OfferVirtualMemory

Author: pavelfatin

llama.cpp

@@ -1546,7 +1546,7 @@ struct llama_mmap {
         }
     }
 
-    void unmap_fragment(size_t first, size_t last) {
+    virtual void unmap_fragment(size_t first, size_t last) {
         // not supported
         GGML_UNUSED(first);
         GGML_UNUSED(last);
@@ -1652,6 +1652,29 @@ struct llama_anonymous_mmap : llama_mmap {
             throw std::runtime_error("unexpectedly reached end of file");
         }
     }
+
+#if _WIN32_WINNT >= 0x603
+    void unmap_fragment(size_t first, size_t last) override {
+        DWORD (WINAPI *pOfferVirtualMemory) (PVOID, SIZE_T, DWORD);
+        HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
+
+        pOfferVirtualMemory = reinterpret_cast<decltype(pOfferVirtualMemory)> (GetProcAddress(hKernel32, "OfferVirtualMemory"));
+
+        if (pOfferVirtualMemory) {
+            SYSTEM_INFO siSysInfo;
+            GetSystemInfo(&siSysInfo);
+            DWORD dwPageSize = siSysInfo.dwPageSize;
+
+            align_to_next_page(&first, dwPageSize);
+            align_to_previous_page(&last, dwPageSize);
+
+            if (pOfferVirtualMemory((char *) addr + first, last - first, 0x00000004 /* VMOfferPriorityNormal */)) {
+                LLAMA_LOG_WARN("warning: OfferVirtualMemory failed: %s\n", llama_format_win_err(GetLastError()).c_str());
+            }
+        }
+    }
+#endif
+
 #else
     llama_anonymous_mmap(struct llama_file * file) {
         GGML_UNUSED(file);