tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 03:55:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml-cpu: switch to vec_xst for 4 element loops also

Browse Source 
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
This commit is contained in:
Aaron Teo
2025-06-21 16:23:23 +08:00
parent f9f6c7e897
commit 6d507bbeb0
1 changed files with 8 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -3175,26 +3175,21 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
uint16x8_t v_xd = vec_convert_from_fp16(v_x, 0);
float32x4_t v_xdh = vec_extend_to_fp32_hi(v_xd, 0);
float32x4_t v_xdl = vec_extend_to_fp32_lo(v_xd, 0);
vec_xst(v_xdh, 0, (float *)(y + i));
vec_xst(v_xdh, 0, (float *)(y + i + 0));
vec_xst(v_xdl, 0, (float *)(y + i + 4));
// y[i + 0] = v_xdh[0];
// y[i + 1] = v_xdh[1];
// y[i + 2] = v_xdh[2];
// y[i + 3] = v_xdh[3];
// y[i + 4] = v_xdl[0];
// y[i + 5] = v_xdl[1];
// y[i + 6] = v_xdl[2];
// y[i + 7] = v_xdl[3];
(void);
}
for (; i + 3 < n; i += 4) {
uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
uint16x8_t nnpa_dfl16 = vec_convert_from_fp16(v_x, 0);
float32x4_t result = vec_extend_to_fp32_hi(nnpa_dfl16, 0);
y[i + 0] = result[0];
y[i + 1] = result[1];
y[i + 2] = result[2];
y[i + 3] = result[3];
vec_xst(result, 0, (float *)(y + i));
(void);
// y[i + 0] = result[0];
// y[i + 1] = result[1];
// y[i + 2] = result[2];
// y[i + 3] = result[3];
}
#endif