fix(sdk): make the install tree usable

CMakeLists.txt

@@ -17,10 +17,19 @@ include(GNUInstallDirs)
 set(VISP_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header files")
 set(VISP_LIB_INSTALL_DIR ${CMAKE_INSTALL_LIBDIR} CACHE PATH "Location of library files")
 set(VISP_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR} CACHE PATH "Location of binary files")
+set(VISP_SHARE_INSTALL_DIR "${CMAKE_INSTALL_DATAROOTDIR}/visioncpp" CACHE PATH "Location of SDK data (generic runner source)")
 if(VISP_INSTALL_MODELS)
   set(VISP_MODEL_INSTALL_DIR "${CMAKE_INSTALL_DATAROOTDIR}/visioncpp" CACHE PATH "Directory to install default models to")
 endif()
 
+# The installed libvisioncpp.so records libggml.so.0 in DT_NEEDED. Without a RUNPATH a
+# consumer links successfully and then fails at startup with
+# "libggml.so.0: cannot open shared object file" -- the linker resolves the path CMake gives
+# it, and the loader searches again at run time with nothing to go on.
+# $ORIGIN is the directory holding the .so itself; ggml installs into the same lib/, so this
+# one line removes the need for LD_LIBRARY_PATH.
+set(CMAKE_INSTALL_RPATH "$ORIGIN")
+
 if(PROJECT_IS_TOP_LEVEL)
   set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
   set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
@@ -169,6 +178,13 @@ install(TARGETS visioncpp
   LIBRARY DESTINATION ${VISP_LIB_INSTALL_DIR}
   ARCHIVE DESTINATION ${VISP_LIB_INSTALL_DIR})
 install(DIRECTORY include/ DESTINATION ${VISP_INCLUDE_INSTALL_DIR})
+# Internal headers that generated code includes; consumers of the install tree need them.
+install(FILES src/visp/nn.h DESTINATION ${VISP_INCLUDE_INSTALL_DIR}/visp)
+install(DIRECTORY src/util/ DESTINATION ${VISP_INCLUDE_INSTALL_DIR}/util
+        FILES_MATCHING PATTERN "*.h")
+
+# Generic runner source. Consumers compile it together with their generated <Arch>.cpp.
+install(FILES tools/verify/backbone/run_dump.cpp DESTINATION ${VISP_SHARE_INSTALL_DIR})
 if(PROJECT_IS_TOP_LEVEL)
   install(FILES README.md LICENSE DESTINATION .)
 endif()
@@ -190,7 +206,7 @@ configure_package_config_file(
   scripts/cmake/visioncpp-config.cmake.in
   ${CMAKE_CURRENT_BINARY_DIR}/visioncpp-config.cmake
   INSTALL_DESTINATION ${VISP_LIB_INSTALL_DIR}/cmake/visioncpp
-  PATH_VARS VISP_INCLUDE_INSTALL_DIR VISP_LIB_INSTALL_DIR VISP_BIN_INSTALL_DIR
+  PATH_VARS VISP_INCLUDE_INSTALL_DIR VISP_LIB_INSTALL_DIR VISP_BIN_INSTALL_DIR VISP_SHARE_INSTALL_DIR
 )
 write_basic_package_version_file(
   visioncpp-version.cmake

scripts/cmake/visioncpp-config.cmake.in

@@ -2,6 +2,11 @@
 
 set_and_check(VISP_INCLUDE_DIR "@PACKAGE_VISP_INCLUDE_INSTALL_DIR@")
 set_and_check(VISP_LIB_DIR "@PACKAGE_VISP_LIB_INSTALL_DIR@")
+set_and_check(VISP_SHARE_DIR "@PACKAGE_VISP_SHARE_INSTALL_DIR@")
+
+# Generic runner source for generated models. For example:
+#   add_executable(run ${VISP_RUNNER} ${GEN}/${ARCH}.cpp)
+set(VISP_RUNNER "${VISP_SHARE_DIR}/run_dump.cpp")
 
 include(CMakeFindDependencyMacro)
 find_dependency(ggml)

tools/verify/backbone/run_dump.cpp

@@ -36,7 +36,13 @@ int main(int argc, char** argv) {
     std::string pref = argv[3];
     const int SZ = argc > 4 ? atoi(argv[4]) : 512;
 
-    backend_device backend = backend_init();
+    // backend_init() picks the best device it can find. When a build includes an
+    // accelerator backend that is not actually present, that choice fails late and
+    // unhelpfully, and numeric comparison wants the CPU anyway.
+    const char* backend_env = std::getenv("VISP_BACKEND");
+    backend_device backend = (backend_env && std::string(backend_env) == "cpu")
+                                 ? backend_init(backend_type::cpu)
+                                 : backend_init();
     model_file file = model_load(gguf);
     model_weights weights = model_init(file.n_tensors());
     model_transfer(file, weights, backend, backend.preferred_float_type(), file.tensor_layout());
@@ -50,6 +56,15 @@ int main(int argc, char** argv) {
     tensor last = FWD(m, input, p);   // 내부에서 compute_graph_output 이 out_i 를 그래프에 등록
     ggml_build_forward_expand(graph, last);
 
+    // VISP_DUMP_NODES=<dir> writes every named intermediate as <name>.bin. Comparing only
+    // the final output says that something diverged, not where.
+    const char* dump_dir = std::getenv("VISP_DUMP_NODES");
+    if (dump_dir) {
+        for (int i = 0; i < ggml_graph_n_nodes(graph.graph); ++i) {
+            ggml_set_output(ggml_graph_node(graph.graph, i));
+        }
+    }
+
     compute_graph_allocate(graph, backend);
     auto in = load_bin(inb, (size_t)3 * SZ * SZ);
     transfer_to_backend(input, std::span<const float>(in.data(), in.size()));
@@ -71,5 +86,22 @@ int main(int argc, char** argv) {
                (long long)t->ne[0], (long long)t->ne[1], (long long)t->ne[2], (long long)t->ne[3]);
     }
     printf("dumped %d outputs → %s.out.*.bin\n", n, pref.c_str());
+
+    if (dump_dir) {
+        int written = 0;
+        for (int i = 0; i < ggml_graph_n_nodes(graph.graph); ++i) {
+            tensor t = ggml_graph_node(graph.graph, i);
+            const char* nm = ggml_get_name(t);
+            if (!nm || !nm[0] || t->type != GGML_TYPE_F32) {
+                continue;
+            }
+            std::vector<float> d(ggml_nelements(t));
+            transfer_from_backend(t, std::span<float>(d.data(), d.size()));
+            std::ofstream of(std::string(dump_dir) + "/" + nm + ".bin", std::ios::binary);
+            of.write(reinterpret_cast<char const*>(d.data()), d.size() * sizeof(float));
+            ++written;
+        }
+        printf("- VISP_DUMP_NODES: %d intermediates → %s\n", written, dump_dir);
+    }
     return 0;
 }