make_trt_net.cpp

/**
 * \file src/tensorrt/test/make_trt_net.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 */

#include "megbrain/opr/blas.h"
#include "megbrain/opr/dnn/convolution.h"
#include "megbrain/opr/io.h"
#include "megbrain/opr/tensor_manip.h"

#include "megbrain/opr/basic_arith.h"
#include "megbrain/plugin/profiler.h"
#include "megbrain/test/helper.h"
#include "megbrain/utils/debug.h"

#if MGB_ENABLE_TENSOR_RT
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
#include "make_trt_net.h"
#include "megbrain/tensorrt/tensorrt_opr.h"

#include <NvInferPlugin.h>
#include <random>

using namespace mgb;
using namespace opr;
using namespace nvinfer1;

intl::SimpleTensorRTNetwork::SimpleTensorRTNetwork() {
    host_x = gen({5, 23, 28, 28});
    host_w = gen({32, 23, 3, 3});
    host_b = gen({1, 32, 1, 1});

    graph = ComputingGraph::make();
    x = Host2DeviceCopy::make(*graph, host_x);
    auto w = Host2DeviceCopy::make(*graph, host_w),
         b = Host2DeviceCopy::make(*graph, host_b), y0 = opr::Convolution::make(x, w);
    y = y0 + b;
}

std::pair<nvinfer1::IBuilder*, INetworkDefinition*> intl::SimpleTensorRTNetwork::
        create_trt_network(bool has_batch_dim) {
    CompNode::load("xpu0").activate();
    Weights wt_filter{DataType::kFLOAT, nullptr, 0},
            wt_bias{DataType::kFLOAT, nullptr, 0};
    wt_filter.type = DataType::kFLOAT;
    wt_bias.type = DataType::kFLOAT;
    wt_filter.values = host_w->raw_ptr();
    wt_bias.values = host_b->raw_ptr();
    wt_filter.count = host_w->shape().total_nr_elems();
    wt_bias.count = host_b->shape().total_nr_elems();
    auto builder = createInferBuilder(TensorRTOpr::Logger::instance());
#if NV_TENSOR_RT_VERSION >= 6001
    nvinfer1::NetworkDefinitionCreationFlags flags;
    ::memset(&flags, 0, sizeof(nvinfer1::NetworkDefinitionCreationFlags));
    if (has_batch_dim)
        flags = 1 << static_cast<int>(
                        nvinfer1::NetworkDefinitionCreationFlag::kEXPLICIT_BATCH);
    auto network = builder->createNetworkV2(flags);
#else
    auto network = builder->createNetwork();
#endif
    nvinfer1::ITensor* data;
#if NV_TENSOR_RT_VERSION >= 6001
    if (has_batch_dim) {
        data = network->addInput("data", DataType::kFLOAT, Dims4{5, 23, 28, 28});
    } else {
        data = network->addInput("data", DataType::kFLOAT, Dims3{23, 28, 28});
    }
    {
        nvinfer1::TensorFormats formats =
                1 << static_cast<int>(nvinfer1::TensorFormat::kLINEAR);
        data->setAllowedFormats(formats);
    }
#else
    if (has_batch_dim) {
        data = network->addInput("data", DataType::kFLOAT, DimsNCHW{5, 23, 28, 28});
    } else {
        data = network->addInput("data", DataType::kFLOAT, DimsCHW{23, 28, 28});
    }
#endif
    mgb_assert(data != nullptr, "data is invalid");
    auto conv1 = network->addConvolution(*data, 32, DimsHW{3, 3}, wt_filter, wt_bias);
    mgb_assert(conv1 != nullptr, "conv1 is invalid");
    conv1->setStride(DimsHW{1, 1});
    conv1->getOutput(0)->setName("prob");
    network->markOutput(*conv1->getOutput(0));
#if NV_TENSOR_RT_VERSION >= 6001
    {
        nvinfer1::TensorFormats formats =
                1 << static_cast<int>(nvinfer1::TensorFormat::kLINEAR);
        conv1->getOutput(0)->setAllowedFormats(formats);
    }
#endif

    return std::make_pair(builder, network);
}

intl::BatchedTensorRTNetwork::BatchedTensorRTNetwork() {
    host_x = gen({23, 28, 28});

