ggml-cpu: optimize ggml_vec_dot_bf16 for Power9 (#18837)

This commit is contained in:
shalinib-ibm
2026-01-15 17:31:18 +08:00
committed by GitHub
parent a7e6ddb8bd
commit 8cc0ba957b
2 changed files with 49 additions and 0 deletions
+31
View File
@@ -654,6 +654,14 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
vec_extract(x[0], 2) + \
vec_extract(x[0], 3); \
}
#define GGML_F32x4_REDUCE_4(res, s0, s1, s2, s3) \
{ \
vector float v = vec_add(vec_add(s0, s1), \
vec_add(s2, s3)); \
v = vec_add(v, vec_sld(v, v, 8)); \
v = vec_add(v, vec_sld(v, v, 4)); \
res += (ggml_float) vec_extract(v, 0); \
}
#define GGML_F32_VEC GGML_F32x4
#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
@@ -690,6 +698,29 @@ static inline unsigned char ggml_endian_byte(int i) {
r[i - GGML_ENDIAN_BYTE(0)]), \
0, p - GGML_F16_EPR)
//BF16 POWER9
#define GGML_BF16_STEP 16
#define GGML_BF16_EPR 8
#define GGML_BF16x8 vector unsigned short
#define GGML_BF16x8_ZERO vec_splats((unsigned short)0)
#define GGML_BF16x8_LOAD(p) vec_xl(0, (const unsigned short *)(p))
#define GGML_BF16_VEC GGML_BF16x8
#define GGML_BF16_VEC_ZERO GGML_BF16x8_ZERO
#define GGML_BF16_VEC_LOAD GGML_BF16x8_LOAD
#if defined(__LITTLE_ENDIAN__)
#define GGML_BF16_TO_F32_LO(v) ((vector float) vec_mergel(GGML_BF16_VEC_ZERO, (v)))
#define GGML_BF16_TO_F32_HI(v) ((vector float) vec_mergeh(GGML_BF16_VEC_ZERO, (v)))
#else
#define GGML_BF16_TO_F32_LO(v) ((vector float) vec_mergel((v), GGML_BF16_VEC_ZERO))
#define GGML_BF16_TO_F32_HI(v) ((vector float) vec_mergeh((v), GGML_BF16_VEC_ZERO))
#endif
#define GGML_BF16_FMA_LO(acc, x, y) \
(acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_LO(x), GGML_BF16_TO_F32_LO(y))
#define GGML_BF16_FMA_HI(acc, x, y) \
(acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_HI(x), GGML_BF16_TO_F32_HI(y))
#elif defined(__wasm_simd128__)
#define GGML_SIMD