working subgroup2 template and funcs

Author: keptsecret

examples_tests

@@ -19,27 +19,27 @@ namespace hlsl
 namespace subgroup2
 {
 
-template<typename Config, class BinOp, int32_t ItemsPerInvocation=1, class device_capabilities=void NBL_PRIMARY_REQUIRES(is_configuration_v<Config>)
+template<typename Config, class BinOp, int32_t ItemsPerInvocation=1, class device_capabilities=void NBL_PRIMARY_REQUIRES(subgroup::is_configuration_v<Config>)
 struct ArithmeticParams
 {
     using config_t = Config;
     using binop_t = BinOp;
-    using type_t = typename BinOp::type_t;
+    using scalar_t = typename BinOp::type_t;    // BinOp should be with scalar type
+    using type_t = vector<scalar_t, ItemsPerInvocation>;
 
-    // static_assert() vector_trait Dimension == ItemsPerInvocation
     NBL_CONSTEXPR_STATIC_INLINE int32_t itemsPerInvocation = ItemsPerInvocation;
     NBL_CONSTEXPR_STATIC_INLINE bool UseNativeIntrinsics = device_capabilities_traits<device_capabilities>::shaderSubgroupArithmetic /*&& /*some heuristic for when its faster*/;
 };
 
 template<typename Params>
-struct reduction : impl::reduction<typename Params::binop_t,Params::UseNativeIntrinsics> {};
+struct reduction : impl::reduction<typename Params::binop_t,typename Params::type_t,Params::UseNativeIntrinsics> {};
 template<typename Params>
-struct inclusive_scan : impl::inclusive_scan<typename Params::binop_t,Params::UseNativeIntrinsics> {};
+struct inclusive_scan : impl::inclusive_scan<typename Params::binop_t,typename Params::type_t,Params::UseNativeIntrinsics> {};
 template<typename Params>
-struct exclusive_scan : impl::exclusive_scan<typename Params::binop_t,Params::UseNativeIntrinsics> {};
+struct exclusive_scan : impl::exclusive_scan<typename Params::binop_t,typename Params::type_t,Params::UseNativeIntrinsics> {};
 
 }
 }
 }
 
-#endif
+#endif

include/nbl/builtin/hlsl/subgroup2/arithmetic_portability.hlsl

@@ -16,44 +16,46 @@ namespace subgroup2
 namespace impl
 {
 
-template<class Binop, bool native>
+template<class Binop, typename T, bool native>
 struct inclusive_scan
 {
-    using type_t = typename Binop::type_t;
-    using scalar_t = typename Binop::scalar_t;
-    using binop_t = Binop<scalar_t>;
-    using binop_par_t = Binop<type_t>;
+    using type_t = T;
+    using scalar_t = typename Binop::type_t;
+    using binop_t = Binop;
     using exclusive_scan_op_t = subgroup::impl::exclusive_scan<binop_t, native>;
 
+    NBL_CONSTEXPR_STATIC_INLINE uint32_t ItemsPerInvocation = vector_traits<T>::Dimension;
+
     type_t operator()(NBL_CONST_REF_ARG(type_t) value)
     {
         binop_t binop;
         type_t retval;
         retval[0] = value[0];
-        [unroll(ItemsPerInvocation-1)]
+        //[unroll(ItemsPerInvocation-1)]
         for (uint32_t i = 1; i < ItemsPerInvocation; i++)
             retval[i] = binop(retval[i-1], value[i]);
 
         exclusive_scan_op_t op;
         scalar_t exclusive = op(retval[ItemsPerInvocation-1]);
 
-        [unroll(ItemsPerInvocation)]
+        //[unroll(ItemsPerInvocation)]
         for (uint32_t i = 0; i < ItemsPerInvocation; i++)
             retval[i] = binop(retval[i], exclusive);
         return retval;
     }
 };
 
-template<class Binop, bool native>
+template<class Binop, typename T, bool native>
 struct exclusive_scan
 {
-    using type_t = typename Binop::type_t;
-    using scalar_t = typename Binop::scalar_t;
-    using binop_t = Binop<scalar_t>;
-    using binop_par_t = Binop<type_t>;
-    using inclusive_scan_op_t = subgroup2::impl::inclusive_scan<binop_par_t, native>;
+    using type_t = T;
+    using scalar_t = typename Binop::type_t;
+    using binop_t = Binop;
+    using inclusive_scan_op_t = subgroup2::impl::inclusive_scan<binop_t, T, native>;
 
-    type_t operator()(NBL_CONST_REF_ARG(type_t) value)
+    NBL_CONSTEXPR_STATIC_INLINE uint32_t ItemsPerInvocation = vector_traits<T>::Dimension;
+
+    type_t operator()(type_t value)
     {
         inclusive_scan_op_t op;
         value = op(value);
@@ -62,32 +64,32 @@ struct exclusive_scan
 
         type_t retval;
         retval[0] = bool(glsl::gl_SubgroupInvocationID()) ? left[ItemsPerInvocation-1] : binop_t::identity;
-        [unroll(ItemsPerInvocation-1)]
+        //[unroll(ItemsPerInvocation-1)]
         for (uint32_t i = 1; i < ItemsPerInvocation; i++)
             retval[i] = value[i-1];
         return retval;
     }
 };
 
-template<class Binop, bool native>
+template<class Binop, typename T, bool native>
 struct reduction
 {
-    using type_t = typename Binop::type_t;
-    using scalar_t = typename Binop::scalar_t;
-    using binop_t = Binop<scalar_t>;
-    using binop_par_t = Binop<type_t>;
-    using op_t = subgroup::impl::reduction<binop_par_t, native>;
+    using type_t = T;   // TODO? assert scalar_type<T> == scalar_t
+    using scalar_t = typename Binop::type_t;
+    using binop_t = Binop;
+    using op_t = subgroup::impl::reduction<binop_t, native>;
+
+    NBL_CONSTEXPR_STATIC_INLINE uint32_t ItemsPerInvocation = vector_traits<T>::Dimension;
 
     scalar_t operator()(NBL_CONST_REF_ARG(type_t) value)
     {
         binop_t binop;
         op_t op;
-        type_t result = op(value);
-        scalar_t retval;
-        [unroll(ItemsPerInvocation-1)]
-        for (uint32_t i = 0; i < ItemsPerInvocation; i++)
-            retval += binop(retval, result[i]);
-        return retval;
+        scalar_t retval = value[0];
+        //[unroll(ItemsPerInvocation-1)]
+        for (uint32_t i = 1; i < ItemsPerInvocation; i++)
+            retval += binop(retval, value[i]);
+        return op(retval);
     }
 };
 
@@ -97,4 +99,4 @@ struct reduction
 }
 }
 
-#endif
+#endif