kernel_bundle.hpp

Go to the documentation of this file.

//==------- kernel_bundle.hpp - SYCL kernel_bundle and free functions ------==//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
 
#pragma once
 
#include <CL/sycl/context.hpp>
#include <CL/sycl/detail/common.hpp>
#include <CL/sycl/detail/kernel_desc.hpp>
#include <CL/sycl/detail/pi.h>
#include <CL/sycl/detail/pi.hpp>
#include <CL/sycl/device.hpp>
#include <CL/sycl/kernel.hpp>
#include <CL/sycl/kernel_bundle_enums.hpp>
 
#include <cassert>
#include <memory>
#include <set>
#include <vector>
 
__SYCL_INLINE_NAMESPACE(cl) {
namespace sycl {
// Forward declaration
template <backend Backend> class backend_traits;
template <backend Backend, class SyclT>
auto get_native(const SyclT &Obj) -> backend_return_t<Backend, SyclT>;
 
namespace detail {
class kernel_id_impl;
}
 
class __SYCL_EXPORT kernel_id {
public:
  kernel_id() = delete;
 
  const char *get_name() const noexcept;
 
  bool operator==(const kernel_id &RHS) const { return impl == RHS.impl; }
 
  bool operator!=(const kernel_id &RHS) const { return !(*this == RHS); }
 
private:
  kernel_id(const char *Name);
 
  kernel_id(const std::shared_ptr<detail::kernel_id_impl> &Impl)
      : impl(std::move(Impl)) {}
 
  std::shared_ptr<detail::kernel_id_impl> impl;
 
  template <class Obj>
  friend decltype(Obj::impl) detail::getSyclObjImpl(const Obj &SyclObject);
 
  template <class T>
  friend T detail::createSyclObjFromImpl(decltype(T::impl) ImplObj);
};
 
namespace detail {
class device_image_impl;
using DeviceImageImplPtr = std::shared_ptr<device_image_impl>;
 
// The class is used as a base for device_image for "untemplating" public
// methods.
class __SYCL_EXPORT device_image_plain {
public:
  device_image_plain(const detail::DeviceImageImplPtr &Impl)
      : impl(std::move(Impl)) {}
 
  bool operator==(const device_image_plain &RHS) const {
    return impl == RHS.impl;
  }
 
  bool operator!=(const device_image_plain &RHS) const {
    return !(*this == RHS);
  }
 
  bool has_kernel(const kernel_id &KernelID) const noexcept;
 
  bool has_kernel(const kernel_id &KernelID, const device &Dev) const noexcept;
 
  pi_native_handle getNative() const;
 
protected:
  detail::DeviceImageImplPtr impl;
 
  template <class Obj>
  friend decltype(Obj::impl) detail::getSyclObjImpl(const Obj &SyclObject);
 
  template <class T>
  friend T detail::createSyclObjFromImpl(decltype(T::impl) ImplObj);
};
} // namespace detail
 
template <sycl::bundle_state State>
class device_image : public detail::device_image_plain {
public:
  device_image() = delete;
 
  bool has_kernel(const kernel_id &KernelID) const noexcept {
    return device_image_plain::has_kernel(KernelID);
  }
 
  bool has_kernel(const kernel_id &KernelID, const device &Dev) const noexcept {
    return device_image_plain::has_kernel(KernelID, Dev);
  }
 
private:
  device_image(detail::DeviceImageImplPtr Impl)
      : device_image_plain(std::move(Impl)) {}
 
  template <class Obj>
  friend decltype(Obj::impl) detail::getSyclObjImpl(const Obj &SyclObject);
 
  template <class T>
  friend T detail::createSyclObjFromImpl(decltype(T::impl) ImplObj);
};
 
namespace detail {
class kernel_bundle_impl;
using KernelBundleImplPtr = std::shared_ptr<detail::kernel_bundle_impl>;
 
