executor.hpp

// SPDX-FileCopyrightText: 2017 - 2024 The Ginkgo authors
//
// SPDX-License-Identifier: BSD-3-Clause
 
#ifndef GKO_PUBLIC_CORE_BASE_EXECUTOR_HPP_
#define GKO_PUBLIC_CORE_BASE_EXECUTOR_HPP_
 
 
#include <array>
#include <atomic>
#include <iostream>
#include <memory>
#include <mutex>
#include <sstream>
#include <string>
#include <tuple>
#include <type_traits>
#include <vector>
 
#include <ginkgo/core/base/device.hpp>
#include <ginkgo/core/base/fwd_decls.hpp>
#include <ginkgo/core/base/machine_topology.hpp>
#include <ginkgo/core/base/memory.hpp>
#include <ginkgo/core/base/scoped_device_id_guard.hpp>
#include <ginkgo/core/base/types.hpp>
#include <ginkgo/core/log/logger.hpp>
#include <ginkgo/core/synthesizer/containers.hpp>
 
 
namespace gko {
 
 
enum class log_propagation_mode {
    never,
    automatic
};
 
 
enum class allocation_mode { device, unified_global, unified_host };
 
 
#ifdef NDEBUG
 
// When in release, prefer device allocations
constexpr allocation_mode default_cuda_alloc_mode = allocation_mode::device;
 
constexpr allocation_mode default_hip_alloc_mode = allocation_mode::device;
 
#else
 
// When in debug, always UM allocations.
constexpr allocation_mode default_cuda_alloc_mode =
    allocation_mode::unified_global;
 
#if (GINKGO_HIP_PLATFORM_HCC == 1)
 
// HIP on AMD GPUs does not support UM, so always prefer device allocations.
constexpr allocation_mode default_hip_alloc_mode = allocation_mode::device;
 
#else
 
// HIP on NVIDIA GPUs supports UM, so prefer UM allocations.
constexpr allocation_mode default_hip_alloc_mode =
    allocation_mode::unified_global;
 
#endif
 
#endif
 
 
}  // namespace gko
 
 
enum class dpcpp_queue_property {
    in_order = 1,
 
    enable_profiling = 2
};
 
GKO_ATTRIBUTES GKO_INLINE dpcpp_queue_property operator|(dpcpp_queue_property a,
                                                         dpcpp_queue_property b)
{
    return static_cast<dpcpp_queue_property>(static_cast<int>(a) |
                                             static_cast<int>(b));
}
 
 
namespace gko {
 
 
#define GKO_FORWARD_DECLARE(_type, ...) class _type
 
GKO_ENABLE_FOR_ALL_EXECUTORS(GKO_FORWARD_DECLARE);
 
#undef GKO_FORWARD_DECLARE
 
 
class ReferenceExecutor;
 
 
namespace detail {
 
 
template <typename>
class ExecutorBase;
 
 
}  // namespace detail
 
 
class Operation {
public:
#define GKO_DECLARE_RUN_OVERLOAD(_type, ...) \
    virtual void run(std::shared_ptr<const _type>) const
 
    GKO_ENABLE_FOR_ALL_EXECUTORS(GKO_DECLARE_RUN_OVERLOAD);
 
#undef GKO_DECLARE_RUN_OVERLOAD
 
    // ReferenceExecutor overload can be defaulted to OmpExecutor's
    virtual void run(std::shared_ptr<const ReferenceExecutor> executor) const;
 
    virtual const char* get_name() const noexcept;
};
 
 
namespace detail {
 
 
template <typename Closure>
class RegisteredOperation : public Operation {
public:
    RegisteredOperation(const char* name, Closure op)
        : name_(name), op_(std::move(op))
    {}
 
    const char* get_name() const noexcept override { return name_; }
 
    void run(std::shared_ptr<const ReferenceExecutor> exec) const override
    {
        op_(exec);
    }
 
    void run(std::shared_ptr<const OmpExecutor> exec) const override
    {
        op_(exec);
    }
 
    void run(std::shared_ptr<const CudaExecutor> exec) const override
    {
        op_(exec);
    }
 
    void run(std::shared_ptr<const HipExecutor> exec) const override
    {
        op_(exec);
    }
 
    void run(std::shared_ptr<const DpcppExecutor> exec) const override
    {
        op_(exec);
    }
 
private:
    const char* name_;
    Closure op_;
};
 
 
template <typename Closure>
RegisteredOperation<Closure> make_register_operation(const char* name,
                                                     Closure op)
{
    return RegisteredOperation<Closure>{name, std::move(op)};
}
 
