[SYCL] Add ext_oneapi_unified_runtime SYCL backend (#7816)

This PR adds new backend to SYCL: ext_oneapi_unified_runtime
E2E test: intel/llvm-test-suite#1450

Signed-off-by: Sergey V Maslov <[email protected]>

Author: smaslov-intel

sycl/include/sycl/backend_types.hpp

@@ -32,6 +32,7 @@ enum class backend : char {
       ext_intel_esimd_emulator,
   ext_oneapi_hip = 6,
   hip __SYCL2020_DEPRECATED("use 'ext_oneapi_hip' instead") = ext_oneapi_hip,
+  ext_oneapi_unified_runtime = 7,
 };
 
 template <backend Backend> class backend_traits;
@@ -63,6 +64,9 @@ inline std::ostream &operator<<(std::ostream &Out, backend be) {
   case backend::ext_oneapi_hip:
     Out << "ext_oneapi_hip";
     break;
+  case backend::ext_oneapi_unified_runtime:
+    Out << "ext_oneapi_unified_runtime";
+    break;
   case backend::all:
     Out << "all";
   }

sycl/include/sycl/detail/pi.hpp

@@ -67,18 +67,21 @@ bool trace(TraceLevel level);
 #define __SYCL_CUDA_PLUGIN_NAME "pi_cuda.dll"
 #define __SYCL_ESIMD_EMULATOR_PLUGIN_NAME "pi_esimd_emulator.dll"
 #define __SYCL_HIP_PLUGIN_NAME "libpi_hip.dll"
+#define __SYCL_UNIFIED_RUNTIME_PLUGIN_NAME "pi_unified_runtime.dll"
 #elif defined(__SYCL_RT_OS_LINUX)
 #define __SYCL_OPENCL_PLUGIN_NAME "libpi_opencl.so"
 #define __SYCL_LEVEL_ZERO_PLUGIN_NAME "libpi_level_zero.so"
 #define __SYCL_CUDA_PLUGIN_NAME "libpi_cuda.so"
 #define __SYCL_ESIMD_EMULATOR_PLUGIN_NAME "libpi_esimd_emulator.so"
+#define __SYCL_UNIFIED_RUNTIME_PLUGIN_NAME "libpi_unified_runtime.so"
 #define __SYCL_HIP_PLUGIN_NAME "libpi_hip.so"
 #elif defined(__SYCL_RT_OS_DARWIN)
 #define __SYCL_OPENCL_PLUGIN_NAME "libpi_opencl.dylib"
 #define __SYCL_LEVEL_ZERO_PLUGIN_NAME "libpi_level_zero.dylib"
 #define __SYCL_CUDA_PLUGIN_NAME "libpi_cuda.dylib"
 #define __SYCL_ESIMD_EMULATOR_PLUGIN_NAME "libpi_esimd_emulator.dylib"
 #define __SYCL_HIP_PLUGIN_NAME "libpi_hip.dylib"
+#define __SYCL_UNIFIED_RUNTIME_PLUGIN_NAME "libpi_unified_runtime.dylib"
 #else
 #error "Unsupported OS"
 #endif

sycl/plugins/level_zero/pi_level_zero.cpp

@@ -198,87 +198,6 @@ template <> ze_result_t zeHostSynchronize(ze_command_queue_handle_t Handle) {
   return zeHostSynchronizeImpl(zeCommandQueueSynchronize, Handle);
 }
 
