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

8
ggml/src/ggml-cpu/vec.h
View File

@@ -1269,6 +1269,14 @@ inline static vfloat32m2_t ggml_v_expf_m2(vfloat32m2_t x, int vl) {
vl);
}
// computes silu x/(1+exp(-x)) in single precision vector
inline static vfloat32m2_t ggml_v_silu_m2(vfloat32m2_t x, int vl) {
const vfloat32m2_t neg_x = __riscv_vfneg_v_f32m2(x, vl);
const vfloat32m2_t exp_neg_x = ggml_v_expf_m2(neg_x, vl);
const vfloat32m2_t one_plus_exp_neg_x = __riscv_vfadd_vf_f32m2(exp_neg_x, 1.0f, vl);
return __riscv_vfdiv_vv_f32m2(x, one_plus_exp_neg_x, vl);
}
#endif // __ARM_NEON / __AVX2__ / __SSE2__ / __riscv_v_intrinsic
inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {