platform_impl.cpp

Go to the documentation of this file.

//==----------- platform_impl.cpp ------------------------------------------==//
//
// 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
//
//===----------------------------------------------------------------------===//
 
#include <detail/allowlist.hpp>
#include <detail/config.hpp>
#include <detail/device_impl.hpp>
#include <detail/global_handler.hpp>
#include <detail/platform_impl.hpp>
#include <detail/platform_info.hpp>
#include <sycl/detail/iostream_proxy.hpp>
#include <sycl/detail/util.hpp>
#include <sycl/device.hpp>
 
#include <algorithm>
#include <cstring>
#include <mutex>
#include <string>
#include <vector>
 
namespace sycl {
__SYCL_INLINE_VER_NAMESPACE(_V1) {
namespace detail {
 
using PlatformImplPtr = std::shared_ptr<platform_impl>;
 
PlatformImplPtr platform_impl::getHostPlatformImpl() {
  static PlatformImplPtr HostImpl = std::make_shared<platform_impl>();
 
  return HostImpl;
}
 
PlatformImplPtr platform_impl::getOrMakePlatformImpl(RT::PiPlatform PiPlatform,
                                                     const plugin &Plugin) {
  PlatformImplPtr Result;
  {
    const std::lock_guard<std::mutex> Guard(
        GlobalHandler::instance().getPlatformMapMutex());
 
    std::vector<PlatformImplPtr> &PlatformCache =
        GlobalHandler::instance().getPlatformCache();
 
    // If we've already seen this platform, return the impl
    for (const auto &PlatImpl : PlatformCache) {
      if (PlatImpl->getHandleRef() == PiPlatform)
        return PlatImpl;
    }
 
    // Otherwise make the impl
    Result = std::make_shared<platform_impl>(PiPlatform, Plugin);
    PlatformCache.emplace_back(Result);
  }
 
  return Result;
}
 
PlatformImplPtr platform_impl::getPlatformFromPiDevice(RT::PiDevice PiDevice,
                                                       const plugin &Plugin) {
  RT::PiPlatform Plt = nullptr; // TODO catch an exception and put it to list
  // of asynchronous exceptions
  Plugin.call<PiApiKind::piDeviceGetInfo>(PiDevice, PI_DEVICE_INFO_PLATFORM,
                                          sizeof(Plt), &Plt, nullptr);
  return getOrMakePlatformImpl(Plt, Plugin);
}
 
static bool IsBannedPlatform(platform Platform) {
  // The NVIDIA OpenCL platform is currently not compatible with DPC++
  // since it is only 1.2 but gets selected by default in many systems
  // There is also no support on the PTX backend for OpenCL consumption,
  // and there have been some internal reports.
  // To avoid problems on default users and deployment of DPC++ on platforms
  // where CUDA is available, the OpenCL support is disabled.
  //
  auto IsNVIDIAOpenCL = [](platform Platform) {
    if (getSyclObjImpl(Platform)->is_host())
      return false;
 
    const bool HasCUDA = Platform.get_info<info::platform::name>().find(
                             "NVIDIA CUDA") != std::string::npos;
    const auto Backend =
        detail::getSyclObjImpl(Platform)->getPlugin().getBackend();
    const bool IsCUDAOCL = (HasCUDA && Backend == backend::opencl);
    if (detail::pi::trace(detail::pi::TraceLevel::PI_TRACE_ALL) && IsCUDAOCL) {
      std::cout << "SYCL_PI_TRACE[all]: "
                << "NVIDIA CUDA OpenCL platform found but is not compatible."
                << std::endl;
    }
    return IsCUDAOCL;
  };
  return IsNVIDIAOpenCL(Platform);
}
 
