EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

ggml-cpu : optimize RVV kernels (#15720)

Browse Source 
* ggml-cpu : optimize rvv ggml_vec_dot_f32

* ggml-cpu : optimize 128-bit rvv ggml_vec_dot_q4_K_q8_K

* ggml-cpu : fix riscv arch flags

* ggml-cpu : add more rvv ops

* ggml-cpu : optimize rvv ggml_vec_dot_q4_K_q8_K

* ggml-cpu : optimize rvv ggml_vec_dot_q6_K_q8_K

* ggml-cpu : minor rvv adjustments

* ggml-cpu : fix riscv include
This commit is contained in:
xctan
2025-09-03 16:16:21 +08:00
committed by GitHub
parent 8c3fdf44ec
commit 05c0380f2a
6 changed files with 289 additions and 114 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
ggml/src/ggml-cpu/ggml-cpu.c
View File

@@ -3221,6 +3221,13 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
}
#elif defined(__riscv_zvfh)
for (int vl; i < n; i += vl) {
vl = __riscv_vsetvl_e32m2(n - i);
vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
vfloat16m1_t vy = __riscv_vfncvt_f_f_w_f16m1(vx, vl);
__riscv_vse16_v_f16m1((_Float16 *)&y[i], vy, vl);
}
#endif
for (; i < n; ++i) {
y[i] = GGML_CPU_FP32_TO_FP16(x[i]);