Merge branch 'main' into steffen/make_ext_func_fail_unsupported

Author: steffenlarsen

source/adapters/hip/device.cpp

@@ -539,8 +539,20 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
     // native asserts are in progress
     std::string SupportedExtensions = "";
     SupportedExtensions += "pi_ext_intel_devicelib_assert ";
-    // Return supported for the UR command-buffer experimental feature
-    SupportedExtensions += "ur_exp_command_buffer ";
+
+    int RuntimeVersion = 0;
+    UR_CHECK_ERROR(hipRuntimeGetVersion(&RuntimeVersion));
+
+    // Return supported for the UR command-buffer experimental feature on
+    // ROCM 5.5.1 and later. This is to workaround HIP driver bug
+    // https://github.com/ROCm/HIP/issues/2450 in older versions.
+    //
+    // The version is returned as (10000000 major + 1000000 minor + patch).
+    const int CmdBufDriverMinVersion = 50530202; // ROCM 5.5.1
+    if (RuntimeVersion >= CmdBufDriverMinVersion) {
+      SupportedExtensions += "ur_exp_command_buffer ";
+    }
+
     SupportedExtensions += " ";
 
     hipDeviceProp_t Props;
@@ -844,9 +856,18 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
     return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
 
   case UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP:
-  case UR_DEVICE_INFO_COMMAND_BUFFER_UPDATE_SUPPORT_EXP:
-    return ReturnValue(true);
+  case UR_DEVICE_INFO_COMMAND_BUFFER_UPDATE_SUPPORT_EXP: {
+    int DriverVersion = 0;
+    UR_CHECK_ERROR(hipDriverGetVersion(&DriverVersion));
 
+    // Return supported for the UR command-buffer experimental feature on
+    // ROCM 5.5.1 and later. This is to workaround HIP driver bug
+    // https://github.com/ROCm/HIP/issues/2450 in older versions.
+    //
+    // The version is returned as (10000000 major + 1000000 minor + patch).
+    const int CmdBufDriverMinVersion = 50530202; // ROCM 5.5.1
+    return ReturnValue(DriverVersion >= CmdBufDriverMinVersion);
+  }
   default:
     break;
   }

source/adapters/level_zero/device.cpp

@@ -1074,7 +1074,7 @@ bool ur_device_handle_t_::useDriverInOrderLists() {
   static const bool UseDriverInOrderLists = [] {
     const char *UrRet = std::getenv("UR_L0_USE_DRIVER_INORDER_LISTS");
     if (!UrRet)
-      return false;
+      return true;
     return std::atoi(UrRet) != 0;
   }();
 

source/adapters/level_zero/memory.cpp

@@ -1833,6 +1833,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urMemBufferCreateWithNativeHandle(
     UR_CALL(
         Buffer->getZeHandle(ZeHandleDst, ur_mem_handle_t_::write_only, Device));
 
+    // Indicate that this buffer has the device buffer mapped to a native buffer
+    // and track the native pointer such that the memory is synced later at
+    // memory free.
+    Buffer->DeviceMappedHostNativePtr = Ptr;
     // zeCommandListAppendMemoryCopy must not be called from simultaneous
     // threads with the same command list handle, so we need exclusive lock.
     std::scoped_lock<ur_mutex> Lock(Context->ImmediateCommandListMutex);
@@ -2169,6 +2173,17 @@ ur_result_t _ur_buffer::free() {
                                                  ? Plt->ContextsMutex
                                                  : UrContext->Mutex);
 
+      // If this memory was allocated as a proxy device buffer, then we must
+      // copy the final memory contents back to the original native pointer
+      // before releasing the buffer memory.
+      if (DeviceMappedHostNativePtr != nullptr) {
+        // zeCommandListAppendMemoryCopy must not be called from simultaneous
+        // threads with the same command list handle, so we need exclusive lock.
+        std::scoped_lock<ur_mutex> Lock(UrContext->ImmediateCommandListMutex);
+        ZE2UR_CALL(zeCommandListAppendMemoryCopy,
+                   (UrContext->ZeCommandListInit, DeviceMappedHostNativePtr,
+                    ZeHandle, Size, nullptr, 0, nullptr));
+      }
       UR_CALL(USMFreeHelper(reinterpret_cast<ur_context_handle_t>(UrContext),
                             ZeHandle));
       break;

source/adapters/level_zero/memory.hpp

@@ -158,6 +158,10 @@ struct _ur_buffer final : ur_mem_handle_t_ {
   // Tells the host allocation to use for buffer map operations.
   char *MapHostPtr{nullptr};
 
+  // Pointer to the original native buffer handle given this memory is a proxy
+  // device buffer.
+  void *DeviceMappedHostNativePtr{nullptr};
+
   // Supplementary data to keep track of the mappings of this buffer
   // created with piEnqueueMemBufferMap.
   struct Mapping {