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

42
ggml-metal.m
View file

@ -2294,13 +2294,13 @@ static void ggml_backend_metal_free_device(void) {
}
}
static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
GGML_CALL static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
return "Metal";
UNUSED(buffer);
}
static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
GGML_CALL static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
for (int i = 0; i < ctx->n_buffers; i++) {
@ -2315,25 +2315,25 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
free(ctx);
}
static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
GGML_CALL static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
return ctx->all_data;
}
static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
GGML_CALL static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
memcpy((char *)tensor->data + offset, data, size);
UNUSED(buffer);
}
static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
GGML_CALL static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
memcpy(data, (const char *)tensor->data + offset, size);
UNUSED(buffer);
}
static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
GGML_CALL static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
if (ggml_backend_buffer_is_host(src->buffer)) {
memcpy(dst->data, src->data, ggml_nbytes(src));
return true;
@ -2343,7 +2343,7 @@ static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, c
UNUSED(buffer);
}
static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
GGML_CALL static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
memset(ctx->all_data, value, ctx->all_size);
@ -2363,13 +2363,13 @@ static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
// default buffer type
static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
GGML_CALL static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "Metal";
UNUSED(buft);
}
static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
const size_t size_page = sysconf(_SC_PAGESIZE);
@ -2421,24 +2421,24 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
}
static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
GGML_CALL static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return 32;
UNUSED(buft);
}
static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
GGML_CALL static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
return ggml_backend_is_metal(backend) || ggml_backend_is_cpu(backend);
UNUSED(buft);
}
static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
GGML_CALL static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
return true;
UNUSED(buft);
}
ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
/* .iface = */ {
/* .get_name = */ ggml_backend_metal_buffer_type_get_name,
@ -2456,7 +2456,7 @@ ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
// buffer from ptr
ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
ctx->all_data = data;
@ -2543,31 +2543,31 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz
// backend
static const char * ggml_backend_metal_name(ggml_backend_t backend) {
GGML_CALL static const char * ggml_backend_metal_name(ggml_backend_t backend) {
return "Metal";
UNUSED(backend);
}
static void ggml_backend_metal_free(ggml_backend_t backend) {
GGML_CALL static void ggml_backend_metal_free(ggml_backend_t backend) {
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
ggml_metal_free(ctx);
free(backend);
}
static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
return ggml_backend_metal_buffer_type();
UNUSED(backend);
}
static bool ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
GGML_CALL static bool ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
return ggml_metal_graph_compute(metal_ctx, cgraph);
}
static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
return ggml_metal_supports_op(metal_ctx, op);
@ -2630,9 +2630,9 @@ bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
}
ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); // silence warning
GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); // silence warning
ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data) {
GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data) {
return ggml_backend_metal_init();
GGML_UNUSED(params);