diff options
Diffstat (limited to 'thirdparty/vulkan/loader/loader.c')
| -rw-r--r-- | thirdparty/vulkan/loader/loader.c | 957 |
1 files changed, 804 insertions, 153 deletions
diff --git a/thirdparty/vulkan/loader/loader.c b/thirdparty/vulkan/loader/loader.c index c7cdb47122..158097a8aa 100644 --- a/thirdparty/vulkan/loader/loader.c +++ b/thirdparty/vulkan/loader/loader.c @@ -70,8 +70,15 @@ #include <initguid.h> #include <devpkey.h> #include <winternl.h> +#include <strsafe.h> +#ifdef __MINGW32__ +#undef strcpy // fix error with redfined strcpy when building with MinGW-w64 +#endif +#include <dxgi1_6.h> #include "adapters.h" -#include "dxgi_loader.h" + +typedef HRESULT (APIENTRY *PFN_CreateDXGIFactory1)(REFIID riid, void **ppFactory); +static PFN_CreateDXGIFactory1 fpCreateDXGIFactory1; #endif // This is a CMake generated file with #defines for any functions/includes @@ -84,6 +91,10 @@ // Generated file containing all the extension data #include "vk_loader_extensions.c" +// Environment Variable information +#define VK_ICD_FILENAMES_ENV_VAR "VK_ICD_FILENAMES" +#define VK_LAYER_PATH_ENV_VAR "VK_LAYER_PATH" + // Override layer information #define VK_OVERRIDE_LAYER_NAME "VK_LAYER_LUNARG_override" @@ -121,12 +132,16 @@ enum loader_data_files_type { // additionally CreateDevice and DestroyDevice needs to be locked loader_platform_thread_mutex loader_lock; loader_platform_thread_mutex loader_json_lock; +loader_platform_thread_mutex loader_preload_icd_lock; -LOADER_PLATFORM_THREAD_ONCE_DECLARATION(once_init); +// A list of ICDs that gets initialized when the loader does its global initialization. This list should never be used by anything +// other than EnumerateInstanceExtensionProperties(), vkDestroyInstance, and loader_release(). This list does not change +// functionality, but the fact that the libraries already been loaded causes any call that needs to load ICD libraries to speed up +// significantly. This can have a huge impact when making repeated calls to vkEnumerateInstanceExtensionProperties and +// vkCreateInstance. +static struct loader_icd_tramp_list scanned_icds; -// This loader supports Vulkan API version 1.1 -uint32_t loader_major_version = 1; -uint32_t loader_minor_version = 2; +LOADER_PLATFORM_THREAD_ONCE_DECLARATION(once_init); void *loader_instance_heap_alloc(const struct loader_instance *instance, size_t size, VkSystemAllocationScope alloc_scope) { void *pMemory = NULL; @@ -238,7 +253,7 @@ void *loader_device_heap_realloc(const struct loader_device *device, void *pMemo } // Environment variables -#if defined(__linux__) || defined(__APPLE__) +#if defined(__linux__) || defined(__APPLE__) || defined(__Fuchsia__) static inline bool IsHighIntegrity() { return geteuid() != getuid() || getegid() != getgid(); @@ -262,6 +277,8 @@ static inline char *loader_secure_getenv(const char *name, const struct loader_i // This algorithm is derived from glibc code that sets an internal // variable (__libc_enable_secure) if the process is running under setuid or setgid. return IsHighIntegrity() ? NULL : loader_getenv(name, inst); +#elif defined(__Fuchsia__) + return loader_getenv(name, inst); #else // Linux #if defined(HAVE_SECURE_GETENV) && !defined(USE_UNSAFE_FILE_SEARCH) @@ -272,13 +289,12 @@ static inline char *loader_secure_getenv(const char *name, const struct loader_i out = __secure_getenv(name); #else out = loader_getenv(name, inst); +#if !defined(USE_UNSAFE_FILE_SEARCH) + loader_log(inst, LOADER_INFO_BIT, 0, "Loader is using non-secure environment variable lookup for %s", name); #endif #endif - if (out == NULL) { - loader_log(inst, LOADER_INFO_BIT, 0, - "Loader is running with elevated permissions. Environment variable %s will be ignored.", name); - } return out; +#endif } static inline void loader_free_getenv(char *val, const struct loader_instance *inst) { @@ -340,8 +356,8 @@ static inline char *loader_getenv(const char *name, const struct loader_instance static inline char *loader_secure_getenv(const char *name, const struct loader_instance *inst) { #if !defined(USE_UNSAFE_FILE_SEARCH) if (IsHighIntegrity()) { - loader_log(inst, LOADER_INFO_BIT, 0, - "Loader is running with elevated permissions. Environment variable %s will be ignored.", name); + loader_log(inst, LOADER_INFO_BIT, 0, "Loader is running with elevated permissions. Environment variable %s will be ignored", + name); return NULL; } #endif @@ -400,7 +416,7 @@ void loader_log(const struct loader_instance *inst, VkFlags msg_type, int32_t ms } else if ((msg_type & LOADER_ERROR_BIT) != 0) { severity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT; } else if ((msg_type & LOADER_DEBUG_BIT) != 0) { - severity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT; + severity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT; } if ((msg_type & LOADER_PERF_BIT) != 0) { @@ -683,7 +699,11 @@ VkResult loaderGetDeviceRegistryFiles(const struct loader_instance *inst, char * LPCSTR value_name) { static const wchar_t *softwareComponentGUID = L"{5c4c3332-344d-483c-8739-259e934c9cc8}"; static const wchar_t *displayGUID = L"{4d36e968-e325-11ce-bfc1-08002be10318}"; +#ifdef CM_GETIDLIST_FILTER_PRESENT const ULONG flags = CM_GETIDLIST_FILTER_CLASS | CM_GETIDLIST_FILTER_PRESENT; +#else + const ULONG flags = 0x300; +#endif wchar_t childGuid[MAX_GUID_STRING_LEN + 2]; // +2 for brackets {} ULONG childGuidSize = sizeof(childGuid); @@ -876,7 +896,7 @@ VkResult loaderGetRegistryFiles(const struct loader_instance *inst, char *locati } if (is_driver) { - HRESULT hres = dyn_CreateDXGIFactory1(&IID_IDXGIFactory1, &dxgi_factory); + HRESULT hres = fpCreateDXGIFactory1(&IID_IDXGIFactory1, (void **)&dxgi_factory); if (hres != S_OK) { loader_log( inst, VK_DEBUG_REPORT_WARNING_BIT_EXT, 0, @@ -1078,11 +1098,11 @@ static size_t loader_platform_combine_path(char *dest, size_t len, ...) { // Given string of three part form "maj.min.pat" convert to a vulkan version number. static uint32_t loader_make_version(char *vers_str) { - uint32_t vers = 0, major = 0, minor = 0, patch = 0; + uint32_t major = 0, minor = 0, patch = 0; char *vers_tok; if (!vers_str) { - return vers; + return 0; } vers_tok = strtok(vers_str, ".\"\n\r"); @@ -1230,6 +1250,10 @@ void loaderDeleteLayerListAndProperties(const struct loader_instance *inst, stru loader_instance_heap_free(inst, layer_list->list[i].override_paths); layer_list->list[i].override_paths = NULL; } + if (NULL != layer_list->list[i].app_key_paths) { + loader_instance_heap_free(inst, layer_list->list[i].app_key_paths); + layer_list->list[i].app_key_paths = NULL; + } loader_destroy_generic_list(inst, (struct loader_generic_list *)&layer_list->list[i].instance_extension_list); dev_ext_list = &layer_list->list[i].device_extension_list; if (dev_ext_list->capacity > 0 && NULL != dev_ext_list->list) { @@ -1253,6 +1277,31 @@ void loaderDeleteLayerListAndProperties(const struct loader_instance *inst, stru } } +void loaderRemoveLayerInList(const struct loader_instance *inst, struct loader_layer_list *layer_list, uint32_t layer_to_remove) { + if (layer_list == NULL || layer_to_remove >= layer_list->count) { + return; + } + if (layer_list->list[layer_to_remove].type_flags & VK_LAYER_TYPE_FLAG_META_LAYER) { + // Delete the component layers + loader_instance_heap_free(inst, layer_list->list[layer_to_remove].component_layer_names); + loader_instance_heap_free(inst, layer_list->list[layer_to_remove].override_paths); + loader_instance_heap_free(inst, layer_list->list[layer_to_remove].blacklist_layer_names); + loader_instance_heap_free(inst, layer_list->list[layer_to_remove].app_key_paths); + } + + // Remove the current invalid meta-layer from the layer list. Use memmove since we are + // overlapping the source and destination addresses. + memmove(&layer_list->list[layer_to_remove], &layer_list->list[layer_to_remove + 1], + sizeof(struct loader_layer_properties) * (layer_list->count - 1 - layer_to_remove)); + + // Make sure to clear out the removed layer, in case new layers are added in the previous location + memset(&layer_list->list[layer_list->count - 1], 0, sizeof(struct