Merge pull request #1192 from maarquitos14/maronas/ext_composite_device

Initial support for sycl_ext_oneapi_composite_device

Author: aarongreig

include/ur_api.h

@@ -1530,6 +1530,10 @@ typedef enum ur_device_info_t {
                                                                     ///< version than older devices.
     UR_DEVICE_INFO_VIRTUAL_MEMORY_SUPPORT = 114,                    ///< [::ur_bool_t] return true if the device supports virtual memory.
     UR_DEVICE_INFO_ESIMD_SUPPORT = 115,                             ///< [::ur_bool_t] return true if the device supports ESIMD.
+    UR_DEVICE_INFO_COMPONENT_DEVICES = 116,                         ///< [::ur_device_handle_t[]] The set of component devices contained by
+                                                                    ///< this composite device.
+    UR_DEVICE_INFO_COMPOSITE_DEVICE = 117,                          ///< [::ur_device_handle_t] The composite device containing this component
+                                                                    ///< device.
     UR_DEVICE_INFO_BINDLESS_IMAGES_SUPPORT_EXP = 0x2000,            ///< [::ur_bool_t] returns true if the device supports the creation of
                                                                     ///< bindless images
     UR_DEVICE_INFO_BINDLESS_IMAGES_SHARED_USM_SUPPORT_EXP = 0x2001, ///< [::ur_bool_t] returns true if the device supports the creation of

include/ur_print.hpp

@@ -2401,6 +2401,12 @@ inline std::ostream &operator<<(std::ostream &os, ur_device_info_t value) {
     case UR_DEVICE_INFO_ESIMD_SUPPORT:
         os << "UR_DEVICE_INFO_ESIMD_SUPPORT";
         break;
+    case UR_DEVICE_INFO_COMPONENT_DEVICES:
+        os << "UR_DEVICE_INFO_COMPONENT_DEVICES";
+        break;
+    case UR_DEVICE_INFO_COMPOSITE_DEVICE:
+        os << "UR_DEVICE_INFO_COMPOSITE_DEVICE";
+        break;
     case UR_DEVICE_INFO_BINDLESS_IMAGES_SUPPORT_EXP:
         os << "UR_DEVICE_INFO_BINDLESS_IMAGES_SUPPORT_EXP";
         break;
@@ -3809,6 +3815,34 @@ inline ur_result_t printTagged(std::ostream &os, const void *ptr, ur_device_info
 
         os << ")";
     } break;
+    case UR_DEVICE_INFO_COMPONENT_DEVICES: {
+
+        const ur_device_handle_t *tptr = (const ur_device_handle_t *)ptr;
+        os << "{";
+        size_t nelems = size / sizeof(ur_device_handle_t);
+        for (size_t i = 0; i < nelems; ++i) {
+            if (i != 0) {
+                os << ", ";
+            }
+
+            ur::details::printPtr(os,
+                                  tptr[i]);
+        }
+        os << "}";
+    } break;
+    case UR_DEVICE_INFO_COMPOSITE_DEVICE: {
+        const ur_device_handle_t *tptr = (const ur_device_handle_t *)ptr;
+        if (sizeof(ur_device_handle_t) > size) {
+            os << "invalid size (is: " << size << ", expected: >=" << sizeof(ur_device_handle_t) << ")";
+            return UR_RESULT_ERROR_INVALID_SIZE;
+        }
+        os << (const void *)(tptr) << " (";
+
+        ur::details::printPtr(os,
+                              *tptr);
+
+        os << ")";
+    } break;
     case UR_DEVICE_INFO_BINDLESS_IMAGES_SUPPORT_EXP: {
         const ur_bool_t *tptr = (const ur_bool_t *)ptr;
         if (sizeof(ur_bool_t) > size) {

