nvidia--tensorrt/plugin/scatterPlugin/scatterPlugin.cpp

/*
 * Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */
#include "scatterPlugin.h"
#include "half.h"
#include <cstring>
#include <cublas_v2.h>
#include <cudnn.h>
#include <iostream>
#include <sstream>

using namespace nvinfer1;

using nvinfer1::plugin::ScatterND;
using nvinfer1::plugin::ScatterNDPluginCreator;

namespace
{

const char* SCATTERND_PLUGIN_VERSION{"1"};
const char* SCATTERND_PLUGIN_NAME{"ScatterND"};
} // namespace

PluginFieldCollection ScatterNDPluginCreator::mFC{};

ScatterND::ScatterND()
{
    
}

int ScatterND::getNbOutputs() const noexcept
{
    // Plugin layer has 1 output
    return 1;
}

DimsExprs ScatterND::getOutputDimensions(int32_t outputIndex, const DimsExprs* inputs, int32_t nbInputs, IExprBuilder& exprBuilder)  noexcept
{
    //output should have same dimensions as data tensor
    DimsExprs ret = inputs[dataTensorIdx];
    return ret;
}

int ScatterND::initialize() noexcept
{
    return 0;
}

void ScatterND::terminate() noexcept
{
}

bool ScatterND::supportsFormatCombination(int32_t pos, const PluginTensorDesc* inOut, int32_t nbInputs, int32_t nbOutputs) noexcept
{
    ASSERT(pos < 4);
    ASSERT(nbInputs == 3);
    ASSERT(nbOutputs == 1);
    const PluginTensorDesc& desc = inOut[pos];
    bool ret = false;
    switch (pos)
    {
    case dataTensorIdx:
    case updateTensorIdx:
        ret = ((desc.type == DataType::kFLOAT || desc.type == DataType::kINT32)
            && desc.format == TensorFormat::kLINEAR);
        break;
    case indexTensorIdx:
        ret = (desc.type == DataType::kINT32 && desc.format == TensorFormat::kLINEAR);
        break;
    case 3:
        ret = ((desc.type == DataType::kFLOAT || desc.type == DataType::kINT32) && desc.format == TensorFormat::kLINEAR);
        break;
    }
    return ret;
}

void ScatterND::configurePlugin(const DynamicPluginTensorDesc* in, int32_t nbInputs, const DynamicPluginTensorDesc* out, int32_t nbOutputs) noexcept
{

}

int32_t ScatterND::calculateNumSlices(Dims indexTensorDims) const noexcept
{
    int32_t nSlices = 1;
    for (int i = 0; i < indexTensorDims.nbDims-1; i++)
    {
        nSlices *= indexTensorDims.d[i];
    }
    return nSlices;
}

size_t ScatterND::getWorkspaceSize(const PluginTensorDesc* inputs, int32_t nbInputs, const PluginTensorDesc* outputs,int32_t nbOutputs) const noexcept
{
    int32_t nSlices = calculateNumSlices(inputs[indexTensorIdx].dims);
    //transformCoeffs + transformed indices
    return outputs[0].dims.MAX_DIMS * sizeof(int32_t) + nSlices * sizeof(int32_t);
}

void ScatterND::calculateTransformCoeff(const Dims& dataTensorDims, int indexRank, int32_t* transformCoeff) const noexcept
{    
    std::vector<int32_t> pitches;    
    for (int32_t i = indexRank - 1, nIndx = 1; i >= 0 ; i--)
    {
        pitches.push_back(nIndx);
        nIndx *= dataTensorDims.d[i];        
    }

    std::reverse(pitches.begin(), pitches.end()); //last dimension pitch is always one (assuming linear mem)

    std::copy(pitches.begin(), pitches.end(), transformCoeff);

}

int32_t ScatterND::calculateCopySize(const Dims& dataDims) const noexcept
{
    int32_t copySize = 1;
    for (int i = 0; i < dataDims.nbDims; i++)
    {
        copySize *= dataDims.d[i];    
    }
    copySize *= sizeof(float);
    return copySize;
}

int32_t ScatterND::enqueue(const PluginTensorDesc* inputDesc, const PluginTensorDesc* outputDesc,
        const void* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream) noexcept
{  
    int32_t transformCoeff[outputDesc[0].dims.MAX_DIMS];
    std::memset(transformCoeff, 0, sizeof(int32_t)*outputDesc[0].dims.MAX_DIMS);
    Dims IndexDims = inputDesc[indexTensorIdx].dims;
    