// The class is used as a base for kernel_bundle to "untemplate" it's methods
class __SYCL_EXPORT kernel_bundle_plain {
public:
  kernel_bundle_plain(const detail::KernelBundleImplPtr &Impl)
      : impl(std::move(Impl)) {}
 
  bool operator==(const kernel_bundle_plain &RHS) const {
    return impl == RHS.impl;
  }
 
  bool operator!=(const kernel_bundle_plain &RHS) const {
    return !(*this == RHS);
  }
 
  bool empty() const noexcept;
 
  backend get_backend() const noexcept;
 
  context get_context() const noexcept;
 
  std::vector<device> get_devices() const noexcept;
 
  bool has_kernel(const kernel_id &KernelID) const noexcept;
 
  bool has_kernel(const kernel_id &KernelID, const device &Dev) const noexcept;
 
  std::vector<kernel_id> get_kernel_ids() const;
 
  bool contains_specialization_constants() const noexcept;
 
  bool native_specialization_constant() const noexcept;
 
protected:
  // \returns a kernel object which represents the kernel identified by
  // kernel_id passed
  kernel get_kernel(const kernel_id &KernelID) const;
 
  // \returns an iterator to the first device image kernel_bundle contains
  const device_image_plain *begin() const;
 
  // \returns an iterator to the last device image kernel_bundle contains
  const device_image_plain *end() const;
 
  bool has_specialization_constant_impl(const char *SpecName) const noexcept;
 
  void set_specialization_constant_impl(const char *SpecName, void *Value,
                                        size_t Size) noexcept;
 
  void get_specialization_constant_impl(const char *SpecName,
                                        void *Value) const noexcept;
 
  bool is_specialization_constant_set(const char *SpecName) const noexcept;
 
  detail::KernelBundleImplPtr impl;
};
 
} // namespace detail
 
template <bundle_state State>
class kernel_bundle : public detail::kernel_bundle_plain {
public:
  using device_image_iterator = const device_image<State> *;
 
  kernel_bundle() = delete;
 
  bool empty() const noexcept { return kernel_bundle_plain::empty(); }
 
  backend get_backend() const noexcept {
    return kernel_bundle_plain::get_backend();
  }
 
  context get_context() const noexcept {
    return kernel_bundle_plain::get_context();
  }
 
  std::vector<device> get_devices() const noexcept {
    return kernel_bundle_plain::get_devices();
  }
 
  bool has_kernel(const kernel_id &KernelID) const noexcept {
    return kernel_bundle_plain::has_kernel(KernelID);
  }
 
  bool has_kernel(const kernel_id &KernelID, const device &Dev) const noexcept {
    return kernel_bundle_plain::has_kernel(KernelID, Dev);
  }
 
  std::vector<kernel_id> get_kernel_ids() const {
    return kernel_bundle_plain::get_kernel_ids();
  }
 
  bool contains_specialization_constants() const noexcept {
    return kernel_bundle_plain::contains_specialization_constants();
  }
 
  bool native_specialization_constant() const noexcept {
    return kernel_bundle_plain::native_specialization_constant();
  }
 
  template <bundle_state _State = State,
            typename = detail::enable_if_t<_State == bundle_state::executable>>
  kernel get_kernel(const kernel_id &KernelID) const {
    return detail::kernel_bundle_plain::get_kernel(KernelID);
  }
 
  // This guard is needed because the libsycl.so can compiled with C++ <=14
  // while the code requires C++17. This code is not supposed to be used by the
  // libsycl.so so it should not be a problem.
#if __cplusplus >= 201703L
  template <auto &SpecName> bool has_specialization_constant() const noexcept {
    const char *SpecSymName = detail::get_spec_constant_symbolic_ID<SpecName>();
    return has_specialization_constant_impl(SpecSymName);
  }
 
  template <auto &SpecName, bundle_state _State = State,
            typename = detail::enable_if_t<_State == bundle_state::input>>
  void set_specialization_constant(
      typename std::remove_reference_t<decltype(SpecName)>::value_type Value) {
    const char *SpecSymName = detail::get_spec_constant_symbolic_ID<SpecName>();
    set_specialization_constant_impl(SpecSymName, &Value,
                                     sizeof(decltype(Value)));
  }
 
  template <auto &SpecName>
  typename std::remove_reference_t<decltype(SpecName)>::value_type
  get_specialization_constant() const {
    const char *SpecSymName = detail::get_spec_constant_symbolic_ID<SpecName>();
    if (!is_specialization_constant_set(SpecSymName))
      return SpecName.getDefaultValue();
 
