#13499: Split runtime_args_data into header and implementation (#13915)

tt_metal/impl/CMakeLists.txt

@@ -7,6 +7,7 @@ set(IMPL_SRC
 	${CMAKE_CURRENT_SOURCE_DIR}/buffers/circular_buffer.cpp
 	${CMAKE_CURRENT_SOURCE_DIR}/buffers/semaphore.cpp
 	${CMAKE_CURRENT_SOURCE_DIR}/kernels/kernel.cpp
+	${CMAKE_CURRENT_SOURCE_DIR}/kernels/runtime_args_data.cpp
 	${CMAKE_CURRENT_SOURCE_DIR}/allocator/algorithms/free_list.cpp
 	${CMAKE_CURRENT_SOURCE_DIR}/allocator/allocator.cpp
 	${CMAKE_CURRENT_SOURCE_DIR}/allocator/basic_allocator.cpp

tt_metal/impl/kernels/runtime_args_data.cpp

@@ -0,0 +1,44 @@
+// SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+
+#include "tt_metal/impl/kernels/runtime_args_data.hpp" // Public API Header
+
+#include "tt_metal/common/assert.hpp"
+
+namespace tt::tt_metal {
+
+std::uint32_t & RuntimeArgsData::operator[](std::size_t index) {
+    TT_ASSERT(index < rt_args_count, "Index specified is larger than runtime args size");
+    return this->rt_args_data[index];
+}
+
+const std::uint32_t & RuntimeArgsData::operator[](std::size_t index) const {
+    TT_ASSERT(index < rt_args_count, "Index specified is larger than runtime args size");
+    return this->rt_args_data[index];
+}
+
+std::uint32_t & RuntimeArgsData::at(std::size_t index) {
+    TT_FATAL(index < rt_args_count, "Index specified is larger than runtime args size");
+    return this->rt_args_data[index];
+}
+
+const std::uint32_t & RuntimeArgsData::at(std::size_t index) const {
+    TT_FATAL(index < rt_args_count, "Index specified is larger than runtime args size");
+    return this->rt_args_data[index];
+}
+
+std::uint32_t * RuntimeArgsData::data() noexcept {
+    return rt_args_data;
+}
+
+const std::uint32_t * RuntimeArgsData::data() const noexcept {
+    return rt_args_data;
+}
+
+std::size_t RuntimeArgsData::size() const noexcept {
+    return rt_args_count;
+}
+
+}
+

tt_metal/impl/kernels/runtime_args_data.hpp

@@ -4,42 +4,31 @@
 
 #pragma once
 
-#include "tt_metal/common/assert.hpp"
+#include <cstddef>
+#include <cstdint>
 
 namespace tt::tt_metal {
 // RuntimeArgsData provides an indirection to the runtime args
 // Prior to generating the cq cmds for the device, this points into a vector within the kernel
 // After generation, this points into the cq cmds so that runtime args API calls
 // update the data directly in the command
 struct RuntimeArgsData {
-    uint32_t * rt_args_data;
-    size_t rt_args_count;
-
-    inline uint32_t & operator[](size_t index) {
-        TT_ASSERT(index < rt_args_count, "Index specified is larger than runtime args size");
-        return this->rt_args_data[index];
-    }
-    inline const uint32_t& operator[](size_t index) const {
-        TT_ASSERT(index < rt_args_count, "Index specified is larger than runtime args size");
-        return this->rt_args_data[index];
-    }
-    inline uint32_t & at(size_t index) {
-        TT_FATAL(index < rt_args_count, "Index specified is larger than runtime args size");
-        return this->rt_args_data[index];
-    }
-    inline const uint32_t& at(size_t index) const {
-        TT_FATAL(index < rt_args_count, "Index specified is larger than runtime args size");
-        return this->rt_args_data[index];
-    }
-    inline uint32_t * data() noexcept {
-        return rt_args_data;
-    }
-    inline const uint32_t * data() const noexcept {
-        return rt_args_data;
-    }
-    inline size_t size() const noexcept{
-        return rt_args_count;
-    }
-};
+    std::uint32_t * rt_args_data;
+    std::size_t rt_args_count;
+
+    std::uint32_t & operator[](std::size_t index);
+
+    const std::uint32_t& operator[](std::size_t index) const;
+
+    std::uint32_t & at(std::size_t index);
 
+    const std::uint32_t& at(std::size_t index) const;
+
+    std::uint32_t * data() noexcept;
+
+    const std::uint32_t * data() const noexcept;
+
+    std::size_t size() const noexcept;
 };
+
+}