root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-11 19:21:46 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

ggml : try fix ppc64 (whisper/0)

Browse Source
This commit is contained in:
Georgi Gerganov 2024-05-12 20:36:31 +03:00
parent 182adefcf3
commit c3c88f296a
2 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml-quants.c
View File

@ -11425,7 +11425,7 @@ void ggml_vec_dot_iq1_s_q8_K (int n, float * restrict s, size_t bs, const void
vector signed short qxh = (vector signed short)vec_sld(vec_splats(qh[1]), vec_splats(qh[0]), 8);
qh += 2;
vector bool short vsel = vec_cmpge(qxh, (vector signed short)v0);
vector __bool short vsel = vec_cmpge(qxh, (vector signed short)v0);
vector signed short q8ysum = vec_sel((vector signed short)vec_xor((vector unsigned short)q8ysums, vsign), q8ysums, vsel);

2
ggml.c
View File

@ -1306,6 +1306,8 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
#define GGML_F16_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F16_VEC_SET1 GGML_F32x4_SET1
#define GGML_F16_VEC_FMA GGML_F32x4_FMA
#define GGML_F16_VEC_ADD GGML_F32x4_ADD
#define GGML_F16_VEC_MUL GGML_F32x4_MUL
#define GGML_F16_VEC_REDUCE GGML_F32x4_REDUCE
// Use vec_xl, not vec_ld, in case the load address is not aligned.
#define GGML_F16_VEC_LOAD(p, i) (i & 0x1) ? \