CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-21 12:16:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml-cpu:add RISC-V RVV (Zvfh) optimization for FP16 vector scaling (#17314)

Browse Source 
* ggml-cpu:add RISC-V RVV (Zvfh) optimization for FP16 vector scaling

Signed-off-by: Wang Yang <yangwang@iscas.ac.cn>

* fix comment

* fix comment 2

---------

Signed-off-by: Wang Yang <yangwang@iscas.ac.cn>
This commit is contained in:
ixgbe
2025-11-20 14:09:18 +08:00
committed by GitHub
parent 7d77f07325
commit 5be353ec4a
1 changed files with 50 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

99
ggml/src/ggml-cpu/vec.h
View File

@@ -698,60 +698,61 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float v) {
}
inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) {
#if defined(GGML_SIMD)
#if defined(__ARM_FEATURE_SVE)
const int sve_register_length = svcntb() * 8;
const int ggml_f16_epr = sve_register_length / 16;
const int ggml_f16_step = 2 * ggml_f16_epr;
#if defined(GGML_SIMD) && defined(__ARM_FEATURE_SVE)
const int sve_register_length = svcntb() * 8;
const int ggml_f16_epr = sve_register_length / 16;
const int ggml_f16_step = 2 * ggml_f16_epr;
GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
const int np = (n & ~(ggml_f16_step - 1));
svfloat16_t ay1, ay2;
GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
const int np = (n & ~(ggml_f16_step - 1));
svfloat16_t ay1, ay2;
for (int i = 0; i < np; i += ggml_f16_step) {
ay1 = GGML_F16x_VEC_LOAD(y + i + 0*ggml_f16_epr, 0);
ay1 = GGML_F16x_VEC_MUL(ay1, vx);
GGML_F16x_VEC_STORE(y + i + 0*ggml_f16_epr, ay1, 0);
for (int i = 0; i < np; i += ggml_f16_step) {
ay1 = GGML_F16x_VEC_LOAD(y + i + 0*ggml_f16_epr, 0);
ay1 = GGML_F16x_VEC_MUL(ay1, vx);
GGML_F16x_VEC_STORE(y + i + 0*ggml_f16_epr, ay1, 0);
ay2 = GGML_F16x_VEC_LOAD(y + i + 1*ggml_f16_epr, 1);
ay2 = GGML_F16x_VEC_MUL(ay2, vx);
GGML_F16x_VEC_STORE(y + i + 1*ggml_f16_epr, ay2, 1);
ay2 = GGML_F16x_VEC_LOAD(y + i + 1*ggml_f16_epr, 1);
ay2 = GGML_F16x_VEC_MUL(ay2, vx);
GGML_F16x_VEC_STORE(y + i + 1*ggml_f16_epr, ay2, 1);
}
// leftovers
// maximum number of leftover elements will be less that ggmlF_16x_epr. Apply predicated svmad on available elements only
if (np < n) {
svbool_t pg = svwhilelt_b16(np, n);
svfloat16_t hy = svld1_f16(pg, (__fp16 *)(y + np));
svfloat16_t out = svmul_f16_m(pg, hy, vx);
svst1_f16(pg, (__fp16 *)(y + np), out);
}
#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfh)
for (int i = 0, vl; i < n; i += vl) {
vl = __riscv_vsetvl_e16m2(n - i);
vfloat16m2_t vy = __riscv_vle16_v_f16m2((_Float16 *)&y[i], vl);
vfloat32m4_t vy32 = __riscv_vfwcvt_f_f_v_f32m4(vy, vl);
vy32 = __riscv_vfmul_vf_f32m4(vy32, v, vl);
vy = __riscv_vfncvt_f_f_w_f16m2(vy32, vl);
__riscv_vse16_v_f16m2((_Float16 *)&y[i], vy, vl);
}
#elif defined(GGML_SIMD)
const int np = (n & ~(GGML_F16_STEP - 1));
GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
GGML_F16_VEC ay[GGML_F16_ARR];
for (int i = 0; i < np; i += GGML_F16_STEP) {
for (int j = 0; j < GGML_F16_ARR; j++) {
ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
}
// leftovers
// maximum number of leftover elements will be less that ggmlF_16x_epr. Apply predicated svmad on available elements only
if (np < n) {
svbool_t pg = svwhilelt_b16(np, n);
svfloat16_t hy = svld1_f16(pg, (__fp16 *)(y + np));
svfloat16_t out = svmul_f16_m(pg, hy, vx);
svst1_f16(pg, (__fp16 *)(y + np), out);
}
#elif defined(__riscv_v_intrinsic)
// todo: RVV impl
// scalar
for (int i = 0; i < n; ++i) {
y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
}
#else
const int np = (n & ~(GGML_F16_STEP - 1));
}
GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
GGML_F16_VEC ay[GGML_F16_ARR];
for (int i = 0; i < np; i += GGML_F16_STEP) {
for (int j = 0; j < GGML_F16_ARR; j++) {
ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
}
}
// leftovers
for (int i = np; i < n; ++i) {
y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
}
#endif
// leftovers
for (int i = np; i < n; ++i) {
y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
}
#else
// scalar
for (int i = 0; i < n; ++i) {