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

opt bytes_from_nibbles_32 and sum_i16_pairs_float

Browse source
This commit is contained in:
Jinyang He 2024-04-09 10:05:51 +08:00 • committed by yala
parent 4cfd8b912e
commit e8ed67052a
1 changed files with 5 additions and 11 deletions
Download patch file Download diff file Expand all files Collapse all files

16
ggml-quants.c
View file

@ -653,21 +653,15 @@ static inline __m256i bytes_from_bits_32(const uint8_t * x) {
// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
{
const __m128i tmp = __lsx_vld((const __m128i *)rsi, 0);
__m128i tmp2 = __lsx_vsrli_h(tmp, 4);
__m128i lowMask = __lsx_vreplgr2vr_b(0xf);
__m128i tmpl = __lsx_vand_v(tmp, lowMask);
__m128i tmph = __lsx_vand_v(tmp2, lowMask);
return MM256_SET_M128I(tmph, tmpl);
const __m128i lo = __lsx_vld((const __m128i *)rsi, 0);
__m128i hi = __lsx_vsrli_h(lo, 4);
return __lasx_xvandi_b(MM256_SET_M128I(hi, lo), 0xf);
}
// add int16_t pairwise and return as float vector
static inline __m256 sum_i16_pairs_float(const __m256i x) {
const __m256i ones = __lasx_xvreplgr2vr_h(1);
__m256i zero256 = __lasx_xvldi(0);
const __m256i tmp1 = __lasx_xvmaddwev_w_h(zero256, ones, x);
const __m256i summed_pairs = __lasx_xvmaddwod_w_h(tmp1, ones, x);
__m256i v = __lasx_xvpackod_h(x, x);
__m256i summed_pairs = __lasx_xvaddwev_w_h(x, v);
return __lasx_xvffint_s_w(summed_pairs);
}