-template <typename T, typename Assign>
-pi_result getInfoImpl(size_t param_value_size, void *param_value,
-                      size_t *param_value_size_ret, T value, size_t value_size,
-                      Assign &&assign_func) {
-
-  if (param_value != nullptr) {
-
-    if (param_value_size < value_size) {
-      return PI_ERROR_INVALID_VALUE;
-    }
-
-    assign_func(param_value, value, value_size);
-  }
-
-  if (param_value_size_ret != nullptr) {
-    *param_value_size_ret = value_size;
-  }
-
-  return PI_SUCCESS;
-}
-
-template <typename T>
-pi_result getInfo(size_t param_value_size, void *param_value,
-                  size_t *param_value_size_ret, T value) {
-
-  auto assignment = [](void *param_value, T value, size_t value_size) {
-    (void)value_size;
-    *static_cast<T *>(param_value) = value;
-  };
-
-  return getInfoImpl(param_value_size, param_value, param_value_size_ret, value,
-                     sizeof(T), assignment);
-}
-
-template <typename T>
-pi_result getInfoArray(size_t array_length, size_t param_value_size,
-                       void *param_value, size_t *param_value_size_ret,
-                       T *value) {
-  return getInfoImpl(param_value_size, param_value, param_value_size_ret, value,
-                     array_length * sizeof(T), memcpy);
-}
-
-template <typename T, typename RetType>
-pi_result getInfoArray(size_t array_length, size_t param_value_size,
-                       void *param_value, size_t *param_value_size_ret,
-                       T *value) {
-  if (param_value) {
-    memset(param_value, 0, param_value_size);
-    for (uint32_t I = 0; I < array_length; I++)
-      ((RetType *)param_value)[I] = (RetType)value[I];
-  }
-  if (param_value_size_ret)
-    *param_value_size_ret = array_length * sizeof(RetType);
-  return PI_SUCCESS;
-}
-
-template <>
-pi_result getInfo<const char *>(size_t param_value_size, void *param_value,
-                                size_t *param_value_size_ret,
-                                const char *value) {
-  return getInfoArray(strlen(value) + 1, param_value_size, param_value,
-                      param_value_size_ret, value);
-}
-
-class ReturnHelper {
-public:
-  ReturnHelper(size_t param_value_size, void *param_value,
-               size_t *param_value_size_ret)
-      : param_value_size(param_value_size), param_value(param_value),
-        param_value_size_ret(param_value_size_ret) {}
-
-  template <class T> pi_result operator()(const T &t) {
-    return getInfo(param_value_size, param_value, param_value_size_ret, t);
-  }
-
-private:
-  size_t param_value_size;
-  void *param_value;
-  size_t *param_value_size_ret;
-};
-
 } // anonymous namespace
 
 // SYCL_PI_LEVEL_ZERO_USE_COMPUTE_ENGINE can be set to an integer (>=0) in
@@ -439,11 +358,6 @@ pi_result _pi_context::decrementUnreleasedEventsInPool(pi_event Event) {
   return PI_SUCCESS;
 }
 
-// Some opencl extensions we know are supported by all Level Zero devices.
-constexpr char ZE_SUPPORTED_EXTENSIONS[] =
-    "cl_khr_il_program cl_khr_subgroups cl_intel_subgroups "
-    "cl_intel_subgroups_short cl_intel_required_subgroup_size ";
-
 // Forward declarations
 static pi_result
 enqueueMemCopyHelper(pi_command_type CommandType, pi_queue Queue, void *Dst,
@@ -2307,51 +2221,17 @@ pi_result piPlatformsGet(pi_uint32 NumEntries, pi_platform *Platforms,
 pi_result piPlatformGetInfo(pi_platform Platform, pi_platform_info ParamName,
                             size_t ParamValueSize, void *ParamValue,
                             size_t *ParamValueSizeRet) {
-
-  PI_ASSERT(Platform, PI_ERROR_INVALID_PLATFORM);
-
   zePrint("==========================\n");
   zePrint("SYCL over Level-Zero %s\n", Platform->ZeDriverVersion.c_str());
   zePrint("==========================\n");
 