    graph = ComputingGraph::make();
    x = Host2DeviceCopy::make(*graph, host_x);
    opr::Reduce::Param param1{Reduce::Mode::SUM, 0, Reduce::Param::DataType::DEFAULT};
    opr::Reduce::Param param2{Reduce::Mode::SUM, 1, Reduce::Param::DataType::DEFAULT};
    auto y0 = opr::Reduce::make(x, param1);
    auto y1 = opr::Reduce::make(y0, param2);
    TensorShape tshp{1, 28};
    y = opr::Reshape::make(y1, tshp);
}

std::pair<nvinfer1::IBuilder*, INetworkDefinition*> intl::BatchedTensorRTNetwork::
        create_trt_network(bool has_batch_dim) {
    CompNode::load("xpu0").activate();
    auto builder = createInferBuilder(TensorRTOpr::Logger::instance());
#if NV_TENSOR_RT_VERSION >= 6001
    nvinfer1::NetworkDefinitionCreationFlags flags;
    ::memset(&flags, 0, sizeof(nvinfer1::NetworkDefinitionCreationFlags));
    if (has_batch_dim)
        flags = 1 << static_cast<int>(
                        nvinfer1::NetworkDefinitionCreationFlag::kEXPLICIT_BATCH);
    auto network = builder->createNetworkV2(flags);
#else
    auto network = builder->createNetwork();
#endif
    nvinfer1::ITensor* data;
#if NV_TENSOR_RT_VERSION >= 6001
    if (has_batch_dim) {
        data = network->addInput("data", DataType::kFLOAT, Dims4{1, 23, 28, 28});
    } else {
        data = network->addInput("data", DataType::kFLOAT, Dims3{23, 28, 28});
    }
    {
        nvinfer1::TensorFormats formats =
                1 << static_cast<int>(nvinfer1::TensorFormat::kLINEAR);
        data->setAllowedFormats(formats);
    }
#else
    if (has_batch_dim) {
        data = network->addInput("data", DataType::kFLOAT, DimsNCHW{1, 23, 28, 28});
    } else {
        data = network->addInput("data", DataType::kFLOAT, DimsCHW{23, 28, 28});
    }
#endif
    mgb_assert(data != nullptr, "data is invalid");
    auto reduce1 = network->addReduce(*data, nvinfer1::ReduceOperation::kSUM, 3, false);
    mgb_assert(reduce1 != nullptr, "reduce1 is invalid");
    reduce1->getOutput(0)->setName("prob");
    network->markOutput(*reduce1->getOutput(0));
#if NV_TENSOR_RT_VERSION >= 6001
    {
        nvinfer1::TensorFormats formats =
                1 << static_cast<int>(nvinfer1::TensorFormat::kLINEAR);
        reduce1->getOutput(0)->setAllowedFormats(formats);
    }
#endif

    return std::make_pair(builder, network);
}

intl::SimpleQuantizedTensorRTNetwork::SimpleQuantizedTensorRTNetwork() {
    host_x = range_gen({32, 8, 28, 28});
    host_w = weight_gen({8, 8, 3, 3});
    host_b = range_gen({1, 8, 1, 1});

    {
        void* w_ptr = host_w->raw_ptr();
        float* ptr = reinterpret_cast<float*>(w_ptr);
        ptr[0] = -127 * 1.1f;
        ptr[1] = 127 * 1.1f;
    }

    graph = ComputingGraph::make();
    auto mkvar = [this](const char* name, const std::shared_ptr<HostTensorND>& host_ts,
                        const DType& dtype) {
        return opr::TypeCvt::make(
                opr::Host2DeviceCopy::make(*graph, host_ts).rename(name), dtype);
    };
    auto mkcvar = [this](const char* name, const std::shared_ptr<HostTensorND>& host_ts,
                         const DType& dtype) {
        return opr::TypeCvt::make(
                opr::SharedDeviceTensor::make(*graph, *host_ts).rename(name), dtype);
    };

    x = mkvar("x", host_x, dtype::Float32());
    quantized_x = mkvar("quantized_x", host_x, dtype::QuantizedS8(1.2f));
    auto float_w = mkcvar("float_w", host_w, dtype::Float32()),
         float_b = mkcvar("float_b", host_b, dtype::Float32()),
         w = opr::TypeCvt::make(float_w, dtype::QuantizedS8(1.1f)),
         b = opr::TypeCvt::make(float_b, dtype::QuantizedS32(1.2f * 1.1f));

    {
        auto xshp = opr::GetVarShape::make(quantized_x);

        auto cv = [this](int v) { return quantized_x.make_scalar(v); };
        auto sub = [&xshp, &cv](int idx) {
            return opr::IndexAt::make(xshp, {{0, cv(idx)}});
        };
        auto tshp = opr::Concat::make({sub(0), sub(1) / 4, cv(4), sub(2), sub(3)}, 0);
        quantized_x = opr::Reshape::make(quantized_x, tshp);
        quantized_x = opr::Dimshuffle::make(quantized_x, {0, 1, 3, 4, 2});
    }

    {
        auto wshp = opr::GetVarShape::make(w);

        auto cv = [&w](int v) { return w.make_scalar(v); };
        auto sub = [&wshp, &cv](int idx) {
            return opr::IndexAt::make(wshp, {{0, cv(idx)}});
        };
        auto tshp = opr::Concat::make({sub(0), sub(1) / 4, cv(4), sub(2), sub(3)}, 0);
        w = opr::Reshape::make(w, tshp);
        w = opr::Dimshuffle::make(w, {0, 1, 3, 4, 2});
    }

    {
        auto bshp = opr::GetVarShape::make(b);

        auto cv = [&b](int v) { return b.make_scalar(v); };
        auto sub = [&bshp, &cv](int idx) {
            return opr::IndexAt::make(bshp, {{0, cv(idx)}});
        };
        auto tshp = opr::Concat::make({sub(0), sub(1) / 4, cv(4), sub(2), sub(3)}, 0);
        b = opr::Reshape::make(b, tshp);
        b = opr::Dimshuffle::make(b, {0, 1, 3, 4, 2});
    }

    opr::ConvBias::Param param;
    param.format = opr::ConvBias::Param::Format::NCHW4;
    param.nonlineMode = opr::ConvBias::Param::NonlineMode::IDENTITY;
    param.stride_h = param.stride_w = 1;
    param.pad_h = param.pad_w = 1;

    quantized_y = opr::ConvBias::make(
            quantized_x, w, b, param, {}, OperatorNodeConfig{dtype::QuantizedS8(1.1f)});
    param.format = opr::ConvBias::Param::Format::NCHW;
    y = opr::ConvBias::make(
            x, float_w, float_b, param, {}, OperatorNodeConfig{dtype::Float32()});

    auto yshp = opr::GetVarShape::make(quantized_y);

    auto cv = [this](int v) { return quantized_y.make_scalar(v); };
    auto sub = [&yshp, &cv](int idx) {
        return opr::IndexAt::make(yshp, {{0, cv(idx)}});
    };
    auto tshp = opr::Concat::make({sub(0), sub(1) * 4, sub(2), sub(3)}, 0);
    quantized_y = opr::Dimshuffle::make(quantized_y, {0, 1, 4, 2, 3});
    quantized_y = opr::Reshape::make(quantized_y, tshp);
    quantized_y = TypeCvt::make(quantized_y, dtype::Float32());
}

std::pair<nvinfer1::IBuilder*, INetworkDefinition*> intl::
        SimpleQuantizedTensorRTNetwork::create_trt_network(bool has_batch_dim) {
    CompNode::load("xpu0").activate();
    Weights wt_filter{DataType::kFLOAT, nullptr, 0},
            wt_bias{DataType::kFLOAT, nullptr, 0};
    wt_filter.type = DataType::kFLOAT;
    wt_bias.type = DataType::kFLOAT;
    wt_filter.values = host_w->raw_ptr();
    wt_bias.values = host_b->raw_ptr();
    wt_filter.count = host_w->shape().total_nr_elems();
    wt_bias.count = host_b->shape().total_nr_elems();
    auto builder = createInferBuilder(TensorRTOpr::Logger::instance());
#if NV_TENSOR_RT_VERSION >= 6001
    nvinfer1::NetworkDefinitionCreationFlags flags;
    ::memset(&flags, 0, sizeof(nvinfer1::NetworkDefinitionCreationFlags));
    if (has_batch_dim)
        flags = 1 << static_cast<int>(
                        nvinfer1::NetworkDefinitionCreationFlag::kEXPLICIT_BATCH);
    auto network = builder->createNetworkV2(flags);
#else
    auto network = builder->createNetwork();
#endif
    nvinfer1::ITensor* data;
#if NV_TENSOR_RT_VERSION >= 6001
    if (has_batch_dim) {
        data = network->addInput("data", DataType::kFLOAT, Dims4{32, 8, 28, 28});
    } else {
        data = network->addInput("data", DataType::kFLOAT, Dims3{8, 28, 28});
    }
    {
        nvinfer1::TensorFormats formats =
                1 << static_cast<int>(nvinfer1::TensorFormat::kLINEAR);
        data->setAllowedFormats(formats);
    }
#else
    if (has_batch_dim) {
        data = network->addInput("data", DataType::kFLOAT, DimsNCHW{32, 8, 28, 28});
    } else {
        data = network->addInput("data", DataType::kFLOAT, DimsCHW{8, 28, 28});
    }
#endif
    data->setDynamicRange(-127.f * 1.2f, 127.f * 1.2f);
    mgb_assert(data != nullptr, "data is invalid");
    auto add_conv = [&](const char* name, nvinfer1::ITensor* inp) {
        auto conv = network->addConvolution(*inp, 8, DimsHW{3, 3}, wt_filter, wt_bias);
        mgb_assert(conv != nullptr, "conv1 is invalid");
        conv->setName(name);
        conv->setStride(DimsHW{1, 1});
        conv->setPadding(DimsHW{1, 1});
        conv->getOutput(0)->setDynamicRange(-127.f * 1.1f, 127.f * 1.1f);
        // conv->setPrecision(nvinfer1::DataType::kINT8);
        return conv->getOutput(0);
    };
    auto out = add_conv("conv1", data);
    out->setName("prob");
#if NV_TENSOR_RT_VERSION >= 6001
    {
        nvinfer1::TensorFormats formats =
                1 << static_cast<int>(nvinfer1::TensorFormat::kLINEAR);
        out->setAllowedFormats(formats);
    }
#endif
    network->markOutput(*out);

