/**
 * Synor Compute SDK - C++ Implementation
 */

#include "synor/compute.hpp"
#include <cmath>
#include <random>
#include <numeric>
#include <algorithm>
#include <stdexcept>
#include <sstream>

namespace synor {

// ============ Tensor Implementation ============

Tensor::Tensor(std::vector<int> shape, std::vector<double> data, Precision dtype)
    : shape_(std::move(shape)), data_(std::move(data)), dtype_(dtype) {
    size_t expected = 1;
    for (int dim : shape_) {
        expected *= dim;
    }
    if (data_.size() != expected) {
        throw std::invalid_argument("Data size does not match shape");
    }
}

Tensor::Tensor(std::vector<int> shape, std::span<const double> data, Precision dtype)
    : shape_(std::move(shape)), data_(data.begin(), data.end()), dtype_(dtype) {
    size_t expected = 1;
    for (int dim : shape_) {
        expected *= dim;
    }
    if (data_.size() != expected) {
        throw std::invalid_argument("Data size does not match shape");
    }
}

Tensor Tensor::zeros(std::vector<int> shape, Precision dtype) {
    size_t size = 1;
    for (int dim : shape) size *= dim;
    return Tensor(std::move(shape), std::vector<double>(size, 0.0), dtype);
}

Tensor Tensor::ones(std::vector<int> shape, Precision dtype) {
    size_t size = 1;
    for (int dim : shape) size *= dim;
    return Tensor(std::move(shape), std::vector<double>(size, 1.0), dtype);
}

Tensor Tensor::rand(std::vector<int> shape, Precision dtype) {
    size_t size = 1;
    for (int dim : shape) size *= dim;

    std::random_device rd;
    std::mt19937 gen(rd());
    std::uniform_real_distribution<> dis(0.0, 1.0);

    std::vector<double> data(size);
    std::generate(data.begin(), data.end(), [&]() { return dis(gen); });

    return Tensor(std::move(shape), std::move(data), dtype);
}

Tensor Tensor::randn(std::vector<int> shape, Precision dtype) {
    size_t size = 1;
    for (int dim : shape) size *= dim;

    std::random_device rd;
    std::mt19937 gen(rd());
    std::normal_distribution<> dis(0.0, 1.0);

    std::vector<double> data(size);
    std::generate(data.begin(), data.end(), [&]() { return dis(gen); });

    return Tensor(std::move(shape), std::move(data), dtype);
}

Tensor Tensor::eye(int n, Precision dtype) {
    std::vector<double> data(n * n, 0.0);
    for (int i = 0; i < n; i++) {
        data[i * n + i] = 1.0;
    }
    return Tensor({n, n}, std::move(data), dtype);
}

Tensor Tensor::arange(double start, double end, double step) {
    int size = static_cast<int>(std::ceil((end - start) / step));
    std::vector<double> data(size);
    for (int i = 0; i < size; i++) {
        data[i] = start + i * step;
    }
    return Tensor({size}, std::move(data));
}

Tensor Tensor::linspace(double start, double end, int num) {
    std::vector<double> data(num);
    double step = (end - start) / (num - 1);
    for (int i = 0; i < num; i++) {
        data[i] = start + i * step;
    }
    return Tensor({num}, std::move(data));
}

Tensor Tensor::reshape(std::vector<int> new_shape) const {
    size_t new_size = 1;
    for (int dim : new_shape) new_size *= dim;
    if (new_size != size()) {
        throw std::invalid_argument("Cannot reshape tensor to incompatible size");
    }
    return Tensor(std::move(new_shape), data_, dtype_);
}

Tensor Tensor::transpose() const {
    if (ndim() != 2) {
        throw std::invalid_argument("Transpose only supported for 2D tensors");
    }
    int rows = shape_[0];
    int cols = shape_[1];
    std::vector<double> transposed(data_.size());
    for (int i = 0; i < rows; i++) {
        for (int j = 0; j < cols; j++) {
            transposed[j * rows + i] = data_[i * cols + j];
        }
    }
    return Tensor({cols, rows}, std::move(transposed), dtype_);
}

double Tensor::mean() const {
    return std::accumulate(data_.begin(), data_.end(), 0.0) / data_.size();
}

double Tensor::sum() const {
    return std::accumulate(data_.begin(), data_.end(), 0.0);
}

double Tensor::std() const {
    double m = mean();
    double sum_sq = std::accumulate(data_.begin(), data_.end(), 0.0,
        [m](double acc, double x) { return acc + (x - m) * (x - m); });
    return std::sqrt(sum_sq / data_.size());
}

double Tensor::max() const {
    return *std::max_element(data_.begin(), data_.end());
}

double Tensor::min() const {
    return *std::min_element(data_.begin(), data_.end());
}

Tensor Tensor::relu() const {
    std::vector<double> result(data_.size());
    std::transform(data_.begin(), data_.end(), result.begin(),
        [](double x) { return std::max(0.0, x); });
    return Tensor(shape_, std::move(result), dtype_);
}

Tensor Tensor::sigmoid() const {
    std::vector<double> result(data_.size());
    std::transform(data_.begin(), data_.end(), result.begin(),
        [](double x) { return 1.0 / (1.0 + std::exp(-x)); });
    return Tensor(shape_, std::move(result), dtype_);
}

