vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

add fall back for axpy mulmat

Browse source
This commit is contained in:
syx 2023-12-12 13:50:25 +08:00
parent a3c295a2ae
commit 6f997d299a
1 changed files with 8 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml.c
View file

@ -14298,7 +14298,7 @@ static void ggml_axpy_normal_f16(const int n, const ggml_fp16_t * vx, const ggml
} }
} }
static void ggml_axpy_avx_f16(const int n, const ggml_fp16_t * restrict vx, const ggml_fp16_t * restrict vy, void* restrict vz, ggml_fp16_t alpha) { static void ggml_axpy_avx_f16(const int n, const ggml_fp16_t * restrict vx, const ggml_fp16_t * restrict vy, void* restrict vz, ggml_fp16_t alpha) {
#if defined(__AVX2__)
float *result = (float *)vz; float *result = (float *)vz;
float alpha_f32 = GGML_FP16_TO_FP32(alpha); float alpha_f32 = GGML_FP16_TO_FP32(alpha);
__m256 scale = _mm256_set1_ps(alpha_f32); // 创建scale向量 __m256 scale = _mm256_set1_ps(alpha_f32); // 创建scale向量
@ -14309,6 +14309,13 @@ static void ggml_axpy_avx_f16(const int n, const ggml_fp16_t * restrict vx, cons
__m256 res = _mm256_fmadd_ps(vx_f32, scale, vy_f32); // 执行向量加法和乘法操作 __m256 res = _mm256_fmadd_ps(vx_f32, scale, vy_f32); // 执行向量加法和乘法操作
_mm256_storeu_ps((float*)(&result[i]), res); // 存储结果 _mm256_storeu_ps((float*)(&result[i]), res); // 存储结果
} }
#else
float *res = (float *)vz;
float alpha_convert = GGML_FP16_TO_FP32(alpha);
for (int i = 0; i < n; i++) {
res[i] = res[i] + (GGML_FP16_TO_FP32(vx[i])*alpha_convert);
}
#endif
} }
atomic_flag g_axpy_dense_lock = ATOMIC_FLAG_INIT; atomic_flag g_axpy_dense_lock = ATOMIC_FLAG_INIT;