Merge branch 'amd-develop' into amd-master

include/ck/utility/amd_inline_asm.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef CK_AMD_INLINE_ASM_HPP
 #define CK_AMD_INLINE_ASM_HPP

include/ck/utility/amd_llvm_intrinsic.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#ifndef CK_AMD_LLVM_INTRINSIC_HPP
-#define CK_AMD_LLVM_INTRINSIC_HPP
-
-#include "data_type.hpp"
-
-namespace ck {
-
-__device__ int32_t llvm_amdgcn_readfirstlane_i32(int32_t i) __asm("llvm.amdgcn.readfirstlane");
-
-} // namespace ck
-#endif

include/ck/utility/amd_wave_read_first_lane.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 <array>
+#include <cstddef>
+#include <cstdint>
+#include <type_traits>
+
+namespace ck {
+namespace detail {
+
+template <unsigned SizeInBytes>
+struct get_carrier;
+
+template <>
+struct get_carrier<1>
+{
+    using type = uint8_t;
+};
+
+template <>
+struct get_carrier<2>
+{
+    using type = uint16_t;
+};
+
+template <>
+struct get_carrier<3>
+{
+    using type = class carrier
+    {
+        using value_type = uint32_t;
+
+        std::array<std::byte, 3> bytes;
+        static_assert(sizeof(bytes) <= sizeof(value_type));
+
+        // replacement of host std::copy_n()
+        template <typename InputIterator, typename Size, typename OutputIterator>
+        __device__ static OutputIterator copy_n(InputIterator from, Size size, OutputIterator to)
+        {
+            if(0 < size)
+            {
+                *to = *from;
+                ++to;
+                for(Size count = 1; count < size; ++count)
+                {
+                    *to = *++from;
+                    ++to;
+                }
+            }
+
+            return to;
+        }
+
+        // method to trigger template substitution failure
+        __device__ carrier(const carrier& other) noexcept
+        {
+            copy_n(other.bytes.begin(), bytes.size(), bytes.begin());
+        }
+
+        public:
+        __device__ carrier& operator=(value_type value) noexcept
+        {
+            copy_n(reinterpret_cast<const std::byte*>(&value), bytes.size(), bytes.begin());
+
+            return *this;
+        }
+
+        __device__ operator value_type() const noexcept
+        {
+            std::byte result[sizeof(value_type)];
+
+            copy_n(bytes.begin(), bytes.size(), result);
+
+            return *reinterpret_cast<const value_type*>(result);
+        }
+    };
+};
+static_assert(sizeof(get_carrier<3>::type) == 3);
+
+template <>
+struct get_carrier<4>
+{
+    using type = uint32_t;
+};
+
+template <unsigned SizeInBytes>
+using get_carrier_t = typename get_carrier<SizeInBytes>::type;
+
+} // namespace detail
+
+__device__ inline int32_t amd_wave_read_first_lane(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 amd_wave_read_first_lane(const Object& obj)
+{
+    using Size                = unsigned;
+    constexpr Size SgprSize   = 4;
+    constexpr Size ObjectSize = sizeof(Object);
+
+    auto* const from_obj = reinterpret_cast<const std::byte*>(&obj);
+    alignas(Object) std::byte to_obj[ObjectSize];
+
+    constexpr Size RemainedSize             = ObjectSize % SgprSize;
+    constexpr Size CompleteSgprCopyBoundary = ObjectSize - RemainedSize;
+    for(Size offset = 0; offset < CompleteSgprCopyBoundary; offset += SgprSize)
+    {
+        using Sgpr = detail::get_carrier_t<SgprSize>;
+
+        *reinterpret_cast<Sgpr*>(to_obj + offset) =
+            amd_wave_read_first_lane(*reinterpret_cast<const Sgpr*>(from_obj + offset));
+    }
+
+    if constexpr(0 < RemainedSize)
+    {
+        using Carrier = detail::get_carrier_t<RemainedSize>;
+
+        *reinterpret_cast<Carrier*>(to_obj + CompleteSgprCopyBoundary) = amd_wave_read_first_lane(
+            *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

include/ck/utility/amd_wmma.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef CK_AMD_WMMA_HPP
 #define CK_AMD_WMMA_HPP

include/ck/utility/amd_xdlops.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef CK_AMD_XDLOPS_HPP
 #define CK_AMD_XDLOPS_HPP
@@ -344,7 +344,7 @@ struct intrin_mfma_f64_16x16x4f64<16, 16>
     template <class FloatC>
     __device__ static void Run(const double& reg_a, const double& reg_b, FloatC& reg_c)
     {
-#if defined(__gfx90a__) || defined(__gfx940__)
+#if defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
         reg_c.template AsType<double4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f64_16x16x4f64(
             reg_a, reg_b, reg_c.template AsType<double4_t>()[Number<0>{}], 0, 0, 0);
 #else