loader_layer_properties)); + + // Decrement the count (because we now have one less) and decrement the loop index since we need to + // re-check this index. + layer_list->count--; +} + // Remove all layers in the layer list that are blacklisted by the override layer. // NOTE: This should only be called if an override layer is found and not expired. void loaderRemoveLayersInBlacklist(const struct loader_instance *inst, struct loader_layer_list *layer_list) { @@ -1925,9 +1974,10 @@ bool loaderAddMetaLayer(const struct loader_instance *inst, const struct loader_ // Search the source_list for any layer with a name that matches the given name and a type // that matches the given type. Add all matching layers to the target_list. // Do not add if found loader_layer_properties is already on the target_list. -void loaderAddLayerNameToList(const struct loader_instance *inst, const char *name, const enum layer_type_flags type_flags, - const struct loader_layer_list *source_list, struct loader_layer_list *target_list, - struct loader_layer_list *expanded_target_list) { +VkResult loaderAddLayerNameToList(const struct loader_instance *inst, const char *name, const enum layer_type_flags type_flags, + const struct loader_layer_list *source_list, struct loader_layer_list *target_list, + struct loader_layer_list *expanded_target_list) { + VkResult res = VK_SUCCESS; bool found = false; for (uint32_t i = 0; i < source_list->count; i++) { struct loader_layer_properties *source_prop = &source_list->list[i]; @@ -1948,9 +1998,17 @@ void loaderAddLayerNameToList(const struct loader_instance *inst, const char *na } } if (!found) { - loader_log(inst, VK_DEBUG_REPORT_WARNING_BIT_EXT, 0, "loaderAddLayerNameToList: Failed to find layer name %s to activate", - name); + if (strcmp(name, "VK_LAYER_LUNARG_standard_validation")) { + loader_log(inst, VK_DEBUG_REPORT_WARNING_BIT_EXT, 0, + "loaderAddLayerNameToList: Failed to find layer name %s to activate", name); + } else { + res = VK_ERROR_LAYER_NOT_PRESENT; + loader_log(inst, VK_DEBUG_REPORT_ERROR_BIT_EXT, 0, + "Layer VK_LAYER_LUNARG_standard_validation has been changed to VK_LAYER_KHRONOS_validation. Please use the " + "new version of the layer."); + } } + return res; } static VkExtensionProperties *get_extension_property(const char *name, const struct loader_extension_list *list) { @@ -2251,13 +2309,20 @@ static VkResult loader_scanned_icd_add(const struct loader_instance *inst, struc PFN_vkGetInstanceProcAddr fp_get_proc_addr; PFN_GetPhysicalDeviceProcAddr fp_get_phys_dev_proc_addr = NULL; PFN_vkNegotiateLoaderICDInterfaceVersion fp_negotiate_icd_version; +#if defined(VK_USE_PLATFORM_WIN32_KHR) + PFN_vk_icdEnumerateAdapterPhysicalDevices fp_enum_dxgi_adapter_phys_devs = NULL; +#endif struct loader_scanned_icd *new_scanned_icd; uint32_t interface_vers; VkResult res = VK_SUCCESS; // TODO implement smarter opening/closing of libraries. For now this // function leaves libraries open and the scanned_icd_clear closes them +#if defined(__Fuchsia__) + handle = loader_platform_open_driver(filename); +#else handle = loader_platform_open_library(filename); +#endif if (NULL == handle) { loader_log(inst, VK_DEBUG_REPORT_ERROR_BIT_EXT, 0, loader_platform_open_library_error(filename)); goto out; @@ -2337,6 +2402,11 @@ static VkResult loader_scanned_icd_add(const struct loader_instance *inst, struc goto out; } fp_get_phys_dev_proc_addr = loader_platform_get_proc_address(handle, "vk_icdGetPhysicalDeviceProcAddr"); +#if defined(VK_USE_PLATFORM_WIN32_KHR) + if (interface_vers >= 6) { + fp_enum_dxgi_adapter_phys_devs = loader_platform_get_proc_address(handle, "vk_icdEnumerateAdapterPhysicalDevices"); + } +#endif } // check for enough capacity @@ -2362,6 +2432,9 @@ static VkResult loader_scanned_icd_add(const struct loader_instance *inst, struc new_scanned_icd->GetPhysicalDeviceProcAddr = fp_get_phys_dev_proc_addr; new_scanned_icd->EnumerateInstanceExtensionProperties = fp_get_inst_ext_props; new_scanned_icd->CreateInstance = fp_create_inst; +#if defined(VK_USE_PLATFORM_WIN32_KHR) + new_scanned_icd->EnumerateAdapterPhysicalDevices = fp_enum_dxgi_adapter_phys_devs; +#endif new_scanned_icd->interface_version = interface_vers; new_scanned_icd->lib_name = (char *)loader_instance_heap_alloc(inst, strlen(filename) + 1, VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE); @@ -2430,7 +2503,7 @@ void loader_initialize(void) { // initialize mutexes loader_platform_thread_create_mutex(&loader_lock); loader_platform_thread_create_mutex(&loader_json_lock); - + loader_platform_thread_create_mutex(&loader_preload_icd_lock); // initialize logging loader_debug_init(); @@ -2444,6 +2517,13 @@ void loader_initialize(void) { // This is needed to ensure that newer APIs are available right away // and not after the first call that has been statically linked LoadLibrary("gdi32.dll"); + + TCHAR systemPath[MAX_PATH] = ""; + GetSystemDirectory(systemPath, MAX_PATH); + StringCchCat(systemPath, MAX_PATH, TEXT("\\dxgi.dll")); + HMODULE dxgi_module = LoadLibrary(systemPath); + fpCreateDXGIFactory1 = dxgi_module == NULL ? NULL : + (PFN_CreateDXGIFactory1)GetProcAddress(dxgi_module, "CreateDXGIFactory1"); #endif } @@ -2454,9 +2534,38 @@ struct loader_data_files { }; void loader_release() { + // Guarantee release of the preloaded ICD libraries. This may have already been called in vkDestroyInstance. + loader_unload_preloaded_icds(); + // release mutexes loader_platform_thread_delete_mutex(&loader_lock); loader_platform_thread_delete_mutex(&loader_json_lock); + loader_platform_thread_delete_mutex(&loader_preload_icd_lock); +} + +// Preload the ICD libraries that are likely to be needed so we don't repeatedly load/unload them later +void loader_preload_icds(void) { + loader_platform_thread_lock_mutex(&loader_preload_icd_lock); + + // Already preloaded, skip loading again. + if (scanned_icds.scanned_list != NULL) { + loader_platform_thread_unlock_mutex(&loader_preload_icd_lock); + return; + } + + memset(&scanned_icds, 0, sizeof(scanned_icds)); + VkResult result = loader_icd_scan(NULL, &scanned_icds); + if (result != VK_SUCCESS) { + loader_scanned_icd_clear(NULL, &scanned_icds); + } + loader_platform_thread_unlock_mutex(&loader_preload_icd_lock); +} + +// Release the ICD libraries that were preloaded +void loader_unload_preloaded_icds(void) { + loader_platform_thread_lock_mutex(&loader_preload_icd_lock); + loader_scanned_icd_clear(NULL, &scanned_icds); + loader_platform_thread_unlock_mutex(&loader_preload_icd_lock); } // Get next file or dirname given a string list or registry key path @@ -2543,7 +2652,12 @@ static VkResult loader_get_json(const struct loader_instance *inst, const char * res = VK_ERROR_INITIALIZATION_FAILED; goto out; } - fseek(file, 0, SEEK_END); + // NOTE: We can't just use fseek(file, 0, SEEK_END) because that isn't guaranteed to be supported on all systems + do { + // We're just seeking the end of the file, so this buffer is never used + char buffer[256]; + fread(buffer, 1, sizeof(buffer), file); + } while (!feof(file)); len = ftell(file); fseek(file, 0, SEEK_SET); json_buf = (char *)loader_stack_alloc(len + 1); @@ -2562,6 +2676,11 @@ static VkResult loader_get_json(const struct loader_instance *inst, const char * } json_buf[len] = '\0'; + // Can't be a valid json if the string is of length zero + if (len == 0) { + res = VK_ERROR_INITIALIZATION_FAILED; + goto out; + } // Parse text from file *json = cJSON_Parse(json_buf); if (*json == NULL) { @@ -2711,6 +2830,9 @@ static void VerifyAllMetaLayers(struct loader_instance *inst, struct loader_laye if (prop->blacklist_layer_names != NULL) { loader_instance_heap_free(inst, prop->blacklist_layer_names); } + if (prop->override_paths != NULL) { + loader_instance_heap_free(inst, prop->override_paths); + } // Remove the current invalid meta-layer from the layer list. Use memmove since we are // overlapping the source and destination addresses. @@ -2727,6 +2849,75 @@ static void VerifyAllMetaLayers(struct loader_instance *inst, struct loader_laye } } +// If the current working directory