vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : introduce GGML_CALL function annotation (#4850)

Browse source 
This change makes it possible to build ggml-cuda.cu and ggml-metal.m as
independent dynamic shared objects, that may be conditionally linked at
runtime in a multiplatform binary. It introduces a GGML_CALL annotation
that documents which functions have a cyclic call relationship, between
the application code and GPU modules.

This change does nothing, unless the build defines -DGGML_MULTIPLATFORM
which causes back-references and function pointers to conform to MS ABI
which is supported by NVCC, ROCm, XCode, GCC and Clang across platforms
This commit is contained in:
Justine Tunney 2024-01-16 03:16:33 -08:00 • committed by GitHub
parent d75c232e1d
commit a0b3ac8c48
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
9 changed files with 244 additions and 235 deletions
Download patch file Download diff file Expand all files Collapse all files

60
ggml-backend-impl.h
View file

@ -16,14 +16,14 @@ extern "C" {
typedef void * ggml_backend_buffer_type_context_t;
struct ggml_backend_buffer_type_i {
const char * (*get_name) (ggml_backend_buffer_type_t buft);
ggml_backend_buffer_t (*alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size);
size_t (*get_alignment) (ggml_backend_buffer_type_t buft); // tensor alignment
size_t (*get_alloc_size) (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
bool (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
const char * (*GGML_CALL get_name) (ggml_backend_buffer_type_t buft);
ggml_backend_buffer_t (*GGML_CALL alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size);
size_t (*GGML_CALL get_alignment) (ggml_backend_buffer_type_t buft); // tensor alignment
size_t (*GGML_CALL get_alloc_size) (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
bool (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
// check if tensor data is in host memory
// should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
bool (*is_host) (ggml_backend_buffer_type_t buft);
bool (*GGML_CALL is_host) (ggml_backend_buffer_type_t buft);
};
struct ggml_backend_buffer_type {
@ -35,15 +35,15 @@ extern "C" {
typedef void * ggml_backend_buffer_context_t;
struct ggml_backend_buffer_i {
const char * (*get_name) (ggml_backend_buffer_t buffer);
void (*free_buffer)(ggml_backend_buffer_t buffer);
void * (*get_base) (ggml_backend_buffer_t buffer);
void (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
void (*set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
bool (*cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
void (*clear) (ggml_backend_buffer_t buffer, uint8_t value);
void (*reset) (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
const char * (*GGML_CALL get_name) (ggml_backend_buffer_t buffer);
void (*GGML_CALL free_buffer)(ggml_backend_buffer_t buffer);
void * (*GGML_CALL get_base) (ggml_backend_buffer_t buffer);
void (*GGML_CALL init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
void (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
bool (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
void (*GGML_CALL clear) (ggml_backend_buffer_t buffer, uint8_t value);
void (*GGML_CALL reset) (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
};
struct ggml_backend_buffer {
@ -54,7 +54,7 @@ extern "C" {
enum ggml_backend_buffer_usage usage;
};
ggml_backend_buffer_t ggml_backend_buffer_init(
GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
ggml_backend_buffer_type_t buft,
struct ggml_backend_buffer_i iface,
ggml_backend_buffer_context_t context,
@ -70,31 +70,31 @@ extern "C" {
typedef void * ggml_backend_context_t;
struct ggml_backend_i {
const char * (*get_name)(ggml_backend_t backend);
const char * (*GGML_CALL get_name)(ggml_backend_t backend);
void (*free)(ggml_backend_t backend);
void (*GGML_CALL free)(ggml_backend_t backend);
// buffer allocation
ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
// (optional) asynchronous tensor data access
void (*set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
bool (*cpy_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
void (*GGML_CALL set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
// (optional) complete all pending operations
void (*synchronize)(ggml_backend_t backend);
void (*GGML_CALL synchronize)(ggml_backend_t backend);
// compute graph with a plan
ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
void (*graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
void (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
void (*GGML_CALL graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
void (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
// compute graph without a plan (async)
bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
bool (*GGML_CALL graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
// check if the backend supports an operation
bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
};
struct ggml_backend {
@ -107,9 +107,9 @@ extern "C" {
// Backend registry
//
typedef ggml_backend_t (*ggml_backend_init_fn)(const char * params, void * user_data);
typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
#ifdef __cplusplus
}