// This routine has the side effect of registering each platform's last device
// id into each plugin, which is used for device counting.
std::vector<platform> platform_impl::get_platforms() {
  std::vector<platform> Platforms;
  std::vector<plugin> &Plugins = RT::initialize();
  for (plugin &Plugin : Plugins) {
    pi_uint32 NumPlatforms = 0;
    // Move to the next plugin if the plugin fails to initialize.
    // This way platforms from other plugins get a chance to be discovered.
    if (Plugin.call_nocheck<PiApiKind::piPlatformsGet>(
            0, nullptr, &NumPlatforms) != PI_SUCCESS)
      continue;
 
    if (NumPlatforms) {
      std::vector<RT::PiPlatform> PiPlatforms(NumPlatforms);
      if (Plugin.call_nocheck<PiApiKind::piPlatformsGet>(
              NumPlatforms, PiPlatforms.data(), nullptr) != PI_SUCCESS)
        return Platforms;
 
      for (const auto &PiPlatform : PiPlatforms) {
        platform Platform = detail::createSyclObjFromImpl<platform>(
            getOrMakePlatformImpl(PiPlatform, Plugin));
        if (IsBannedPlatform(Platform)) {
          continue; // bail as early as possible, otherwise banned platforms may
                    // mess up device counting
        }
 
        {
          std::lock_guard<std::mutex> Guard(*Plugin.getPluginMutex());
          // insert PiPlatform into the Plugin
          Plugin.getPlatformId(PiPlatform);
        }
        if (!Platform.get_devices(info::device_type::all).empty())
          Platforms.push_back(Platform);
      }
    }
  }
 
  // Register default context release handler after plugins have been loaded and
  // after the first calls to each plugin. This initializes a function-local
  // variable that should be destroyed before any global variables in the
  // plugins are destroyed. This is done after the first call to the backends to
  // ensure any lazy-loaded dependencies are loaded prior to the handler
  // variable's initialization. Note: The default context release handler is not
  // guaranteed to be destroyed before function-local static variables as they
  // may be initialized after.
  GlobalHandler::registerDefaultContextReleaseHandler();
 
  return Platforms;
}
 
// Filter out the devices that are not compatible with SYCL_DEVICE_FILTER or
// ONEAPI_DEVICE_SELECTOR This function matches devices in the order of backend,
// device_type, and device_num. The device_filter and ods_target structs pun for
// each other, as do device_filter_list and ods_target_list.
// Since ONEAPI_DEVICE_SELECTOR admits negative filters, we use type traits
// to distinguish the case where we are working with ONEAPI_DEVICE_SELECTOR
// in the places where the functionality diverges between these two
// environment variables.
// The return value is a vector that represents the indices of the chosen 
// devices.
template <typename ListT, typename FilterT>
static std::vector<int> filterDeviceFilter(std::vector<RT::PiDevice> &PiDevices,
                              RT::PiPlatform Platform, ListT *FilterList) {
 
  constexpr bool is_ods_target = std::is_same_v<FilterT, ods_target>;
  // There are some differences in implementation between SYCL_DEVICE_FILTER
  // and ONEAPI_DEVICE_SELECTOR so we use if constexpr to select the
  // appropriate execution path if we are dealing with the latter variable.
 
  if constexpr (is_ods_target) {
 
    // Since we are working with ods_target filters ,which can be negative,
    // we sort the filters so that all the negative filters appear before
    // all the positive filters.  This enables us to have the full list of
    // blacklisted devices by the time we get to the positive filters
    // so that if a positive filter matches a blacklisted device we do
    // not add it to the list of available devices.
    std::sort(FilterList->get().begin(), FilterList->get().end(),
              [](const ods_target &filter1, const ods_target &filter2) {
                std::ignore = filter1;
                if (filter2.IsNegativeTarget)
                  return false;
                return true;
              });
  }
 
  // this map keeps track of devices discarded by negative filters, it is only
  // used in the ONEAPI_DEVICE_SELECTOR implemenation. It cannot be placed
  // in the if statement above because it will then be out of scope in the rest
  // of the function
  std::map<int, bool> Blacklist;
  // original indices keeps track of the device numbers of the chosen
  // devices and is whats returned by the function
  std::vector<int> original_indices;
 