    return std::make_pair(builder, network);
}

intl::ConcatConvTensorRTNetwork::ConcatConvTensorRTNetwork() {
    host_x0 = gen({5, 23, 14, 28});
    host_x1 = gen({5, 23, 14, 28});
    host_w = gen({32, 46, 3, 3});
    host_b = gen({1, 32, 1, 1});

    graph = ComputingGraph::make();
    x0 = Host2DeviceCopy::make(*graph, host_x0);
    x1 = Host2DeviceCopy::make(*graph, host_x1);
    auto y0 = opr::Concat::make({x0, x1}, 1), w = Host2DeviceCopy::make(*graph, host_w),
         b = Host2DeviceCopy::make(*graph, host_b), y1 = opr::Convolution::make(y0, w);
    y = y1 + b;
}

std::pair<nvinfer1::IBuilder*, INetworkDefinition*> intl::ConcatConvTensorRTNetwork::
        create_trt_network(bool has_batch_dim) {
    CompNode::load("xpu0").activate();
    auto builder = createInferBuilder(TensorRTOpr::Logger::instance());
#if NV_TENSOR_RT_VERSION >= 6001
    nvinfer1::NetworkDefinitionCreationFlags flags;
    ::memset(&flags, 0, sizeof(nvinfer1::NetworkDefinitionCreationFlags));
    if (has_batch_dim)
        flags = 1 << static_cast<int>(
                        nvinfer1::NetworkDefinitionCreationFlag::kEXPLICIT_BATCH);
    auto network = builder->createNetworkV2(flags);
#else
    auto network = builder->createNetwork();
#endif
    ITensor *data0, *data1;
#if NV_TENSOR_RT_VERSION >= 6001
    if (has_batch_dim) {
        data0 = network->addInput("x0", DataType::kFLOAT, Dims4{5, 23, 14, 28});
        data1 = network->addInput("x1", DataType::kFLOAT, Dims4{5, 23, 14, 28});
    } else {
        data0 = network->addInput("x0", DataType::kFLOAT, Dims3{23, 14, 28});
        data1 = network->addInput("x1", DataType::kFLOAT, Dims3{23, 14, 28});
    }
    {
        nvinfer1::TensorFormats formats =
                1 << static_cast<int>(nvinfer1::TensorFormat::kLINEAR);
        data0->setAllowedFormats(formats);
        data1->setAllowedFormats(formats);
    }
#else
    if (has_batch_dim) {
        data0 = network->addInput("x0", DataType::kFLOAT, DimsNCHW{5, 23, 14, 28});
        data1 = network->addInput("x1", DataType::kFLOAT, DimsNCHW{5, 23, 14, 28});
    } else {
        data0 = network->addInput("x0", DataType::kFLOAT, DimsCHW{23, 14, 28});
        data1 = network->addInput("x1", DataType::kFLOAT, DimsCHW{23, 14, 28});
    }
#endif
    ITensor* inputTensors[] = {data0, data1};
    auto concat = network->addConcatenation(inputTensors, 2);
    mgb_assert(concat != nullptr, "concat is null!");
    concat->setName("concat0");
    if (has_batch_dim) {
        concat->setAxis(1);
    } else {
        concat->setAxis(0);
    }

