Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml : add some lsx support (#23798)

Browse Source 
* loongarch : optimize LSX fp16 load/store with native intrinsics

Use __lsx_vfcvtl_s_h and __lsx_vfcvt_h_s instead of scalar loops in
__lsx_f16x4_load and __lsx_f16x4_store.

* loongarch : add LSX implementation for q8_0 dot product

* loongarch : add LSX implementation for q6_K dot product

* loongarch : add LSX implementation for iq4_xs dot product

* Improve reduce ops when sun int16 pairs to int32
This commit is contained in:
Jinyang He
2026-05-30 11:53:26 +03:00
committed by GitHub
parent 6e093b80ea
commit d48a56effb
2 changed files with 154 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/simd-mappings.h
+3 -16
View File
@@ -1125,25 +1125,12 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
#define GGML_F16_EPR 4
static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) {
float tmp[4];
tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]);
tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]);
tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]);
tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]);
return (__m128)__lsx_vld(tmp, 0);
return __lsx_vfcvtl_s_h(__lsx_vld((const void *)x, 0));
}
static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
float arr[4];
__lsx_vst(y, arr, 0);
x[0] = GGML_CPU_FP32_TO_FP16(arr[0]);
x[1] = GGML_CPU_FP32_TO_FP16(arr[1]);
x[2] = GGML_CPU_FP32_TO_FP16(arr[2]);
x[3] = GGML_CPU_FP32_TO_FP16(arr[3]);
__m128i a = __lsx_vfcvt_h_s(y, y);
memcpy(x, &a, sizeof(ggml_fp16_t) * 4);
}
#define GGML_F32Cx4 __m128