mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 35 Packages Projects Releases Wiki Activity

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

Browse Source
This commit is contained in:
shalinib-ibm
2026-01-15 15:01:18 +05:30
committed by GitHub
parent a7e6ddb8bd
commit 8cc0ba957b
2 changed files with 49 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/vec.cpp
+18
View File
@@ -237,6 +237,24 @@ void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t *
sumf += __riscv_vfmv_f_s_f32m1_f32(redsum);
#endif
#if defined(__POWER9_VECTOR__)
const int np = (n & ~(GGML_BF16_STEP - 1));
if (np > 0) {
GGML_F32_VEC sum[4] = {GGML_F32_VEC_ZERO};
for (; i < np; i += GGML_BF16_STEP) {
GGML_BF16_VEC vx0 = GGML_BF16_VEC_LOAD(x + i);
GGML_BF16_VEC vx1 = GGML_BF16_VEC_LOAD(x + i + 8);
GGML_BF16_VEC vy0 = GGML_BF16_VEC_LOAD(y + i);
GGML_BF16_VEC vy1 = GGML_BF16_VEC_LOAD(y + i + 8);
GGML_BF16_FMA_LO(sum[0], vx0, vy0);
GGML_BF16_FMA_HI(sum[1], vx0, vy0);
GGML_BF16_FMA_LO(sum[2], vx1, vy1);
GGML_BF16_FMA_HI(sum[3], vx1, vy1);
}
GGML_F32x4_REDUCE_4(sumf, sum[0], sum[1], sum[2], sum[3]);
}
#endif
for (; i < n; ++i) {
sumf += (ggml_float)(GGML_BF16_TO_FP32(x[i]) *
GGML_BF16_TO_FP32(y[i]));