    using SCType =
        typename std::remove_reference_t<decltype(SpecName)>::value_type;
 
    std::array<char *, sizeof(SCType)> RetValue;
 
    get_specialization_constant_impl(SpecSymName, RetValue.data());
 
    return *reinterpret_cast<SCType *>(RetValue.data());
  }
#endif
 
  device_image_iterator begin() const {
    return reinterpret_cast<device_image_iterator>(
        kernel_bundle_plain::begin());
  }
 
  device_image_iterator end() const {
    return reinterpret_cast<device_image_iterator>(kernel_bundle_plain::end());
  }
 
  template <backend Backend>
  __SYCL_DEPRECATED("Use SYCL 2020 sycl::get_native free function")
  backend_return_t<Backend, kernel_bundle<State>> get_native() const {
    return getNative<Backend>();
  }
 
private:
  kernel_bundle(detail::KernelBundleImplPtr Impl)
      : kernel_bundle_plain(std::move(Impl)) {}
 
  template <class Obj>
  friend decltype(Obj::impl) detail::getSyclObjImpl(const Obj &SyclObject);
 
  template <class T>
  friend T detail::createSyclObjFromImpl(decltype(T::impl) ImplObj);
 
  template <backend Backend, class SyclT>
  friend auto get_native(const SyclT &Obj) -> backend_return_t<Backend, SyclT>;
 
  template <backend Backend>
  backend_return_t<Backend, kernel_bundle<State>> getNative() const {
    // NOTE: implementation assumes that the return type is a
    // derivative of std::vector.
    backend_return_t<Backend, kernel_bundle<State>> ReturnValue;
    ReturnValue.reserve(std::distance(begin(), end()));
 
    for (const device_image<State> &DevImg : *this) {
      ReturnValue.push_back(
          detail::pi::cast<typename decltype(ReturnValue)::value_type>(
              DevImg.getNative()));
    }
 
    return ReturnValue;
  }
};
 
// get_kernel_id API
 
namespace detail {
// Internal non-template versions of get_kernel_id API which is used by public
// onces
__SYCL_EXPORT kernel_id get_kernel_id_impl(std::string KernelName);
} // namespace detail
 
template <typename KernelName> kernel_id get_kernel_id() {
  using KI = sycl::detail::KernelInfo<KernelName>;
  return detail::get_kernel_id_impl(KI::getName());
}
 
__SYCL_EXPORT std::vector<kernel_id> get_kernel_ids();
 
// get_kernel_bundle API
 
namespace detail {
 
// Internal non-template versions of get_kernel_bundle API which is used by
// public onces
__SYCL_EXPORT detail::KernelBundleImplPtr
get_kernel_bundle_impl(const context &Ctx, const std::vector<device> &Devs,
                       bundle_state State);
 
inline auto getDeviceComparisonLambda() {
  return [](device a, device b) { return a.getNative() != b.getNative(); };
}
 
inline const std::vector<device>
removeDuplicateDevices(const std::vector<device> &Devs) {
  auto compareDevices = getDeviceComparisonLambda();
  std::set<device, decltype(compareDevices)> UniqueDeviceSet(
      Devs.begin(), Devs.end(), compareDevices);
  std::vector<device> UniqueDevices(UniqueDeviceSet.begin(),
                                    UniqueDeviceSet.end());
  return UniqueDevices;
}
 
} // namespace detail
 
template <bundle_state State>
kernel_bundle<State> get_kernel_bundle(const context &Ctx,
                                       const std::vector<device> &Devs) {
  std::vector<device> UniqueDevices = detail::removeDuplicateDevices(Devs);
 
  detail::KernelBundleImplPtr Impl =
      detail::get_kernel_bundle_impl(Ctx, UniqueDevices, State);
 
  return detail::createSyclObjFromImpl<kernel_bundle<State>>(Impl);
}
 
template <bundle_state State>
kernel_bundle<State> get_kernel_bundle(const context &Ctx) {
  return get_kernel_bundle<State>(Ctx, Ctx.get_devices());
}
 
