metal : refactor context

ggml-ci

Author: ggerganov

ggml/include/ggml-metal.h

@@ -39,6 +39,7 @@ extern "C" {
 // user-code should use only these functions
 //
 
+// TODO: remove in the future
 GGML_BACKEND_API ggml_backend_t ggml_backend_metal_init(void);
 
 GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);

ggml/src/ggml-metal/CMakeLists.txt

@@ -9,6 +9,8 @@ ggml_add_backend_library(ggml-metal
                          ggml-metal-device.m
                          ggml-metal-device.cpp
                          ggml-metal-common.cpp
+                         ggml-metal-context.m
+                         ggml-metal-ops.cpp
                         )
 
 target_link_libraries(ggml-metal PRIVATE

ggml/src/ggml-metal/ggml-metal-common.cpp

@@ -375,7 +375,7 @@ static std::vector<int> ggml_metal_graph_optimize_reorder(const std::vector<node
     return res;
 }
 
-void ggml_metal_graph_optimize(ggml_cgraph * gf) {
+void ggml_graph_optimize(ggml_cgraph * gf) {
     constexpr int MAX_FUSE = 16;
 
     const int n = gf->n_nodes;

ggml/src/ggml-metal/ggml-metal-common.h

@@ -45,7 +45,7 @@ bool ggml_mem_ranges_check(ggml_mem_ranges_t mrs, const struct ggml_tensor * ten
 //
 // note: this implementation is generic and not specific to metal
 //       if it proves to work well, we can start using it for other backends in the future
-void ggml_metal_graph_optimize(struct ggml_cgraph * gf);
+void ggml_graph_optimize(struct ggml_cgraph * gf);
 
 #ifdef __cplusplus
 }

ggml/src/ggml-metal/ggml-metal-context.h

@@ -0,0 +1,52 @@
+#pragma once
+
+#include "ggml-metal-device.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//
+// MTLFunctionConstantValues wrapper
+//
+
+typedef struct ggml_metal_cv * ggml_metal_cv_t;
+
+ggml_metal_cv_t ggml_metal_cv_init(void);
+void ggml_metal_cv_free(ggml_metal_cv_t cv);
+
+void ggml_metal_cv_set_int32(ggml_metal_cv_t cv, int32_t value, int32_t idx);
+void ggml_metal_cv_set_bool (ggml_metal_cv_t cv, bool    value, int32_t idx);
+
+//
+// backend
+//
+
+typedef struct ggml_metal * ggml_metal_t;
+
+ggml_metal_t ggml_metal_init(ggml_metal_device_t ctx_dev);
+void ggml_metal_free(ggml_metal_t ctx);
+
+typedef void * ggml_metal_pipeline_t;
+
+ggml_metal_pipeline_t ggml_metal_get_pipeline(ggml_metal_t ctx, const char * name);
+
+ggml_metal_pipeline_t ggml_metal_compile_pipeline(ggml_metal_t ctx, const char * base, const char * name, ggml_metal_cv_t cv);
+
+void ggml_metal_synchronize(ggml_metal_t ctx);
+
+void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
+
+enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph * gf);
+
+void ggml_metal_graph_optimize(ggml_metal_t ctx, struct ggml_cgraph * gf);
+
+void ggml_metal_set_n_cb            (ggml_metal_t ctx, int n_cb);
+void ggml_metal_set_abort_callback  (ggml_metal_t ctx, ggml_abort_callback abort_callback, void * user_data);
+bool ggml_metal_supports_family     (ggml_metal_t ctx, int family);
+void ggml_metal_capture_next_compute(ggml_metal_t ctx);
+
+#ifdef __cplusplus
+}
+#endif