-  ReturnHelper ReturnValue(ParamValueSize, ParamValue, ParamValueSizeRet);
-
-  switch (ParamName) {
-  case PI_PLATFORM_INFO_NAME:
-    // TODO: Query Level Zero driver when relevant info is added there.
+  // To distinguish this L0 platform from Unified Runtime one.
+  if (ParamName == PI_PLATFORM_INFO_NAME) {
+    ReturnHelper ReturnValue(ParamValueSize, ParamValue, ParamValueSizeRet);
     return ReturnValue("Intel(R) Level-Zero");
-  case PI_PLATFORM_INFO_VENDOR:
-    // TODO: Query Level Zero driver when relevant info is added there.
-    return ReturnValue("Intel(R) Corporation");
-  case PI_PLATFORM_INFO_EXTENSIONS:
-    // Convention adopted from OpenCL:
-    //     "Returns a space-separated list of extension names (the extension
-    // names themselves do not contain any spaces) supported by the platform.
-    // Extensions defined here must be supported by all devices associated
-    // with this platform."
-    //
-    // TODO: Check the common extensions supported by all connected devices and
-    // return them. For now, hardcoding some extensions we know are supported by
-    // all Level Zero devices.
-    return ReturnValue(ZE_SUPPORTED_EXTENSIONS);
-  case PI_PLATFORM_INFO_PROFILE:
-    // TODO: figure out what this means and how is this used
-    return ReturnValue("FULL_PROFILE");
-  case PI_PLATFORM_INFO_VERSION:
-    // TODO: this should query to zeDriverGetDriverVersion
-    // but we don't yet have the driver handle here.
-    //
-    // From OpenCL 2.1: "This version string has the following format:
-    // OpenCL<space><major_version.minor_version><space><platform-specific
-    // information>. Follow the same notation here.
-    //
-    return ReturnValue(Platform->ZeDriverApiVersion.c_str());
-  default:
-    zePrint("piPlatformGetInfo: unrecognized ParamName\n");
-    return PI_ERROR_INVALID_VALUE;
   }
-
-  return PI_SUCCESS;
+  return pi2ur::piPlatformGetInfo(Platform, ParamName, ParamValueSize,
+                                  ParamValue, ParamValueSizeRet);
 }
 
 pi_result piextPlatformGetNativeHandle(pi_platform Platform,
@@ -3068,10 +2948,9 @@ pi_result piDeviceGetInfo(pi_device Device, pi_device_info ParamName,
   case PI_DEVICE_INFO_SUB_GROUP_SIZES_INTEL: {
     // ze_device_compute_properties.subGroupSizes is in uint32_t whereas the
     // expected return is size_t datatype. size_t can be 8 bytes of data.
-    return getInfoArray<uint32_t, size_t>(
-        Device->ZeDeviceComputeProperties->numSubGroupSizes, ParamValueSize,
-        ParamValue, ParamValueSizeRet,
-        Device->ZeDeviceComputeProperties->subGroupSizes);
+    return ReturnValue.template operator()<size_t>(
+        Device->ZeDeviceComputeProperties->subGroupSizes,
+        Device->ZeDeviceComputeProperties->numSubGroupSizes);
   }
   case PI_DEVICE_INFO_IL_VERSION: {
     // Set to a space separated list of IL version strings of the form
@@ -3463,8 +3342,7 @@ pi_result piContextGetInfo(pi_context Context, pi_context_info ParamName,
   ReturnHelper ReturnValue(ParamValueSize, ParamValue, ParamValueSizeRet);
   switch (ParamName) {
   case PI_CONTEXT_INFO_DEVICES:
-    return getInfoArray(Context->Devices.size(), ParamValueSize, ParamValue,
-                        ParamValueSizeRet, &Context->Devices[0]);
+    return ReturnValue(&Context->Devices[0], Context->Devices.size());
   case PI_CONTEXT_INFO_NUM_DEVICES:
     return ReturnValue(pi_uint32(Context->Devices.size()));
   case PI_CONTEXT_INFO_REFERENCE_COUNT:
@@ -5375,7 +5253,8 @@ pi_result piKernelGetGroupInfo(pi_kernel Kernel, pi_device Device,
     return ReturnValue(WorkSize);
   }
   case PI_KERNEL_GROUP_INFO_WORK_GROUP_SIZE: {
-    // As of right now, L0 is missing API to query kernel and device specific max work group size. 
+    // As of right now, L0 is missing API to query kernel and device specific
+    // max work group size.
     return ReturnValue(
         pi_uint64{Device->ZeDeviceComputeProperties->maxTotalGroupSize});
   }

