// ExecutionPlan / Device / PrecisionPolicy — how and where a computation runs.
#pragma once
#include <string>
#include <vector>
#include <nlohmann/json.hpp>
#include "raid/core/types.hpp"

namespace raid {

struct Device {
  std::string name;
  Backend kind{Backend::cpu};
  u64 memory_bytes{0};
  int compute_units{0};        // CPU cores or GPU cores (if known)
  bool unified_memory{true};
  bool supports_fp64{true};    // false for every Apple GPU (verified 2026-11-02)
  u64 max_buffer_bytes{0};
  std::string details;
  nlohmann::json to_json() const;
};

struct PrecisionPolicy {
  Precision requested{Precision::fp64};
  /// Maximum reduction length accumulated in FP32 on the GPU before a FP64 CPU combine step.
  usize fp32_chunk_max{4096};
  /// Use compensated (Neumaier) accumulation on GPU partial sums.
  bool compensated{true};
  /// Tolerance (relative) used by auto validation of GPU vs CPU reference.
  f64 validate_rtol{1e-5};
  nlohmann::json to_json() const;
};

struct ExecutionPlan {
  Backend backend{Backend::automatic};
  PrecisionPolicy precision{};
  bool deterministic{true};        // fixed-order reductions (bitwise reproducible across runs on the same device)
  unsigned threads{0};             // 0 = all
  usize chunk_rows{0};             // 0 = automatic
  u64 max_memory_bytes{0};         // 0 = no explicit bound (device-reported limits still apply)
  bool allow_fallback{true};       // Metal unsupported → CPU
  u64 seed{20261102};
  nlohmann::json to_json() const;
};

}  // namespace raid