Merge branch 'develop' into ref_fp8

src/dynamic_loader.cpp

@@ -130,6 +130,30 @@ struct dynamic_loader_impl
     tmp_dir temp;
 };
 
+fs::path dynamic_loader::path(void* address)
+{
+    HMODULE module = nullptr;
+    if(GetModuleHandleEx(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS |
+                             GET_MODULE_HANDLE_EX_FLAG_UNCHANGED_REFCOUNT,
+                         static_cast<LPCSTR>(address),
+                         &module) == 0)
+    {
+        auto err = GetLastError();
+        MIGRAPHX_THROW("Unable to obtain module handle, error = " + std::to_string(err));
+    }
+    TCHAR buffer[MAX_PATH];
+    if(GetModuleFileName(module, buffer, sizeof(buffer)) == 0)
+    {
+        auto err = GetLastError();
+        MIGRAPHX_THROW("Unable to read module file path, error = " + std::to_string(err));
+    }
+    if(GetLastError() == ERROR_INSUFFICIENT_BUFFER)
+    {
+        MIGRAPHX_THROW("Buffer too small (" + std::to_string(MAX_PATH) + ") to hold the path");
+    }
+    return {buffer};
+}
+
 #endif
 
 optional<dynamic_loader> dynamic_loader::try_load(const fs::path& p)

src/include/migraphx/dynamic_loader.hpp

@@ -38,15 +38,12 @@ struct dynamic_loader_impl;
 
 struct MIGRAPHX_EXPORT dynamic_loader
 {
-#ifndef _WIN32
     template <class T>
     static fs::path path(T* address)
     {
         return path(reinterpret_cast<void*>(address));
     }
     static fs::path path(void* address);
-#endif
-
     static optional<dynamic_loader> try_load(const fs::path& p);
 
     dynamic_loader() = default;

src/targets/gpu/compile_hip.cpp

@@ -251,10 +251,21 @@ compile_hip_src(const std::vector<src_file>& srcs, std::string params, const std
             std::cout << std::string(src.content) << std::endl;
         }
     }
+    auto fname = fs::path{"migraphx-hiprtc-driver"};
+#ifdef _WIN32
+    fname.replace_extension(".exe");
+#endif
     auto p      = dynamic_loader::path(&compile_hip_src_with_hiprtc);
-    auto driver = p.parent_path().parent_path() / "bin" / "migraphx-hiprtc-driver";
+    auto driver = p.parent_path() / fname;
+
+    bool found = fs::exists(driver);
+    if(not found)
+    {
+        driver = p.parent_path().parent_path() / "bin" / fname;
+        found  = fs::exists(driver);
+    }
 
-    if(fs::exists(driver))
+    if(found)
     {
         value v;
         v["srcs"]   = to_value(hsrcs);