 
}  // namespace detail
 
 
#define GKO_REGISTER_OPERATION(_name, _kernel)                                 \
    template <typename... Args>                                                \
    auto make_##_name(Args&&... args)                                          \
    {                                                                          \
        return ::gko::detail::make_register_operation(                         \
            #_kernel, [&args...](auto exec) {                                  \
                using exec_type = decltype(exec);                              \
                if (std::is_same<                                              \
                        exec_type,                                             \
                        std::shared_ptr<const ::gko::ReferenceExecutor>>::     \
                        value) {                                               \
                    ::gko::kernels::reference::_kernel(                        \
                        std::dynamic_pointer_cast<                             \
                            const ::gko::ReferenceExecutor>(exec),             \
                        std::forward<Args>(args)...);                          \
                } else if (std::is_same<                                       \
                               exec_type,                                      \
                               std::shared_ptr<const ::gko::OmpExecutor>>::    \
                               value) {                                        \
                    ::gko::kernels::omp::_kernel(                              \
                        std::dynamic_pointer_cast<const ::gko::OmpExecutor>(   \
                            exec),                                             \
                        std::forward<Args>(args)...);                          \
                } else if (std::is_same<                                       \
                               exec_type,                                      \
                               std::shared_ptr<const ::gko::CudaExecutor>>::   \
                               value) {                                        \
                    ::gko::kernels::cuda::_kernel(                             \
                        std::dynamic_pointer_cast<const ::gko::CudaExecutor>(  \
                            exec),                                             \
                        std::forward<Args>(args)...);                          \
                } else if (std::is_same<                                       \
                               exec_type,                                      \
                               std::shared_ptr<const ::gko::HipExecutor>>::    \
                               value) {                                        \
                    ::gko::kernels::hip::_kernel(                              \
                        std::dynamic_pointer_cast<const ::gko::HipExecutor>(   \
                            exec),                                             \
                        std::forward<Args>(args)...);                          \
                } else if (std::is_same<                                       \
                               exec_type,                                      \
                               std::shared_ptr<const ::gko::DpcppExecutor>>::  \
                               value) {                                        \
                    ::gko::kernels::dpcpp::_kernel(                            \
                        std::dynamic_pointer_cast<const ::gko::DpcppExecutor>( \
                            exec),                                             \
                        std::forward<Args>(args)...);                          \
                } else {                                                       \
                    GKO_NOT_IMPLEMENTED;                                       \
                }                                                              \
            });                                                                \
    }                                                                          \
    static_assert(true,                                                        \
                  "This assert is used to counter the false positive extra "   \
                  "semi-colon warnings")
 
 
#define GKO_REGISTER_HOST_OPERATION(_name, _kernel)                          \
    template <typename... Args>                                              \
    auto make_##_name(Args&&... args)                                        \
    {                                                                        \
        return ::gko::detail::make_register_operation(                       \
            #_kernel,                                                        \
            [&args...](auto) { _kernel(std::forward<Args>(args)...); });     \
    }                                                                        \
    static_assert(true,                                                      \
                  "This assert is used to counter the false positive extra " \
                  "semi-colon warnings")
 
 
#define GKO_DECLARE_EXECUTOR_FRIEND(_type, ...) friend class _type
 
class Executor : public log::EnableLogging<Executor> {
    template <typename T>
    friend class detail::ExecutorBase;
 
    GKO_ENABLE_FOR_ALL_EXECUTORS(GKO_DECLARE_EXECUTOR_FRIEND);
    friend class ReferenceExecutor;
 
public:
    virtual ~Executor() = default;
 
    Executor() = default;
    Executor(Executor&) = delete;
    Executor(Executor&&) = delete;
    Executor& operator=(Executor&) = delete;
    Executor& operator=(Executor&&) = delete;
 
    virtual void run(const Operation& op) const = 0;
 
    template <typename ClosureOmp, typename ClosureCuda, typename ClosureHip,
              typename ClosureDpcpp>
    void run(const ClosureOmp& op_omp, const ClosureCuda& op_cuda,
             const ClosureHip& op_hip, const ClosureDpcpp& op_dpcpp) const
    {
        LambdaOperation<ClosureOmp, ClosureCuda, ClosureHip, ClosureDpcpp> op(
            op_omp, op_cuda, op_hip, op_dpcpp);
        this->run(op);
    }
 
    template <typename T>
    T* alloc(size_type num_elems) const
    {
        this->template log<log::Logger::allocation_started>(
            this, num_elems * sizeof(T));
        T* allocated = static_cast<T*>(this->raw_alloc(num_elems * sizeof(T)));
        this->template log<log::Logger::allocation_completed>(
            this, num_elems * sizeof(T), reinterpret_cast<uintptr>(allocated));
        return allocated;
    }
 
