Make various updates and fixes:

- Add support for legacy CUDA versions; now compatible with CUDA 12.3 and newer
- Add support for NVRTC compilation
- Other fixes and code refactoring

csrc/utils/exception.hpp

@@ -35,6 +35,16 @@ do { \
 #define DG_HOST_UNREACHABLE(reason) (throw DGException("Assertion", __FILE__, __LINE__, reason))
 #endif
 
+#ifndef DG_NVRTC_CHECK
+#define DG_NVRTC_CHECK(cmd) \
+do { \
+    const auto& e = (cmd); \
+    if (e != NVRTC_SUCCESS) { \
+        throw DGException("NVRTC", __FILE__, __LINE__, nvrtcGetErrorString(e)); \
+    } \
+} while (0)
+#endif
+
 #ifndef DG_CUDA_DRIVER_CHECK
 #define DG_CUDA_DRIVER_CHECK(cmd) \
 do { \

csrc/utils/lazy_init.hpp

@@ -0,0 +1,27 @@
+#pragma once
+
+#include <functional>
+#include <memory>
+
+#define DG_DECLARE_STATIC_VAR_IN_CLASS(cls, name) decltype(cls::name) cls::name
+
+namespace deep_gemm {
+
+template <typename T>
+class LazyInit {
+public:
+    explicit LazyInit(std::function<std::shared_ptr<T>()> factory)
+        : factory(std::move(factory)) {}
+
+    T* operator -> () {
+        if (ptr == nullptr)
+            ptr = factory();
+        return ptr.get();
+    }
+
+private:
+    std::shared_ptr<T> ptr;
+    std::function<std::shared_ptr<T>()> factory;
+};
+
+} // namespace deep_gemm

csrc/utils/system.hpp

@@ -38,8 +38,8 @@ static std::tuple<int, std::string> call_external_command(std::string command) {
     std::string output;
     while (fgets(buffer.data(), buffer.size(), pipe.get()))
         output += buffer.data();
-    const auto exit_code = pclose(pipe.release());
-    return {WEXITSTATUS(exit_code), output};
+    const auto& exit_code = WEXITSTATUS(pclose(pipe.release()));
+    return {exit_code, output};
 }
 
 static std::vector<std::filesystem::path> collect_files(const std::filesystem::path& root) {