scaleTensor(scale->get_dim(),
                                                              scale->get_stride());
    hipdnn_data_sdk::utilities::Tensor<InputType> outTensor(convOutput->get_dim(),
                                                            convOutput->get_stride());
    hipdnn_data_sdk::utilities::Tensor<InputType> sumTensor(sum->get_dim(), sum->get_stride());
    hipdnn_data_sdk::utilities::Tensor<InputType> sqSumTensor(sqSum->get_dim(),
                                                              sqSum->get_stride());

    std::unordered_map<int64_t, void*> variantPack;
    variantPack[input->get_uid()] = inputTensor.memory().deviceData();
    variantPack[filter->get_uid()] = wTensor.memory().deviceData();
    variantPack[bias->get_uid()] = biasTensor.memory().deviceData();
    variantPack[scale->get_uid()] = scaleTensor.memory().deviceData();
    variantPack[convOutput->get_uid()] = outTensor.memory().deviceData();
    variantPack[sum->get_uid()] = sumTensor.memory().deviceData();
    variantPack[sqSum->get_uid()] = sqSumTensor.memory().deviceData();

    int64_t workspaceSize = 0;
    HIPDNN_FE_CHECK(graph->get_workspace_size(workspaceSize));
    const hipdnn_data_sdk::utilities::Workspace workspace(static_cast<size_t>(workspaceSize));

    HIPDNN_FE_CHECK(graph->execute(handle, variantPack, workspace.get()));

    std::cout << "Pointwise_conv_genstats graph execution complete. \n";

    HIPDNN_CHECK(backend->destroy(handle));
    return 0;
}