    Weights wt_filter{DataType::kFLOAT, host_w->raw_ptr(), 0},
            wt_bias{DataType::kFLOAT, host_b->raw_ptr(), 0};
    wt_filter.count = host_w->shape().total_nr_elems();
    wt_bias.count = host_b->shape().total_nr_elems();
    auto conv1 = network->addConvolution(
            *concat->getOutput(0), 32, DimsHW{3, 3}, wt_filter, wt_bias);
    mgb_assert(conv1 != nullptr, "conv1 is invalid");
    conv1->setName("conv1");
    conv1->setStride(DimsHW{1, 1});
    conv1->getOutput(0)->setName("convOut");
    network->markOutput(*conv1->getOutput(0));
#if NV_TENSOR_RT_VERSION >= 6001
    {
        nvinfer1::TensorFormats formats =
                1 << static_cast<int>(nvinfer1::TensorFormat::kLINEAR);
        conv1->getOutput(0)->setAllowedFormats(formats);
    }
#endif
    return std::make_pair(builder, network);
}

intl::ReshapeConcatTensorRTNetwork::ReshapeConcatTensorRTNetwork() {
    host_x0 = gen({2, 2, 2, 2});
    host_y0 = gen({2, 3, 2, 2});

    graph = ComputingGraph::make();
    x0 = Host2DeviceCopy::make(*graph, host_x0);
    y0 = Host2DeviceCopy::make(*graph, host_y0);
    auto x1 = opr::Reshape::make(x0, {2, 8, 1, 1}),
         y1 = opr::Reshape::make(y0, {2, 12, 1, 1});
    z = opr::Concat::make({x1, y1}, 1);
}

std::pair<nvinfer1::IBuilder*, INetworkDefinition*> intl::ReshapeConcatTensorRTNetwork::
        create_trt_network(bool has_batch_dim) {
    initLibNvInferPlugins(&TensorRTOpr::Logger::instance(), "");

    CompNode::load("xpu0").activate();
    auto builder = createInferBuilder(TensorRTOpr::Logger::instance());
#if NV_TENSOR_RT_VERSION >= 6001
    nvinfer1::NetworkDefinitionCreationFlags flags;
    ::memset(&flags, 0, sizeof(nvinfer1::NetworkDefinitionCreationFlags));
    if (has_batch_dim)
        flags = 1 << static_cast<int>(
                        nvinfer1::NetworkDefinitionCreationFlag::kEXPLICIT_BATCH);
    auto network = builder->createNetworkV2(flags);
#else
    auto network = builder->createNetwork();
#endif
    nvinfer1::ITensor *data0, *data1;
#if NV_TENSOR_RT_VERSION >= 6001
    if (has_batch_dim) {
        data0 = network->addInput("x0", DataType::kFLOAT, Dims4{2, 2, 2, 2});
        data1 = network->addInput("y0", DataType::kFLOAT, Dims4{2, 3, 2, 2});
    } else {
        data0 = network->addInput("x0", DataType::kFLOAT, Dims3{2, 2, 2});
        data1 = network->addInput("y0", DataType::kFLOAT, Dims3{3, 2, 2});
    }
    {
        nvinfer1::TensorFormats formats =
                1 << static_cast<int>(nvinfer1::TensorFormat::kLINEAR);
        data0->setAllowedFormats(formats);
        data1->setAllowedFormats(formats);
    }
#else
    if (has_batch_dim) {
        data0 = network->addInput("x0", DataType::kFLOAT, DimsNCHW{2, 2, 2, 2});
        data1 = network->addInput("y0", DataType::kFLOAT, DimsNCHW{2, 3, 2, 2});
    } else {
        data0 = network->addInput("x0", DataType::kFLOAT, DimsCHW{2, 2, 2});
        data1 = network->addInput("y0", DataType::kFLOAT, DimsCHW{3, 2, 2});
    }
#endif
    int axis = 1;
    bool ignoreBatch = false;
    nvinfer1::PluginField fields[2] = {
            nvinfer1::PluginField{"axis", &axis, nvinfer1::PluginFieldType::kINT32, 1},
            nvinfer1::PluginField{
                    "ignoreBatch", &ignoreBatch, nvinfer1::PluginFieldType::kINT32, 1},
    };
    nvinfer1::PluginFieldCollection fc{2, fields};

