alibaba--mnn/source/backend/cpu/CPURandomUniform.cpp

//
//  CPURandomUniform.cpp
//  MNN
//
//  Created by MNN on 2020/8/14.
//  Copyright © 2018, Alibaba Group Holding Limited
//

#include <random>
#include "backend/cpu/CPURandomUniform.hpp"
#include "core/Macro.h"
#include "backend/cpu/CPUBackend.hpp"

namespace MNN {
ErrorCode CPURandomUniform::onResize(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
    return NO_ERROR;
}

ErrorCode CPURandomUniform::onExecute(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
    MNN_ASSERT(outputs.size() == 1);
    auto output = outputs[0];
    int size = output->elementSize();
    if (size <= 0) {
        return NO_ERROR;
    }
    auto parameter = mOp->main_as_RandomUniform();
    float low = parameter->low();
    float high = parameter->high();
    if (low >= high) {
        MNN_ERROR("RandomUniform requires low < high, got low=%f, high=%f\n", low, high);
        return INPUT_DATA_ERROR;
    }
    auto dtype = output->getType();
    std::uniform_real_distribution<float> distribution(low, high);
    int seed = parameter->seed();
    int seed1 = parameter->seed2();
    if (dtype.code == halide_type_float) {
        auto outputPtr = output->host<float>();
        if (seed || seed1) {
            std::mt19937 generator(seed || seed1);
            for (int i = 0; i < size; i++) {
                outputPtr[i] = distribution(generator);
            }
        } else {
            std::default_random_engine generator;
            for (int i = 0; i < size; i++) {
                outputPtr[i] = distribution(generator);
            }
        }
    } else if (dtype.code == halide_type_int && dtype.bits == 32) {
        auto outputPtr = output->host<int32_t>();
        if (seed || seed1) {
            std::mt19937 generator(seed || seed1);
            for (int i = 0; i < size; i++) {
                outputPtr[i] = static_cast<int32_t>(distribution(generator));
            }
        } else {
            std::default_random_engine generator;
            for (int i = 0; i < size; i++) {
                outputPtr[i] = static_cast<int32_t>(distribution(generator));
            }
        }
    } else if (dtype.code == halide_type_uint && dtype.bits == 8) {
        auto outputPtr = output->host<uint8_t>();
        if (seed || seed1) {
            std::mt19937 generator(seed || seed1);
            for (int i = 0; i < size; i++) {
                outputPtr[i] = static_cast<uint8_t>(distribution(generator));
            }
        } else {
            std::default_random_engine generator;
            for (int i = 0; i < size; i++) {
                outputPtr[i] = static_cast<uint8_t>(distribution(generator));
            }
        }
    } else {
        // Fallback: treat as float (original behavior)
        auto outputPtr = output->host<float>();
        if (seed || seed1) {
            std::mt19937 generator(seed || seed1);
            for (int i = 0; i < size; i++) {
                outputPtr[i] = distribution(generator);
            }
        } else {
            std::default_random_engine generator;
            for (int i = 0; i < size; i++) {
                outputPtr[i] = distribution(generator);
            }
        }
    }
    return NO_ERROR;
}

ErrorCode CPURandomNormal::onResize(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
    return NO_ERROR;
}

ErrorCode CPURandomNormal::onExecute(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
    MNN_ASSERT(outputs.size() == 1);
    auto output = outputs[0];
    int size = output->elementSize();
    auto parameter = mOp->main_as_RandomUniform();
    auto outputPtr = output->host<float>();
    // RandomUniform and RandomNormal use same param table. low -> mean, high -> scale
    std::normal_distribution<float> distribution(parameter->low(),parameter->high());
    int seed = parameter->seed();
    int seed1 = parameter->seed2();
    if (seed || seed1) {
        std::mt19937 generator(seed || seed1);
        for (int i = 0; i < size; i++) {
            outputPtr[i] = distribution(generator);
        }
    } else {
        std::default_random_engine generator;
        for (int i = 0; i < size; i++) {
            outputPtr[i] = distribution(generator);
        }
    }
    return NO_ERROR;
}

class CPURandomCreator : public CPUBackend::Creator {
public:
    virtual Execution *onCreate(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs,
                                const MNN::Op *op, Backend *backend) const override {
        if (op->type() == OpType_RandomUniform) {
            return new CPURandomUniform(backend, op);
        } else {
            return new CPURandomNormal(backend, op);
        }
    }
};
REGISTER_CPU_OP_CREATOR(CPURandomCreator, OpType_RandomUniform);
REGISTER_CPU_OP_CREATOR(CPURandomCreator, OpType_RandomNormal);
} // namespace MNN