[UR] Check for memory alignment before calling clEnqueueMemFillINTEL_fn in OpenCL (#18423)

Fixes oneapi-src/unified-runtime#2440

[clEnqueueMemFillINTEL returns CL_INVALID_VALUE if dst_ptr is NULL, or
if dst_ptr is not aligned to pattern_size bytes.
](https://registry.khronos.org/OpenCL/extensions/intel/cl_intel_unified_shared_memory.html)

This PR adds a memory alignment check before calling
`clEnqueueMemFillINTEL` to ensure we can safely call it, otherwise take
the host side copy path.

Re-enables the `USM/fill_any_size.cpp` and also moves
`checkUSMImplAlignment` from Hip into a common header so it can be used
in both Hip and OpenCL.

Author: martygrant

sycl/test-e2e/USM/fill_any_size.cpp

@@ -1,8 +1,6 @@
 // RUN: %{build} -o %t1.out
 // RUN: %{run} %t1.out
 // clang-format off
-// UNSUPPORTED: opencl
-// UNSUPPORTED-TRACKER: https://github.com/oneapi-src/unified-runtime/issues/2440
 // clang-format on
 
 /**

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

@@ -144,7 +144,7 @@ ur_result_t USMDeviceAllocImpl(void **ResultPtr, ur_context_handle_t,
     return Err;
   }
 
-  assert(checkUSMImplAlignment(Alignment, ResultPtr));
+  assert(isPointerAlignedTo(Alignment, *ResultPtr));
   return UR_RESULT_SUCCESS;
 }
 
@@ -160,7 +160,7 @@ ur_result_t USMSharedAllocImpl(void **ResultPtr, ur_context_handle_t,
     return Err;
   }
 
-  assert(checkUSMImplAlignment(Alignment, ResultPtr));
+  assert(isPointerAlignedTo(Alignment, *ResultPtr));
   return UR_RESULT_SUCCESS;
 }
 
@@ -174,7 +174,7 @@ ur_result_t USMHostAllocImpl(void **ResultPtr,
     return Err;
   }
 
-  assert(checkUSMImplAlignment(Alignment, ResultPtr));
+  assert(isPointerAlignedTo(Alignment, *ResultPtr));
   return UR_RESULT_SUCCESS;
 }
 
@@ -485,11 +485,6 @@ bool checkUSMAlignment(uint32_t &alignment, const ur_usm_desc_t *pUSMDesc) {
           (alignment == 0 || ((alignment & (alignment - 1)) == 0)));
 }
 
-bool checkUSMImplAlignment(uint32_t Alignment, void **ResultPtr) {
-  return Alignment == 0 ||
-         reinterpret_cast<std::uintptr_t>(*ResultPtr) % Alignment == 0;
-}
-
 UR_APIEXPORT ur_result_t UR_APICALL urUSMPoolCreateExp(ur_context_handle_t,
                                                        ur_device_handle_t,
                                                        ur_usm_pool_desc_t *,

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

@@ -120,5 +120,3 @@ ur_result_t USMHostAllocImpl(void **ResultPtr, ur_context_handle_t Context,
                              uint32_t Alignment);
 
 bool checkUSMAlignment(uint32_t &alignment, const ur_usm_desc_t *pUSMDesc);
-
-bool checkUSMImplAlignment(uint32_t Alignment, void **ResultPtr);

unified-runtime/source/adapters/opencl/usm.cpp

@@ -269,7 +269,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueUSMFill(
   // Have to look up the context from the kernel
   cl_context CLContext = hQueue->Context->CLContext;
 
-  if (patternSize <= 128 && isPowerOf2(patternSize)) {
+  if (patternSize <= 128 && isPowerOf2(patternSize) &&
+      isPointerAlignedTo(patternSize, ptr)) {
     clEnqueueMemFillINTEL_fn EnqueueMemFill = nullptr;
     UR_RETURN_ON_FAILURE(
         cl_ext::getExtFuncFromContext<clEnqueueMemFillINTEL_fn>(

unified-runtime/source/common/ur_util.hpp

@@ -548,4 +548,9 @@ static inline std::string groupDigits(Numeric numeric) {
 
 template <typename T> Spinlock<Rc<T>> AtomicSingleton<T>::instance;
 
+inline bool isPointerAlignedTo(uint32_t Alignment, void *Ptr) {
+  return Alignment == 0 ||
+         reinterpret_cast<std::uintptr_t>(Ptr) % Alignment == 0;
+}
+
 #endif /* UR_UTIL_H */