scripts/core/device.yml

@@ -396,6 +396,10 @@ etors:
       desc: "[$x_bool_t] return true if the device supports virtual memory."
     - name: ESIMD_SUPPORT
       desc: "[$x_bool_t] return true if the device supports ESIMD."
+    - name: COMPONENT_DEVICES
+      desc: "[$x_device_handle_t[]] The set of component devices contained by this composite device."
+    - name: COMPOSITE_DEVICE
+      desc: "[$x_device_handle_t] The composite device containing this component device."
 --- #--------------------------------------------------------------------------
 type: function
 desc: "Retrieves various information about device"

source/adapters/cuda/device.cpp

@@ -1022,6 +1022,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
     return ReturnValue(true);
   case UR_DEVICE_INFO_ESIMD_SUPPORT:
     return ReturnValue(false);
+  case UR_DEVICE_INFO_COMPONENT_DEVICES:
+  case UR_DEVICE_INFO_COMPOSITE_DEVICE:
+    // These two are exclusive of L0.
+    return ReturnValue(0);
   case UR_DEVICE_INFO_MAX_READ_WRITE_IMAGE_ARGS:
   case UR_DEVICE_INFO_GPU_EU_COUNT:
   case UR_DEVICE_INFO_GPU_EU_SIMD_WIDTH:

source/adapters/hip/device.cpp

@@ -823,6 +823,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
     return ReturnValue(false);
   case UR_DEVICE_INFO_ESIMD_SUPPORT:
     return ReturnValue(false);
+  case UR_DEVICE_INFO_COMPONENT_DEVICES:
+  case UR_DEVICE_INFO_COMPOSITE_DEVICE:
+    // These two are exclusive of L0.
+    return ReturnValue(0);
 
   // TODO: Investigate if this information is available on HIP.
   case UR_DEVICE_INFO_GPU_EU_COUNT:

source/adapters/level_zero/device.cpp

@@ -42,6 +42,34 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGet(
   // Filter available devices based on input DeviceType.
   std::vector<ur_device_handle_t> MatchedDevices;
   std::shared_lock<ur_shared_mutex> Lock(Platform->URDevicesCacheMutex);
+  // We need to filter out composite devices when
+  // ZE_FLAT_DEVICE_HIERARCHY=COMBINED. We can know if we are in combined
+  // mode depending on the return value of zeDeviceGetRootDevice:
+  //   - If COMPOSITE, L0 returns cards as devices. Since we filter out
+  //     subdevices early, zeDeviceGetRootDevice must return nullptr, because we
+  //     only query for root-devices and they don't have any device higher up in
+  //     the hierarchy.
+  //   - If FLAT,  according to L0 spec, zeDeviceGetRootDevice always returns
+  //     nullptr in this mode.
+  //   - If COMBINED, L0 returns tiles as devices, and zeDeviceGetRootdevice
+  //     returns the card containing a given tile.
+  bool isCombinedMode =
+      std::any_of(Platform->URDevicesCache.begin(),
+                  Platform->URDevicesCache.end(), [](const auto &D) {
+                    if (D->isSubDevice())
+                      return false;
+                    ze_device_handle_t RootDev = nullptr;
+                    // Query Root Device for root-devices.
+                    // We cannot use ZE2UR_CALL because under some circumstances
+                    // this call may return ZE_RESULT_ERROR_UNSUPPORTED_FEATURE,
+                    // and ZE2UR_CALL will abort because it's not
+                    // UR_RESULT_SUCCESS. Instead, we use ZE_CALL_NOCHECK and we
+                    // check manually that the result is either
+                    // ZE_RESULT_SUCCESS or ZE_RESULT_ERROR_UNSUPPORTED_FEATURE.
+                    auto errc = ZE_CALL_NOCHECK(zeDeviceGetRootDevice,
+                                                (D->ZeDevice, &RootDev));
+                    return (errc == ZE_RESULT_SUCCESS && RootDev != nullptr);
+                  });
   for (auto &D : Platform->URDevicesCache) {
     // Only ever return root-devices from urDeviceGet, but the
     // devices cache also keeps sub-devices.
@@ -71,8 +99,14 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGet(
       urPrint("Unknown device type");
       break;
     }
-    if (Matched)
-      MatchedDevices.push_back(D.get());
+
+    if (Matched) {
+      bool isComposite =
+          isCombinedMode && (D->ZeDeviceProperties->flags &
+                             ZE_DEVICE_PROPERTY_FLAG_SUBDEVICE) == 0;
+      if (!isComposite)
+        MatchedDevices.push_back(D.get());
+    }
   }
 
   uint32_t ZeDeviceCount = MatchedDevices.size();
@@ -826,6 +860,62 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(
     return ReturnValue(result);
   }
 
