alibaba--mnn/tools/converter/source/onnx/MatMulOnnx.cpp

//
//  MatMulOnnx.cpp
//  MNNConverter
//
//  Created by MNN on 2019/01/31.
//  Copyright © 2018, Alibaba Group Holding Limited
//

#include <stdio.h>
#include "onnxOpConverter.hpp"

DECLARE_OP_CONVERTER(GemmOnnx);

MNN::OpType GemmOnnx::opType() {
    return MNN::OpType_InnerProduct;
}
MNN::OpParameter GemmOnnx::type() {
    return MNN::OpParameter_InnerProduct;
}

void GemmOnnx::run(MNN::OpT* dstOp, const onnx::NodeProto* onnxNode, OnnxScope* scope) {
    // get input initializers of gemm
    std::vector<const onnx::TensorProto*> initializers;
    for (int k = 0; k < onnxNode->input_size(); ++k) {
        const auto& inputName = onnxNode->input(k);
        const auto it         = scope->mInitializers.find(inputName);
        if (it != scope->mInitializers.end()) {
            initializers.push_back(it->second);
        }
    }
    const int size = initializers.size();
    DCHECK(size <= 2 && size >= 1) << "Gemm Input ERROR!";
    auto gemmParam = new MNN::InnerProductT;

    bool transA = false;
    bool transB = false;
    float alpha = 1.0f;
    float beta  = 1.0f;

    for (int i = 0; i < onnxNode->attribute_size(); ++i) {
        const auto& attributeProto = onnxNode->attribute(i);
        const auto& attributeName  = attributeProto.name();
        if (attributeName == "transA") {
            DCHECK(attributeProto.type() == ::onnx::AttributeProto_AttributeType_INT) << "Node Attribute ERROR";
            transA = static_cast<bool>(attributeProto.i());
        } else if (attributeName == "transB") {
            DCHECK(attributeProto.type() == ::onnx::AttributeProto_AttributeType_INT) << "Node Attribute ERROR";
            transB = static_cast<bool>(attributeProto.i());
        } else if (attributeName == "alpha") {
            DCHECK(attributeProto.type() == ::onnx::AttributeProto_AttributeType_FLOAT) << "Node Attribute ERROR";
            alpha = attributeProto.f();
        } else if (attributeName == "beta") {
            DCHECK(attributeProto.type() == ::onnx::AttributeProto_AttributeType_FLOAT) << "Node Attribute ERROR";
            beta = attributeProto.f();
        }
    }

    // TODO, MNN implement (alpha * A * B + beta * C), now (A * B + C)
    DCHECK(1 == alpha);
    DCHECK(1 == beta);

    DCHECK(!transA) << "Now GEMM not support transpose the intput tensor!";

    int weightSize         = 1;
    const auto weightProto = initializers[0];
    DCHECK(2 == weightProto->dims_size()) << "Gemm weight dimensions should be 2";
    int bK = weightProto->dims(0);
    int bN = weightProto->dims(1);

    if (transB) {
        int temp = bK;
        bK       = bN;
        bN       = temp;
    }
    for (int i = 0; i < weightProto->dims_size(); ++i) {
        weightSize *= weightProto->dims(i);
    }

    std::vector<float> weightContainer(weightSize);
    auto weightPtr = weightContainer.data();

    if (weightProto->float_data_size() != 0) {
        for (int i = 0; i < weightSize; ++i) {
            weightPtr[i] = weightProto->float_data(i);
        }
    } else if (weightProto->raw_data().data()) {
        ::memcpy(weightPtr, reinterpret_cast<const float*>(weightProto->raw_data().data()), weightSize * sizeof(float));
    } else {
        DLOG(ERROR) << "ERROR";
    }

    auto weightBlob = new MNN::BlobT;
    // tranpose weight if necessary
    weightBlob->dims.resize(2);
    weightBlob->dims[0] = bK;
    weightBlob->dims[1] = bN;
    if (transB) {
        auto& weightVector = weightBlob->float32s;
        weightVector.resize(weightSize);
        for (int i = 0; i < bK; ++i) {
            for (int j = 0; j < bN; ++j) {
                weightVector[i * bN + j] = weightContainer[j * bK + i];
            }
        }
    } else {
        weightBlob->float32s = weightContainer;
    }
    gemmParam->weight = weightContainer;

    // bias
    std::vector<float> biasContainer(bN);
    const auto biasProto = size == 2 ? initializers[1] : nullptr;
    if (biasProto) {
        int biasSize = 1;
        DCHECK(1 == biasProto->dims_size()) << "Gemm bias dimension should be 1";
        for (int i = 0; i < biasProto->dims_size(); ++i) {
            biasSize *= biasProto->dims(i);
        }
        // TODO, MNN support broadcast add for( + C)
        DCHECK(bN == biasSize) << "Gemm Now not support for broadcast mode(+ C)";
        auto biasPtr = biasContainer.data();
        if (biasProto->float_data_size() != 0) {
            for (int i = 0; i < biasSize; ++i) {
                biasPtr[i] = biasProto->float_data(i);
            }
        } else if (biasProto->raw_data().data()) {
            ::memcpy(biasPtr, reinterpret_cast<const float*>(biasProto->raw_data().data()), biasSize * sizeof(float));
        } else {
            DLOG(ERROR) << "ERROR";
        }
    }
    gemmParam->bias = biasContainer;

    gemmParam->outputCount = bN;
    gemmParam->axis        = 1;
    gemmParam->transpose   = false;
    gemmParam->biasTerm    = 1;

    dstOp->main.value = gemmParam;
}

//REGISTER_CONVERTER(GemmOnnx, Gemm);


DECLARE_OP_CONVERTER(MatMulOnnx);

MNN::OpType MatMulOnnx::opType(){
    return MNN::OpType_MatMul;
}

MNN::OpParameter MatMulOnnx::type(){
    return MNN::OpParameter_MatMul;
}

void MatMulOnnx::run(MNN::OpT *dstOp, const onnx::NodeProto *onnxNode, OnnxScope* scope){
    
    CHECK(2 == onnxNode->input_size()) << "ONNX Matmul input error!";
    auto param = new MNN::MatMulT;
    param->T = MNN::DataType_DT_FLOAT;
    
    dstOp->main.value = param;
}

REGISTER_CONVERTER(MatMulOnnx, MatMul);