    auto creator = getPluginRegistry()->getPluginCreator("FlattenConcat_TRT", "1", "");
    TensorRTUniquePtr<nvinfer1::IPluginV2> plugin(
            creator->createPlugin("FlattenConcat_TRT", &fc));
    ITensor* inputTensors[] = {data0, data1};
    auto flt_cct = network->addPluginV2(inputTensors, 2, *plugin);
    mgb_assert(flt_cct != nullptr, "FlattenConcat_TRT is invalid");
    network->markOutput(*flt_cct->getOutput(0));
#if NV_TENSOR_RT_VERSION >= 6001
    {
        nvinfer1::TensorFormats formats =
                1 << static_cast<int>(nvinfer1::TensorFormat::kLINEAR);
        flt_cct->getOutput(0)->setAllowedFormats(formats);
    }
#endif
    return std::make_pair(builder, network);
}

#if NV_TENSOR_RT_VERSION >= 6001
intl::DynamicShapeTensorRTNetwork::DynamicShapeTensorRTNetwork(
        size_t n, size_t c, size_t h, size_t w) {
    host_x = gen({n, c, h, w});
    host_w1 = gen({32, 23, 3, 3});
    host_b1 = gen({1, 32, 1, 1});

    graph = ComputingGraph::make();
    x = Host2DeviceCopy::make(*graph, host_x);
    auto w1 = Host2DeviceCopy::make(*graph, host_w1),
         b1 = Host2DeviceCopy::make(*graph, host_b1),
         y01 = opr::Convolution::make(x, w1);
    y1 = y01 + b1;
}

TensorRTUniquePtr<ICudaEngine> intl::DynamicShapeTensorRTNetwork::create_trt_network() {
    CompNode::load("xpu0").activate();
    Weights wt_filter_1{DataType::kFLOAT, nullptr, 0},
            wt_bias_1{DataType::kFLOAT, nullptr, 0};
    wt_filter_1.type = DataType::kFLOAT;
    wt_bias_1.type = DataType::kFLOAT;
    wt_filter_1.values = host_w1->raw_ptr();
    wt_bias_1.values = host_b1->raw_ptr();
    wt_filter_1.count = host_w1->shape().total_nr_elems();
    wt_bias_1.count = host_b1->shape().total_nr_elems();
    auto builder = createInferBuilder(TensorRTOpr::Logger::instance());

    auto network = builder->createNetworkV2(
            1 << static_cast<int>(
                    nvinfer1::NetworkDefinitionCreationFlag::kEXPLICIT_BATCH));

    nvinfer1::ITensor* data;

    data = network->addInput("data", DataType::kFLOAT, Dims4{-1, 23, -1, -1});

    nvinfer1::IBuilderConfig* config = builder->createBuilderConfig();
    nvinfer1::IOptimizationProfile* profile = builder->createOptimizationProfile();
    profile->setDimensions(
            "data", nvinfer1::OptProfileSelector::kMIN, Dims4(3, 23, 16, 16));
    profile->setDimensions(
            "data", nvinfer1::OptProfileSelector::kOPT, Dims4(4, 23, 24, 24));
    profile->setDimensions(
            "data", nvinfer1::OptProfileSelector::kMAX, Dims4(5, 23, 28, 28));
    config->addOptimizationProfile(profile);

    {
        nvinfer1::TensorFormats formats =
                1 << static_cast<int>(nvinfer1::TensorFormat::kLINEAR);
        data->setAllowedFormats(formats);
    }

    mgb_assert(data != nullptr, "data is invalid");
    auto conv1 =
            network->addConvolution(*data, 32, DimsHW{3, 3}, wt_filter_1, wt_bias_1);
    mgb_assert(conv1 != nullptr, "conv1 is invalid");
    conv1->setStride(DimsHW{1, 1});
    conv1->getOutput(0)->setName("prob1");
    network->markOutput(*conv1->getOutput(0));

    {
        nvinfer1::TensorFormats formats =
                1 << static_cast<int>(nvinfer1::TensorFormat::kLINEAR);
        conv1->getOutput(0)->setAllowedFormats(formats);
    }

    TensorRTUniquePtr<ICudaEngine> cuda_engine{
            builder->buildEngineWithConfig(*network, *config)};

    return cuda_engine;
}
#endif

#pragma GCC diagnostic pop
#endif  // MGB_ENABLE_TENSOR_RT

// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}