    void free(void* ptr) const noexcept
    {
        this->template log<log::Logger::free_started>(
            this, reinterpret_cast<uintptr>(ptr));
        this->raw_free(ptr);
        this->template log<log::Logger::free_completed>(
            this, reinterpret_cast<uintptr>(ptr));
    }
 
    template <typename T>
    void copy_from(ptr_param<const Executor> src_exec, size_type num_elems,
                   const T* src_ptr, T* dest_ptr) const
    {
        const auto src_loc = reinterpret_cast<uintptr>(src_ptr);
        const auto dest_loc = reinterpret_cast<uintptr>(dest_ptr);
        this->template log<log::Logger::copy_started>(
            src_exec.get(), this, src_loc, dest_loc, num_elems * sizeof(T));
        if (this != src_exec.get()) {
            src_exec->template log<log::Logger::copy_started>(
                src_exec.get(), this, src_loc, dest_loc, num_elems * sizeof(T));
        }
        try {
            this->raw_copy_from(src_exec.get(), num_elems * sizeof(T), src_ptr,
                                dest_ptr);
        } catch (NotSupported&) {
#if (GKO_VERBOSE_LEVEL >= 1) && !defined(NDEBUG)
            // Unoptimized copy. Try to go through the masters.
            // output to log when verbose >= 1 and debug build
            std::clog << "Not direct copy. Try to copy data from the masters."
                      << std::endl;
#endif
            auto src_master = src_exec->get_master().get();
            if (num_elems > 0 && src_master != src_exec.get()) {
                auto* master_ptr = src_exec->get_master()->alloc<T>(num_elems);
                src_master->copy_from<T>(src_exec, num_elems, src_ptr,
                                         master_ptr);
                this->copy_from<T>(src_master, num_elems, master_ptr, dest_ptr);
                src_master->free(master_ptr);
            }
        }
        this->template log<log::Logger::copy_completed>(
            src_exec.get(), this, src_loc, dest_loc, num_elems * sizeof(T));
        if (this != src_exec.get()) {
            src_exec->template log<log::Logger::copy_completed>(
                src_exec.get(), this, src_loc, dest_loc, num_elems * sizeof(T));
        }
    }
 
    template <typename T>
    void copy(size_type num_elems, const T* src_ptr, T* dest_ptr) const
    {
        this->copy_from(this, num_elems, src_ptr, dest_ptr);
    }
 
    template <typename T>
    T copy_val_to_host(const T* ptr) const
    {
        T out{};
        this->get_master()->copy_from(this, 1, ptr, &out);
        return out;
    }
 
    virtual std::shared_ptr<Executor> get_master() noexcept = 0;
 
    virtual std::shared_ptr<const Executor> get_master() const noexcept = 0;
 
    virtual void synchronize() const = 0;
 
    void add_logger(std::shared_ptr<const log::Logger> logger) override
    {
        this->propagating_logger_refcount_.fetch_add(
            logger->needs_propagation() ? 1 : 0);
        this->EnableLogging<Executor>::add_logger(logger);
    }
 
    void remove_logger(const log::Logger* logger) override
    {
        this->propagating_logger_refcount_.fetch_sub(
            logger->needs_propagation() ? 1 : 0);
        this->EnableLogging<Executor>::remove_logger(logger);
    }
 
    using EnableLogging<Executor>::remove_logger;
 
    void set_log_propagation_mode(log_propagation_mode mode)
    {
        log_propagation_mode_ = mode;
    }
 
    bool should_propagate_log() const
    {
        return this->propagating_logger_refcount_.load() > 0 &&
               log_propagation_mode_ == log_propagation_mode::automatic;
    }
 
    bool memory_accessible(const std::shared_ptr<const Executor>& other) const
    {
        return this->verify_memory_from(other.get());
    }
 
    virtual scoped_device_id_guard get_scoped_device_id_guard() const = 0;
 
    virtual std::string get_description() const = 0;
 
protected:
    struct exec_info {
        int device_id = -1;
 
        std::string device_type;
 
        int numa_node = -1;
 
        int num_computing_units = -1;
 
        int num_pu_per_cu = -1;
 
        std::vector<int> subgroup_sizes{};
 
        int max_subgroup_size = -1;
 
        std::vector<int> max_workitem_sizes{};
 
        int max_workgroup_size;
 
        int major = -1;
 
        int minor = -1;
 
        std::string pci_bus_id = std::string(13, 'x');
 
        std::vector<int> closest_pu_ids{};
    };
 
    const exec_info& get_exec_info() const { return this->exec_info_; }
 
    virtual void* raw_alloc(size_type size) const = 0;
 
    virtual void raw_free(void* ptr) const noexcept = 0;
 
    virtual void raw_copy_from(const Executor* src_exec, size_type n_bytes,
                               const void* src_ptr, void* dest_ptr) const = 0;
 
