happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

ggml : remove useless and error-prone variadic macros (#17399) 

Signed-off-by: Adrien Gallouët <angt@huggingface.co>
This commit is contained in:
Adrien Gallouët 2025-11-20 11:18:27 +01:00 committed by GitHub
parent 3ae282a06f
commit 79bb743512
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 16 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

30
ggml/src/ggml-cpu/simd-mappings.h
View File

@ -160,18 +160,18 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
#define GGML_F32xt svfloat32_t
#define GGML_F32xt_ZERO svdup_n_f32(0.0f)
#define GGML_F32xt_SET1(x) svdup_n_f32(x)
#define GGML_F32xt_LOAD_IMPL(pg, a, ...) svld1_f32(pg, a)
#define GGML_F32xt_LOAD(...) GGML_F32xt_LOAD_IMPL(DEFAULT_PG, __VA_ARGS__)
#define GGML_F32xt_STORE_IMPL(pg,a,b) svst1_f32(pg, a, b)
#define GGML_F32xt_STORE(...) GGML_F32xt_STORE_IMPL(DEFAULT_PG, __VA_ARGS__)
#define GGML_F32xt_LOAD_IMPL(pg, a) svld1_f32(pg, a)
#define GGML_F32xt_LOAD(a) GGML_F32xt_LOAD_IMPL(DEFAULT_PG, a)
#define GGML_F32xt_STORE_IMPL(pg, a, b) svst1_f32(pg, a, b)
#define GGML_F32xt_STORE(a, b) GGML_F32xt_STORE_IMPL(DEFAULT_PG, a, b)
#define GGML_F32xt_FMA_IMPL(pg, a, b, c) svmad_f32_m(pg, b, c, a)
#define GGML_F32xt_FMA(...) GGML_F32xt_FMA_IMPL(DEFAULT_PG, __VA_ARGS__)
#define GGML_F32xt_FMA(a, b, c) GGML_F32xt_FMA_IMPL(DEFAULT_PG, a, b, c)
#define GGML_F32xt_ADD_IMPL(pg, a, b) svadd_f32_m(pg, a, b)
#define GGML_F32xt_ADD(...) GGML_F32xt_ADD_IMPL(DEFAULT_PG, __VA_ARGS__)
#define GGML_F32xt_ADD(a, b) GGML_F32xt_ADD_IMPL(DEFAULT_PG, a, b)
#define GGML_F32xt_MUL_IMPL(pg, a, b) svmul_f32_m(pg, a, b)
#define GGML_F32xt_MUL(...) GGML_F32xt_MUL_IMPL(DEFAULT_PG, __VA_ARGS__)
#define GGML_F32xt_MUL(a, b) GGML_F32xt_MUL_IMPL(DEFAULT_PG, a, b)
#define GGML_F32xt_REDUCE_ONE_IMPL(pg, a) svaddv(pg, a)
#define GGML_F32xt_REDUCE_ONE(...) GGML_F32xt_REDUCE_ONE_IMPL(DEFAULT_PG, __VA_ARGS__)
#define GGML_F32xt_REDUCE_ONE(a) GGML_F32xt_REDUCE_ONE_IMPL(DEFAULT_PG, a)
#define GGML_F32xt_REDUCE_IMPL(pg, res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8) \
{ \
sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum2); \
@ -183,7 +183,8 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum5); \
(res) = (ggml_float) GGML_F32xt_REDUCE_ONE(sum1); \
}
#define GGML_F32xt_REDUCE(...) GGML_F32xt_REDUCE_IMPL(DEFAULT_PG, __VA_ARGS__)
#define GGML_F32xt_REDUCE(res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8) \
GGML_F32xt_REDUCE_IMPL(DEFAULT_PG, res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8)
#define GGML_F32_VEC GGML_F32xt
#define GGML_F32_VEC_ZERO GGML_F32xt_ZERO
@ -206,11 +207,11 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
#define GGML_F32Cxt_STORE(dst_ptr, src_vec) svst1_f16(DEFAULT_PG16, (__fp16 *)(dst_ptr), (src_vec))
#define GGML_F32Cxt_FMA_IMPL(pg, a, b, c) svmad_f16_x(pg, b, c, a)
#define GGML_F32Cxt_FMA(...) GGML_F32Cxt_FMA_IMPL(DEFAULT_PG16, __VA_ARGS__)
#define GGML_F32Cxt_FMA(a, b, c) GGML_F32Cxt_FMA_IMPL(DEFAULT_PG16, a, b, c)
#define GGML_F32Cxt_ADD_IMPL(pg, a, b) svadd_f16_x(pg, a, b)
#define GGML_F32Cxt_ADD(...) GGML_F32Cxt_ADD_IMPL(DEFAULT_PG16, __VA_ARGS__)
#define GGML_F32Cxt_ADD(a, b) GGML_F32Cxt_ADD_IMPL(DEFAULT_PG16, a, b)
#define GGML_F32Cxt_MUL_IMPL(pg, a, b) svmul_f16_x(pg, a, b)
#define GGML_F32Cxt_MUL(...) GGML_F32Cxt_MUL_IMPL(DEFAULT_PG16, __VA_ARGS__)
#define GGML_F32Cxt_MUL(a, b) GGML_F32Cxt_MUL_IMPL(DEFAULT_PG16, a, b)
#define GGML_F32Cxt_REDUCE GGML_F16xt_REDUCE_MIXED
#define GGML_F16x_VEC GGML_F32Cxt
@ -224,7 +225,7 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
#define GGML_F16x_VEC_REDUCE GGML_F32Cxt_REDUCE
#define GGML_F16xt_REDUCE_ONE_IMPL(pg, a) svaddv_f16(pg, a)
#define GGML_F16xt_REDUCE_ONE(...) GGML_F16xt_REDUCE_ONE_IMPL(DEFAULT_PG16, __VA_ARGS__)
#define GGML_F16xt_REDUCE_ONE(a) GGML_F16xt_REDUCE_ONE_IMPL(DEFAULT_PG16, a)
#define GGML_F16xt_REDUCE_MIXED_IMPL(pg16, res, sum1, sum2, sum3, sum4) \
{ \
@ -234,7 +235,8 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
__fp16 sum_f16 = svaddv_f16(pg16, sum1); \
(res) = (ggml_float) sum_f16; \
}
#define GGML_F16xt_REDUCE_MIXED(...) GGML_F16xt_REDUCE_MIXED_IMPL(DEFAULT_PG16, __VA_ARGS__)
#define GGML_F16xt_REDUCE_MIXED(res, sum1, sum2, sum3, sum4) \
GGML_F16xt_REDUCE_MIXED_IMPL(DEFAULT_PG16, res, sum1, sum2, sum3, sum4)
// F16 NEON