  std::vector<plugin> &Plugins = RT::initialize();
  auto It =
      std::find_if(Plugins.begin(), Plugins.end(), [Platform](plugin &Plugin) {
        return Plugin.containsPiPlatform(Platform);
      });
  if (It == Plugins.end()) {
    return original_indices;
  }
  plugin &Plugin = *It;
  backend Backend = Plugin.getBackend();
  int InsertIDx = 0;
  // DeviceIds should be given consecutive numbers across platforms in the same
  // backend
  std::lock_guard<std::mutex> Guard(*Plugin.getPluginMutex());
  int DeviceNum = Plugin.getStartingDeviceId(Platform);
  for (RT::PiDevice Device : PiDevices) {
    RT::PiDeviceType PiDevType;
    Plugin.call<PiApiKind::piDeviceGetInfo>(Device, PI_DEVICE_INFO_TYPE,
                                            sizeof(RT::PiDeviceType),
                                            &PiDevType, nullptr);
    // Assumption here is that there is 1-to-1 mapping between PiDevType and
    // Sycl device type for GPU, CPU, and ACC.
    info::device_type DeviceType = pi::cast<info::device_type>(PiDevType);
 
    for (const FilterT &Filter : FilterList->get()) {
      backend FilterBackend = Filter.Backend.value_or(backend::all);
      // First, match the backend entry
      if (FilterBackend == Backend || FilterBackend == backend::all) {
        info::device_type FilterDevType =
            Filter.DeviceType.value_or(info::device_type::all);
        // Next, match the device_type entry
        if (FilterDevType == info::device_type::all) {
          // Last, match the device_num entry
          if (!Filter.DeviceNum || DeviceNum == Filter.DeviceNum.value()) {
            if constexpr (is_ods_target) {      // dealing with ODS filters
              if (!Blacklist[DeviceNum]) {      // ensure it is not blacklisted
                if (!Filter.IsNegativeTarget) { // is filter positive?
                  PiDevices[InsertIDx++] = Device;
                  original_indices.push_back(DeviceNum);
                } else {
                  // Filter is negative and the device matches the filter so
                  // blacklist the device.
                  Blacklist[DeviceNum] = true;
                }
              }
            } else { // dealing with SYCL_DEVICE_FILTER
              PiDevices[InsertIDx++] = Device;
              original_indices.push_back(DeviceNum);
            }
            break;
          }
 
        } else if (FilterDevType == DeviceType) {
          if (!Filter.DeviceNum || DeviceNum == Filter.DeviceNum.value()) {
            if constexpr (is_ods_target) {
              if (!Blacklist[DeviceNum]) {
                if (!Filter.IsNegativeTarget) {
                  PiDevices[InsertIDx++] = Device;
                  original_indices.push_back(DeviceNum);
                } else {
                  // Filter is negative and the device matches the filter so
                  // blacklist the device.
                  Blacklist[DeviceNum] = true;
                }
              }
            } else {
              PiDevices[InsertIDx++] = Device;
              original_indices.push_back(DeviceNum);
            }
            break;
          }
        }
      }
    }
    DeviceNum++;
  }
  PiDevices.resize(InsertIDx);
  // remember the last backend that has gone through this filter function
  // to assign a unique device id number across platforms that belong to
  // the same backend. For example, opencl:cpu:0, opencl:acc:1, opencl:gpu:2
  Plugin.setLastDeviceId(Platform, DeviceNum);
  return original_indices;
}
 
std::shared_ptr<device_impl>
platform_impl::getDeviceImpl(RT::PiDevice PiDevice) {
  const std::lock_guard<std::mutex> Guard(MDeviceMapMutex);
  return getDeviceImplHelper(PiDevice);
}
 
std::shared_ptr<device_impl> platform_impl::getOrMakeDeviceImpl(
    RT::PiDevice PiDevice, const std::shared_ptr<platform_impl> &PlatformImpl) {
  const std::lock_guard<std::mutex> Guard(MDeviceMapMutex);
  // If we've already seen this device, return the impl
  std::shared_ptr<device_impl> Result = getDeviceImplHelper(PiDevice);
  if (Result)
    return Result;
 
