Work around compiler vectorization issue

With uint8_t types, the icpx compiler fails to vectorize even when calling begin() on our range within a kernel to pull out a raw pointer. To work around this issue, begin() needs to be called on the host and passed to the kernel Signed-off-by: Matthew Michel <[email protected]>
uxlfoundation · Jan 6, 2025 · 1336735 · 1336735
1 parent 32612a1
commit 1336735
Show file tree

Hide file tree

Showing 2 changed files with 151 additions and 165 deletions.
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_for.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_for.h
@@ -192,8 +192,20 @@ __parallel_for(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&&
     {
         if (__count >= __large_submitter::__estimate_best_start_size(__exec, __brick))
         {
-            return __large_submitter{}(std::forward<_ExecutionPolicy>(__exec), __brick, __count,
-                                       std::forward<_Ranges>(__rngs)...);
+            // Passing begin() of each range is needed for the icpx compiler to vectorize. The indirection introduced
+            // by our all / guard views interfere with compiler vectorization. At this point, we have ensured that
+            // input is contiguous and can be operated on directly. The begin() function for these views will return a
+            // pointer which is passed to the kernel.
+            if constexpr (_Fp::__can_vectorize)
+            {
+                return __large_submitter{}(std::forward<_ExecutionPolicy>(__exec), __brick, __count,
+                                           std::forward<_Ranges>(__rngs).begin()...);
+            }
+            else
+            {
+                return __large_submitter{}(std::forward<_ExecutionPolicy>(__exec), __brick, __count,
+                                           std::forward<_Ranges>(__rngs)...);
+            }
         }
     }
     return __small_submitter{}(std::forward<_ExecutionPolicy>(__exec), __brick, __count,