matches any app_key_path of the layers, remove all other override layers. +// Otherwise if no matching app_key was found, remove all but the global override layer, which has no app_key_path. +static void RemoveAllNonValidOverrideLayers(struct loader_instance *inst, struct loader_layer_list *instance_layers) { + if (instance_layers == NULL) { + return; + } + + char cur_path[MAX_STRING_SIZE]; + char *ret = loader_platform_executable_path(cur_path, sizeof(cur_path)); + if (ret == NULL) { + loader_log(inst, VK_DEBUG_REPORT_ERROR_BIT_EXT, 0, + "RemoveAllNonValidOverrideLayers: Failed to get executable path and name"); + return; + } + + // Find out if there is an override layer with same the app_key_path as the path to the current executable. + // If more than one is found, remove it and use the first layer + // Remove any layers which aren't global and do not have the same app_key_path as the path to the current executable. + bool found_active_override_layer = false; + int global_layer_index = -1; + for (uint32_t i = 0; i < instance_layers->count; i++) { + struct loader_layer_properties *props = &instance_layers->list[i]; + if (strcmp(props->info.layerName, VK_OVERRIDE_LAYER_NAME) == 0) { + if (props->num_app_key_paths > 0) { // not the global layer + for (uint32_t j = 0; j < props->num_app_key_paths; j++) { + if (strcmp(props->app_key_paths[j], cur_path) == 0) { + if (!found_active_override_layer) { + found_active_override_layer = true; + } else { + loader_log(inst, VK_DEBUG_REPORT_WARNING_BIT_EXT, 0, + "RemoveAllNonValidOverrideLayers: Multiple override layers where the same" + "path in app_keys was found. Using the first layer found"); + + // Remove duplicate active override layers that have the same app_key_path + loaderRemoveLayerInList(inst, instance_layers, i); + i--; + } + } + } + if (!found_active_override_layer) { + // Remove non-global override layers that don't have an app_key that matches cur_path + loaderRemoveLayerInList(inst, instance_layers, i); + i--; + } + } else { + if (global_layer_index == -1) { + global_layer_index = i; + } else { + loader_log(inst, VK_DEBUG_REPORT_WARNING_BIT_EXT, 0, + "RemoveAllNonValidOverrideLayers: Multiple global override layers " + "found. Using the first global layer found"); + loaderRemoveLayerInList(inst, instance_layers, i); + i--; + } + } + } + } + // Remove global layer if layer with same the app_key_path as the path to the current executable is found + if (found_active_override_layer && global_layer_index >= 0) { + loaderRemoveLayerInList(inst, instance_layers, global_layer_index); + } + // Should be at most 1 override layer in the list now. + if (found_active_override_layer) { + loader_log(inst, VK_DEBUG_REPORT_INFORMATION_BIT_EXT, 0, "Using the override layer for app key %s", cur_path); + } else if (global_layer_index >= 0) { + loader_log(inst, VK_DEBUG_REPORT_INFORMATION_BIT_EXT, 0, "Using the global override layer"); + } +} + // This structure is used to store the json file version // in a more manageable way. typedef struct { @@ -3008,30 +3199,31 @@ static VkResult loaderReadLayerJson(const struct loader_instance *inst, struct l name); } else { props->num_blacklist_layers = cJSON_GetArraySize(blacklisted_layers); - - // Allocate the blacklist array - props->blacklist_layer_names = loader_instance_heap_alloc( - inst, sizeof(char[MAX_STRING_SIZE]) * props->num_blacklist_layers, VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE); - if (props->blacklist_layer_names == NULL) { - result = VK_ERROR_OUT_OF_HOST_MEMORY; - goto out; - } - - // Copy the blacklisted layers into the array - for (i = 0; i < (int)props->num_blacklist_layers; ++i) { - cJSON *black_layer = cJSON_GetArrayItem(blacklisted_layers, i); - if (black_layer == NULL) { - continue; - } - temp = cJSON_Print(black_layer); - if (temp == NULL) { + if (props->num_blacklist_layers > 0) { + // Allocate the blacklist array + props->blacklist_layer_names = loader_instance_heap_alloc( + inst, sizeof(char[MAX_STRING_SIZE]) * props->num_blacklist_layers, VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE); + if (props->blacklist_layer_names == NULL) { result = VK_ERROR_OUT_OF_HOST_MEMORY; goto out; } - temp[strlen(temp) - 1] = '\0'; - strncpy(props->blacklist_layer_names[i], temp + 1, MAX_STRING_SIZE - 1); - props->blacklist_layer_names[i][MAX_STRING_SIZE - 1] = '\0'; - cJSON_Free(temp); + + // Copy the blacklisted layers into the array + for (i = 0; i < (int)props->num_blacklist_layers; ++i) { + cJSON *black_layer = cJSON_GetArrayItem(blacklisted_layers, i); + if (black_layer == NULL) { + continue; + } + temp = cJSON_Print(black_layer); + if (temp == NULL) { + result = VK_ERROR_OUT_OF_HOST_MEMORY; + goto out; + } + temp[strlen(temp) - 1] = '\0'; + strncpy(props->blacklist_layer_names[i], temp + 1, MAX_STRING_SIZE - 1); + props->blacklist_layer_names[i][MAX_STRING_SIZE - 1] = '\0'; + cJSON_Free(temp); + } } } } @@ -3045,28 +3237,29 @@ static VkResult loaderReadLayerJson(const struct loader_instance *inst, struct l } int count = cJSON_GetArraySize(override_paths); props->num_override_paths = count; + if (count > 0) { + // Allocate buffer for override paths + props->override_paths = + loader_instance_heap_alloc(inst, sizeof(char[MAX_STRING_SIZE]) * count, VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE); + if (NULL == props->override_paths) { + result = VK_ERROR_OUT_OF_HOST_MEMORY; + goto out; + } - // Allocate buffer for override paths - props->override_paths = - loader_instance_heap_alloc(inst, sizeof(char[MAX_STRING_SIZE]) * count, VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE); - if (NULL == props->override_paths) { - result = VK_ERROR_OUT_OF_HOST_MEMORY; - goto out; - } - - // Copy the override paths into the array - for (i = 0; i < count; i++) { - cJSON *override_path = cJSON_GetArrayItem(override_paths, i); - if (NULL != override_path) { - temp = cJSON_Print(override_path); - if (NULL == temp) { - result = VK_ERROR_OUT_OF_HOST_MEMORY; - goto out; + // Copy the override paths into the array + for (i = 0; i < count; i++) { + cJSON *override_path = cJSON_GetArrayItem(override_paths, i); + if (NULL != override_path) { + temp = cJSON_Print(override_path); + if (NULL == temp) { + result = VK_ERROR_OUT_OF_HOST_MEMORY; + goto out; + } + temp[strlen(temp) - 1] = '\0'; + strncpy(props->override_paths[i], temp + 1, MAX_STRING_SIZE - 1); + props->override_paths[i][MAX_STRING_SIZE - 1] = '\0'; + cJSON_Free(temp); } - temp[strlen(temp) - 1] = '\0'; - strncpy(props->override_paths[i], temp + 1, MAX_STRING_SIZE - 1); - props->override_paths[i][MAX_STRING_SIZE - 1] = '\0'; - cJSON_Free(temp); } } } @@ -3128,6 +3321,7 @@ static VkResult loaderReadLayerJson(const struct loader_instance *inst, struct l char *vkNegotiateLoaderLayerInterfaceVersion = NULL; char *spec_version = NULL; char **entry_array = NULL; + cJSON *app_keys = NULL; // Layer interface functions // vkGetInstanceProcAddr @@ -3316,6 +3510,45 @@ static VkResult loaderReadLayerJson(const struct loader_instance *inst, struct l } } + props->num_app_key_paths = 0; + props->app_key_paths = NULL; + app_keys = cJSON_GetObjectItem(layer_node, "app_keys"); + if (app_keys != NULL) { + if (strcmp(name, VK_OVERRIDE_LAYER_NAME)) { + loader_log(inst, VK_DEBUG_REPORT_WARNING_BIT_EXT, 0, + "Layer %s contains app_keys, but any app_keys can only be provided by the override metalayer. " + "These will be ignored.", + name); + } else { + props->num_app_key_paths = cJSON_GetArraySize(app_keys); + + // Allocate the blacklist array + props->app_key_paths = loader_instance_heap_alloc(inst, sizeof(char[MAX_STRING_SIZE]) * props->num_app_key_paths, + VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE); + if (props->app_key_paths == NULL) { + result = VK_ERROR_OUT_OF_HOST_MEMORY; + goto out; + } + + // Copy the app_key_paths into the array + for (i = 0; i < (int)props->num_app_key_paths; ++i) { + cJSON *app_key_path = cJSON_GetArrayItem(app_keys, i); + if (app_key_path == NULL) { + continue; + } + temp = cJSON_Print(app_key_path); + if (temp == NULL) { + result = VK_ERROR_OUT_OF_HOST_MEMORY; + goto out; + } + temp[strlen(temp) - 1] = '\0'; + strncpy(props->app_key_paths[i], temp + 1, MAX_STRING_SIZE - 1); + props->app_key_paths[i][MAX_STRING_SIZE - 1] = '\0'; + cJSON_Free(temp); + } + } + } + result = VK_SUCCESS; out: @@ -3332,12 +3565,17 @@ out: if (NULL != props->override_paths) { loader_instance_heap_free(inst, props->override_paths); } + if (NULL != props->app_key_paths) { + loader_instance_heap_free(inst, props->app_key_paths); + } props->num_blacklist_layers = 0; props->blacklist_layer_names = NULL; props->num_component_layers = 0; props->component_layer_names = NULL; props->num_override_paths = 0; props->override_paths = NULL; + props->num_app_key_paths = 0; + props->app_key_paths = NULL; } return result; @@ -3380,6 +3618,10 @@ static VkResult loaderAddLayerProperties(const struct loader_instance *inst, str layer_json_version json_version = {0, 0, 0}; char *vers_tok; cJSON *disable_environment = NULL; + // Make sure sure the top level json value is an object + if (!json || json->type != 6) { + goto out; + } item = cJSON_GetObjectItem(json, "file_format_version"); if (item == NULL) { goto out; @@ -3596,7 +3838,7 @@ out: } static VkResult AddDataFilesInPath(const struct loader_instance *inst, char *search_path, bool is_directory_list, - struct loader_data_files *out_files) { + struct loader_data_files *out_files, bool use_first_found_manifest) { VkResult vk_result = VK_SUCCESS; DIR *dir_stream = NULL; struct dirent *dir_entry; @@ -3678,6 +3920,9 @@ static VkResult AddDataFilesInPath(const struct loader_instance *inst, char *sea break; } } + if (use_first_found_manifest && out_files->count > 0) { + break; + } } out: @@ -3699,6 +3944,7 @@ static VkResult ReadDataFilesInSearchPaths(const struct loader_instance *inst, e char *search_path = NULL; char *cur_path_ptr = NULL; size_t rel_size = 0; + bool use_first_found_manifest = false; #ifndef _WIN32 bool xdgconfig_alloc = true; bool xdgdata_alloc = true; @@ -3719,12 +3965,14 @@ static VkResult ReadDataFilesInSearchPaths(const struct loader_instance *inst, e if (xdgdatadirs == NULL) { xdgdata_alloc = false; } +#if !defined(__Fuchsia__) if (xdgconfdirs == NULL || xdgconfdirs[0] == '\0') { xdgconfdirs = FALLBACK_CONFIG_DIRS; } if (xdgdatadirs == NULL || xdgdatadirs[0] == '\0') { xdgdatadirs = FALLBACK_DATA_DIRS; } +#endif // Only use HOME if XDG_DATA_HOME is not present on the system if (NULL == xdgdatahome) { @@ -3828,6 +4076,10 @@ static VkResult ReadDataFilesInSearchPaths(const struct loader_instance *inst, e memcpy(cur_path_ptr, relative_location, rel_size); cur_path_ptr += rel_size; *cur_path_ptr++ = PATH_SEPARATOR; + // only for ICD manifests + if (env_override != NULL && strcmp(VK_ICD_FILENAMES_ENV_VAR, env_override) == 0) { + use_first_found_manifest = true; + } } CFRelease(ref); } @@ -3853,6 +4105,39 @@ static VkResult ReadDataFilesInSearchPaths(const struct loader_instance *inst, e #endif } + // Remove duplicate paths, or it would result in duplicate extensions, duplicate devices, etc. + // This uses minimal memory, but is O(N^2) on the number of paths. Expect only a few paths. + char path_sep_str[2] = {PATH_SEPARATOR, '\0'}; + size_t search_path_updated_size = strlen(search_path); + for (size_t first = 0; first < search_path_updated_size;) { + // If this is an empty path, erase it + if (search_path[first] == PATH_SEPARATOR) { + memmove(&search_path[first], &search_path[first + 1], search_path_updated_size - first + 1); + search_path_updated_size -= 1; + continue; + } + + size_t first_end = first + 1; + first_end += strcspn(&search_path[first_end], path_sep_str); + for (size_t second = first_end + 1; second < search_path_updated_size;) { + size_t second_end = second + 1; + second_end += strcspn(&search_path[second_end], path_sep_str); + if (first_end - first == second_end - second && + !strncmp(&search_path[first], &search_path[second], second_end - second)) { + // Found duplicate. Include PATH_SEPARATOR in second_end, then erase it from search_path. + if (search_path[second_end] == PATH_SEPARATOR) { + second_end++; + } + memmove(&search_path[second], &search_path[second_end], search_path_updated_size - second_end + 1); + search_path_updated_size -= second_end - second; + } else { + second = second_end + 1; + } + } + first = first_end + 1; + } + search_path_size = search_path_updated_size; + // Print out the paths being searched if debugging is enabled if (search_path_size > 0) { loader_log(inst, VK_DEBUG_REPORT_DEBUG_BIT_EXT, 0, @@ -3860,7 +4145,7 @@ static VkResult ReadDataFilesInSearchPaths(const struct loader_instance *inst, e } // Now, parse the paths and add any manifest files found in them. - vk_result = AddDataFilesInPath(inst, search_path, is_directory_list, out_files); + vk_result = AddDataFilesInPath(inst, search_path, is_directory_list, out_files, use_first_found_manifest); if (NULL != override_path) { *override_active = true; @@ -4035,7 +4320,6 @@ out: static VkResult ReadDataFilesInRegistry(const struct loader_instance *inst, enum loader_data_files_type data_file_type, bool warn_if_not_present, char *registry_location, struct loader_data_files *out_files) { VkResult vk_result = VK_SUCCESS; - bool is_icd = (data_file_type == LOADER_DATA_FILE_MANIFEST_ICD); char *search_path = NULL; // These calls look at the PNP/Device section of the registry. @@ -4096,7 +4380,7 @@ static VkResult ReadDataFilesInRegistry(const struct loader_instance *inst, enum } // Now, parse the paths and add any manifest files found in them. - vk_result = AddDataFilesInPath(inst, search_path, false, out_files); + vk_result = AddDataFilesInPath(inst, search_path, false, out_files, false); out: @@ -4216,14 +4500,12 @@ VkResult loader_icd_scan(const struct loader_instance *inst, struct loader_icd_t if (VK_SUCCESS != res) { goto out; } - // Get a list of manifest files for ICDs - res = loaderGetDataFiles(inst, LOADER_DATA_FILE_MANIFEST_ICD, true, "VK_ICD_FILENAMES", NULL, VK_DRIVERS_INFO_REGISTRY_LOC, - VK_DRIVERS_INFO_RELATIVE_DIR, &manifest_files); + res = loaderGetDataFiles(inst, LOADER_DATA_FILE_MANIFEST_ICD, true, VK_ICD_FILENAMES_ENV_VAR, NULL, + VK_DRIVERS_INFO_REGISTRY_LOC, VK_DRIVERS_INFO_RELATIVE_DIR, &manifest_files); if (VK_SUCCESS != res || manifest_files.count == 0) { goto out; } - loader_platform_thread_lock_mutex(&loader_json_lock); lockedMutex = true; for (uint32_t i = 0; i < manifest_files.count; i++) { @@ -4244,6 +4526,7 @@ VkResult loader_icd_scan(const struct loader_instance *inst, struct loader_icd_t res = temp_res; } if (temp_res == VK_ERROR_OUT_OF_HOST_MEMORY) { + res = VK_ERROR_OUT_OF_HOST_MEMORY; break; } else { continue; @@ -4486,12 +4769,16 @@ void loaderScanForLayers(struct loader_instance *inst, struct loader_layer_list VkResult local_res = loaderAddLayerProperties(inst, instance_layers, json, true, file_str); cJSON_Delete(json); - if (VK_SUCCESS != local_res) { + // If the error is anything other than out of memory we still want to try to load the other layers + if (VK_ERROR_OUT_OF_HOST_MEMORY == local_res) { goto out; } } } + // Remove any extraneous override layers. + RemoveAllNonValidOverrideLayers(inst, instance_layers); + // Check to see if the override layer is present, and use it's override paths. for (int32_t i = 0; i < (int32_t)instance_layers->count; i++) { struct loader_layer_properties *prop = &instance_layers->list[i]; @@ -4519,7 +4806,7 @@ void loaderScanForLayers(struct loader_instance *inst, struct loader_layer_list } // Get a list of manifest files for explicit layers - if (VK_SUCCESS != loaderGetDataFiles(inst, LOADER_DATA_FILE_MANIFEST_LAYER, true, "VK_LAYER_PATH", override_paths, + if (VK_SUCCESS != loaderGetDataFiles(inst, LOADER_DATA_FILE_MANIFEST_LAYER, true, VK_LAYER_PATH_ENV_VAR, override_paths, VK_ELAYERS_INFO_REGISTRY_LOC, VK_ELAYERS_INFO_RELATIVE_DIR, &manifest_files)) { goto out; } @@ -4631,6 +4918,9 @@ void loaderScanForImplicitLayers(struct loader_instance *inst, struct loader_lay } } + // Remove any extraneous override layers. + RemoveAllNonValidOverrideLayers(inst, instance_layers); + // Check