  // Otherwise make the impl
  Result = std::make_shared<device_impl>(PiDevice, PlatformImpl);
  MDeviceCache.emplace_back(Result);
 
  return Result;
}
 
static bool supportsAffinityDomain(const device &dev,
                                   info::partition_property partitionProp,
                                   info::partition_affinity_domain domain) {
  if (partitionProp != info::partition_property::partition_by_affinity_domain) {
    return true;
  }
  auto supported = dev.get_info<info::device::partition_affinity_domains>();
  auto It = std::find(std::begin(supported), std::end(supported), domain);
  return It != std::end(supported);
}
 
static bool supportsPartitionProperty(const device &dev,
                                      info::partition_property partitionProp) {
  auto supported = dev.get_info<info::device::partition_properties>();
  auto It =
      std::find(std::begin(supported), std::end(supported), partitionProp);
  return It != std::end(supported);
}
 
static std::vector<device> amendDeviceAndSubDevices(
    backend PlatformBackend, std::vector<device> &DeviceList,
    ods_target_list *OdsTargetList, const std::vector<int>& original_indices,
    PlatformImplPtr PlatformImpl) {
  constexpr info::partition_property partitionProperty =
      info::partition_property::partition_by_affinity_domain;
  constexpr info::partition_affinity_domain affinityDomain =
      info::partition_affinity_domain::next_partitionable;
 
  std::vector<device> FinalResult;
  // (Only) when amending sub-devices for ONEAPI_DEVICE_SELECTOR, all
  // sub-devices are treated as root.
  TempAssignGuard<bool> TAG(PlatformImpl->MAlwaysRootDevice, true);
 
  for (unsigned i = 0; i < DeviceList.size(); i++) {
    // device has already been screened. The question is whether it should be a
    // top level device and/or is expected to add its sub-devices to the list.
    device &dev = DeviceList[i];
    bool deviceAdded = false;
    for (ods_target target : OdsTargetList->get()) {
      backend TargetBackend = target.Backend.value_or(backend::all);
      if (PlatformBackend == TargetBackend || TargetBackend == backend::all) {
        bool deviceMatch = target.HasDeviceWildCard; // opencl:*
        if (target.DeviceType) {                     // opencl:gpu
          deviceMatch = ((target.DeviceType == info::device_type::all) ||
                         (dev.get_info<info::device::device_type>() ==
                          target.DeviceType));
 
        } else if (target.DeviceNum) { // opencl:0
          deviceMatch =
              (target.DeviceNum.value() == original_indices[i]);
        }
 
        if (deviceMatch) {
          // Top level matches. Do we add it, or subdevices, or sub-sub-devices?
          bool wantSubDevice =
              target.SubDeviceNum || target.HasSubDeviceWildCard;
          bool supportsSubPartitioning =
              (supportsPartitionProperty(dev, partitionProperty) &&
               supportsAffinityDomain(dev, partitionProperty, affinityDomain));
          bool wantSubSubDevice =
              target.SubSubDeviceNum || target.HasSubSubDeviceWildCard;
 
          // -- Add top level device.
          if (!wantSubDevice) {
            if (!deviceAdded) {
              FinalResult.push_back(dev);
              deviceAdded = true;
            }
          } else {
            if (!supportsSubPartitioning) {
              if (target.DeviceNum ||
                  (target.DeviceType &&
                   (target.DeviceType.value() != info::device_type::all))) {
                // This device was specifically requested and yet is not
                // partitionable.
                std::cout << "device is not partitionable: " << target
                          << std::endl;
              }
              continue;
            }
            // -- Add sub sub device.
            if (wantSubSubDevice) {
 