namespace detail {
 
// Internal non-template versions of get_kernel_bundle API which is used by
// public onces
__SYCL_EXPORT detail::KernelBundleImplPtr
get_kernel_bundle_impl(const context &Ctx, const std::vector<device> &Devs,
                       const std::vector<kernel_id> &KernelIDs,
                       bundle_state State);
} // namespace detail
 
template <bundle_state State>
kernel_bundle<State>
get_kernel_bundle(const context &Ctx, const std::vector<device> &Devs,
                  const std::vector<kernel_id> &KernelIDs) {
  std::vector<device> UniqueDevices = detail::removeDuplicateDevices(Devs);
 
  detail::KernelBundleImplPtr Impl =
      detail::get_kernel_bundle_impl(Ctx, UniqueDevices, KernelIDs, State);
  return detail::createSyclObjFromImpl<kernel_bundle<State>>(Impl);
}
 
template <bundle_state State>
kernel_bundle<State>
get_kernel_bundle(const context &Ctx, const std::vector<kernel_id> &KernelIDs) {
  return get_kernel_bundle<State>(Ctx, Ctx.get_devices(), KernelIDs);
}
 
template <typename KernelName, bundle_state State>
kernel_bundle<State> get_kernel_bundle(const context &Ctx) {
  return get_kernel_bundle<State>(Ctx, Ctx.get_devices(),
                                  {get_kernel_id<KernelName>()});
}
 
template <typename KernelName, bundle_state State>
kernel_bundle<State> get_kernel_bundle(const context &Ctx,
                                       const std::vector<device> &Devs) {
  return get_kernel_bundle<State>(Ctx, Devs, {get_kernel_id<KernelName>()});
}
 
namespace detail {
 
// Stable selector function type for passing thru library boundaries
using DevImgSelectorImpl =
    std::function<bool(const detail::DeviceImageImplPtr &DevImgImpl)>;
 
// Internal non-template versions of get_kernel_bundle API which is used by
// public onces
__SYCL_EXPORT detail::KernelBundleImplPtr
get_kernel_bundle_impl(const context &Ctx, const std::vector<device> &Devs,
                       bundle_state State, const DevImgSelectorImpl &Selector);
 
// Internal non-template versions of get_empty_interop_kernel_bundle API which
// is used by public onces
__SYCL_EXPORT detail::KernelBundleImplPtr
get_empty_interop_kernel_bundle_impl(const context &Ctx,
                                     const std::vector<device> &Devs);
 
template <bundle_state State>
kernel_bundle<State> get_empty_interop_kernel_bundle(const context &Ctx) {
  detail::KernelBundleImplPtr Impl =
      detail::get_empty_interop_kernel_bundle_impl(Ctx, Ctx.get_devices());
  return detail::createSyclObjFromImpl<sycl::kernel_bundle<State>>(Impl);
}
} // namespace detail
 
template <bundle_state State, typename SelectorT>
kernel_bundle<State> get_kernel_bundle(const context &Ctx,
                                       const std::vector<device> &Devs,
                                       SelectorT Selector) {
  std::vector<device> UniqueDevices = detail::removeDuplicateDevices(Devs);
 
  detail::DevImgSelectorImpl SelectorWrapper =
      [Selector](const detail::DeviceImageImplPtr &DevImg) {
        return Selector(
            detail::createSyclObjFromImpl<sycl::device_image<State>>(DevImg));
      };
 
  detail::KernelBundleImplPtr Impl = detail::get_kernel_bundle_impl(
      Ctx, UniqueDevices, State, SelectorWrapper);
 
  return detail::createSyclObjFromImpl<sycl::kernel_bundle<State>>(Impl);
}
 
template <bundle_state State, typename SelectorT>
kernel_bundle<State> get_kernel_bundle(const context &Ctx, SelectorT Selector) {
  return get_kernel_bundle<State>(Ctx, Ctx.get_devices(), Selector);
}
 
// has_kernel_bundle API
 
namespace detail {
 
__SYCL_EXPORT bool has_kernel_bundle_impl(const context &Ctx,
                                          const std::vector<device> &Devs,
                                          bundle_state State);
 
