minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[CPU] improve silu performance by replacing fp32 div with rcp14 (#31304)

Browse Source
This commit is contained in:
Ma Mingfei
2026-07-17 12:02:37 +08:00
committed by GitHub
parent 27a52d2530
commit 486a56be56
15 changed files with 509 additions and 557 deletions
Download Patch File Download Diff File Expand all files Collapse all files
sgl-kernel/csrc/cpu/decode.cpp
+4 -4
View File
@@ -150,8 +150,9 @@ inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ acc,
int64_t d = 0;
#pragma GCC unroll 4
for (; d <= size - kVecSize; d += kVecSize) {
fVec a_fvec0 = fVec::loadu(acc + d) * s_fvec;
fVec a_fvec1 = fVec::loadu(acc + d + fVec::size()) * s_fvec;
auto [a_fvec0, a_fvec1] = load_float_vec2(acc + d);
a_fvec0 = a_fvec0 * s_fvec;
a_fvec1 = a_fvec1 * s_fvec;
bVec out_bvec = convert_from_float_ext<scalar_t>(a_fvec0, a_fvec1);
out_bvec.store(out + d);
}
@@ -186,8 +187,7 @@ inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ inpu
constexpr int col = i % COLS;
// for COLS = 2, 4 use 512bit store
if constexpr (col % 2 == 0) {
fVec a_fvec0 = fVec::loadu(input + col * 16);
fVec a_fvec1 = fVec::loadu(input + col * 16 + 16);
auto [a_fvec0, a_fvec1] = load_float_vec2(input + col * 16);
bVec out_bvec = convert_from_float_ext<scalar_t>(a_fvec0, a_fvec1);
out_bvec.store(out + col * 16);
}