Merge pull request #2259 from nrspruit/immediate_dg2_or_newer_default

[L0] Enable Immediate Command Lists by default given Intel GPU DG2 or newer

Author: pbalcer

source/adapters/level_zero/device.cpp

@@ -1053,9 +1053,7 @@ ur_result_t urDeviceGetInfo(
   case UR_DEVICE_INFO_COMMAND_BUFFER_EVENT_SUPPORT_EXP:
     return ReturnValue(false);
   case UR_DEVICE_INFO_BINDLESS_IMAGES_SUPPORT_EXP: {
-    bool DeviceIsDG2OrNewer =
-        Device->ZeDeviceIpVersionExt->ipVersion >= 0x030dc000;
-    return ReturnValue(DeviceIsDG2OrNewer &&
+    return ReturnValue(Device->isIntelDG2OrNewer() &&
                        Device->ZeDeviceImageProperties->maxImageDims1D > 0 &&
                        Device->ZeDeviceImageProperties->maxImageDims2D > 0 &&
                        Device->ZeDeviceImageProperties->maxImageDims3D > 0);
@@ -1065,15 +1063,11 @@ ur_result_t urDeviceGetInfo(
     return ReturnValue(false);
   }
   case UR_DEVICE_INFO_BINDLESS_IMAGES_1D_USM_SUPPORT_EXP: {
-    bool DeviceIsDG2OrNewer =
-        Device->ZeDeviceIpVersionExt->ipVersion >= 0x030dc000;
-    return ReturnValue(DeviceIsDG2OrNewer &&
+    return ReturnValue(Device->isIntelDG2OrNewer() &&
                        Device->ZeDeviceImageProperties->maxImageDims1D > 0);
   }
   case UR_DEVICE_INFO_BINDLESS_IMAGES_2D_USM_SUPPORT_EXP: {
-    bool DeviceIsDG2OrNewer =
-        Device->ZeDeviceIpVersionExt->ipVersion >= 0x030dc000;
-    return ReturnValue(DeviceIsDG2OrNewer &&
+    return ReturnValue(Device->isIntelDG2OrNewer() &&
                        Device->ZeDeviceImageProperties->maxImageDims2D > 0);
   }
   case UR_DEVICE_INFO_IMAGE_PITCH_ALIGN_EXP:
@@ -1409,13 +1403,35 @@ ur_result_t urDeviceRelease(ur_device_handle_t Device) {
 }
 } // namespace ur::level_zero
 
-// Whether immediate commandlists will be used for kernel launches and copies.
-// The default is standard commandlists. Setting 1 or 2 specifies use of
-// immediate commandlists. Note: when immediate commandlists are used then
-// device-only events must be either AllHostVisible or OnDemandHostVisibleProxy.
-// (See env var UR_L0_DEVICE_SCOPE_EVENTS).
-
-// Get value of immediate commandlists env var setting or -1 if unset
+/**
+ * @brief Determines the mode of immediate command lists to be used.
+ *
+ * This function checks environment variables and device properties to decide
+ * the mode of immediate command lists. The mode can be influenced by the
+ * following environment variables:
+ * - `UR_L0_USE_IMMEDIATE_COMMANDLISTS`
+ * - `SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS`
+ *
+ * If neither environment variable is set, the function defaults to using the
+ * device's properties to determine the mode.
+ *
+ * @return The mode of immediate command lists, which can be one of the
+ * following:
+ * - `NotUsed`: Immediate command lists are not used.
+ * - `PerQueue`: Immediate command lists are used per queue.
+ * - `PerThreadPerQueue`: Immediate command lists are used per thread per queue.
+ *
+ * The decision process is as follows:
+ * 1. If the environment variables are not set, the function checks if the
+ * device is Intel DG2 or newer and if the driver version is supported. If both
+ *    conditions are met, or if the device is PVC, it returns `PerQueue`.
+ *    Otherwise, it returns `NotUsed`.
+ * 2. If the environment variable is set, it returns the corresponding mode:
+ *    - `0`: `NotUsed`
+ *    - `1`: `PerQueue`
+ *    - `2`: `PerThreadPerQueue`
+ *    - Any other value: `NotUsed`
+ */
 ur_device_handle_t_::ImmCmdlistMode
 ur_device_handle_t_::useImmediateCommandLists() {
   // If immediate commandlist setting is not explicitly set, then use the device
@@ -1433,9 +1449,10 @@ ur_device_handle_t_::useImmediateCommandLists() {
   }();
 
   if (ImmediateCommandlistsSetting == -1) {
+    bool isDG2OrNewer = this->isIntelDG2OrNewer();
     bool isDG2SupportedDriver =
         this->Platform->isDriverVersionNewerOrSimilar(1, 5, 30820);
-    if ((isDG2SupportedDriver && isDG2()) || isPVC()) {
+    if ((isDG2SupportedDriver && isDG2OrNewer) || isPVC()) {
       return PerQueue;
     } else {
       return NotUsed;

source/adapters/level_zero/device.hpp

@@ -180,10 +180,6 @@ struct ur_device_handle_t_ : _ur_object {
 
   bool isSubDevice() { return RootDevice != nullptr; }
 
-  // Is this a Data Center GPU Max series (aka PVC)?
-  // TODO: change to use
-  // https://spec.oneapi.io/level-zero/latest/core/api.html#ze-device-ip-version-ext-t
-  // when that is stable.
   bool isPVC() {
     return (ZeDeviceProperties->deviceId & 0xff0) == 0xbd0 ||
            (ZeDeviceProperties->deviceId & 0xff0) == 0xb60;
@@ -192,6 +188,11 @@ struct ur_device_handle_t_ : _ur_object {
   // Checks if this GPU is an Intel Flex GPU or Intel Arc Alchemist
   bool isDG2() { return (ZeDeviceProperties->deviceId & 0xff00) == 0x5600; }
 
+  bool isIntelDG2OrNewer() {
+    return (ZeDeviceProperties->vendorId == 0x8086 &&
+            ZeDeviceIpVersionExt->ipVersion >= 0x030dc000);
+  }
+
   bool isIntegrated() {
     return (ZeDeviceProperties->flags & ZE_DEVICE_PROPERTY_FLAG_INTEGRATED);
   }