metal : fix memory leaks due to missing autoreleasepools (#27758)
This commit is contained in:
Niklas Wenzel
@@ -84,106 +84,108 @@ struct ggml_metal {
|
||||
ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
|
||||
GGML_LOG_INFO("%s: allocating\n", __func__);
|
||||
|
||||
@autoreleasepool {
|
||||
#if TARGET_OS_OSX && !GGML_METAL_NDEBUG
|
||||
// Show all the Metal device instances in the system
|
||||
NSArray * devices = MTLCopyAllDevices();
|
||||
for (id<MTLDevice> device in devices) {
|
||||
GGML_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]);
|
||||
}
|
||||
[devices release]; // since it was created by a *Copy* C method
|
||||
// Show all the Metal device instances in the system
|
||||
NSArray * devices = MTLCopyAllDevices();
|
||||
for (id<MTLDevice> device in devices) {
|
||||
GGML_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]);
|
||||
}
|
||||
[devices release]; // since it was created by a *Copy* C method
|
||||
#endif
|
||||
|
||||
// init context
|
||||
ggml_metal_t res = calloc(1, sizeof(struct ggml_metal));
|
||||
// init context
|
||||
ggml_metal_t res = calloc(1, sizeof(struct ggml_metal));
|
||||
|
||||
id<MTLDevice> device = ggml_metal_device_get_obj(dev);
|
||||
id<MTLDevice> device = ggml_metal_device_get_obj(dev);
|
||||
|
||||
GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
|
||||
|
||||
// TODO: would it be better to have one queue for the backend and one queue for the device?
|
||||
// the graph encoders and async ops would use the backend queue while the sync ops would use the device queue?
|
||||
//res->queue = [device newCommandQueue]; [TAG_QUEUE_PER_BACKEND]
|
||||
id<MTLCommandQueue> queue = ggml_metal_device_get_queue(dev);
|
||||
if (queue == nil) {
|
||||
GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
res->dev = dev;
|
||||
res->lib = ggml_metal_device_get_library(dev);
|
||||
if (res->lib == NULL) {
|
||||
GGML_LOG_WARN("%s: the device does not have a precompiled Metal library - this is unexpected\n", __func__);
|
||||
GGML_LOG_WARN("%s: will try to compile it on the fly\n", __func__);
|
||||
|
||||
res->lib = ggml_metal_library_init(dev);
|
||||
if (res->lib == NULL) {
|
||||
GGML_LOG_ERROR("%s: error: failed to initialize the Metal library\n", __func__);
|
||||
|
||||
free(res);
|
||||
GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
|
||||
|
||||
// TODO: would it be better to have one queue for the backend and one queue for the device?
|
||||
// the graph encoders and async ops would use the backend queue while the sync ops would use the device queue?
|
||||
//res->queue = [device newCommandQueue]; [TAG_QUEUE_PER_BACKEND]
|
||||
id<MTLCommandQueue> queue = ggml_metal_device_get_queue(dev);
|
||||
if (queue == nil) {
|
||||
GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__);
|
||||
return NULL;
|
||||
}
|
||||
}
|
||||
|
||||
res->ev_cpy = ggml_metal_device_event_init(dev);
|
||||
res->dev = dev;
|
||||
res->lib = ggml_metal_device_get_library(dev);
|
||||
if (res->lib == NULL) {
|
||||
GGML_LOG_WARN("%s: the device does not have a precompiled Metal library - this is unexpected\n", __func__);
|
||||
GGML_LOG_WARN("%s: will try to compile it on the fly\n", __func__);
|
||||
|
||||
const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev);
|
||||
res->lib = ggml_metal_library_init(dev);
|
||||
if (res->lib == NULL) {
|
||||
GGML_LOG_ERROR("%s: error: failed to initialize the Metal library\n", __func__);
|
||||
|
||||
snprintf(res->name, sizeof(res->name), "%s", props_dev->name);
|
||||
free(res);
|
||||
|
||||
res->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
|
||||
|
||||
