feat: 添加昇腾ElementWise算子组件

src/infiniop/elementwise/ascend/elementwise_ascend.h

@@ -0,0 +1,54 @@
+#ifndef __INFINIOP_ELEMENTWISE_ASCEND_H__
+#define __INFINIOP_ELEMENTWISE_ASCEND_H__
+
+#include "../../devices/ascend/common_ascend.h"
+#include "../elementwise.h"
+#include <tuple>
+#include <utility>
+
+namespace op::elementwise::ascend {
+// template <typename... TensorDescs>
+class DeviceImpl final {
+    struct Opaque;
+    std::shared_ptr<Opaque> _opaque;
+
+    DeviceImpl(std::shared_ptr<Opaque> opaque) : _opaque(std::move(opaque)) {}
+
+public:
+    ~DeviceImpl() = default;
+
+    template <typename... Args>
+    static utils::Result<DeviceImpl *> create(Args &&...args);
+}
+
+template <typename... TensorDescs,
+          std::enable_if_t<(std::is_same_v<TensorDescs, aclnnTensorDescriptor_t> && ...), int> = 0>
+struct Opaque {
+    mutable aclOpExecutor *executor;
+    size_t workspaceSize;
+    TensorDescs outTensorDesc;
+    std::tuple<TensorDescs...> inTensorDescs;
+
+    explicit Opaque(aclOpExecutor *exec, size_t wsSize, TensorDescs outDesc, TensorDescs... descs)
+        : executor(exec), workspaceSize(wsSize), outTensorDesc(outDesc), inTensorDescs(std::forward<TensorDescs>(descs)...) {}
+
+    ~Opaque() {
+        aclDestroyAclOpExecutor(executor);
+        delete outDesc;
+        // 遍历元组并释放每个 Tensor 描述符
+        std::apply([](auto &&...args) {
+            (..., (delete args));
+        },
+                   inTensorDescs);
+    }
+
+    // 获取输出 Tensor 描述符
+    template <size_t N>
+    auto getInTensor() -> decltype(auto) {
+        return std::get<N>(inTensorDescs);
+    }
+}
+
+} // namespace op::elementwise::ascend
+
+#endif // __INFINIOP_ELEMENTWISE_ASCEND_H__