to see if either the override layer is present, or another implicit meta-layer. // Each of these may require explicit layers to be enabled at this time. for (int32_t i = 0; i < (int32_t)instance_layers->count; i++) { @@ -4667,7 +4957,7 @@ void loaderScanForImplicitLayers(struct loader_instance *inst, struct loader_lay // explicit layer info as well. Not to worry, though, all explicit layers not included // in the override layer will be removed below in loaderRemoveLayersInBlacklist(). if (override_layer_valid || implicit_metalayer_present) { - if (VK_SUCCESS != loaderGetDataFiles(inst, LOADER_DATA_FILE_MANIFEST_LAYER, true, "VK_LAYER_PATH", override_paths, + if (VK_SUCCESS != loaderGetDataFiles(inst, LOADER_DATA_FILE_MANIFEST_LAYER, true, VK_LAYER_PATH_ENV_VAR, override_paths, VK_ELAYERS_INFO_REGISTRY_LOC, VK_ELAYERS_INFO_RELATIVE_DIR, &manifest_files)) { goto out; } @@ -4689,6 +4979,7 @@ void loaderScanForImplicitLayers(struct loader_instance *inst, struct loader_lay res = loaderAddLayerProperties(inst, instance_layers, json, true, file_str); loader_instance_heap_free(inst, file_str); + manifest_files.filename_list[i] = NULL; cJSON_Delete(json); if (VK_ERROR_OUT_OF_HOST_MEMORY == res) { @@ -4713,6 +5004,11 @@ out: if (NULL != override_paths) { loader_instance_heap_free(inst, override_paths); } + for (uint32_t i = 0; i < manifest_files.count; i++) { + if (NULL != manifest_files.filename_list[i]) { + loader_instance_heap_free(inst, manifest_files.filename_list[i]); + } + } if (NULL != manifest_files.filename_list) { loader_instance_heap_free(inst, manifest_files.filename_list); } @@ -5349,9 +5645,11 @@ static void loaderAddImplicitLayers(const struct loader_instance *inst, struct l // Get the layer name(s) from the env_name environment variable. If layer is found in // search_list then add it to layer_list. But only add it to layer_list if type_flags matches. -static void loaderAddEnvironmentLayers(struct loader_instance *inst, const enum layer_type_flags type_flags, const char *env_name, - struct loader_layer_list *target_list, struct loader_layer_list *expanded_target_list, - const struct loader_layer_list *source_list) { +static VkResult loaderAddEnvironmentLayers(struct loader_instance *inst, const enum layer_type_flags type_flags, + const char *env_name, struct loader_layer_list *target_list, + struct loader_layer_list *expanded_target_list, + const struct loader_layer_list *source_list) { + VkResult res = VK_SUCCESS; char *next, *name; char *layer_env = loader_getenv(env_name, inst); if (layer_env == NULL) { @@ -5365,7 +5663,10 @@ static void loaderAddEnvironmentLayers(struct loader_instance *inst, const enum while (name && *name) { next = loader_get_next_path(name); - loaderAddLayerNameToList(inst, name, type_flags, source_list, target_list, expanded_target_list); + res = loaderAddLayerNameToList(inst, name, type_flags, source_list, target_list, expanded_target_list); + if (res != VK_SUCCESS) { + goto out; + } name = next; } @@ -5375,7 +5676,7 @@ out: loader_free_getenv(layer_env, inst); } - return; + return res; } VkResult loaderEnableInstanceLayers(struct loader_instance *inst, const VkInstanceCreateInfo *pCreateInfo, @@ -5404,8 +5705,11 @@ VkResult loaderEnableInstanceLayers(struct loader_instance *inst, const VkInstan loaderAddImplicitLayers(inst, &inst->app_activated_layer_list, &inst->expanded_activated_layer_list, instance_layers); // Add any layers specified via environment variable next - loaderAddEnvironmentLayers(inst, VK_LAYER_TYPE_FLAG_EXPLICIT_LAYER, "VK_INSTANCE_LAYERS", &inst->app_activated_layer_list, - &inst->expanded_activated_layer_list, instance_layers); + err = loaderAddEnvironmentLayers(inst, VK_LAYER_TYPE_FLAG_EXPLICIT_LAYER, "VK_INSTANCE_LAYERS", &inst->app_activated_layer_list, + &inst->expanded_activated_layer_list, instance_layers); + if (err != VK_SUCCESS) { + goto out; + } // Add layers specified by the application err = loaderAddLayerNamesToList(inst, &inst->app_activated_layer_list, &inst->expanded_activated_layer_list, @@ -5427,6 +5731,7 @@ VkResult loaderEnableInstanceLayers(struct loader_instance *inst, const VkInstan } } +out: return err; } @@ -5749,28 +6054,46 @@ VkResult loader_create_instance_chain(const VkInstanceCreateInfo *pCreateInfo, c } } + VkLoaderFeatureFlags feature_flags = 0; +#if defined(_WIN32) + IDXGIFactory6* dxgi_factory = NULL; + HRESULT hres = fpCreateDXGIFactory1(&IID_IDXGIFactory6, (void **)&dxgi_factory); + if (hres == S_OK) { + feature_flags |= VK_LOADER_FEATURE_PHYSICAL_DEVICE_SORTING; + dxgi_factory->lpVtbl->Release(dxgi_factory); + } +#endif + PFN_vkCreateInstance fpCreateInstance = (PFN_vkCreateInstance)next_gipa(*created_instance, "vkCreateInstance"); if (fpCreateInstance) { - VkLayerInstanceCreateInfo create_info_disp; - - create_info_disp.sType = VK_STRUCTURE_TYPE_LOADER_INSTANCE_CREATE_INFO; - create_info_disp.function = VK_LOADER_DATA_CALLBACK; - - create_info_disp.u.pfnSetInstanceLoaderData = vkSetInstanceDispatch; - - create_info_disp.pNext = loader_create_info.pNext; - loader_create_info.pNext = &create_info_disp; - - VkLayerInstanceCreateInfo create_info_disp2; - - create_info_disp2.sType = VK_STRUCTURE_TYPE_LOADER_INSTANCE_CREATE_INFO; - create_info_disp2.function = VK_LOADER_LAYER_CREATE_DEVICE_CALLBACK; - - create_info_disp2.u.layerDevice.pfnLayerCreateDevice = loader_layer_create_device; - create_info_disp2.u.layerDevice.pfnLayerDestroyDevice = loader_layer_destroy_device; - - create_info_disp2.pNext = loader_create_info.pNext; - loader_create_info.pNext = &create_info_disp2; + const VkLayerInstanceCreateInfo instance_dispatch = { + .sType = VK_STRUCTURE_TYPE_LOADER_INSTANCE_CREATE_INFO, + .pNext = loader_create_info.pNext, + .function = VK_LOADER_DATA_CALLBACK, + .u = { + .pfnSetInstanceLoaderData = vkSetInstanceDispatch, + }, + }; + const VkLayerInstanceCreateInfo device_callback = { + .sType = VK_STRUCTURE_TYPE_LOADER_INSTANCE_CREATE_INFO, + .pNext = &instance_dispatch, + .function = VK_LOADER_LAYER_CREATE_DEVICE_CALLBACK, + .u = { + .layerDevice = { + .pfnLayerCreateDevice = loader_layer_create_device, + .pfnLayerDestroyDevice = loader_layer_destroy_device, + }, + }, + }; + const VkLayerInstanceCreateInfo loader_features = { + .sType = VK_STRUCTURE_TYPE_LOADER_INSTANCE_CREATE_INFO, + .pNext = &device_callback, + .function = VK_LOADER_FEATURES, + .u = { + .loaderFeatures = feature_flags, + }, + }; + loader_create_info.pNext = &loader_features; res = fpCreateInstance(&loader_create_info, pAllocator, created_instance); } else { @@ -6011,8 +6334,11 @@ VkResult loader_validate_instance_extensions(struct loader_instance *inst, const // Build the lists of active layers (including metalayers) and expanded layers (with metalayers resolved to their components) loaderAddImplicitLayers(inst, &active_layers, &expanded_layers, instance_layers); - loaderAddEnvironmentLayers(inst, VK_LAYER_TYPE_FLAG_EXPLICIT_LAYER, ENABLED_LAYERS_ENV, &active_layers, &expanded_layers, - instance_layers); + res = loaderAddEnvironmentLayers(inst, VK_LAYER_TYPE_FLAG_EXPLICIT_LAYER, ENABLED_LAYERS_ENV, &active_layers, &expanded_layers, + instance_layers); + if (res != VK_SUCCESS) { + goto out; + } res = loaderAddLayerNamesToList(inst, &active_layers, &expanded_layers, pCreateInfo->enabledLayerCount, pCreateInfo->ppEnabledLayerNames, instance_layers); if (VK_SUCCESS != res) { @@ -6579,7 +6905,7 @@ VKAPI_ATTR VkResult VKAPI_CALL terminator_CreateDevice(VkPhysicalDevice physical if (!dev->extensions.khr_device_group_enabled) { VkPhysicalDeviceProperties properties; icd_term->dispatch.GetPhysicalDeviceProperties(phys_dev_term->phys_dev, &properties); - if (properties.apiVersion >= VK_MAKE_VERSION(1, 1, 0)) { + if (properties.apiVersion >= VK_API_VERSION_1_1) { dev->extensions.khr_device_group_enabled = true; } } @@ -6744,11 +7070,143 @@ out: return res; } +struct LoaderSortedPhysicalDevice { + uint32_t device_count; + VkPhysicalDevice* physical_devices; + uint32_t icd_index; + struct loader_icd_term* icd_term; +}; + +// This function allocates an array