              auto subDevicesToPartition =
                  dev.create_sub_devices<partitionProperty>(affinityDomain);
              if (target.SubDeviceNum) {
                if (subDevicesToPartition.size() >
                    target.SubDeviceNum.value()) {
                  subDevicesToPartition[0] =
                      subDevicesToPartition[target.SubDeviceNum.value()];
                  subDevicesToPartition.resize(1);
                } else {
                  std::cout << "subdevice index out of bounds: " << target
                            << std::endl;
                  continue;
                }
              }
              for (device subDev : subDevicesToPartition) {
                bool supportsSubSubPartitioning =
                    (supportsPartitionProperty(subDev, partitionProperty) &&
                     supportsAffinityDomain(subDev, partitionProperty,
                                            affinityDomain));
                if (!supportsSubSubPartitioning) {
                  if (target.SubDeviceNum) {
                    // Parent subdevice was specifically requested, yet is not
                    // partitionable.
                    std::cout << "sub-device is not partitionable: " << target
                              << std::endl;
                  }
                  continue;
                }
                // Allright, lets get them sub-sub-devices.
                auto subSubDevices =
                    subDev.create_sub_devices<partitionProperty>(
                        affinityDomain);
                if (target.HasSubSubDeviceWildCard) {
                  FinalResult.insert(FinalResult.end(), subSubDevices.begin(),
                                     subSubDevices.end());
                } else {
                  if (subSubDevices.size() > target.SubSubDeviceNum.value()) {
                    FinalResult.push_back(
                        subSubDevices[target.SubSubDeviceNum.value()]);
                  } else {
                    std::cout
                        << "sub-sub-device index out of bounds: " << target
                        << std::endl;
                  }
                }
              }
            } else if (wantSubDevice) {
              auto subDevices = dev.create_sub_devices<
                  info::partition_property::partition_by_affinity_domain>(
                  affinityDomain);
              if (target.HasSubDeviceWildCard) {
                FinalResult.insert(FinalResult.end(), subDevices.begin(),
                                   subDevices.end());
              } else {
                if (subDevices.size() > target.SubDeviceNum.value()) {
                  FinalResult.push_back(
                      subDevices[target.SubDeviceNum.value()]);
                } else {
                  std::cout << "subdevice index out of bounds: " << target
                            << std::endl;
                }
              }
            }
          }
        } // /if deviceMatch
      }
    } // /for
  }   // /for
  return FinalResult;
}
 
std::vector<device>
platform_impl::get_devices(info::device_type DeviceType) const {
  std::vector<device> Res;
  // Will we be filtering with SYCL_DEVICE_FILTER or ONEAPI_DEVICE_SELECTOR ?
  // We do NOT attempt to support both simultaneously.
  ods_target_list *OdsTargetList = SYCLConfig<ONEAPI_DEVICE_SELECTOR>::get();
  device_filter_list *FilterList = SYCLConfig<SYCL_DEVICE_FILTER>::get();
 
  if (is_host() && (DeviceType == info::device_type::host ||
                    DeviceType == info::device_type::all)) {
    Res.push_back(
        createSyclObjFromImpl<device>(device_impl::getHostDeviceImpl()));
  }
 
  // If any DeviceType other than host was requested for host platform,
  // an empty vector will be returned.
  if (is_host() || DeviceType == info::device_type::host)
    return Res;
 
  pi_uint32 NumDevices = 0;
  const detail::plugin &Plugin = getPlugin();
  Plugin.call<PiApiKind::piDevicesGet>(
      MPlatform, pi::cast<RT::PiDeviceType>(DeviceType),
      0, // CP info::device_type::all
      pi::cast<RT::PiDevice *>(nullptr), &NumDevices);
  const backend Backend = Plugin.getBackend();
 
  if (NumDevices == 0) {
    // If platform doesn't have devices (even without filter)
    // LastDeviceIds[PlatformId] stay 0 that affects next platform devices num
    // analysis. Doing adjustment by simple copy of last device num from
    // previous platform.
    // Needs non const plugin reference.
    std::vector<plugin> &Plugins = RT::initialize();
    auto It = std::find_if(Plugins.begin(), Plugins.end(),
                           [&Platform = MPlatform](plugin &Plugin) {
                             return Plugin.containsPiPlatform(Platform);
                           });
    if (It != Plugins.end()) {
      std::lock_guard<std::mutex> Guard(*(It->getPluginMutex()));
      (*It).adjustLastDeviceId(MPlatform);
    }
    return Res;
  }
 
  std::vector<RT::PiDevice> PiDevices(NumDevices);
  // TODO catch an exception and put it to list of asynchronous exceptions
  Plugin.call<PiApiKind::piDevicesGet>(
      MPlatform,
      pi::cast<RT::PiDeviceType>(DeviceType), // CP info::device_type::all
      NumDevices, PiDevices.data(), nullptr);
 
