tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-05 17:23:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml-cpu: move nnpa fp16->fp32 and fp32->fp16 to simd-mappings

Browse Source 
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
This commit is contained in:
Aaron Teo
2025-06-23 17:52:28 +08:00
parent 5004e4395b
commit 5834dee1fc
2 changed files with 39 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
ggml/src/ggml-cpu/simd-mappings.h
View File

@ -1013,6 +1013,45 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
#define GGML_F16_VEC_MUL GGML_F32x4_MUL
#define GGML_F16_VEC_REDUCE GGML_F32x4_REDUCE
#if defined(__NNPA__)
#ifdef GGML_FP16_TO_FP32
#undef GGML_FP16_TO_FP32
#endif
#ifdef GGML_FP32_TO_FP16
#undef GGML_FP32_TO_FP16
#endif
#ifdef GGML_COMPUTE_FP16_TO_FP32
#undef GGML_COMPUTE_FP16_TO_FP32
#endif
#ifdef GGML_COMPUTE_FP32_TO_FP16
#undef GGML_COMPUTE_FP32_TO_FP16
#endif
#define GGML_COMPUTE_FP16_TO_FP32(x) nnpa_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) nnpa_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 nnpa_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 nnpa_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);
}
#endif // __NNPA__
#endif
// GGML_F32_ARR / GGML_F16_ARR

37
ggml/src/ggml-impl.h
View File

@ -417,43 +417,6 @@ 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_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#elif defined(__NNPA__)
/*
* Note: This functionality is ready for use, but the compiler macros
* defined for the s390x platform are defined in ggml-cpu while
* this file is 1 step behind, in ggml-src. I currently have no
* idea how to fix this, so I am leaving it as is.
*
* CMake chain: ggml -> ggml-src -> ggml-cpu
* ^^^^^^^^ ^^^^^^^^
* | | ggml-cpu defines the macros
* | | needed for s390x detection.
* | this file is here, where the s390x
* | detection macros are not defined.
*
* TODO: Fix s390x platform detection in this file.
*/
#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
// FP16 <-> FP32