sycl/plugins/unified_runtime/pi2ur.hpp

@@ -15,13 +15,7 @@
 
 // Map of UR error codes to PI error codes
 static pi_result ur2piResult(zer_result_t urResult) {
-
-  // TODO: replace "global lifetime" objects with a non-trivial d'tor with
-  // either pointers to such objects (which would be allocated and dealocated
-  // during init and teardown) or objects with trivial d'tor.
-  // E.g. for this case we could have an std::array with sorted values.
-  //
-  static std::unordered_map<zer_result_t, pi_result> ErrorMapping = {
+  std::unordered_map<zer_result_t, pi_result> ErrorMapping = {
       {ZER_RESULT_SUCCESS, PI_SUCCESS},
       {ZER_RESULT_ERROR_UNKNOWN, PI_ERROR_UNKNOWN},
       {ZER_RESULT_ERROR_DEVICE_LOST, PI_ERROR_DEVICE_NOT_FOUND},
@@ -50,6 +44,24 @@ static pi_result ur2piResult(zer_result_t urResult) {
   if (auto Result = urCall)                                                    \
     return ur2piResult(Result);
 
+// A version of return helper that returns pi_result and not zer_result_t
+class ReturnHelper : public UrReturnHelper {
+public:
+  using UrReturnHelper::UrReturnHelper;
+
+  template <class T> pi_result operator()(const T &t) {
+    return ur2piResult(UrReturnHelper::operator()(t));
+  }
+  // Array return value
+  template <class T> pi_result operator()(const T *t, size_t s) {
+    return ur2piResult(UrReturnHelper::operator()(t, s));
+  }
+  // Array return value where element type is differrent from T
+  template <class RetType, class T> pi_result operator()(const T *t, size_t s) {
+    return ur2piResult(UrReturnHelper::operator()<RetType>(t, s));
+  }
+};
+
 namespace pi2ur {
 inline pi_result piPlatformsGet(pi_uint32 num_entries, pi_platform *platforms,
                                 pi_uint32 *num_platforms) {
@@ -66,14 +78,31 @@ inline pi_result piPlatformsGet(pi_uint32 num_entries, pi_platform *platforms,
 }
 
 inline pi_result piPlatformGetInfo(pi_platform platform,
-                                   pi_platform_info param_name,
-                                   size_t param_value_size, void *param_value,
-                                   size_t *param_value_size_ret) {
-  (void)platform;
-  (void)param_name;
-  (void)param_value_size;
-  (void)param_value;
-  (void)param_value_size_ret;
-  die("Unified Runtime: piPlatformGetInfo is not implemented");
+                                   pi_platform_info ParamName,
+                                   size_t ParamValueSize, void *ParamValue,
+                                   size_t *ParamValueSizeRet) {
+
+  static std::unordered_map<pi_platform_info, zer_platform_info_t> InfoMapping =
+      {
+          {PI_PLATFORM_INFO_EXTENSIONS, ZER_PLATFORM_INFO_NAME},
+          {PI_PLATFORM_INFO_NAME, ZER_PLATFORM_INFO_NAME},
+          {PI_PLATFORM_INFO_PROFILE, ZER_PLATFORM_INFO_PROFILE},
+          {PI_PLATFORM_INFO_VENDOR, ZER_PLATFORM_INFO_VENDOR_NAME},
+          {PI_PLATFORM_INFO_VERSION, ZER_PLATFORM_INFO_VERSION},
+      };
+
+  auto InfoType = InfoMapping.find(ParamName);
+  if (InfoType == InfoMapping.end()) {
+    return PI_ERROR_UNKNOWN;
+  }
+
+  size_t SizeInOut = ParamValueSize;
+  auto hPlatform = reinterpret_cast<zer_platform_handle_t>(platform);
+  HANDLE_ERRORS(
+      zerPlatformGetInfo(hPlatform, InfoType->second, &SizeInOut, ParamValue));
+  if (ParamValueSizeRet) {
+    *ParamValueSizeRet = SizeInOut;
+  }
+  return PI_SUCCESS;
 }
 } // namespace pi2ur