#define GKO_ENABLE_RAW_COPY_TO(_exec_type, ...)                              \
    virtual void raw_copy_to(const _exec_type* dest_exec, size_type n_bytes, \
                             const void* src_ptr, void* dest_ptr) const = 0
 
    GKO_ENABLE_FOR_ALL_EXECUTORS(GKO_ENABLE_RAW_COPY_TO);
 
#undef GKO_ENABLE_RAW_COPY_TO
 
    virtual bool verify_memory_from(const Executor* src_exec) const = 0;
 
#define GKO_ENABLE_VERIFY_MEMORY_TO(_exec_type, ...) \
    virtual bool verify_memory_to(const _exec_type* dest_exec) const = 0
 
    GKO_ENABLE_FOR_ALL_EXECUTORS(GKO_ENABLE_VERIFY_MEMORY_TO);
 
    GKO_ENABLE_VERIFY_MEMORY_TO(ReferenceExecutor, ref);
 
#undef GKO_ENABLE_VERIFY_MEMORY_TO
 
    virtual void populate_exec_info(const machine_topology* mach_topo) = 0;
 
    exec_info& get_exec_info() { return this->exec_info_; }
 
    exec_info exec_info_;
 
    log_propagation_mode log_propagation_mode_{log_propagation_mode::automatic};
 
    std::atomic<int> propagating_logger_refcount_{};
 
private:
    template <typename ClosureOmp, typename ClosureCuda, typename ClosureHip,
              typename ClosureDpcpp>
    class LambdaOperation : public Operation {
    public:
        LambdaOperation(const ClosureOmp& op_omp, const ClosureCuda& op_cuda,
                        const ClosureHip& op_hip, const ClosureDpcpp& op_dpcpp)
            : op_omp_(op_omp),
              op_cuda_(op_cuda),
              op_hip_(op_hip),
              op_dpcpp_(op_dpcpp)
        {}
 
        void run(std::shared_ptr<const OmpExecutor>) const override
        {
            op_omp_();
        }
 
        void run(std::shared_ptr<const ReferenceExecutor>) const override
        {
            op_omp_();
        }
 
        void run(std::shared_ptr<const CudaExecutor>) const override
        {
            op_cuda_();
        }
 
        void run(std::shared_ptr<const HipExecutor>) const override
        {
            op_hip_();
        }
 
        void run(std::shared_ptr<const DpcppExecutor>) const override
        {
            op_dpcpp_();
        }
 
    private:
        ClosureOmp op_omp_;
        ClosureCuda op_cuda_;
        ClosureHip op_hip_;
        ClosureDpcpp op_dpcpp_;
    };
};
 
 
template <typename T>
class executor_deleter {
public:
    using pointer = T*;
 
    explicit executor_deleter(std::shared_ptr<const Executor> exec)
        : exec_{exec}
    {}
 
    void operator()(pointer ptr) const
    {
        if (exec_) {
            exec_->free(ptr);
        }
    }
 
private:
    std::shared_ptr<const Executor> exec_;
};
 
// a specialization for arrays
template <typename T>
class executor_deleter<T[]> {
public:
    using pointer = T[];
 
    explicit executor_deleter(std::shared_ptr<const Executor> exec)
        : exec_{exec}
    {}
 
    void operator()(pointer ptr) const
    {
        if (exec_) {
            exec_->free(ptr);
        }
    }
 
private:
    std::shared_ptr<const Executor> exec_;
};
 
 
namespace detail {
 
 
template <typename ConcreteExecutor>
class ExecutorBase : public Executor {
    GKO_ENABLE_FOR_ALL_EXECUTORS(GKO_DECLARE_EXECUTOR_FRIEND);
    friend class ReferenceExecutor;
 
public:
    using Executor::run;
 
    void run(const Operation& op) const override
    {
        this->template log<log::Logger::operation_launched>(this, &op);
        auto scope_guard = get_scoped_device_id_guard();
        op.run(self()->shared_from_this());
        this->template log<log::Logger::operation_completed>(this, &op);
    }
 
protected:
    void raw_copy_from(const Executor* src_exec, size_type n_bytes,
                       const void* src_ptr, void* dest_ptr) const override
    {
        src_exec->raw_copy_to(self(), n_bytes, src_ptr, dest_ptr);
    }
 
    virtual bool verify_memory_from(const Executor* src_exec) const override
    {
        return src_exec->verify_memory_to(self());
    }
 
private:
    ConcreteExecutor* self() noexcept
    {
        return static_cast<ConcreteExecutor*>(this);
    }
 
    const ConcreteExecutor* self() const noexcept
    {
        return static_cast<const ConcreteExecutor*>(this);
    }
};
 
#undef GKO_DECLARE_EXECUTOR_FRIEND
 
 
class EnableDeviceReset {
public:
    GKO_DEPRECATED(
        "device_reset is no longer supported, call "
        "cudaDeviceReset/hipDeviceReset manually")
    void set_device_reset(bool device_reset) {}
 
    GKO_DEPRECATED(
        "device_reset is no longer supported, call "
        "cudaDeviceReset/hipDeviceReset manually")
    bool get_device_reset() { return false; }
 
protected:
    EnableDeviceReset() {}
 
    GKO_DEPRECATED(
        "device_reset is no longer supported, call "
        "cudaDeviceReset/hipDeviceReset manually")
    EnableDeviceReset(bool device_reset) {}
};
 
 
}  // namespace detail
 
 
#define GKO_OVERRIDE_RAW_COPY_TO(_executor_type, ...)                    \
    void raw_copy_to(const _executor_type* dest_exec, size_type n_bytes, \
                     const void* src_ptr, void* dest_ptr) const override
 
 
#define GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(dest_, bool_)                     \
    virtual bool verify_memory_to(const dest_* other) const override         \
    {                                                                        \
        return bool_;                                                        \
    }                                                                        \
    static_assert(true,                                                      \
                  "This assert is used to counter the false positive extra " \
                  "semi-colon warnings")
 
 
class OmpExecutor : public detail::ExecutorBase<OmpExecutor>,
                    public std::enable_shared_from_this<OmpExecutor> {
    friend class detail::ExecutorBase<OmpExecutor>;
 
public:
    static std::shared_ptr<OmpExecutor> create(
        std::shared_ptr<CpuAllocatorBase> alloc =
            std::make_shared<CpuAllocator>())
    {
        return std::shared_ptr<OmpExecutor>(new OmpExecutor(std::move(alloc)));
    }
 
    std::shared_ptr<Executor> get_master() noexcept override;
 
    std::shared_ptr<const Executor> get_master() const noexcept override;
 
    void synchronize() const override;
 
    int get_num_cores() const
    {
        return this->get_exec_info().num_computing_units;
    }
 
    int get_num_threads_per_core() const
    {
        return this->get_exec_info().num_pu_per_cu;
    }
 
    static int get_num_omp_threads();
 
    scoped_device_id_guard get_scoped_device_id_guard() const override;
 
    std::string get_description() const override;
 
protected:
    OmpExecutor(std::shared_ptr<CpuAllocatorBase> alloc)
        : alloc_{std::move(alloc)}
    {
        this->OmpExecutor::populate_exec_info(machine_topology::get_instance());
    }
 
    void populate_exec_info(const machine_topology* mach_topo) override;
 
    void* raw_alloc(size_type size) const override;
 
    void raw_free(void* ptr) const noexcept override;
 
    GKO_ENABLE_FOR_ALL_EXECUTORS(GKO_OVERRIDE_RAW_COPY_TO);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(OmpExecutor, true);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(ReferenceExecutor, false);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(HipExecutor, false);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(CudaExecutor, false);
 
    bool verify_memory_to(const DpcppExecutor* dest_exec) const override;
 
    std::shared_ptr<CpuAllocatorBase> alloc_;
};
 
 
namespace kernels {
namespace omp {
using DefaultExecutor = OmpExecutor;
}  // namespace omp
}  // namespace kernels
 
 
class ReferenceExecutor : public OmpExecutor {
public:
    static std::shared_ptr<ReferenceExecutor> create(
        std::shared_ptr<CpuAllocatorBase> alloc =
            std::make_shared<CpuAllocator>())
    {
        return std::shared_ptr<ReferenceExecutor>(
            new ReferenceExecutor(std::move(alloc)));
    }
 
    scoped_device_id_guard get_scoped_device_id_guard() const override
    {
        return {this, 0};
    }
 
    std::string get_description() const override { return "ReferenceExecutor"; }
 
    void run(const Operation& op) const override
    {
        this->template log<log::Logger::operation_launched>(this, &op);
        op.run(std::static_pointer_cast<const ReferenceExecutor>(
            this->shared_from_this()));
        this->template log<log::Logger::operation_completed>(this, &op);
    }
 
protected:
    ReferenceExecutor(std::shared_ptr<CpuAllocatorBase> alloc)
        : OmpExecutor{std::move(alloc)}
    {
        this->ReferenceExecutor::populate_exec_info(
            machine_topology::get_instance());
    }
 