Tensor Tensor::softmax() const {
    double max_val = max();
    std::vector<double> exp_vals(data_.size());
    std::transform(data_.begin(), data_.end(), exp_vals.begin(),
        [max_val](double x) { return std::exp(x - max_val); });
    double sum = std::accumulate(exp_vals.begin(), exp_vals.end(), 0.0);
    std::transform(exp_vals.begin(), exp_vals.end(), exp_vals.begin(),
        [sum](double x) { return x / sum; });
    return Tensor(shape_, std::move(exp_vals), dtype_);
}

double Tensor::operator()(std::initializer_list<int> indices) const {
    if (indices.size() != shape_.size()) {
        throw std::invalid_argument("Index dimensions must match tensor dimensions");
    }
    size_t idx = 0;
    size_t stride = 1;
    auto it = indices.end();
    for (int i = static_cast<int>(shape_.size()) - 1; i >= 0; i--) {
        --it;
        idx += *it * stride;
        stride *= shape_[i];
    }
    return data_[idx];
}

bool Tensor::operator==(const Tensor& other) const {
    return shape_ == other.shape_ && data_ == other.data_ && dtype_ == other.dtype_;
}

// ============ ModelInfo Implementation ============

std::string ModelInfo::formatted_parameters() const {
    if (!parameters) return "Unknown";
    int64_t p = *parameters;
    if (p >= 1'000'000'000) {
        return std::to_string(p / 1'000'000'000) + "B";
    } else if (p >= 1'000'000) {
        return std::to_string(p / 1'000'000) + "M";
    } else if (p >= 1'000) {
        return std::to_string(p / 1'000) + "K";
    }
    return std::to_string(p);
}

// ============ Utility Functions ============

std::string_view to_string(ProcessorType type) {
    switch (type) {
        case ProcessorType::CPU: return "cpu";
        case ProcessorType::GPU: return "gpu";
        case ProcessorType::TPU: return "tpu";
        case ProcessorType::NPU: return "npu";
        case ProcessorType::LPU: return "lpu";
        case ProcessorType::FPGA: return "fpga";
        case ProcessorType::DSP: return "dsp";
        case ProcessorType::WebGPU: return "webgpu";
        case ProcessorType::WASM: return "wasm";
        case ProcessorType::Auto: return "auto";
    }
    return "unknown";
}

std::string_view to_string(Precision precision) {
    switch (precision) {
        case Precision::FP64: return "fp64";
        case Precision::FP32: return "fp32";
        case Precision::FP16: return "fp16";
        case Precision::BF16: return "bf16";
        case Precision::INT8: return "int8";
        case Precision::INT4: return "int4";
    }
    return "unknown";
}

std::string_view to_string(Priority priority) {
    switch (priority) {
        case Priority::Critical: return "critical";
        case Priority::High: return "high";
        case Priority::Normal: return "normal";
        case Priority::Low: return "low";
        case Priority::Background: return "background";
    }
    return "unknown";
}

std::string_view to_string(JobStatus status) {
    switch (status) {
        case JobStatus::Pending: return "pending";
        case JobStatus::Queued: return "queued";
        case JobStatus::Running: return "running";
        case JobStatus::Completed: return "completed";
        case JobStatus::Failed: return "failed";
        case JobStatus::Cancelled: return "cancelled";
    }
    return "unknown";
}

std::string_view to_string(ModelCategory category) {
    switch (category) {
        case ModelCategory::LLM: return "llm";
        case ModelCategory::Embedding: return "embedding";
        case ModelCategory::ImageGeneration: return "image_generation";
        case ModelCategory::ImageClassification: return "image_classification";
        case ModelCategory::ObjectDetection: return "object_detection";
        case ModelCategory::SpeechToText: return "speech_to_text";
        case ModelCategory::TextToSpeech: return "text_to_speech";
        case ModelCategory::Code: return "code";
        case ModelCategory::Custom: return "custom";
    }
    return "unknown";
}

std::optional<ProcessorType> processor_from_string(std::string_view str) {
    if (str == "cpu") return ProcessorType::CPU;
    if (str == "gpu") return ProcessorType::GPU;
    if (str == "tpu") return ProcessorType::TPU;
    if (str == "npu") return ProcessorType::NPU;
    if (str == "lpu") return ProcessorType::LPU;
    if (str == "fpga") return ProcessorType::FPGA;
    if (str == "dsp") return ProcessorType::DSP;
    if (str == "webgpu") return ProcessorType::WebGPU;
    if (str == "wasm") return ProcessorType::WASM;
    if (str == "auto") return ProcessorType::Auto;
    return std::nullopt;
}

std::optional<Precision> precision_from_string(std::string_view str) {
    if (str == "fp64") return Precision::FP64;
    if (str == "fp32") return Precision::FP32;
    if (str == "fp16") return Precision::FP16;
    if (str == "bf16") return Precision::BF16;
    if (str == "int8") return Precision::INT8;
    if (str == "int4") return Precision::INT4;
    return std::nullopt;
}

} // namespace synor