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

add GGML_PAD

Browse source
This commit is contained in:
slaren 2023-07-26 12:49:18 +02:00
parent 156d99abde
commit 7949dcaaf7
2 changed files with 5 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml.c
View file

@ -4383,7 +4383,8 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
return NULL; return NULL;
} }
const size_t mem_size = (params.mem_size + GGML_MEM_ALIGN - 1) & ~(GGML_MEM_ALIGN - 1); // REVIEW: this was probably a bug - the size was always padded even if the user provided the buffer
const size_t mem_size = params.mem_buffer ? params.mem_size : GGML_PAD(params.mem_size, GGML_MEM_ALIGN);
*ctx = (struct ggml_context) { *ctx = (struct ggml_context) {
/*.mem_size =*/ mem_size, /*.mem_size =*/ mem_size,
@ -4519,10 +4520,8 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml
const size_t cur_size = obj_cur == NULL ? 0 : obj_cur->size; const size_t cur_size = obj_cur == NULL ? 0 : obj_cur->size;
const size_t cur_end = cur_offs + cur_size; const size_t cur_end = cur_offs + cur_size;
size_t size_needed = size;
// align to GGML_MEM_ALIGN // align to GGML_MEM_ALIGN
size_needed = ((size_needed + GGML_MEM_ALIGN - 1)/GGML_MEM_ALIGN)*GGML_MEM_ALIGN; size_t size_needed = GGML_PAD(size, GGML_MEM_ALIGN);
char * const mem_buffer = ctx->mem_buffer; char * const mem_buffer = ctx->mem_buffer;
struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end); struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end);
@ -15858,7 +15857,7 @@ struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct gg
} }
size_t ggml_graph_overhead(void) { size_t ggml_graph_overhead(void) {
return GGML_OBJECT_SIZE + ((GGML_GRAPH_SIZE + GGML_MEM_ALIGN - 1)/GGML_MEM_ALIGN)*GGML_MEM_ALIGN; return GGML_OBJECT_SIZE + GGML_PAD(GGML_GRAPH_SIZE, GGML_MEM_ALIGN);
} }
// //

1
ggml.h
View file

@ -208,6 +208,7 @@
#define GGML_UNUSED(x) (void)(x) #define GGML_UNUSED(x) (void)(x)
#define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
#define GGML_ASSERT(x) \ #define GGML_ASSERT(x) \
do { \ do { \