    void populate_exec_info(const machine_topology*) override
    {
        this->get_exec_info().device_id = -1;
        this->get_exec_info().num_computing_units = 1;
        this->get_exec_info().num_pu_per_cu = 1;
    }
 
    bool verify_memory_from(const Executor* src_exec) const override
    {
        return src_exec->verify_memory_to(this);
    }
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(ReferenceExecutor, true);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(OmpExecutor, false);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(DpcppExecutor, false);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(CudaExecutor, false);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(HipExecutor, false);
};
 
 
namespace kernels {
namespace reference {
using DefaultExecutor = ReferenceExecutor;
}  // namespace reference
}  // namespace kernels
 
 
class CudaExecutor : public detail::ExecutorBase<CudaExecutor>,
                     public std::enable_shared_from_this<CudaExecutor>,
                     public detail::EnableDeviceReset {
    friend class detail::ExecutorBase<CudaExecutor>;
 
public:
    GKO_DEPRECATED(
        "calling this CudaExecutor::create method is deprecated, because"
        "device_reset no longer has an effect"
        "call CudaExecutor::create("
        "  int device_id, std::shared_ptr<Executor> master,"
        "  std::shared_ptr<CudaAllocatorBase> alloc,"
        "  CUstream_st* stream);"
        "instead")
    static std::shared_ptr<CudaExecutor> create(
        int device_id, std::shared_ptr<Executor> master, bool device_reset,
        allocation_mode alloc_mode = default_cuda_alloc_mode,
        CUstream_st* stream = nullptr);
 
    static std::shared_ptr<CudaExecutor> create(
        int device_id, std::shared_ptr<Executor> master,
        std::shared_ptr<CudaAllocatorBase> alloc =
            std::make_shared<CudaAllocator>(),
        CUstream_st* stream = nullptr);
 
    std::shared_ptr<Executor> get_master() noexcept override;
 
    std::shared_ptr<const Executor> get_master() const noexcept override;
 
    void synchronize() const override;
 
    scoped_device_id_guard get_scoped_device_id_guard() const override;
 
    std::string get_description() const override;
 
    int get_device_id() const noexcept
    {
        return this->get_exec_info().device_id;
    }
 
    static int get_num_devices();
 
    int get_num_warps_per_sm() const noexcept
    {
        return this->get_exec_info().num_pu_per_cu;
    }
 
    int get_num_multiprocessor() const noexcept
    {
        return this->get_exec_info().num_computing_units;
    }
 
    int get_num_warps() const noexcept
    {
        return this->get_exec_info().num_computing_units *
               this->get_exec_info().num_pu_per_cu;
    }
 
    int get_warp_size() const noexcept
    {
        return this->get_exec_info().max_subgroup_size;
    }
 
    int get_major_version() const noexcept
    {
        return this->get_exec_info().major;
    }
 
    int get_minor_version() const noexcept
    {
        return this->get_exec_info().minor;
    }
 
    GKO_DEPRECATED("use get_blas_handle() instead")
    cublasContext* get_cublas_handle() const { return get_blas_handle(); }
 
    cublasContext* get_blas_handle() const { return cublas_handle_.get(); }
 
    GKO_DEPRECATED("use get_sparselib_handle() instead")
    cusparseContext* get_cusparse_handle() const
    {
        return get_sparselib_handle();
    }
 
    cusparseContext* get_sparselib_handle() const
    {
        return cusparse_handle_.get();
    }
 
    std::vector<int> get_closest_pus() const
    {
        return this->get_exec_info().closest_pu_ids;
    }
 
    int get_closest_numa() const { return this->get_exec_info().numa_node; }
 
    CUstream_st* get_stream() const { return stream_; }
 
protected:
    void set_gpu_property();
 
    void init_handles();
 
    CudaExecutor(int device_id, std::shared_ptr<Executor> master,
                 std::shared_ptr<CudaAllocatorBase> alloc, CUstream_st* stream)
        : alloc_{std::move(alloc)}, master_(master), stream_{stream}
    {
        this->get_exec_info().device_id = device_id;
        this->get_exec_info().num_computing_units = 0;
        this->get_exec_info().num_pu_per_cu = 0;
        this->CudaExecutor::populate_exec_info(
            machine_topology::get_instance());
        this->set_gpu_property();
        this->init_handles();
    }
 
    void* raw_alloc(size_type size) const override;
 
    void raw_free(void* ptr) const noexcept override;
 
