Add fastpath for loading/storing SimdN<T, 1, ...>

alpaka-group · Jan 2, 2024 · e698d3f · e698d3f
1 parent 729d1ea
commit e698d3f
Showing 1 changed file with 8 additions and 2 deletions.
diff --git a/include/llama/Simd.hpp b/include/llama/Simd.hpp
@@ -311,8 +311,11 @@ namespace llama
         // structured dstSimd type and record reference
         if constexpr(isRecordRef<Simd> && isRecordRef<T>)
         {
-            forEachLeafCoord<typename Simd::AccessibleRecordDim>([&](auto rc) LLAMA_LAMBDA_INLINE
-                                                                 { internal::loadSimdFromField(srcRef, dstSimd, rc); });
+            if constexpr(simdLanes<Simd> == simdLanes<T>) // fast path mainly for scalar SimdN<T, 1, ...>
+                dstSimd = srcRef;
+            else
+                forEachLeafCoord<typename Simd::AccessibleRecordDim>(
+                    [&](auto rc) LLAMA_LAMBDA_INLINE { internal::loadSimdFromField(srcRef, dstSimd, rc); });
         }
         // unstructured dstSimd and reference type
         else if constexpr(!isRecordRef<Simd> && !isRecordRef<T>)
@@ -339,6 +342,9 @@ namespace llama
         // structured Simd type and record reference
         if constexpr(isRecordRef<Simd> && isRecordRef<T>)
         {
+            if constexpr(simdLanes<Simd> == simdLanes<T>) // fast path mainly for scalar SimdN<T, 1, ...>
+                dstRef = srcSimd;
+            else
             forEachLeafCoord<typename T::AccessibleRecordDim>([&](auto rc) LLAMA_LAMBDA_INLINE
                                                               { internal::storeSimdToField(srcSimd, dstRef, rc); });
         }