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

ggml : use posix_memalign on non-Windows env

Browse source
This commit is contained in:
Georgi Gerganov 2023-04-15 14:25:45 +03:00
parent c12b14b77f
commit aa485cee33
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 16 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

23
ggml.c
View file

@ -118,7 +118,16 @@ typedef void* thread_ret_t;
#define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN) #define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN)
#define GGML_ALIGNED_FREE(ptr) _aligned_free(ptr) #define GGML_ALIGNED_FREE(ptr) _aligned_free(ptr)
#else #else
#define GGML_ALIGNED_MALLOC(size) aligned_alloc(GGML_MEM_ALIGN, size) inline static void* ggml_aligned_malloc(size_t size) {
void* aligned_memory = NULL;
int result = posix_memalign(&aligned_memory, GGML_MEM_ALIGN, size);
if (result != 0) {
// Handle allocation failure
return NULL;
}
return aligned_memory;
}
#define GGML_ALIGNED_MALLOC(size) ggml_aligned_malloc(size)
#define GGML_ALIGNED_FREE(ptr) free(ptr) #define GGML_ALIGNED_FREE(ptr) free(ptr)
#endif #endif
@ -531,31 +540,31 @@ inline static float vaddvq_f32(float32x4_t v) {
return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3); return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
} }
inline float vminvq_f32(float32x4_t v) { float vminvq_f32(float32x4_t v) {
return return
MIN(MIN(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)), MIN(MIN(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
MIN(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3))); MIN(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
} }
inline float vmaxvq_f32(float32x4_t v) { float vmaxvq_f32(float32x4_t v) {
return return
MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)), MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3))); MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
} }
inline int8x8_t vzip1_s8(int8x8_t a, int8x8_t b) { int8x8_t vzip1_s8(int8x8_t a, int8x8_t b) {
return vget_low_s8(vcombine_s8(a, b)); return vget_low_s8(vcombine_s8(a, b));
} }
inline int8x8_t vzip2_s8(int8x8_t a, int8x8_t b) { int8x8_t vzip2_s8(int8x8_t a, int8x8_t b) {
return vget_high_s8(vcombine_s8(a, b)); return vget_high_s8(vcombine_s8(a, b));
} }
inline uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) { uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
return vget_low_u8(vcombine_u8(a, b)); return vget_low_u8(vcombine_u8(a, b));
} }
inline uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) { uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
return vget_high_u8(vcombine_u8(a, b)); return vget_high_u8(vcombine_u8(a, b));
} }