tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-09 10:32:38 -04:00
Code Issues Packages Projects Releases Wiki Activity

ggml-cpu: switch fp16->fp32 to inline asm and test

Browse Source 
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
This commit is contained in:
Aaron Teo
2025-06-20 21:01:37 +08:00
parent ae9c5f928a
commit a88843aa10
1 changed files with 11 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
ggml/src/ggml-impl.h
View File

@@ -428,9 +428,17 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size);
// TODO: Determine if inline assembly is faster
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
uint16x8_t v_h = vec_splats(h);
uint16x8_t nnpa_dlf16 = vec_convert_from_fp16(v_h, 0);
return vec_extend_to_fp32_hi(nnpa_dlf16, 0)[0];
float f;
__asm__ (
"vlvgp %%v0, %1, %1\n"
"vreph %%v0, %%v0, 3\n"
"vcnf %%v0, %%v0, 0, 1\n"
"vclfnh %%v0, %%v0, 2, 0\n"
"ler %0, %%f0\n" :
/* out */ "=f"(f) :
/* in */ "r"(h) :
/* clobber */ "v0", "f0");
return f;
}
// TODO: Determine if inline assembly is faster