+  case UR_DEVICE_INFO_COMPONENT_DEVICES: {
+    ze_device_handle_t DevHandle = Device->ZeDevice;
+    uint32_t SubDeviceCount = 0;
+    // First call to get SubDeviceCount.
+    ZE2UR_CALL(zeDeviceGetSubDevices, (DevHandle, &SubDeviceCount, nullptr));
+    if (SubDeviceCount == 0)
+      return ReturnValue(0);
+
+    std::vector<ze_device_handle_t> SubDevs(SubDeviceCount);
+    // Second call to get the actual list of devices.
+    ZE2UR_CALL(zeDeviceGetSubDevices,
+               (DevHandle, &SubDeviceCount, SubDevs.data()));
+
+    size_t SubDeviceCount_s{SubDeviceCount};
+    auto ResSize =
+        std::min(SubDeviceCount_s, propSize / sizeof(ur_device_handle_t));
+    std::vector<ur_device_handle_t> Res;
+    for (const auto &d : SubDevs) {
+      // We can only reach this code if ZE_FLAT_DEVICE_HIERARCHY != FLAT,
+      // because in flat mode we directly get tiles, and those don't have any
+      // further divisions, so zeDeviceGetSubDevices always will return an empty
+      // list. Thus, there's only two options left: (a) composite mode, and (b)
+      // combined mode. In (b), zeDeviceGet returns tiles as devices, and those
+      // are presented as root devices (i.e. isSubDevice() returns false). In
+      // contrast, in (a), zeDeviceGet returns cards as devices, so tiles are
+      // not root devices (i.e. isSubDevice() returns true). Since we only reach
+      // this code if there are tiles returned by zeDeviceGetSubDevices, we
+      // can know if we are in (a) or (b) by checking if a tile is root device
+      // or not.
+      ur_device_handle_t URDev = Device->Platform->getDeviceFromNativeHandle(d);
+      if (URDev->isSubDevice())
+        // We are in COMPOSITE mode, return an empty list.
+        return ReturnValue(0);
+
+      Res.push_back(URDev);
+    }
+    if (pSize)
+      *pSize = SubDeviceCount * sizeof(ur_device_handle_t);
+    if (ParamValue) {
+      return ReturnValue(Res.data(), ResSize);
+    }
+    return UR_RESULT_SUCCESS;
+  }
+  case UR_DEVICE_INFO_COMPOSITE_DEVICE: {
+    ur_device_handle_t UrRootDev = nullptr;
+    ze_device_handle_t DevHandle = Device->ZeDevice;
+    ze_device_handle_t RootDev;
+    // Query Root Device.
+    auto errc = ZE_CALL_NOCHECK(zeDeviceGetRootDevice, (DevHandle, &RootDev));
+    UrRootDev = Device->Platform->getDeviceFromNativeHandle(RootDev);
+    if (errc != ZE_RESULT_SUCCESS &&
+        errc != ZE_RESULT_ERROR_UNSUPPORTED_FEATURE)
+      return ze2urResult(errc);
+    return ReturnValue(UrRootDev);
+  }
+
   default:
     urPrint("Unsupported ParamName in urGetDeviceInfo\n");
     urPrint("ParamName=%d(0x%x)\n", ParamName, ParamName);

source/adapters/level_zero/platform.cpp

@@ -337,6 +337,35 @@ ur_result_t ur_platform_handle_t_::populateDeviceCacheIfNeeded() {
       }
       delete[] ZeSubdevices;
 