__SYCL_EXPORT bool
has_kernel_bundle_impl(const context &Ctx, const std::vector<device> &Devs,
                       const std::vector<kernel_id> &kernelIds,
                       bundle_state State);
} // namespace detail
 
template <bundle_state State>
bool has_kernel_bundle(const context &Ctx, const std::vector<device> &Devs) {
  return detail::has_kernel_bundle_impl(Ctx, Devs, State);
}
 
template <bundle_state State>
bool has_kernel_bundle(const context &Ctx, const std::vector<device> &Devs,
                       const std::vector<kernel_id> &KernelIDs) {
  return detail::has_kernel_bundle_impl(Ctx, Devs, KernelIDs, State);
}
 
template <bundle_state State> bool has_kernel_bundle(const context &Ctx) {
  return has_kernel_bundle<State>(Ctx, Ctx.get_devices());
}
 
template <bundle_state State>
bool has_kernel_bundle(const context &Ctx,
                       const std::vector<kernel_id> &KernelIDs) {
  return has_kernel_bundle<State>(Ctx, Ctx.get_devices(), KernelIDs);
}
 
template <typename KernelName, bundle_state State>
bool has_kernel_bundle(const context &Ctx) {
  return has_kernel_bundle<State>(Ctx, {get_kernel_id<KernelName>()});
}
 
template <typename KernelName, bundle_state State>
bool has_kernel_bundle(const context &Ctx, const std::vector<device> &Devs) {
  return has_kernel_bundle<State>(Ctx, Devs, {get_kernel_id<KernelName>()});
}
 
// is_compatible API
 
bool is_compatible(const std::vector<kernel_id> &KernelIDs, const device &Dev);
 
template <typename KernelName> bool is_compatible(const device &Dev) {
  return is_compatible({get_kernel_id<KernelName>()}, Dev);
}
 
// join API
 
namespace detail {
 
// TODO: This is no longer in use. Remove when ABI break is allowed.
__SYCL_EXPORT std::shared_ptr<detail::kernel_bundle_impl>
join_impl(const std::vector<detail::KernelBundleImplPtr> &Bundles);
 
__SYCL_EXPORT std::shared_ptr<detail::kernel_bundle_impl>
join_impl(const std::vector<detail::KernelBundleImplPtr> &Bundles,
          bundle_state State);
}
 
template <sycl::bundle_state State>
sycl::kernel_bundle<State>
join(const std::vector<sycl::kernel_bundle<State>> &Bundles) {
  // Convert kernel_bundle<State> to impls to abstract template parameter away
  std::vector<detail::KernelBundleImplPtr> KernelBundleImpls;
  KernelBundleImpls.reserve(Bundles.size());
  for (const sycl::kernel_bundle<State> &Bundle : Bundles)
    KernelBundleImpls.push_back(detail::getSyclObjImpl(Bundle));
 
  std::shared_ptr<detail::kernel_bundle_impl> Impl =
      detail::join_impl(KernelBundleImpls, State);
  return detail::createSyclObjFromImpl<kernel_bundle<State>>(Impl);
}
 
// compile API
 
namespace detail {
 
__SYCL_EXPORT std::shared_ptr<detail::kernel_bundle_impl>
compile_impl(const kernel_bundle<bundle_state::input> &InputBundle,
             const std::vector<device> &Devs, const property_list &PropList);
}
 
inline kernel_bundle<bundle_state::object>
compile(const kernel_bundle<bundle_state::input> &InputBundle,
        const std::vector<device> &Devs, const property_list &PropList = {}) {
  std::vector<device> UniqueDevices = detail::removeDuplicateDevices(Devs);
 
  detail::KernelBundleImplPtr Impl =
      detail::compile_impl(InputBundle, UniqueDevices, PropList);
  return detail::createSyclObjFromImpl<
      kernel_bundle<sycl::bundle_state::object>>(Impl);
}
 
inline kernel_bundle<bundle_state::object>
compile(const kernel_bundle<bundle_state::input> &InputBundle,
        const property_list &PropList = {}) {
  return compile(InputBundle, InputBundle.get_devices(), PropList);
}
 
