ggml : add metal backend registry / device

ggml/include/ggml-backend.h

@@ -127,6 +127,8 @@ extern "C" {
         bool async;
         // pinned host buffer
         bool host_buffer;
+        // creating buffers from host ptr
+        bool buffer_from_host_ptr;
         // event synchronization
         bool events;
     };

ggml/include/ggml-metal.h

@@ -43,7 +43,9 @@ GGML_API ggml_backend_t ggml_backend_metal_init(void);
 
 GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
+GGML_DEPRECATED(
+        GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
+        "obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713");
 
 GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 
@@ -57,6 +59,8 @@ GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int fam
 // capture all command buffers committed the next time `ggml_backend_graph_compute` is called
 GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
 
+GGML_API ggml_backend_reg_t ggml_backend_metal_reg(void);
+
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-backend.cpp

@@ -463,6 +463,7 @@ enum ggml_backend_dev_type ggml_backend_dev_type(ggml_backend_dev_t device) {
 }
 
 void ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props) {
+    memset(props, 0, sizeof(*props));
     device->iface.get_props(device, props);
 }
 
@@ -479,6 +480,10 @@ ggml_backend_buffer_type_t ggml_backend_dev_buffer_type(ggml_backend_dev_t devic
 }
 
 ggml_backend_buffer_type_t ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device) {
+    if (device->iface.get_host_buffer_type == NULL) {
+        return NULL;
+    }
+
     return device->iface.get_host_buffer_type(device);
 }
 
@@ -525,6 +530,10 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
 #include "ggml-cuda.h"
 #endif
 
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
 struct ggml_backend_registry {
     std::vector<ggml_backend_reg_t> backends;
     std::vector<ggml_backend_dev_t> devices;
@@ -533,10 +542,13 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_CUDA
         register_backend(ggml_backend_cuda_reg());
 #endif
+#ifdef GGML_USE_METAL
+        register_backend(ggml_backend_metal_reg());
+#endif
 
         register_backend(ggml_backend_cpu_reg());
 
-        // TODO: sycl, metal, vulkan, kompute, cann
+        // TODO: sycl, vulkan, kompute, cann
     }
 
     void register_backend(ggml_backend_reg_t reg) {
@@ -1118,9 +1130,10 @@ static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggm
     props->type        = ggml_backend_cpu_device_get_type(dev);
     ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
     props->caps = {
-        /* async       */ false,
-        /* host_buffer */ false,
-        /* events      */ false,
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ true,
+        /* .events                = */ false,
     };
 }
 

ggml/src/ggml-cuda.cu

@@ -2920,9 +2920,10 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
 #endif
 
     props->caps = {
-        /* async       */ true,
-        /* host_buffer */ host_buffer,
-        /* events      */ events,
+        /* .async                 = */ true,
+        /* .host_buffer           = */ host_buffer,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ events,
     };
 }