include/ck/utility/array.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef CK_ARRAY_HPP
 #define CK_ARRAY_HPP

include/ck/utility/array_multi_index.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef CK_ARRAY_MULTI_INDEX_HPP
 #define CK_ARRAY_MULTI_INDEX_HPP

include/ck/utility/c_style_pointer_cast.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef CK_C_STYLE_POINTER_CAST_HPP
 #define CK_C_STYLE_POINTER_CAST_HPP

include/ck/utility/common_header.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -33,6 +33,7 @@
 #include "ck/utility/debug.hpp"
 
 #include "ck/utility/amd_buffer_addressing.hpp"
+#include "ck/utility/amd_wave_read_first_lane.hpp"
 #include "ck/utility/generic_memory_space_atomic.hpp"
 #include "ck/utility/get_id.hpp"
 #include "ck/utility/thread_group.hpp"

include/ck/utility/container_element_picker.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef CK_CONTAINER_ELEMENT_PICKER_HPP
 #define CK_CONTAINER_ELEMENT_PICKER_HPP

include/ck/utility/container_helper.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef CK_CONTAINER_HELPER_HPP
 #define CK_CONTAINER_HELPER_HPP

include/ck/utility/debug.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef UTILITY_DEBUG_HPP
 #define UTILITY_DEBUG_HPP

include/ck/utility/dynamic_buffer.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -19,7 +19,8 @@ namespace ck {
 template <AddressSpaceEnum BufferAddressSpace,
           typename T,
           typename ElementSpaceSize,
-          bool InvalidElementUseNumericalZeroValue>
+          bool InvalidElementUseNumericalZeroValue,
+          AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence>
 struct DynamicBuffer
 {
     using type = T;
@@ -77,13 +78,16 @@ struct DynamicBuffer
 
             if constexpr(InvalidElementUseNumericalZeroValue)
             {
-                return amd_buffer_load_invalid_element_return_zero<remove_cvref_t<T>, t_per_x>(
+                return amd_buffer_load_invalid_element_return_zero<remove_cvref_t<T>,
+                                                                   t_per_x,
+                                                                   coherence>(
                     p_data_, i, is_valid_element, element_space_size_);
             }
             else
             {
                 return amd_buffer_load_invalid_element_return_customized_value<remove_cvref_t<T>,
-                                                                               t_per_x>(
+                                                                               t_per_x,
+                                                                               coherence>(
                     p_data_, i, is_valid_element, element_space_size_, invalid_element_value_);
             }
         }
@@ -173,7 +177,7 @@ struct DynamicBuffer
         {
             constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
 
-            amd_buffer_store<remove_cvref_t<T>, t_per_x>(
+            amd_buffer_store<remove_cvref_t<T>, t_per_x, coherence>(
                 x, p_data_, i, is_valid_element, element_space_size_);
         }
         else if constexpr(GetAddressSpace() == AddressSpaceEnum::Lds &&
@@ -376,14 +380,19 @@ struct DynamicBuffer
     __host__ __device__ static constexpr bool IsDynamicBuffer() { return true; }
 };
 
-template <AddressSpaceEnum BufferAddressSpace, typename T, typename ElementSpaceSize>
+template <AddressSpaceEnum BufferAddressSpace,
+          AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence,
+          typename T,
+          typename ElementSpaceSize>
 __host__ __device__ constexpr auto make_dynamic_buffer(T* p, ElementSpaceSize element_space_size)
 {
-    return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, true>{p, element_space_size};
+    return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, true, coherence>{
+        p, element_space_size};
 }
 
 template <
     AddressSpaceEnum BufferAddressSpace,
+    AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence,
     typename T,
     typename ElementSpaceSize,
     typename X,
@@ -391,7 +400,7 @@ template <
 __host__ __device__ constexpr auto
 make_dynamic_buffer(T* p, ElementSpaceSize element_space_size, X invalid_element_value)
 {
-    return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, false>{
+    return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, false, coherence>{
         p, element_space_size, invalid_element_value};
 }
 

include/ck/utility/enable_if.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

include/ck/utility/functional.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

include/ck/utility/functional2.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

include/ck/utility/functional3.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

include/ck/utility/functional4.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef CK_FUNCTIONAL4_HPP
 #define CK_FUNCTIONAL4_HPP

include/ck/utility/generic_memory_space_atomic.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 #include "data_type.hpp"

include/ck/utility/get_id.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once