// link API
 
namespace detail {
__SYCL_EXPORT std::vector<sycl::device> find_device_intersection(
    const std::vector<kernel_bundle<bundle_state::object>> &ObjectBundles);
 
__SYCL_EXPORT std::shared_ptr<detail::kernel_bundle_impl>
link_impl(const std::vector<kernel_bundle<bundle_state::object>> &ObjectBundles,
          const std::vector<device> &Devs, const property_list &PropList);
} // namespace detail
 
inline kernel_bundle<bundle_state::executable>
link(const std::vector<kernel_bundle<bundle_state::object>> &ObjectBundles,
     const std::vector<device> &Devs, const property_list &PropList = {}) {
  std::vector<device> UniqueDevices = detail::removeDuplicateDevices(Devs);
 
  detail::KernelBundleImplPtr Impl =
      detail::link_impl(ObjectBundles, UniqueDevices, PropList);
  return detail::createSyclObjFromImpl<
      kernel_bundle<sycl::bundle_state::executable>>(Impl);
}
 
inline kernel_bundle<bundle_state::executable>
link(const kernel_bundle<bundle_state::object> &ObjectBundle,
     const property_list &PropList = {}) {
  return link(std::vector<kernel_bundle<bundle_state::object>>{ObjectBundle},
              ObjectBundle.get_devices(), PropList);
}
 
inline kernel_bundle<bundle_state::executable>
link(const std::vector<kernel_bundle<bundle_state::object>> &ObjectBundles,
     const property_list &PropList = {}) {
  std::vector<sycl::device> IntersectDevices =
      find_device_intersection(ObjectBundles);
  return link(ObjectBundles, IntersectDevices, PropList);
}
 
inline kernel_bundle<bundle_state::executable>
link(const kernel_bundle<bundle_state::object> &ObjectBundle,
     const std::vector<device> &Devs, const property_list &PropList = {}) {
  return link(std::vector<kernel_bundle<bundle_state::object>>{ObjectBundle},
              Devs, PropList);
}
 
// build API
 
namespace detail {
__SYCL_EXPORT std::shared_ptr<detail::kernel_bundle_impl>
build_impl(const kernel_bundle<bundle_state::input> &InputBundle,
           const std::vector<device> &Devs, const property_list &PropList);
}
 
inline kernel_bundle<bundle_state::executable>
build(const kernel_bundle<bundle_state::input> &InputBundle,
      const std::vector<device> &Devs, const property_list &PropList = {}) {
  std::vector<device> UniqueDevices = detail::removeDuplicateDevices(Devs);
 
  detail::KernelBundleImplPtr Impl =
      detail::build_impl(InputBundle, UniqueDevices, PropList);
  return detail::createSyclObjFromImpl<
      kernel_bundle<sycl::bundle_state::executable>>(Impl);
}
 
inline kernel_bundle<bundle_state::executable>
build(const kernel_bundle<bundle_state::input> &InputBundle,
      const property_list &PropList = {}) {
  return build(InputBundle, InputBundle.get_devices(), PropList);
}
 
} // namespace sycl
} // __SYCL_INLINE_NAMESPACE(cl)
 
namespace std {
template <> struct hash<cl::sycl::kernel_id> {
  size_t operator()(const cl::sycl::kernel_id &KernelID) const {
    return hash<std::shared_ptr<cl::sycl::detail::kernel_id_impl>>()(
        cl::sycl::detail::getSyclObjImpl(KernelID));
  }
};
 
template <cl::sycl::bundle_state State>
struct hash<cl::sycl::device_image<State>> {
  size_t operator()(const cl::sycl::device_image<State> &DeviceImage) const {
    return hash<std::shared_ptr<cl::sycl::detail::device_image_impl>>()(
        cl::sycl::detail::getSyclObjImpl(DeviceImage));
  }
};
 
template <cl::sycl::bundle_state State>
struct hash<cl::sycl::kernel_bundle<State>> {
  size_t operator()(const cl::sycl::kernel_bundle<State> &KernelBundle) const {
    return hash<std::shared_ptr<cl::sycl::detail::kernel_bundle_impl>>()(
        cl::sycl::detail::getSyclObjImpl(KernelBundle));
  }
};
} // namespace std