Merge branch 'feature/support-readfirstlane-for-object-types' into...

Merge branch 'feature/support-readfirstlane-for-object-types' into feature/simplify-karg-for-device-gemm-xdl-improved

include/ck/utility/common_header.hpp

@@ -40,6 +40,7 @@
 #include "ck/utility/amd_address_space.hpp"
 #include "ck/utility/static_buffer.hpp"
 #include "ck/utility/dynamic_buffer.hpp"
+#include "ck/utility/readfirstlane.hpp"
 
 // TODO: remove this
 #if CK_USE_AMD_INLINE_ASM

include/ck/utility/readfirstlane.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/functional2.hpp"
+#include "ck/utility/math.hpp"
+
+#include <cstddef>
+#include <cstdint>
+#include <type_traits>
+
+namespace ck {
+namespace detail {
+
+template <std::size_t Size>
+struct get_signed_int;
+
+template <>
+struct get_signed_int<1>
+{
+    using type = std::int8_t;
+};
+
+template <>
+struct get_signed_int<2>
+{
+    using type = std::int16_t;
+};
+
+template <>
+struct get_signed_int<4>
+{
+    using type = std::int32_t;
+};
+
+template <std::size_t Size>
+using get_signed_int_t = typename get_signed_int<Size>::type;
+
+} // namespace detail
+
+__device__ inline std::int32_t readfirstlane(std::int32_t value)
+{
+    return __builtin_amdgcn_readfirstlane(value);
+}
+
+template <
+    typename Object,
+    typename = std::enable_if_t<std::is_class_v<Object> && std::is_trivially_copyable_v<Object>>>
+__device__ auto readfirstlane(const Object& obj)
+{
+    constexpr std::size_t SgprSize   = 4;
+    constexpr std::size_t ObjectSize = sizeof(Object);
+
+    using Sgpr = detail::get_signed_int_t<SgprSize>;
+
+    alignas(Object) std::byte to_obj[ObjectSize];
+
+    auto* const from_obj = reinterpret_cast<const std::byte*>(&obj);
+
+    constexpr std::size_t RemainedSize             = ObjectSize % SgprSize;
+    constexpr std::size_t CompleteSgprCopyBoundary = ObjectSize - RemainedSize;
+    static_for<0, CompleteSgprCopyBoundary, SgprSize>{}([&](auto offset) {
+        *reinterpret_cast<Sgpr*>(to_obj + offset) =
+            readfirstlane(*reinterpret_cast<const Sgpr*>(from_obj + offset));
+    });
+
+    if constexpr(0 < RemainedSize)
+    {
+        using Carrier = detail::get_signed_int_t<RemainedSize>;
+
+        *reinterpret_cast<Carrier>(to_obj + CompleteSgprCopyBoundary) =
+            readfirstlane(*reinterpret_cast<const Carrier*>(from_obj + CompleteSgprCopyBoundary));
+    }
+
+    /// NOTE: Implicitly start object lifetime. It's better to use
+    ///        std::start_lifetime_at() in this scenario
+    return *reinterpret_cast<Object*>(to_obj);
+}
+
+} // namespace ck