[SYCL] Add forward declarations for Subgroup Shuffle functions (#8450)

The main reason for the change is being able to call Generic overloading
of Shuffle from Vector overloading of Shuffle.

Author: maksimsab

sycl/include/sycl/detail/spirv.hpp

@@ -537,6 +537,49 @@ using EnableIfVectorShuffle =
     std::enable_if_t<detail::is_vector_arithmetic<T>::value, T>;
 #endif // ifndef __NVPTX__
 
+// Bitcast shuffles can be implemented using a single SubgroupShuffle
+// intrinsic, but require type-punning via an appropriate integer type
+#ifndef __NVPTX__
+template <typename T>
+using EnableIfBitcastShuffle =
+    std::enable_if_t<!detail::is_arithmetic<T>::value &&
+                         (std::is_trivially_copyable_v<T> &&
+                          (sizeof(T) == 1 || sizeof(T) == 2 || sizeof(T) == 4 ||
+                           sizeof(T) == 8)),
+                     T>;
+#else
+template <typename T>
+using EnableIfBitcastShuffle =
+    std::enable_if_t<!(std::is_integral_v<T> &&
+                       (sizeof(T) <= sizeof(int32_t))) &&
+                         !detail::is_vector_arithmetic<T>::value &&
+                         (std::is_trivially_copyable_v<T> &&
+                          (sizeof(T) == 1 || sizeof(T) == 2 || sizeof(T) == 4)),
+                     T>;
+#endif // ifndef __NVPTX__
+
+// Generic shuffles may require multiple calls to SubgroupShuffle
+// intrinsics, and should use the fewest shuffles possible:
+// - Loop over 64-bit chunks until remaining bytes < 64-bit
+// - At most one 32-bit, 16-bit and 8-bit chunk left over
+#ifndef __NVPTX__
+template <typename T>
+using EnableIfGenericShuffle =
+    std::enable_if_t<!detail::is_arithmetic<T>::value &&
+                         !(std::is_trivially_copyable_v<T> &&
+                           (sizeof(T) == 1 || sizeof(T) == 2 ||
+                            sizeof(T) == 4 || sizeof(T) == 8)),
+                     T>;
+#else
+template <typename T>
+using EnableIfGenericShuffle = std::enable_if_t<
+    !(std::is_integral<T>::value && (sizeof(T) <= sizeof(int32_t))) &&
+        !detail::is_vector_arithmetic<T>::value &&
+        !(std::is_trivially_copyable_v<T> &&
+          (sizeof(T) == 1 || sizeof(T) == 2 || sizeof(T) == 4)),
+    T>;
+#endif
+
 #ifdef __NVPTX__
 inline uint32_t membermask() {
   // use a full mask as sync operations are required to be convergent and exited
@@ -545,6 +588,31 @@ inline uint32_t membermask() {
 }
 #endif
 
+// Forward declarations for template overloadings
+template <typename T>
+EnableIfBitcastShuffle<T> SubgroupShuffle(T x, id<1> local_id);
+
+template <typename T>
+EnableIfBitcastShuffle<T> SubgroupShuffleXor(T x, id<1> local_id);
+
+template <typename T>
+EnableIfBitcastShuffle<T> SubgroupShuffleDown(T x, id<1> local_id);
+
+template <typename T>
+EnableIfBitcastShuffle<T> SubgroupShuffleUp(T x, id<1> local_id);
+
+template <typename T>
+EnableIfGenericShuffle<T> SubgroupShuffle(T x, id<1> local_id);
+
+template <typename T>
+EnableIfGenericShuffle<T> SubgroupShuffleXor(T x, id<1> local_id);
+
+template <typename T>
+EnableIfGenericShuffle<T> SubgroupShuffleDown(T x, id<1> local_id);
+
+template <typename T>
+EnableIfGenericShuffle<T> SubgroupShuffleUp(T x, id<1> local_id);
+
 template <typename T>
 EnableIfNativeShuffle<T> SubgroupShuffle(T x, id<1> local_id) {
 #ifndef __NVPTX__
@@ -623,26 +691,6 @@ EnableIfVectorShuffle<T> SubgroupShuffleUp(T x, uint32_t delta) {
   return result;
 }
 
-// Bitcast shuffles can be implemented using a single SubgroupShuffle
-// intrinsic, but require type-punning via an appropriate integer type
-#ifndef __NVPTX__
-template <typename T>
-using EnableIfBitcastShuffle =
-    detail::enable_if_t<!detail::is_arithmetic<T>::value &&
-                            (std::is_trivially_copyable<T>::value &&
-                             (sizeof(T) == 1 || sizeof(T) == 2 ||
-                              sizeof(T) == 4 || sizeof(T) == 8)),
-                        T>;
-#else
-template <typename T>
-using EnableIfBitcastShuffle = detail::enable_if_t<
-    !(std::is_integral<T>::value && (sizeof(T) <= sizeof(int32_t))) &&
-        !detail::is_vector_arithmetic<T>::value &&
-        (std::is_trivially_copyable<T>::value &&
-         (sizeof(T) == 1 || sizeof(T) == 2 || sizeof(T) == 4)),
-    T>;
-#endif
-
 template <typename T>
 using ConvertToNativeShuffleType_t = select_cl_scalar_integral_unsigned_t<T>;
 
@@ -699,28 +747,6 @@ EnableIfBitcastShuffle<T> SubgroupShuffleUp(T x, uint32_t delta) {
   return bit_cast<T>(Result);
 }
 
-// Generic shuffles may require multiple calls to SubgroupShuffle
-// intrinsics, and should use the fewest shuffles possible:
-// - Loop over 64-bit chunks until remaining bytes < 64-bit
-// - At most one 32-bit, 16-bit and 8-bit chunk left over
-#ifndef __NVPTX__
-template <typename T>
-using EnableIfGenericShuffle =
-    detail::enable_if_t<!detail::is_arithmetic<T>::value &&
-                            !(std::is_trivially_copyable<T>::value &&
-                              (sizeof(T) == 1 || sizeof(T) == 2 ||
-                               sizeof(T) == 4 || sizeof(T) == 8)),
-                        T>;
-#else
-template <typename T>
-using EnableIfGenericShuffle = detail::enable_if_t<
-    !(std::is_integral<T>::value && (sizeof(T) <= sizeof(int32_t))) &&
-        !detail::is_vector_arithmetic<T>::value &&
-        !(std::is_trivially_copyable<T>::value &&
-          (sizeof(T) == 1 || sizeof(T) == 2 || sizeof(T) == 4)),
-    T>;
-#endif
-
 template <typename T>
 EnableIfGenericShuffle<T> SubgroupShuffle(T x, id<1> local_id) {
   T Result;