in sorted_devices which must be freed by the caller if not null +VkResult ReadSortedPhysicalDevices(struct loader_instance *inst, struct LoaderSortedPhysicalDevice **sorted_devices, uint32_t* sorted_count) +{ + VkResult res = VK_SUCCESS; + +#if defined(_WIN32) + uint32_t sorted_alloc = 0; + struct loader_icd_term *icd_term = NULL; + IDXGIFactory6* dxgi_factory = NULL; + HRESULT hres = fpCreateDXGIFactory1(&IID_IDXGIFactory6, (void **)&dxgi_factory); + if (hres != S_OK) { + loader_log(inst, VK_DEBUG_REPORT_INFORMATION_BIT_EXT, 0, "Failed to create DXGI factory 6. Physical devices will not be sorted"); + } + else { + sorted_alloc = 16; + *sorted_devices = loader_instance_heap_alloc(inst, sorted_alloc * sizeof(struct LoaderSortedPhysicalDevice), VK_SYSTEM_ALLOCATION_SCOPE_COMMAND); + if (*sorted_devices == NULL) { + res = VK_ERROR_OUT_OF_HOST_MEMORY; + goto out; + } + + memset(*sorted_devices, 0, sorted_alloc * sizeof(struct LoaderSortedPhysicalDevice)); + + *sorted_count = 0; + for (uint32_t i = 0; ; ++i) { + IDXGIAdapter1* adapter; + hres = dxgi_factory->lpVtbl->EnumAdapterByGpuPreference(dxgi_factory, i, DXGI_GPU_PREFERENCE_UNSPECIFIED, &IID_IDXGIAdapter1, (void **)&adapter); + if (hres == DXGI_ERROR_NOT_FOUND) { + break; // No more adapters + } + else if (hres != S_OK) { + loader_log(inst, VK_DEBUG_REPORT_WARNING_BIT_EXT, 0, "Failed to enumerate adapters by GPU preference at index %u. This adapter will not be sorted", i); + break; + } + + DXGI_ADAPTER_DESC1 description; + hres = adapter->lpVtbl->GetDesc1(adapter, &description); + if (hres != S_OK) { + loader_log(inst, VK_DEBUG_REPORT_WARNING_BIT_EXT, 0, "Failed to get adapter LUID index %u. This adapter will not be sorted", i); + continue; + } + + if (sorted_alloc <= i) { + uint32_t old_size = sorted_alloc * sizeof(struct LoaderSortedPhysicalDevice); + *sorted_devices = loader_instance_heap_realloc(inst, *sorted_devices, old_size, 2 * old_size, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND); + if (*sorted_devices == NULL) { + adapter->lpVtbl->Release(adapter); + res = VK_ERROR_OUT_OF_HOST_MEMORY; + goto out; + } + sorted_alloc *= 2; + } + struct LoaderSortedPhysicalDevice *sorted_array = *sorted_devices; + sorted_array[*sorted_count].device_count = 0; + sorted_array[*sorted_count].physical_devices = NULL; + //*sorted_count = i; + + icd_term = inst->icd_terms; + for (uint32_t icd_idx = 0; NULL != icd_term; icd_term = icd_term->next, icd_idx++) { + // This is the new behavior, which cannot be run unless the ICD provides EnumerateAdapterPhysicalDevices + if (icd_term->scanned_icd->EnumerateAdapterPhysicalDevices == NULL) { + continue; + } + + uint32_t count; + VkResult vkres = icd_term->scanned_icd->EnumerateAdapterPhysicalDevices(icd_term->instance, description.AdapterLuid, &count, NULL); + if (vkres == VK_ERROR_INCOMPATIBLE_DRIVER) { + continue; // This driver doesn't support the adapter + } else if (vkres == VK_ERROR_OUT_OF_HOST_MEMORY) { + res = VK_ERROR_OUT_OF_HOST_MEMORY; + goto out; + } else if (vkres != VK_SUCCESS) { + loader_log(inst, VK_DEBUG_REPORT_WARNING_BIT_EXT, 0, "Failed to convert DXGI adapter into Vulkan physical device with unexpected error code"); + continue; + } + + // Get the actual physical devices + if (0 != count) + { + do { + sorted_array[*sorted_count].physical_devices = loader_instance_heap_realloc(inst, sorted_array[*sorted_count].physical_devices, sorted_array[*sorted_count].device_count * sizeof(VkPhysicalDevice), count * sizeof(VkPhysicalDevice), VK_SYSTEM_ALLOCATION_SCOPE_COMMAND); + if (sorted_array[*sorted_count].physical_devices == NULL) { + res = VK_ERROR_OUT_OF_HOST_MEMORY; + break; + } + sorted_array[*sorted_count].device_count = count; + } while ((vkres = icd_term->scanned_icd->EnumerateAdapterPhysicalDevices(icd_term->instance, description.AdapterLuid, &count, sorted_array[*sorted_count].physical_devices)) == VK_INCOMPLETE); + } + + if (vkres != VK_SUCCESS) { + loader_instance_heap_free(inst, sorted_array[*sorted_count].physical_devices); + sorted_array[*sorted_count].physical_devices = NULL; + if (vkres == VK_ERROR_OUT_OF_HOST_MEMORY) { + res = VK_ERROR_OUT_OF_HOST_MEMORY; + goto out; + } else { + loader_log(inst, VK_DEBUG_REPORT_WARNING_BIT_EXT, 0, + "Failed to convert DXGI adapter into Vulkan physical device"); + continue; + } + } + inst->total_gpu_count += (sorted_array[*sorted_count].device_count = count); + sorted_array[*sorted_count].icd_index = icd_idx; + sorted_array[*sorted_count].icd_term = icd_term; + ++(*sorted_count); + } + + adapter->lpVtbl->Release(adapter); + } + + dxgi_factory->lpVtbl->Release(dxgi_factory); + } + +out: +#endif + + if (*sorted_count == 0 && *sorted_devices != NULL) { + loader_instance_heap_free(inst, *sorted_devices); + *sorted_devices = NULL; + } + return res; +} + VkResult setupLoaderTermPhysDevs(struct loader_instance *inst) { VkResult res = VK_SUCCESS; struct loader_icd_term *icd_term; struct loader_phys_dev_per_icd *icd_phys_dev_array = NULL; struct loader_physical_device_term **new_phys_devs = NULL; + struct LoaderSortedPhysicalDevice *sorted_phys_dev_array = NULL; + uint32_t sorted_count = 0; inst->total_gpu_count = 0; @@ -6765,11 +7223,29 @@ VkResult setupLoaderTermPhysDevs(struct loader_instance *inst) { goto out; } memset(icd_phys_dev_array, 0, sizeof(struct loader_phys_dev_per_icd) * inst->total_icd_count); - icd_term = inst->icd_terms; + + // Get the physical devices supported by platform sorting mechanism into a separate list + res = ReadSortedPhysicalDevices(inst, &sorted_phys_dev_array, &sorted_count); + if (VK_SUCCESS != res) { + goto out; + } // For each ICD, query the number of physical devices, and then get an // internal value for those physical devices. + icd_term = inst->icd_terms; for (uint32_t icd_idx = 0; NULL != icd_term; icd_term = icd_term->next, icd_idx++) { + icd_phys_dev_array[icd_idx].count = 0; + icd_phys_dev_array[icd_idx].phys_devs = NULL; + icd_phys_dev_array[icd_idx].this_icd_term = NULL; + + // This is the legacy behavior which should be skipped if EnumerateAdapterPhysicalDevices is available + // and we successfully enumerated sorted adapters using ReadSortedPhysicalDevices. +#if defined(VK_USE_PLATFORM_WIN32_KHR) + if (sorted_count && icd_term->scanned_icd->EnumerateAdapterPhysicalDevices != NULL) { + continue; + } +#endif + res = icd_term->dispatch.EnumeratePhysicalDevices(icd_term->instance, &icd_phys_dev_array[icd_idx].count, NULL); if (VK_SUCCESS != res) { loader_log(inst, VK_DEBUG_REPORT_ERROR_BIT_EXT, 0, @@ -6822,6 +7298,48 @@ VkResult setupLoaderTermPhysDevs(struct loader_instance *inst) { // Copy or create everything to fill the new array of physical devices uint32_t idx = 0; + +#if defined(_WIN32) + // Copy over everything found through sorted enumeration + for (uint32_t i = 0; i < sorted_count; ++i) { + for (uint32_t j = 0; j < sorted_phys_dev_array[i].device_count; ++j) { + // Check if this physical device is already in the old buffer + if (NULL != inst->phys_devs_term) { + for (uint32_t old_idx = 0; old_idx < inst->phys_dev_count_term; old_idx++) { + if (sorted_phys_dev_array[i].physical_devices[j] == inst->phys_devs_term[old_idx]->phys_dev) { + new_phys_devs[idx] = inst->phys_devs_term[old_idx]; + break; + } + } + } + + // If this physical device isn't in the old buffer, then we need to create it. + if (NULL == new_phys_devs[idx]) { + new_phys_devs[idx] = loader_instance_heap_alloc(inst, sizeof(struct loader_physical_device_term), + VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE); + if (NULL == new_phys_devs[idx]) { + loader_log(inst, VK_DEBUG_REPORT_ERROR_BIT_EXT, 0, + "setupLoaderTermPhysDevs: Failed to allocate " + "physical device terminator object %d", + idx); + inst->total_gpu_count = idx; + res = VK_ERROR_OUT_OF_HOST_MEMORY; + goto out; + } + + loader_set_dispatch((void *)new_phys_devs[idx], inst->disp); + new_phys_devs[idx]->this_icd_term = sorted_phys_dev_array[i].icd_term; + new_phys_devs[idx]->icd_index = (uint8_t)(sorted_phys_dev_array[i].icd_index); + new_phys_devs[idx]->phys_dev = sorted_phys_dev_array[i].physical_devices[j]; + } + + // Increment the count of new physical devices + idx++; + } + } +#endif + + // Copy over everything found through EnumeratePhysicalDevices for (uint32_t icd_idx = 0; icd_idx < inst->total_icd_count; icd_idx++) { for (uint32_t pd_idx = 0; pd_idx < icd_phys_dev_array[icd_idx].count; pd_idx++) { // Check if this physical device is already in the old buffer @@ -6893,6 +7411,15 @@ out: inst->phys_devs_term = new_phys_devs; } + if (sorted_phys_dev_array != NULL) { + for (uint32_t i = 0; i < sorted_count; ++i) { + if (sorted_phys_dev_array[i].device_count > 0 && sorted_phys_dev_array[i].physical_devices != NULL) { + loader_instance_heap_free(inst, sorted_phys_dev_array[i].physical_devices); + } + } + loader_instance_heap_free(inst, sorted_phys_dev_array); + } + return res; } @@ -7036,7 +7563,6 @@ VKAPI_ATTR VkResult VKAPI_CALL terminator_EnumerateDeviceExtensionProperties(VkP if (pProperties == NULL) { *pPropertyCount = count; loader_destroy_generic_list(inst, (struct loader_generic_list *)&local_ext_list); - loader_platform_thread_unlock_mutex(&loader_lock); return VK_SUCCESS; } @@ -7048,14 +7574,12 @@ VKAPI_ATTR VkResult VKAPI_CALL terminator_EnumerateDeviceExtensionProperties(VkP loader_destroy_generic_list(inst, (struct loader_generic_list *)&local_ext_list); if (copy_size < count) { - loader_platform_thread_unlock_mutex(&loader_lock); return VK_INCOMPLETE; } } else { loader_log(inst, VK_DEBUG_REPORT_ERROR_BIT_EXT, 0, "vkEnumerateDeviceExtensionProperties: pLayerName " "is too long or is badly formed"); - loader_platform_thread_unlock_mutex(&loader_lock); return VK_ERROR_EXTENSION_NOT_PRESENT; } @@ -7065,10 +7589,26 @@ VKAPI_ATTR VkResult VKAPI_CALL terminator_EnumerateDeviceExtensionProperties(VkP // This case is during the call down the instance chain with pLayerName == NULL struct loader_icd_term *icd_term = phys_dev_term->this_icd_term; uint32_t icd_ext_count = *pPropertyCount; + VkExtensionProperties *icd_props_list = pProperties; VkResult res; - // Get the available device extensions - res = icd_term->dispatch.EnumerateDeviceExtensionProperties(phys_dev_term->phys_dev, NULL, &icd_ext_count, pProperties); + if (NULL == icd_props_list) { + // We need to find the count without duplicates. This requires querying the driver for the names of the extensions. + // A small amount of storage is then needed to facilitate the de-duplication. + res = icd_term->dispatch.EnumerateDeviceExtensionProperties(phys_dev_term->phys_dev, NULL, &icd_ext_count, NULL); + if (res != VK_SUCCESS) { + goto out; + } + icd_props_list = loader_instance_heap_alloc(icd_term->this_instance, sizeof(VkExtensionProperties) * icd_ext_count, + VK_SYSTEM_ALLOCATION_SCOPE_COMMAND); + if (NULL == icd_props_list) { + res = VK_ERROR_OUT_OF_HOST_MEMORY; + goto out; + } + } + + // Get the available device extension count, and if pProperties is not NULL, the extensions as well + res = icd_term->dispatch.EnumerateDeviceExtensionProperties(phys_dev_term->phys_dev, NULL, &icd_ext_count, icd_props_list); if (res != VK_SUCCESS) { goto out; } @@ -7079,36 +7619,36 @@ VKAPI_ATTR VkResult VKAPI_CALL terminator_EnumerateDeviceExtensionProperties(VkP } loaderAddImplicitLayers(icd_term->this_instance, &implicit_layer_list, NULL, &icd_term->this_instance->instance_layer_list); + + // Initialize dev_extension list within the physicalDevice object + res = loader_init_device_extensions(icd_term->this_instance, phys_dev_term, icd_ext_count, icd_props_list, &icd_exts); + if (res != VK_SUCCESS) { + goto out; + } + // We need to determine which implicit layers are active, and then add their extensions. This can't be cached as // it depends on results of environment variables (which can change). - if (pProperties != NULL) { - // Initialize dev_extension list within the physicalDevice object - res = loader_init_device_extensions(icd_term->this_instance, phys_dev_term, icd_ext_count, pProperties, &icd_exts); - if (res != VK_SUCCESS) { - goto out; - } - - // We need to determine which implicit layers are active, and then add their extensions. This can't be cached as - // it depends on results of environment variables (which can change). - res = loader_add_to_ext_list(icd_term->this_instance, &all_exts, icd_exts.count, icd_exts.list); - if (res != VK_SUCCESS) { - goto out; - } + res = loader_add_to_ext_list(icd_term->this_instance, &all_exts, icd_exts.count, icd_exts.list); + if (res != VK_SUCCESS) { + goto out; + } - loaderAddImplicitLayers(icd_term->this_instance, &implicit_layer_list, NULL, &icd_term->this_instance->instance_layer_list); + loaderAddImplicitLayers(icd_term->this_instance, &implicit_layer_list, NULL, &icd_term->this_instance->instance_layer_list); - for (uint32_t i = 0; i < implicit_layer_list.count; i++) { - for (uint32_t j = 0; j < implicit_layer_list.list[i].device_extension_list.count; j++) { - res = loader_add_to_ext_list(icd_term->this_instance, &all_exts, 1, - &implicit_layer_list.list[i].device_extension_list.list[j].props); - if (res != VK_SUCCESS) { - goto out; - } + for (uint32_t i = 0; i < implicit_layer_list.count; i++) { + for (uint32_t j = 0; j < implicit_layer_list.list[i].device_extension_list.count; j++) { + res = loader_add_to_ext_list(icd_term->this_instance, &all_exts, 1, + &implicit_layer_list.list[i].device_extension_list.list[j].props); + if (res != VK_SUCCESS) { + goto out; } } - uint32_t capacity = *pPropertyCount; - VkExtensionProperties *props = pProperties; + } + uint32_t capacity = *pPropertyCount; + VkExtensionProperties *props = pProperties; + res = VK_SUCCESS; + if (NULL != pProperties) { for (uint32_t i = 0; i < all_exts.count && i < capacity; i++) { props[i] = all_exts.list[i]; } @@ -7120,14 +7660,7 @@ VKAPI_ATTR VkResult VKAPI_CALL terminator_EnumerateDeviceExtensionProperties(VkP *pPropertyCount = all_exts.count; } } else { - // Just return the count; need to add in the count of implicit layer extensions - // don't worry about duplicates being added in the count - *pPropertyCount = icd_ext_count; - - for (uint32_t i = 0; i < implicit_layer_list.count; i++) { - *pPropertyCount += implicit_layer_list.list[i].device_extension_list.count; - } - res = VK_SUCCESS; + *pPropertyCount = all_exts.count; } out: @@ -7141,7 +7674,9 @@ out: if (NULL != icd_exts.list) { loader_destroy_generic_list(icd_term->this_instance, (struct loader_generic_list *)&icd_exts); } - + if (NULL == pProperties && NULL != icd_props_list) { + loader_instance_heap_free(icd_term->this_instance, icd_props_list); + } return res; } @@ -7201,7 +7736,7 @@ VKAPI_ATTR VkResult VKAPI_CALL terminator_EnumerateInstanceVersion(const VkEnumerateInstanceVersionChain *chain, uint32_t* pApiVersion) { // NOTE: The Vulkan WG doesn't want us checking pApiVersion for NULL, but instead // prefers us crashing. - *pApiVersion = VK_MAKE_VERSION(loader_major_version, loader_minor_version, VK_HEADER_VERSION); + *pApiVersion = VK_HEADER_VERSION_COMPLETE; return VK_SUCCESS; } @@ -7238,10 +7773,14 @@ terminator_EnumerateInstanceExtensionProperties(const VkEnumerateInstanceExtensi } } } else { + // Preload ICD libraries so subsequent calls to EnumerateInstanceExtensionProperties don't have to load them + loader_preload_icds(); + // Scan/discover all ICD libraries memset(&icd_tramp_list, 0, sizeof(icd_tramp_list)); res = loader_icd_scan(NULL, &icd_tramp_list); - if (VK_SUCCESS != res) { + // EnumerateInstanceExtensionProperties can't return anything other than OOM or VK_ERROR_LAYER_NOT_PRESENT + if ((VK_SUCCESS != res && icd_tramp_list.count > 0) || res == VK_ERROR_OUT_OF_HOST_MEMORY) { goto out; } // Get extensions from all ICD's, merge so no duplicates @@ -7362,7 +7901,10 @@ VkResult setupLoaderTermPhysDevGroups(struct loader_instance *inst) { uint32_t cur_icd_group_count = 0; VkPhysicalDeviceGroupPropertiesKHR **new_phys_dev_groups = NULL; VkPhysicalDeviceGroupPropertiesKHR *local_phys_dev_groups = NULL; + bool *local_phys_dev_group_sorted = NULL; PFN_vkEnumeratePhysicalDeviceGroups fpEnumeratePhysicalDeviceGroups = NULL; + struct LoaderSortedPhysicalDevice* sorted_phys_dev_array = NULL; + uint32_t sorted_count = 0; if (0 == inst->phys_dev_count_term) { loader_log(inst, VK_DEBUG_REPORT_ERROR_BIT_EXT, 0, @@ -7426,7 +7968,8 @@ VkResult setupLoaderTermPhysDevGroups(struct loader_instance *inst) { // Create a temporary array (on the stack) to keep track of the // returned VkPhysicalDevice values. local_phys_dev_groups = loader_stack_alloc(sizeof(VkPhysicalDeviceGroupProperties) * total_count); - if (NULL == local_phys_dev_groups) { + local_phys_dev_group_sorted = loader_stack_alloc(sizeof(bool) * total_count); + if (NULL == local_phys_dev_groups || NULL == local_phys_dev_group_sorted) { loader_log(inst, VK_DEBUG_REPORT_ERROR_BIT_EXT, 0, "setupLoaderTermPhysDevGroups: Failed to allocate local " "physical device group array of size %d", @@ -7436,17 +7979,31 @@ VkResult setupLoaderTermPhysDevGroups(struct loader_instance *inst) { } // Initialize the memory to something valid memset(local_phys_dev_groups, 0, sizeof(VkPhysicalDeviceGroupProperties) * total_count); + memset(local_phys_dev_group_sorted, 0, sizeof(bool) * total_count); for (uint32_t group = 0; group < total_count; group++) { local_phys_dev_groups[group].