+      // When using ZE_FLAT_DEVICE_HIERARCHY=COMBINED, zeDeviceGet will
+      // return tiles as devices, but we can get the card device handle
+      // through zeDeviceGetRootDevice. We need to cache the card device
+      // handle too, such that it is readily visible to the
+      // urDeviceCreateWithNativeHandle.
+      ze_device_handle_t RootDevice = nullptr;
+      // We cannot use ZE2UR_CALL because under some circumstances this call may
+      // return ZE_RESULT_ERROR_UNSUPPORTED_FEATURE, and ZE2UR_CALL will abort
+      // because it's not UR_RESULT_SUCCESS. Instead, we use ZE_CALL_NOCHECK and
+      // we check manually that the result is either ZE_RESULT_SUCCESS or
+      // ZE_RESULT_ERROR_UNSUPPORTED_FEATURE.
+      auto errc = ZE_CALL_NOCHECK(zeDeviceGetRootDevice,
+                                  (Device->ZeDevice, &RootDevice));
+      if (errc != ZE_RESULT_SUCCESS &&
+          errc != ZE_RESULT_ERROR_UNSUPPORTED_FEATURE)
+        return ze2urResult(errc);
+
+      if (RootDevice) {
+        if (std::find_if(URDevicesCache.begin(), URDevicesCache.end(),
+                         [&](auto &Dev) {
+                           return Dev->ZeDevice == RootDevice;
+                         }) == URDevicesCache.end()) {
+          std::unique_ptr<ur_device_handle_t_> UrRootDevice(
+              new ur_device_handle_t_(RootDevice, (ur_platform_handle_t)this));
+          UR_CALL(UrRootDevice->initialize());
+          URDevicesCache.push_back(std::move(UrRootDevice));
+        }
+      }
+
       // Save the root device in the cache for future uses.
       URDevicesCache.push_back(std::move(Device));
     }

source/adapters/native_cpu/device.cpp

@@ -300,6 +300,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
   }
   case UR_DEVICE_INFO_ESIMD_SUPPORT:
     return ReturnValue(false);
+  case UR_DEVICE_INFO_COMPONENT_DEVICES:
+  case UR_DEVICE_INFO_COMPOSITE_DEVICE:
+    // These two are exclusive of L0.
+    return ReturnValue(0);
 
     CASE_UR_UNSUPPORTED(UR_DEVICE_INFO_MAX_MEMORY_BANDWIDTH);
   case UR_DEVICE_INFO_VIRTUAL_MEMORY_SUPPORT:

source/adapters/opencl/device.cpp

@@ -923,6 +923,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
     }
     return ReturnValue(SupportedExtensions.c_str());
   }
+  case UR_DEVICE_INFO_COMPONENT_DEVICES:
+  case UR_DEVICE_INFO_COMPOSITE_DEVICE:
+    // These two are exclusive of L0.
+    return ReturnValue(0);
   /* TODO: Check regularly to see if support is enabled in OpenCL. Intel GPU
    * EU device-specific information extensions. Some of the queries are
    * enabled by cl_intel_device_attribute_query extension, but it's not yet in

test/conformance/device/urDeviceGetInfo.cpp

@@ -112,6 +112,8 @@ static std::unordered_map<ur_device_info_t, size_t> device_info_size_map = {
     {UR_DEVICE_INFO_MEM_CHANNEL_SUPPORT, sizeof(ur_bool_t)},
     {UR_DEVICE_INFO_HOST_PIPE_READ_WRITE_SUPPORTED, sizeof(ur_bool_t)},
     {UR_DEVICE_INFO_MAX_REGISTERS_PER_WORK_GROUP, sizeof(uint32_t)},
+    {UR_DEVICE_INFO_COMPONENT_DEVICES, sizeof(uint32_t)},
+    {UR_DEVICE_INFO_COMPOSITE_DEVICE, sizeof(ur_device_handle_t)},
 };
 
 struct urDeviceGetInfoTest : uur::urAllDevicesTest,