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

PR #4766

Browse source
This commit is contained in:
Jared Van Bortel 2024-01-10 11:29:04 -05:00
parent 3773e1afe7
commit 1eb8804c18
18 changed files with 2183 additions and 2081 deletions
Download patch file Download diff file Expand all files Collapse all files

39
ggml-metal.m
View file

@ -2482,10 +2482,10 @@ static void ggml_backend_metal_free_device(void) {
}
}
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;
static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
return "Metal";
return ctx->all_data;
UNUSED(buffer);
}
static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
@ -2503,6 +2503,12 @@ 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) {
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) {
memcpy((char *)tensor->data + offset, data, size);
@ -2515,13 +2521,13 @@ static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, c
UNUSED(buffer);
}
static void ggml_backend_metal_buffer_cpy_tensor_from(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
static void ggml_backend_metal_buffer_cpy_tensor_from(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
UNUSED(buffer);
}
static void ggml_backend_metal_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
static void ggml_backend_metal_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
UNUSED(buffer);
@ -2534,6 +2540,7 @@ static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_
}
static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
/* .get_name = */ ggml_backend_metal_buffer_get_name,
/* .free_buffer = */ ggml_backend_metal_buffer_free_buffer,
/* .get_base = */ ggml_backend_metal_buffer_get_base,
/* .init_tensor = */ NULL,
@ -2542,10 +2549,17 @@ static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
/* .cpy_tensor_from = */ ggml_backend_metal_buffer_cpy_tensor_from,
/* .cpy_tensor_to = */ ggml_backend_metal_buffer_cpy_tensor_to,
/* .clear = */ ggml_backend_metal_buffer_clear,
/* .reset = */ NULL,
};
// default buffer type
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) {
struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
@ -2618,6 +2632,7 @@ static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t bu
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,
/* .alloc_buffer = */ ggml_backend_metal_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_metal_buffer_type_get_alignment,
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
@ -2641,6 +2656,14 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz
ctx->n_buffers = 0;
const size_t size_page = sysconf(_SC_PAGESIZE);
// page-align the data ptr
{
const uintptr_t offs = (uintptr_t) data % size_page;
data = (void *) ((char *) data - offs);
size += offs;
}
size_t size_aligned = size;
if ((size_aligned % size_page) != 0) {
size_aligned += (size_page - (size_aligned % size_page));
@ -2741,7 +2764,7 @@ static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct
UNUSED(backend);
}
static struct ggml_backend_i metal_backend_i = {
static struct ggml_backend_i ggml_backend_metal_i = {
/* .get_name = */ ggml_backend_metal_name,
/* .free = */ ggml_backend_metal_free,
/* .get_default_buffer_type = */ ggml_backend_metal_get_default_buffer_type,
@ -2767,7 +2790,7 @@ ggml_backend_t ggml_backend_metal_init(void) {
ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend));
*metal_backend = (struct ggml_backend) {
/* .interface = */ metal_backend_i,
/* .interface = */ ggml_backend_metal_i,
/* .context = */ ctx,
};
@ -2775,7 +2798,7 @@ ggml_backend_t ggml_backend_metal_init(void) {
}
bool ggml_backend_is_metal(ggml_backend_t backend) {
return backend->iface.get_name == ggml_backend_metal_name;
return backend && backend->iface.get_name == ggml_backend_metal_name;
}
void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {