tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 12:05:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml-cpu: activate nnpa fp32->fp16 or fp16->fp32 compute

Browse Source 
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
This commit is contained in:
Aaron Teo
2025-06-21 17:51:55 +08:00
parent 8ef51b9055
commit f1b1d98e8d
1 changed files with 22 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
ggml/src/ggml-impl.h
View File

@ -417,6 +417,28 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size);
#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#elif defined(__NNPA__)
#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_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
uint16x8_t v_h = vec_splats(h);
uint16x8_t v_hd = vec_convert_from_fp16(v_h, 0);
return vec_extend_to_fp32_hi(v_hd, 0)[0];
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
float32x4_t v_f = vec_splats(f);
float32x4_t v_zero = vec_splats(0.0f);
uint16x8_t v_hd = vec_round_from_fp32(v_f, v_zero, 0);
uint16x8_t v_h = vec_convert_to_fp16(v_hd, 0);
return vec_extract(v_h, 0);
}
#else #else
// FP16 <-> FP32 // FP16 <-> FP32