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

test

Browse source
This commit is contained in:
John 2023-06-20 04:01:09 +02:00
parent dd80fb5320
commit c09786bac5
2 changed files with 10 additions and 30 deletions
Download patch file Download diff file Expand all files Collapse all files

33
ggml.c
View file

@ -45,8 +45,7 @@
#include <windows.h> #include <windows.h>
typedef volatile LONG atomic_int;
typedef atomic_int atomic_bool;
static void atomic_store(atomic_int* ptr, LONG val) { static void atomic_store(atomic_int* ptr, LONG val) {
InterlockedExchange(ptr, val); InterlockedExchange(ptr, val);
@ -3761,32 +3760,6 @@ struct ggml_context_container {
struct ggml_context context; struct ggml_context context;
}; };
//
// compute types
//
enum ggml_task_type {
GGML_TASK_INIT = 0,
GGML_TASK_COMPUTE,
GGML_TASK_FINALIZE,
};
struct ggml_compute_params {
enum ggml_task_type type;
int ith, nth;
// work buffer for all threads
size_t wsize;
void * wdata;
// atomic counter used to distribute chunks of work
atomic_int * aic;
};
//
// ggml state
//
struct ggml_state { struct ggml_state {
struct ggml_context_container contexts[GGML_MAX_CONTEXTS]; struct ggml_context_container contexts[GGML_MAX_CONTEXTS];
@ -10791,8 +10764,8 @@ static void ggml_compute_forward_mul_mat_q_f32(
// parallelize by src0 rows using ggml_vec_dot_q // parallelize by src0 rows using ggml_vec_dot_q
const int nr = ggml_nrows(src0); const int nr = ggml_nrows(src0); // nr stands for number of rows
const int dr = (nr + 8*nth - 1)/(8*nth); const int dr = (nr + 8*nth - 1)/(8*nth); // dr stands for delta rows
while (true) { while (true) {
const int ir0 = atomic_fetch_add(params->aic, dr); const int ir0 = atomic_fetch_add(params->aic, dr);

7
ggml.h
View file

@ -219,6 +219,9 @@ extern "C" {
#else #else
typedef uint16_t ggml_fp16_t; typedef uint16_t ggml_fp16_t;
#endif #endif
typedef long LONG;
typedef volatile LONG atomic_int;
typedef atomic_int atomic_bool;
// convert FP16 <-> FP32 // convert FP16 <-> FP32
GGML_API float ggml_fp16_to_fp32(ggml_fp16_t x); GGML_API float ggml_fp16_to_fp32(ggml_fp16_t x);
@ -454,8 +457,12 @@ extern "C" {
// work buffer for all threads // work buffer for all threads
size_t wsize; size_t wsize;
void * wdata; void * wdata;
// atomic counter used to distribute chunks of work
atomic_int * aic;
}; };
// misc // misc
GGML_API void ggml_time_init(void); // call this once at the beginning of the program GGML_API void ggml_time_init(void); // call this once at the beginning of the program