  // Filter out devices that are not present in the SYCL_DEVICE_ALLOWLIST
  if (SYCLConfig<SYCL_DEVICE_ALLOWLIST>::get())
    applyAllowList(PiDevices, MPlatform, Plugin);
 
  // The first step is to filter out devices that are not compatible with
  // SYCL_DEVICE_FILTER or ONEAPI_DEVICE_SELECTOR. This is also the mechanism by
  // which top level device ids are assigned.
  std::vector<int> PlatformDeviceIndices;
  if (OdsTargetList) {
    if (FilterList) {
      throw sycl::exception(sycl::make_error_code(errc::invalid),
                            "ONEAPI_DEVICE_SELECTOR cannot be used in "
                            "conjunction with SYCL_DEVICE_FILTER");
    }
    PlatformDeviceIndices = filterDeviceFilter<ods_target_list, ods_target>(
        PiDevices, MPlatform, OdsTargetList);
  } else if (FilterList) {
    PlatformDeviceIndices = filterDeviceFilter<device_filter_list, device_filter>(
        PiDevices, MPlatform, FilterList);
  }
 
  // The next step is to inflate the filtered PIDevices into SYCL Device
  // objects.
  PlatformImplPtr PlatformImpl = getOrMakePlatformImpl(MPlatform, Plugin);
  std::transform(
      PiDevices.begin(), PiDevices.end(), std::back_inserter(Res),
      [PlatformImpl](const RT::PiDevice &PiDevice) -> device {
        return detail::createSyclObjFromImpl<device>(
            PlatformImpl->getOrMakeDeviceImpl(PiDevice, PlatformImpl));
      });
 
  // If we aren't using ONEAPI_DEVICE_SELECTOR, then we are done.
  // and if there are no devices so far, there won't be any need to replace them
  // with subdevices.
  if (!OdsTargetList || Res.size() == 0)
    return Res;
 
  // Otherwise, our last step is to revisit the devices, possibly replacing
  // them with subdevices (which have been ignored until now)
  return amendDeviceAndSubDevices(Backend, Res, OdsTargetList,
                                  PlatformDeviceIndices, PlatformImpl);
}
 
bool platform_impl::has_extension(const std::string &ExtensionName) const {
  if (is_host())
    return false;
 
  std::string AllExtensionNames = get_platform_info_string_impl(
      MPlatform, getPlugin(),
      detail::PiInfoCode<info::platform::extensions>::value);
  return (AllExtensionNames.find(ExtensionName) != std::string::npos);
}
 
pi_native_handle platform_impl::getNative() const {
  const auto &Plugin = getPlugin();
  pi_native_handle Handle;
  Plugin.call<PiApiKind::piextPlatformGetNativeHandle>(getHandleRef(), &Handle);
  return Handle;
}
 
template <typename Param>
typename Param::return_type platform_impl::get_info() const {
  if (is_host())
    return get_platform_info_host<Param>();
 
  return get_platform_info<Param>(this->getHandleRef(), getPlugin());
}
 
// All devices on the platform must have the given aspect.
bool platform_impl::has(aspect Aspect) const {
  for (const auto &dev : get_devices()) {
    if (dev.has(Aspect) == false) {
      return false;
    }
  }
  return true;
}
 
std::shared_ptr<device_impl>
platform_impl::getDeviceImplHelper(RT::PiDevice PiDevice) {
  for (const std::weak_ptr<device_impl> &DeviceWP : MDeviceCache) {
    if (std::shared_ptr<device_impl> Device = DeviceWP.lock()) {
      if (Device->getHandleRef() == PiDevice)
        return Device;
    }
  }
  return nullptr;
}
 
#define __SYCL_PARAM_TRAITS_SPEC(DescType, Desc, ReturnT, PiCode)              \
  template ReturnT platform_impl::get_info<info::platform::Desc>() const;
 
#include <sycl/info/platform_traits.def>
#undef __SYCL_PARAM_TRAITS_SPEC
 
} // namespace detail
} // __SYCL_INLINE_VER_NAMESPACE(_V1)
} // namespace sycl