Change env var to UR_L0_USE_OPTIMIZED_32BIT_ACCESS

This is closer to the actual functionality being selected, and also
aligns with having large allocations by default.

Signed-off-by: Jaime Arteaga <[email protected]>

Author: Jaime Arteaga

source/adapters/level_zero/device.cpp

@@ -267,9 +267,9 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(
     return ReturnValue(uint32_t{64});
   }
   case UR_DEVICE_INFO_MAX_MEM_ALLOC_SIZE:
-    // if using large allocations, then return total size in the device.
-    // if not, then return L0's maxMemAllocSize.
-    if (Device->useLargeAllocations()) {
+    // if not optimized for 32-bit access, return total memory size.
+    // otherwise, return only maximum allocatable size.
+    if (Device->useOptimized32bitAccess() == 0) {
       return ReturnValue(uint64_t{getGlobalMemSize(Device)});
     } else {
       return ReturnValue(uint64_t{Device->ZeDeviceProperties->maxMemAllocSize});
@@ -910,15 +910,20 @@ ur_device_handle_t_::useImmediateCommandLists() {
   }
 }
 
-bool ur_device_handle_t_::useLargeAllocations() {
-  static const bool UseLargeAllocations = [this] {
-    const char *UrRet = std::getenv("UR_L0_ALLOW_LARGE_ALLOCATIONS");
+int32_t ur_device_handle_t_::useOptimized32bitAccess() {
+  static const int32_t Optimize32bitAccessMode = [this] {
+    // If device is Intel(R) Data Center GPU Max,
+    // use default provided by L0 driver.
+    // TODO: Use IP versioning to select based on range of devices
+    if (this->isPVC())
+      return -1;
+    const char *UrRet = std::getenv("UR_L0_USE_OPTIMIZED_32BIT_ACCESS");
     if (!UrRet)
-      return (this->isPVC() ? true : false);
-    return std::atoi(UrRet) != 0;
+      return 0;
+    return std::atoi(UrRet);
   }();
 
-  return UseLargeAllocations;
+  return Optimize32bitAccessMode;
 }
 
 ur_result_t ur_device_handle_t_::initialize(int SubSubDeviceOrdinal,

source/adapters/level_zero/device.hpp

@@ -155,7 +155,7 @@ struct ur_device_handle_t_ : _ur_object {
   // provide support for only one, like for Intel(R)
   // Data Center GPU Max, for which L0 driver only
   // supports stateless.
-  bool useLargeAllocations();
+  int32_t useOptimized32bitAccess();
 
   bool isSubDevice() { return RootDevice != nullptr; }
 

source/adapters/level_zero/program.cpp

@@ -161,7 +161,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urProgramBuildExp(
     ZeBuildOptions += pOptions;
   }
 
-  if (phDevices[0]->useLargeAllocations()) {
+  if (phDevices[0]->useOptimized32bitAccess() == 0) {
     ZeBuildOptions += " -ze-opt-greater-than-4GB-buffer-required";
   }
 
@@ -256,7 +256,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urProgramCompile(
     // ze-opt-greater-than-4GB-buffer-required to disable
     // stateful optimizations and be able to use larger than
     // 4GB allocations on these kernels.
-    if (Context->Devices[0]->useLargeAllocations()) {
+    if (Context->Devices[0]->useOptimized32bitAccess() == 0) {
       Program->BuildFlags += " -ze-opt-greater-than-4GB-buffer-required";
     }
   }

source/adapters/level_zero/usm.cpp

@@ -178,11 +178,11 @@ static ur_result_t USMDeviceAllocImpl(void **ResultPtr,
   ZeDesc.flags = 0;
   ZeDesc.ordinal = 0;
 
-  ZeStruct<ze_relaxed_allocation_limits_exp_desc_t> RelaxedDesc;
-  if (Device->useLargeAllocations() &&
+  if (Device->useOptimized32bitAccess() == 0 &&
       (Size > Device->ZeDeviceProperties->maxMemAllocSize)) {
     // Tell Level-Zero to accept Size > maxMemAllocSize if
     // large allocations are used.
+    ZeStruct<ze_relaxed_allocation_limits_exp_desc_t> RelaxedDesc;
     RelaxedDesc.flags = ZE_RELAXED_ALLOCATION_LIMITS_EXP_FLAG_MAX_SIZE;
     ZeDesc.pNext = &RelaxedDesc;
   }