    Dims dataDims = inputDesc[dataTensorIdx].dims;

    int32_t indexRank = IndexDims.d[IndexDims.nbDims-1];
    ASSERT(indexRank <= dataDims.nbDims);

    int32_t nSlices = calculateNumSlices(IndexDims);
    int32_t rowSize = 1;
    int32_t copySize = calculateCopySize(dataDims);
    int32_t elementSizeInBytes = 1;
    switch (inputDesc->type)
    {
    case DataType::kFLOAT:
    case DataType::kINT32:
        elementSizeInBytes = 4;
        break;
    case DataType::kHALF:
        elementSizeInBytes = 2;
        break;
    case DataType::kINT8:
    case DataType::kBOOL:
        elementSizeInBytes = 1;
        break;
    }
    
    for (int i = indexRank; i < dataDims.nbDims; i++)
    {
        rowSize *= dataDims.d[i];
    }
    
    calculateTransformCoeff(dataDims, indexRank, transformCoeff);

    scatterNDInference(stream, transformCoeff, 
    dataDims.nbDims, 
    indexRank, 
    nSlices, 
    rowSize,  
    copySize, 
    elementSizeInBytes,  
    inputs[indexTensorIdx],
    inputs[updateTensorIdx],
    inputs[dataTensorIdx],
    outputs[0],
    workspace );    

    return 0;
}

size_t ScatterND::getSerializationSize() const noexcept
{
    
    return 0;
}

void ScatterND::serialize(void* buffer) const noexcept
{
    return;
}



// Set plugin namespace
void ScatterND::setPluginNamespace(const char* pluginNamespace) noexcept
{
    mPluginNamespace = pluginNamespace;
}

const char* ScatterND::getPluginNamespace() const noexcept
{
    return mPluginNamespace.c_str();
}

// Return the DataType of the plugin output at the requested index
DataType ScatterND::getOutputDataType(int index, const nvinfer1::DataType* inputTypes, int nbInputs) const noexcept
{
    ASSERT(index == 0);
    return inputTypes[dataTensorIdx];
}

// Attach the plugin object to an execution context and grant the plugin the access to some context resource.
void ScatterND::attachToContext(cudnnContext* cudnn, cublasContext* cublas, IGpuAllocator* gpuAllocator) noexcept
{
    return;    
}

// Detach the plugin object from its execution context.
void ScatterND::detachFromContext() noexcept {}

const char* ScatterND::getPluginType() const noexcept
{
    return SCATTERND_PLUGIN_NAME;
}

const char* ScatterND::getPluginVersion() const noexcept
{
    return SCATTERND_PLUGIN_VERSION;
}

void ScatterND::destroy() noexcept
{
    delete this;
}

// Clone the plugin
IPluginV2DynamicExt* ScatterND::clone() const noexcept
{
    // Create a new instance
    ScatterND* plugin = new ScatterND();
    plugin->setPluginNamespace(mPluginNamespace.c_str());
    return plugin;
}

ScatterNDPluginCreator::ScatterNDPluginCreator()
{
    mFC.nbFields = 0;
}

const char* ScatterNDPluginCreator::getPluginName() const noexcept
{
    return SCATTERND_PLUGIN_NAME;
}

const char* ScatterNDPluginCreator::getPluginVersion() const noexcept
{
    return SCATTERND_PLUGIN_VERSION;
}

const PluginFieldCollection* ScatterNDPluginCreator::getFieldNames() noexcept
{
    return &mFC;
}

IPluginV2Ext* ScatterNDPluginCreator::createPlugin(const char* name, const PluginFieldCollection* fc) noexcept
{    
    ScatterND* obj = new ScatterND();
    obj->setPluginNamespace(mNamespace.c_str());
    return obj;
}

IPluginV2Ext* ScatterNDPluginCreator::deserializePlugin(const char* name, const void* serialData, size_t serialLength) noexcept
{
    // This object will be deleted when the network is destroyed, which will
    // call Normalize::destroy()
    ScatterND* obj = new ScatterND();
    obj->setPluginNamespace(mNamespace.c_str());
    return obj;
}