ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-16 13:58:09 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

ggml : refactoring (llama/#0)

Browse Source 
- d6a04f87
- 23e0d70b
This commit is contained in:
Georgi Gerganov
2024-09-20 21:24:06 +03:00
parent d245d7aec7
commit 34291099fb
14 changed files with 768 additions and 706 deletions
Download Patch File Download Diff File Expand all files Collapse all files

113
ggml/src/ggml.c
View File

@ -3,6 +3,7 @@
#include "ggml-backend.h"
#include "ggml-impl.h"
#include "ggml-cpu-impl.h"
#include "ggml-quants.h"
#include "ggml.h"
#include "ggml-aarch64.h"
@ -288,6 +289,7 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) {
#define GGML_DEBUG 0
#define GGML_GELU_FP16
#define GGML_GELU_QUICK_FP16
#define GGML_N_TASKS_MAX (-1)
#define GGML_SOFT_MAX_UNROLL 4
#define GGML_VEC_DOT_UNROLL 2
@ -1121,21 +1123,21 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
#define GGML_F32x4_ADD vaddq_f32
#define GGML_F32x4_MUL vmulq_f32
#define GGML_F32x4_REDUCE_ONE(x) vaddvq_f32(x)
#define GGML_F32x4_REDUCE(res, x) \
{ \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vaddq_f32(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vaddq_f32(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vaddq_f32(x[i], x[offset+i]); \
} \
res = GGML_F32x4_REDUCE_ONE(x[0]); \
#define GGML_F32x4_REDUCE(res, x) \
{ \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
(x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
(x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
(x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
} \
(res) = GGML_F32x4_REDUCE_ONE((x)[0]); \
}
#define GGML_F32_VEC GGML_F32x4
@ -1162,30 +1164,30 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
#define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c)
#define GGML_F16x8_ADD vaddq_f16
#define GGML_F16x8_MUL vmulq_f16
#define GGML_F16x8_REDUCE(res, x) \
do { \
int offset = GGML_F16_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vaddq_f16(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vaddq_f16(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vaddq_f16(x[i], x[offset+i]); \
} \
const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 (x[0])); \
const float32x4_t t1 = vcvt_f32_f16(vget_high_f16(x[0])); \
res = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1)); \
#define GGML_F16x8_REDUCE(res, x) \
do { \
int offset = GGML_F16_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
(x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
(x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
(x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
} \
const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 ((x)[0])); \
const float32x4_t t1 = vcvt_f32_f16(vget_high_f16((x)[0])); \
(res) = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1)); \
} while (0)
#define GGML_F16_VEC GGML_F16x8
#define GGML_F16_VEC_ZERO GGML_F16x8_ZERO
#define GGML_F16_VEC_SET1 GGML_F16x8_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F16x8_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((ggml_fp16_internal_t *)(p), r[i])
#define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((ggml_fp16_internal_t *)(p), (r)[i])
#define GGML_F16_VEC_FMA GGML_F16x8_FMA
#define GGML_F16_VEC_ADD GGML_F16x8_ADD
#define GGML_F16_VEC_MUL GGML_F16x8_MUL
@ -1894,6 +1896,23 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
#define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR)
#endif
//
// ggml object
//
struct ggml_object {
size_t offs;
size_t size;
struct ggml_object * next;
enum ggml_object_type type;
char padding[4];
};
static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
//
// ggml context
//
@ -19408,6 +19427,34 @@ void ggml_graph_clear(struct ggml_cgraph * cgraph) {
ggml_hash_set_reset(&cgraph->visited_hash_set);
}
int ggml_graph_size(struct ggml_cgraph * cgraph) {
return cgraph->size;
}
struct ggml_tensor * ggml_graph_node(struct ggml_cgraph * cgraph, int i) {
if (i < 0) {
GGML_ASSERT(cgraph->n_nodes + i >= 0);
return cgraph->nodes[cgraph->n_nodes + i];
}
GGML_ASSERT(i < cgraph->n_nodes);
return cgraph->nodes[i];
}
struct ggml_tensor ** ggml_graph_nodes(struct ggml_cgraph * cgraph) {
return cgraph->nodes;
}
int ggml_graph_n_nodes(struct ggml_cgraph * cgraph) {
return cgraph->n_nodes;
}
void ggml_graph_add_node(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor) {
GGML_ASSERT(cgraph->size > cgraph->n_nodes);
cgraph->nodes[cgraph->n_nodes] = tensor;
cgraph->n_nodes++;
}
// Android's libc implementation "bionic" does not support setting affinity
#if defined(__gnu_linux__)
static void set_numa_thread_affinity(int thread_n) {