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_GROUP_PROPERTIES_KHR; local_phys_dev_groups[group].pNext = NULL; local_phys_dev_groups[group].subsetAllocation = false; } + // Get the physical devices supported by platform sorting mechanism into a separate list + res = ReadSortedPhysicalDevices(inst, &sorted_phys_dev_array, &sorted_count); + if (VK_SUCCESS != res) { + goto out; + } + cur_icd_group_count = 0; icd_term = inst->icd_terms; for (uint32_t icd_idx = 0; NULL != icd_term; icd_term = icd_term->next, icd_idx++) { uint32_t count_this_time = total_count - cur_icd_group_count; + // Check if this group can be sorted +#if defined(VK_USE_PLATFORM_WIN32_KHR) + bool icd_sorted = sorted_count && (icd_term->scanned_icd->EnumerateAdapterPhysicalDevices != NULL); +#else + bool icd_sorted = false; +#endif + // Get the function pointer to use to call into the ICD. This could be the core or KHR version if (inst->enabled_known_extensions.khr_device_group_creation) { fpEnumeratePhysicalDeviceGroups = icd_term->dispatch.EnumeratePhysicalDeviceGroupsKHR; @@ -7478,10 +8035,14 @@ VkResult setupLoaderTermPhysDevGroups(struct loader_instance *inst) { for (uint32_t indiv_gpu = 0; indiv_gpu < count_this_time; indiv_gpu++) { local_phys_dev_groups[indiv_gpu + cur_icd_group_count].physicalDeviceCount = 1; local_phys_dev_groups[indiv_gpu + cur_icd_group_count].physicalDevices[0] = phys_dev_array[indiv_gpu]; + local_phys_dev_group_sorted[indiv_gpu + cur_icd_group_count] = icd_sorted; } } else { res = fpEnumeratePhysicalDeviceGroups(icd_term->instance, &count_this_time, &local_phys_dev_groups[cur_icd_group_count]); + for (uint32_t group = 0; group < count_this_time; ++group) { + local_phys_dev_group_sorted[group + cur_icd_group_count] = icd_sorted; + } if (VK_SUCCESS != res) { loader_log(inst, VK_DEBUG_REPORT_ERROR_BIT_EXT, 0, "setupLoaderTermPhysDevGroups: Failed during dispatch call of " @@ -7516,8 +8077,87 @@ VkResult setupLoaderTermPhysDevGroups(struct loader_instance *inst) { } } + uint32_t idx = 0; + +#if defined(_WIN32) + // Copy over everything found through sorted enumeration + for (uint32_t i = 0; i < sorted_count; ++i) { + + // Find the VkPhysicalDeviceGroupProperties object in local_phys_dev_groups + VkPhysicalDeviceGroupProperties *group_properties = NULL; + for (uint32_t group = 0; group < total_count; group++) { + if (sorted_phys_dev_array[i].device_count != local_phys_dev_groups[group].physicalDeviceCount) { + continue; + } + + bool match = true; + for (uint32_t group_gpu = 0; group_gpu < local_phys_dev_groups[group].physicalDeviceCount; group_gpu++) { + if (sorted_phys_dev_array[i].physical_devices[group_gpu] != ((struct loader_physical_device_term*) local_phys_dev_groups[group].physicalDevices[group_gpu])->phys_dev) { + match = false; + break; + } + } + + if (match) { + group_properties = &local_phys_dev_groups[group]; + } + } + + // Check if this physical device group with the same contents is already in the old buffer + for (uint32_t old_idx = 0; old_idx < inst->phys_dev_group_count_term; old_idx++) { + if (NULL != group_properties && group_properties->physicalDeviceCount == inst->phys_dev_groups_term[old_idx]->physicalDeviceCount) { + bool found_all_gpus = true; + for (uint32_t old_gpu = 0; old_gpu < inst->phys_dev_groups_term[old_idx]->physicalDeviceCount; old_gpu++) { + bool found_gpu = false; + for (uint32_t new_gpu = 0; new_gpu < group_properties->physicalDeviceCount; new_gpu++) { + if (group_properties->physicalDevices[new_gpu] == inst->phys_dev_groups_term[old_idx]->physicalDevices[old_gpu]) { + found_gpu = true; + break; + } + } + + if (!found_gpu) { + found_all_gpus = false; + break; + } + } + if (!found_all_gpus) { + continue; + } + else { + new_phys_dev_groups[idx] = inst->phys_dev_groups_term[old_idx]; + break; + } + } + } + + // If this physical device group isn't in the old buffer, create it + if (group_properties != NULL && NULL == new_phys_dev_groups[idx]) { + new_phys_dev_groups[idx] = (VkPhysicalDeviceGroupPropertiesKHR*)loader_instance_heap_alloc( + inst, sizeof(VkPhysicalDeviceGroupPropertiesKHR), VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE); + if (NULL == new_phys_dev_groups[idx]) { + loader_log(inst, VK_DEBUG_REPORT_ERROR_BIT_EXT, 0, + "setupLoaderTermPhysDevGroups: Failed to allocate " + "physical device group Terminator object %d", + idx); + total_count = idx; + res = VK_ERROR_OUT_OF_HOST_MEMORY; + goto out; + } + memcpy(new_phys_dev_groups[idx], group_properties, sizeof(VkPhysicalDeviceGroupPropertiesKHR)); + } + + ++idx; + } +#endif + // Copy or create everything to fill the new array of physical device groups for (uint32_t new_idx = 0; new_idx < total_count; new_idx++) { + // Skip groups which have been included through sorting + if (local_phys_dev_group_sorted[new_idx] || local_phys_dev_groups[new_idx].physicalDeviceCount == 0) { + continue; + } + // Check if this physical device group with the same contents is already in the old buffer for (uint32_t old_idx = 0; old_idx < inst->phys_dev_group_count_term; old_idx++) { if (local_phys_dev_groups[new_idx].physicalDeviceCount == inst->phys_dev_groups_term[old_idx]->physicalDeviceCount) { @@ -7539,28 +8179,30 @@ VkResult setupLoaderTermPhysDevGroups(struct loader_instance *inst) { if (!found_all_gpus) { continue; } else { - new_phys_dev_groups[new_idx] = inst->phys_dev_groups_term[old_idx]; + new_phys_dev_groups[idx] = inst->phys_dev_groups_term[old_idx]; break; } } } // If this physical device group isn't in the old buffer, create it - if (NULL == new_phys_dev_groups[new_idx]) { - new_phys_dev_groups[new_idx] = (VkPhysicalDeviceGroupPropertiesKHR *)loader_instance_heap_alloc( + if (NULL == new_phys_dev_groups[idx]) { + new_phys_dev_groups[idx] = (VkPhysicalDeviceGroupPropertiesKHR *)loader_instance_heap_alloc( inst, sizeof(VkPhysicalDeviceGroupPropertiesKHR), VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE); - if (NULL == new_phys_dev_groups[new_idx]) { + if (NULL == new_phys_dev_groups[idx]) { loader_log(inst, VK_DEBUG_REPORT_ERROR_BIT_EXT, 0, "setupLoaderTermPhysDevGroups: Failed to allocate " "physical device group Terminator object %d", - new_idx); - total_count = new_idx; + idx); + total_count = idx; res = VK_ERROR_OUT_OF_HOST_MEMORY; goto out; } - memcpy(new_phys_dev_groups[new_idx], &local_phys_dev_groups[new_idx], + memcpy(new_phys_dev_groups[idx], &local_phys_dev_groups[new_idx], sizeof(VkPhysicalDeviceGroupPropertiesKHR)); } + + ++idx; } out: @@ -7597,6 +8239,15 @@ out: inst->phys_dev_groups_term = new_phys_dev_groups; } + if (sorted_phys_dev_array != NULL) { + for (uint32_t i = 0; i < sorted_count; ++i) { + if (sorted_phys_dev_array[i].device_count > 0 && sorted_phys_dev_array[i].physical_devices != NULL) { + loader_instance_heap_free(inst, sorted_phys_dev_array[i].physical_devices); + } + } + loader_instance_heap_free(inst, sorted_phys_dev_array); + } + return res; } |