sycl/plugins/unified_runtime/pi_unified_runtime.cpp

@@ -6,12 +6,79 @@
 //
 //===------------------------------------------------------------------===//
 
+#include <cstring>
+
 #include <pi2ur.hpp>
+#include <pi_unified_runtime.hpp>
+
+// Stub function to where all not yet supported PI API are bound
+static void DieUnsupported() {
+  die("Unified Runtime: functionality is not supported");
+}
 
+// All PI API interfaces are C interfaces
 extern "C" {
 __SYCL_EXPORT pi_result piPlatformsGet(pi_uint32 num_entries,
                                        pi_platform *platforms,
                                        pi_uint32 *num_platforms) {
   return pi2ur::piPlatformsGet(num_entries, platforms, num_platforms);
 }
+
+__SYCL_EXPORT pi_result piPlatformGetInfo(pi_platform Platform,
+                                          pi_platform_info ParamName,
+                                          size_t ParamValueSize,
+                                          void *ParamValue,
+                                          size_t *ParamValueSizeRet) {
+  return pi2ur::piPlatformGetInfo(Platform, ParamName, ParamValueSize,
+                                  ParamValue, ParamValueSizeRet);
+}
+
+__SYCL_EXPORT pi_result piDevicesGet(pi_platform Platform,
+                                     pi_device_type DeviceType,
+                                     pi_uint32 NumEntries, pi_device *Devices,
+                                     pi_uint32 *NumDevices) {
+  // Report no devices, stab to have a minimal SYCL test running
+  if (NumDevices) {
+    *NumDevices = 0;
+  }
+  return PI_SUCCESS;
+}
+
+// This interface is not in Unified Runtime currently
+__SYCL_EXPORT pi_result piTearDown(void *) { return PI_SUCCESS; }
+
+// This interface is not in Unified Runtime currently
+__SYCL_EXPORT pi_result piPluginInit(pi_plugin *PluginInit) {
+  PI_ASSERT(PluginInit, PI_ERROR_INVALID_VALUE);
+
+  const char SupportedVersion[] = _PI_UNIFIED_RUNTIME_PLUGIN_VERSION_STRING;
+
+  // Check that the major version matches in PiVersion and SupportedVersion
+  _PI_PLUGIN_VERSION_CHECK(PluginInit->PiVersion, SupportedVersion);
+
+  // TODO: handle versioning/targets properly.
+  size_t PluginVersionSize = sizeof(PluginInit->PluginVersion);
+
+  PI_ASSERT(strlen(_PI_UNIFIED_RUNTIME_PLUGIN_VERSION_STRING) <
+                PluginVersionSize,
+            PI_ERROR_INVALID_VALUE);
+
+  strncpy(PluginInit->PluginVersion, SupportedVersion, PluginVersionSize);
+
+  // Bind interfaces that are already supported and "die" for unsupported ones
+#define _PI_API(api)                                                           \
+  (PluginInit->PiFunctionTable).api = (decltype(&::api))(&DieUnsupported);
+#include <sycl/detail/pi.def>
+
+#define _PI_API(api)                                                           \
+  (PluginInit->PiFunctionTable).api = (decltype(&::api))(&api);
+
+  _PI_API(piPlatformsGet)
+  _PI_API(piPlatformGetInfo)
+  _PI_API(piDevicesGet)
+  _PI_API(piTearDown)
+
+  return PI_SUCCESS;
+}
+
 } // extern "C