res->use_fusion = getenv("GGML_METAL_FUSION_DISABLE") == nil;
|
||||
res->use_concurrency = getenv("GGML_METAL_CONCURRENCY_DISABLE") == nil;
|
||||
|
||||
{
|
||||
const char * val = getenv("GGML_METAL_GRAPH_DEBUG");
|
||||
res->debug_graph = val ? atoi(val) : 0;
|
||||
}
|
||||
|
||||
{
|
||||
const char * val = getenv("GGML_METAL_FUSION_DEBUG");
|
||||
res->debug_fusion = val ? atoi(val) : 0;
|
||||
}
|
||||
|
||||
res->use_graph_optimize = true;
|
||||
|
||||
if (getenv("GGML_METAL_GRAPH_OPTIMIZE_DISABLE") != NULL) {
|
||||
res->use_graph_optimize = false;
|
||||
}
|
||||
|
||||
memset(res->fuse_cnt, 0, sizeof(res->fuse_cnt));
|
||||
|
||||
GGML_LOG_INFO("%s: use fusion = %s\n", __func__, res->use_fusion ? "true" : "false");
|
||||
GGML_LOG_INFO("%s: use concurrency = %s\n", __func__, res->use_concurrency ? "true" : "false");
|
||||
GGML_LOG_INFO("%s: use graph optimize = %s\n", __func__, res->use_graph_optimize ? "true" : "false");
|
||||
|
||||
res->capture_compute = 0;
|
||||
res->capture_started = false;
|
||||
res->capture_scope = nil;
|
||||
|
||||
{
|
||||
const char * val = getenv("GGML_METAL_CAPTURE_COMPUTE");
|
||||
if (val) {
|
||||
res->capture_compute = atoi(val);
|
||||
return NULL;
|
||||
}
|
||||
}
|
||||
|
||||
res->ev_cpy = ggml_metal_device_event_init(dev);
|
||||
|
||||
const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev);
|
||||
|
||||
snprintf(res->name, sizeof(res->name), "%s", props_dev->name);
|
||||
|
||||
res->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
|
||||
|
||||
res->use_fusion = getenv("GGML_METAL_FUSION_DISABLE") == nil;
|
||||
res->use_concurrency = getenv("GGML_METAL_CONCURRENCY_DISABLE") == nil;
|
||||
|
||||
{
|
||||
const char * val = getenv("GGML_METAL_GRAPH_DEBUG");
|
||||
res->debug_graph = val ? atoi(val) : 0;
|
||||
}
|
||||
|
||||
{
|
||||
const char * val = getenv("GGML_METAL_FUSION_DEBUG");
|
||||
res->debug_fusion = val ? atoi(val) : 0;
|
||||
}
|
||||
|
||||
res->use_graph_optimize = true;
|
||||
|
||||
if (getenv("GGML_METAL_GRAPH_OPTIMIZE_DISABLE") != NULL) {
|
||||
res->use_graph_optimize = false;
|
||||
}
|
||||
|
||||
memset(res->fuse_cnt, 0, sizeof(res->fuse_cnt));
|
||||
|
||||
GGML_LOG_INFO("%s: use fusion = %s\n", __func__, res->use_fusion ? "true" : "false");
|
||||
GGML_LOG_INFO("%s: use concurrency = %s\n", __func__, res->use_concurrency ? "true" : "false");
|
||||
GGML_LOG_INFO("%s: use graph optimize = %s\n", __func__, res->use_graph_optimize ? "true" : "false");
|
||||
|
||||
res->capture_compute = 0;
|
||||
res->capture_started = false;
|
||||
res->capture_scope = nil;
|
||||
|
||||
{
|
||||
const char * val = getenv("GGML_METAL_CAPTURE_COMPUTE");
|
||||
if (val) {
|
||||
res->capture_compute = atoi(val);
|
||||
}
|
||||
}
|
||||
|
||||
res->has_error = false;
|
||||
|
||||
res->gf = nil;
|
||||
res->encode_async = nil;
|
||||
for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) {
|
||||
res->cmd_bufs[i].obj = nil;
|
||||
}
|
||||
|
||||
res->cmd_bufs_ext = [[NSMutableArray alloc] init];
|
||||
|
||||
res->cmd_buf_last = nil;
|
||||
|
||||
res->pipelines_ext = ggml_metal_pipelines_init();
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
res->has_error = false;
|
||||
|
||||
res->gf = nil;
|
||||
res->encode_async = nil;
|
||||
for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) {
|
||||
res->cmd_bufs[i].obj = nil;
|
||||
}
|
||||
|
||||
res->cmd_bufs_ext = [[NSMutableArray alloc] init];
|
||||
|
||||
res->cmd_buf_last = nil;
|
||||
|
||||
res->pipelines_ext = ggml_metal_pipelines_init();
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
void ggml_metal_free(ggml_metal_t ctx) {
|
||||
|
||||
Reference in New Issue
Block a user