ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-25 07:34:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq4_xs_r4: NEON

Browse Source 
We get PP-512(LLaMA-3.1-8B) = 115.6 t/s on M2-Max,
up from 68.2 t/s for iq4_xs!
This commit is contained in:
Iwan Kawrakow
2024-12-04 14:52:22 +01:00
parent 525ad3b4c1
commit 59a3e12508
1 changed files with 65 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

65
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -7300,6 +7300,60 @@ void mul_mat_iq4_nl_x4_q8_0(int n, const void * vx, size_t bx, const DataInfo& i
}
}
template <int nrc_y>
void mul_mat_iq4_xs_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
GGML_ASSERT(nrc_x%4 == 0);
Q8<nrc_y, block_q8_K> q8(info);
auto m4 = vdupq_n_u8(0xf);
auto values = vld1q_s8(iq4k_values);
int nbl = n / QK_K;
int8x16_t qx[8];
float32x4_t acc[nrc_y] = {};
for (int ix = 0; ix < nrc_x; ix += 4) {
const block_iq4_xs_r4 * iq4 = (const block_iq4_xs_r4 *)((const char *)vx + ix*bx);
for (int ibl = 0; ibl < nbl; ++ibl) {
const uint32_t * scales_l = (const uint32_t *)iq4[ibl].scales_l;
const uint32_t * scales_h = (const uint32_t *)iq4[ibl].scales_h;
auto d4 = vcvt_f32_f16(vld1_f16((const float16_t *)iq4[ibl].d));
for (int ib = 0; ib < QK_K/32; ++ib) {
auto ul = (scales_l[ib%4] >> 4*(ib/4)) & 0x0f0f0f0f;
auto uh = (scales_h[ib%2] >> 2*(ib/2)) & 0x03030303;
auto sl8 = vsub_s8(vreinterpret_s8_s32(vdup_n_s32(ul | (uh << 4))), vdup_n_s8(32));
auto sl16 = vmovl_s8(sl8);
auto sl32 = vmovl_s16(vget_low_s16(sl16));
auto scales = vmulq_f32(d4, vcvtq_f32_s32(sl32));
auto bits = vld1q_u8_x4(iq4[ibl].qs + 64*ib);
qx[0] = vqtbl1q_s8(values, vandq_u8(bits.val[0], m4)); // 0...3 from the 4 rows
qx[1] = vqtbl1q_s8(values, vandq_u8(bits.val[1], m4)); // 16..19
qx[2] = vqtbl1q_s8(values, vandq_u8(bits.val[2], m4)); // 4...7
qx[3] = vqtbl1q_s8(values, vandq_u8(bits.val[3], m4)); // 20..23
qx[4] = vqtbl1q_s8(values, vshrq_n_u8(bits.val[0], 4)); // 8..11
qx[5] = vqtbl1q_s8(values, vshrq_n_u8(bits.val[1], 4)); // 24..27
qx[6] = vqtbl1q_s8(values, vshrq_n_u8(bits.val[2], 4)); // 12..15
qx[7] = vqtbl1q_s8(values, vshrq_n_u8(bits.val[3], 4)); // 28..31
for (int iy = 0; iy < nrc_y; ++iy) {
auto y = vld1q_s8_x2(q8.y[iy][ibl].qs+32*ib);
auto sumi = vdupq_n_s32(0);
sumi = vdotq_laneq_s32(sumi, qx[0], y.val[0], 0);
sumi = vdotq_laneq_s32(sumi, qx[1], y.val[1], 0);
sumi = vdotq_laneq_s32(sumi, qx[2], y.val[0], 1);
sumi = vdotq_laneq_s32(sumi, qx[3], y.val[1], 1);
sumi = vdotq_laneq_s32(sumi, qx[4], y.val[0], 2);
sumi = vdotq_laneq_s32(sumi, qx[5], y.val[1], 2);
sumi = vdotq_laneq_s32(sumi, qx[6], y.val[0], 3);
sumi = vdotq_laneq_s32(sumi, qx[7], y.val[1], 3);
auto d4d8 = vmulq_f32(scales, vdupq_n_f32(q8.scale(iy, ibl)));
acc[iy] = vfmaq_f32(acc[iy], d4d8, vcvtq_f32_s32(sumi));
}
}
}
for (int iy = 0; iy < nrc_y; ++iy) {
info.store(ix, iy, acc[iy]);
acc[iy] = vdupq_n_f32(0.f);
}
}
}
void mul_mat_iq4_nl_x4_q8_0_1(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
GGML_ASSERT(nrc_x%4 == 0);
Q8<1, block_q8_0_x4> q8(info);
@@ -7707,6 +7761,17 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& m, int /*Ny*/) {
m.funcs[7] = mul_mat_iq4_nl_x4_q8_0<8>;
expected_Btype = GGML_TYPE_Q8_0;
break;
case GGML_TYPE_IQ4_XS_R4:
m.funcs[0] = mul_mat_iq4_nl_x4_q8_0_1;
m.funcs[1] = mul_mat_iq4_xs_r4_q8_k<2>;
m.funcs[2] = mul_mat_iq4_xs_r4_q8_k<3>;
m.funcs[3] = mul_mat_iq4_xs_r4_q8_k<4>;
m.funcs[4] = mul_mat_iq4_xs_r4_q8_k<5>;
m.funcs[5] = mul_mat_iq4_xs_r4_q8_k<6>;
m.funcs[6] = mul_mat_iq4_xs_r4_q8_k<7>;
m.funcs[7] = mul_mat_iq4_xs_r4_q8_k<8>;
expected_Btype = GGML_TYPE_Q8_K;
break;
case GGML_TYPE_Q4_0_R4:
m.funcs[0] = mul_mat_q4_0_r4_q8_0<1>;
m.funcs[1] = mul_mat_q4_0_r4_q8_0<2>;