    GKO_ENABLE_FOR_ALL_EXECUTORS(GKO_OVERRIDE_RAW_COPY_TO);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(OmpExecutor, false);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(ReferenceExecutor, false);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(DpcppExecutor, false);
 
    bool verify_memory_to(const HipExecutor* dest_exec) const override;
 
    bool verify_memory_to(const CudaExecutor* dest_exec) const override;
 
    void populate_exec_info(const machine_topology* mach_topo) override;
 
private:
    std::shared_ptr<Executor> master_;
 
    template <typename T>
    using handle_manager = std::unique_ptr<T, std::function<void(T*)>>;
    handle_manager<cublasContext> cublas_handle_;
    handle_manager<cusparseContext> cusparse_handle_;
    std::shared_ptr<CudaAllocatorBase> alloc_;
    CUstream_st* stream_;
};
 
 
namespace kernels {
namespace cuda {
using DefaultExecutor = CudaExecutor;
}  // namespace cuda
}  // namespace kernels
 
 
class HipExecutor : public detail::ExecutorBase<HipExecutor>,
                    public std::enable_shared_from_this<HipExecutor>,
                    public detail::EnableDeviceReset {
    friend class detail::ExecutorBase<HipExecutor>;
 
public:
    GKO_DEPRECATED(
        "device_reset is deprecated entirely, call hipDeviceReset directly. "
        "alloc_mode was replaced by the Allocator type "
        "hierarchy.")
    static std::shared_ptr<HipExecutor> create(
        int device_id, std::shared_ptr<Executor> master, bool device_reset,
        allocation_mode alloc_mode = default_hip_alloc_mode,
        GKO_HIP_STREAM_STRUCT* stream = nullptr);
 
    static std::shared_ptr<HipExecutor> create(
        int device_id, std::shared_ptr<Executor> master,
        std::shared_ptr<HipAllocatorBase> alloc =
            std::make_shared<HipAllocator>(),
        GKO_HIP_STREAM_STRUCT* stream = nullptr);
 
    std::shared_ptr<Executor> get_master() noexcept override;
 
    std::shared_ptr<const Executor> get_master() const noexcept override;
 
    void synchronize() const override;
 
    scoped_device_id_guard get_scoped_device_id_guard() const override;
 
    std::string get_description() const override;
 
    int get_device_id() const noexcept
    {
        return this->get_exec_info().device_id;
    }
 
    static int get_num_devices();
 
    int get_num_warps_per_sm() const noexcept
    {
        return this->get_exec_info().num_pu_per_cu;
    }
 
    int get_num_multiprocessor() const noexcept
    {
        return this->get_exec_info().num_computing_units;
    }
 
    int get_major_version() const noexcept
    {
        return this->get_exec_info().major;
    }
 
    int get_minor_version() const noexcept
    {
        return this->get_exec_info().minor;
    }
 
    int get_num_warps() const noexcept
    {
        return this->get_exec_info().num_computing_units *
               this->get_exec_info().num_pu_per_cu;
    }
 
    int get_warp_size() const noexcept
    {
        return this->get_exec_info().max_subgroup_size;
    }
 
    GKO_DEPRECATED("use get_blas_handle() instead")
    hipblasContext* get_hipblas_handle() const { return get_blas_handle(); }
 
    hipblasContext* get_blas_handle() const { return hipblas_handle_.get(); }
 
    GKO_DEPRECATED("use get_sparselib_handle() instead")
    hipsparseContext* get_hipsparse_handle() const
    {
        return get_sparselib_handle();
    }
 
    hipsparseContext* get_sparselib_handle() const
    {
        return hipsparse_handle_.get();
    }
 
    int get_closest_numa() const { return this->get_exec_info().numa_node; }
 
    std::vector<int> get_closest_pus() const
    {
        return this->get_exec_info().closest_pu_ids;
    }
 
    GKO_HIP_STREAM_STRUCT* get_stream() const { return stream_; }
 
protected:
    void set_gpu_property();
 
    void init_handles();
 
    HipExecutor(int device_id, std::shared_ptr<Executor> master,
                std::shared_ptr<HipAllocatorBase> alloc,
                GKO_HIP_STREAM_STRUCT* stream)
        : master_{std::move(master)}, alloc_{std::move(alloc)}, stream_{stream}
    {
        this->get_exec_info().device_id = device_id;
        this->get_exec_info().num_computing_units = 0;
        this->get_exec_info().num_pu_per_cu = 0;
        this->HipExecutor::populate_exec_info(machine_topology::get_instance());
        this->set_gpu_property();
        this->init_handles();
    }
 
    void* raw_alloc(size_type size) const override;
 
    void raw_free(void* ptr) const noexcept override;
 
