[UR][CUDA][HIP] Use bool to check max local size

Author: npmiller

unified-runtime/source/adapters/cuda/device.cpp

@@ -505,7 +505,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
     // OpenCL's "local memory" maps most closely to CUDA's "shared memory".
     // CUDA has its own definition of "local memory", which maps to OpenCL's
     // "private memory".
-    if (hDevice->getMaxChosenLocalMem()) {
+    if (hDevice->hasMaxChosenLocalMem()) {
       return ReturnValue(
           static_cast<uint64_t>(hDevice->getMaxChosenLocalMem()));
     } else {

unified-runtime/source/adapters/cuda/device.hpp

@@ -34,6 +34,7 @@ struct ur_device_handle_t_ : ur::cuda::handle_base {
   int MaxRegsPerBlock{0};
   int MaxCapacityLocalMem{0};
   int MaxChosenLocalMem{0};
+  bool HasMaxChosenLocalMem{false};
   uint32_t NumComputeUnits{0};
   std::once_flag NVMLInitFlag;
   std::optional<nvmlDevice_t> NVMLDevice;
@@ -73,7 +74,8 @@ struct ur_device_handle_t_ : ur::cuda::handle_base {
 
     if (LocalMemSizePtr) {
       MaxChosenLocalMem = std::atoi(LocalMemSizePtr);
-      if (MaxChosenLocalMem <= 0) {
+      HasMaxChosenLocalMem = true;
+      if (MaxChosenLocalMem < 0) {
         setErrorMessage(LocalMemSizePtrUR ? "Invalid value specified for "
                                             "UR_CUDA_MAX_LOCAL_MEM_SIZE"
                                           : "Invalid value specified for "
@@ -157,6 +159,8 @@ struct ur_device_handle_t_ : ur::cuda::handle_base {
 
   int getMaxChosenLocalMem() const noexcept { return MaxChosenLocalMem; };
 
+  int hasMaxChosenLocalMem() const noexcept { return HasMaxChosenLocalMem; };
+
   uint32_t getNumComputeUnits() const noexcept { return NumComputeUnits; };
 
   // bookkeeping for mipmappedArray leaks in Mapping external Memory

unified-runtime/source/adapters/hip/device.cpp

@@ -382,7 +382,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
     // OpenCL's "local memory" maps most closely to HIP's "shared memory".
     // HIP has its own definition of "local memory", which maps to OpenCL's
     // "private memory".
-    if (hDevice->getMaxChosenLocalMem()) {
+    if (hDevice->hasMaxChosenLocalMem()) {
       return ReturnValue(
           static_cast<uint64_t>(hDevice->getMaxChosenLocalMem()));
     } else {

unified-runtime/source/adapters/hip/device.hpp

@@ -33,6 +33,7 @@ struct ur_device_handle_t_ : ur::hip::handle_base {
   int MaxBlockDimZ{0};
   int DeviceMaxLocalMem{0};
   int MaxChosenLocalMem{0};
+  bool HasMaxChosenLocalMem{false};
   int ManagedMemSupport{0};
   int ConcurrentManagedAccess{0};
   bool HardwareImageSupport{false};
@@ -77,7 +78,8 @@ struct ur_device_handle_t_ : ur::hip::handle_base {
 
     if (LocalMemSzPtr) {
       MaxChosenLocalMem = std::atoi(LocalMemSzPtr);
-      if (MaxChosenLocalMem <= 0 || MaxChosenLocalMem > DeviceMaxLocalMem) {
+      HasMaxChosenLocalMem = true;
+      if (MaxChosenLocalMem < 0 || MaxChosenLocalMem > DeviceMaxLocalMem) {
         setErrorMessage(LocalMemSzPtrUR ? "Invalid value specified for "
                                           "UR_HIP_MAX_LOCAL_MEM_SIZE"
                                         : "Invalid value specified for "
@@ -114,6 +116,8 @@ struct ur_device_handle_t_ : ur::hip::handle_base {
 
   int getMaxChosenLocalMem() const noexcept { return MaxChosenLocalMem; };
 
+  int hasMaxChosenLocalMem() const noexcept { return HasMaxChosenLocalMem; };
+
   int getManagedMemSupport() const noexcept { return ManagedMemSupport; };
 
   int getConcurrentManagedAccess() const noexcept {

unified-runtime/source/adapters/hip/enqueue.cpp

@@ -1818,7 +1818,7 @@ setKernelParams(const ur_device_handle_t Device, const uint32_t WorkDim,
       Kernel->setImplicitOffsetArg(sizeof(ImplicitOffset), ImplicitOffset);
     }
 
-    if (Device->getMaxChosenLocalMem()) {
+    if (Device->hasMaxChosenLocalMem()) {
       UR_CHECK_ERROR(hipFuncSetAttribute(
           HIPFunc, hipFuncAttributeMaxDynamicSharedMemorySize,
           Device->getMaxChosenLocalMem()));