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

ggml : cont

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-02-22 17:12:12 +02:00
parent bf74c0fdf2
commit 3c5cc30023
No known key found for this signature in database
GPG key ID: BF970631944C16B7
1 changed files with 6 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml-impl.h
View file

@ -56,6 +56,8 @@ extern "C" {
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#define GGML_FP16_TO_FP32(x) ((float)(*(const __fp16 *)&(x)))
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
__fp16 tmp; __fp16 tmp;
memcpy(&tmp, &h, sizeof(ggml_fp16_t)); memcpy(&tmp, &h, sizeof(ggml_fp16_t));
@ -224,8 +226,7 @@ extern float ggml_table_f32_f16[1 << 16];
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32, // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON. // so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
// This is also true for POWER9. // This is also true for POWER9.
#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16) #if !defined(GGML_FP16_TO_FP32)
inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) { inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
uint16_t s; uint16_t s;
memcpy(&s, &f, sizeof(uint16_t)); memcpy(&s, &f, sizeof(uint16_t));
@ -233,8 +234,10 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
} }
#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x) #define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) #endif
#if !defined(GGML_FP32_TO_FP16)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#endif #endif
#define GGML_HASHTABLE_FULL ((size_t)-1) #define GGML_HASHTABLE_FULL ((size_t)-1)