    GKO_ENABLE_FOR_ALL_EXECUTORS(GKO_OVERRIDE_RAW_COPY_TO);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(OmpExecutor, false);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(ReferenceExecutor, false);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(DpcppExecutor, false);
 
    bool verify_memory_to(const CudaExecutor* dest_exec) const override;
 
    bool verify_memory_to(const HipExecutor* dest_exec) const override;
 
    void populate_exec_info(const machine_topology* mach_topo) override;
 
private:
    std::shared_ptr<Executor> master_;
 
    template <typename T>
    using handle_manager = std::unique_ptr<T, std::function<void(T*)>>;
    handle_manager<hipblasContext> hipblas_handle_;
    handle_manager<hipsparseContext> hipsparse_handle_;
    std::shared_ptr<HipAllocatorBase> alloc_;
    GKO_HIP_STREAM_STRUCT* stream_;
};
 
 
namespace kernels {
namespace hip {
using DefaultExecutor = HipExecutor;
}  // namespace hip
}  // namespace kernels
 
 
class DpcppExecutor : public detail::ExecutorBase<DpcppExecutor>,
                      public std::enable_shared_from_this<DpcppExecutor> {
    friend class detail::ExecutorBase<DpcppExecutor>;
 
public:
    static std::shared_ptr<DpcppExecutor> create(
        int device_id, std::shared_ptr<Executor> master,
        std::string device_type = "all",
        dpcpp_queue_property property = dpcpp_queue_property::in_order);
 
    std::shared_ptr<Executor> get_master() noexcept override;
 
    std::shared_ptr<const Executor> get_master() const noexcept override;
 
    void synchronize() const override;
 
    scoped_device_id_guard get_scoped_device_id_guard() const override;
 
    std::string get_description() const override;
 
    int get_device_id() const noexcept
    {
        return this->get_exec_info().device_id;
    }
 
    sycl::queue* get_queue() const { return queue_.get(); }
 
    static int get_num_devices(std::string device_type);
 
    const std::vector<int>& get_subgroup_sizes() const noexcept
    {
        return this->get_exec_info().subgroup_sizes;
    }
 
    int get_num_computing_units() const noexcept
    {
        return this->get_exec_info().num_computing_units;
    }
 
    int get_num_subgroups() const noexcept
    {
        return this->get_exec_info().num_computing_units *
               this->get_exec_info().num_pu_per_cu;
    }
 
    const std::vector<int>& get_max_workitem_sizes() const noexcept
    {
        return this->get_exec_info().max_workitem_sizes;
    }
 
    int get_max_workgroup_size() const noexcept
    {
        return this->get_exec_info().max_workgroup_size;
    }
 
    int get_max_subgroup_size() const noexcept
    {
        return this->get_exec_info().max_subgroup_size;
    }
 
    std::string get_device_type() const noexcept
    {
        return this->get_exec_info().device_type;
    }
 
protected:
    void set_device_property(
        dpcpp_queue_property property = dpcpp_queue_property::in_order);
 
    DpcppExecutor(
        int device_id, std::shared_ptr<Executor> master,
        std::string device_type = "all",
        dpcpp_queue_property property = dpcpp_queue_property::in_order)
        : master_(master)
    {
        std::for_each(device_type.begin(), device_type.end(),
                      [](char& c) { c = std::tolower(c); });
        this->get_exec_info().device_type = std::string(device_type);
        this->get_exec_info().device_id = device_id;
        this->set_device_property(property);
    }
 
    void populate_exec_info(const machine_topology* mach_topo) override;
 
    void* raw_alloc(size_type size) const override;
 
    void raw_free(void* ptr) const noexcept override;
 
    GKO_ENABLE_FOR_ALL_EXECUTORS(GKO_OVERRIDE_RAW_COPY_TO);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(CudaExecutor, false);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(HipExecutor, false);
 
    GKO_DEFAULT_OVERRIDE_VERIFY_MEMORY(ReferenceExecutor, false);
 
    bool verify_memory_to(const OmpExecutor* dest_exec) const override;
 
    bool verify_memory_to(const DpcppExecutor* dest_exec) const override;
 
private:
    std::shared_ptr<Executor> master_;
 
    template <typename T>
    using queue_manager = std::unique_ptr<T, std::function<void(T*)>>;
    queue_manager<sycl::queue> queue_;
};
 
 
namespace kernels {
namespace dpcpp {
using DefaultExecutor = DpcppExecutor;
}  // namespace dpcpp
}  // namespace kernels
 
 
#undef GKO_OVERRIDE_RAW_COPY_TO
 
 
}  // namespace gko
 
 
#endif  // GKO_PUBLIC_CORE_BASE_EXECUTOR_HPP_