ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-25 15:44:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

q6_k_r4: 1st NEON version

Browse Source 
PP-512(LLaMA-3.1-8B) = 78 t/s vs 57.6 t/s for q6_K.
TG-128 is slightly lower rthan q6_K for low number of threads,
becomes very slightly better at 8 threads.
This commit is contained in:
Iwan Kawrakow
2024-12-10 08:32:24 +01:00
parent a5db08118a
commit 99d79116ee
1 changed files with 77 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

77
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -7955,6 +7955,19 @@ IQK_ALWAYS_INLINE int32x4_t interleaved_dotq(const int8x16_t * qx, const int8x16
return sumi;
}
IQK_ALWAYS_INLINE int32x4x2_t interleaved_dotq_b16(const int8x16_t * qx, const int8x16x2_t& y) {
int32x4x2_t sumi = { vdupq_n_s32(0), vdupq_n_s32(0) };
sumi.val[0] = vdotq_laneq_s32(sumi.val[0], qx[0], y.val[0], 0);
sumi.val[1] = vdotq_laneq_s32(sumi.val[1], qx[1], y.val[1], 0);
sumi.val[0] = vdotq_laneq_s32(sumi.val[0], qx[2], y.val[0], 1);
sumi.val[1] = vdotq_laneq_s32(sumi.val[1], qx[3], y.val[1], 1);
sumi.val[0] = vdotq_laneq_s32(sumi.val[0], qx[4], y.val[0], 2);
sumi.val[1] = vdotq_laneq_s32(sumi.val[1], qx[5], y.val[1], 2);
sumi.val[0] = vdotq_laneq_s32(sumi.val[0], qx[6], y.val[0], 3);
sumi.val[1] = vdotq_laneq_s32(sumi.val[1], qx[7], y.val[1], 3);
return sumi;
}
IQK_ALWAYS_INLINE void prepare_iq4_nl_quants(const int8x16_t& values, const uint8x16_t& m4, const uint8x16x4_t& bits, int8x16_t * qx) {
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
@@ -8101,6 +8114,66 @@ void mul_mat_q4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& inf
}
}
template <int nrc_y>
void mul_mat_q6_k_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 mf = vdupq_n_u8(0x0f);
auto m3 = vdupq_n_u8(0x30);
auto m32 = vdupq_n_s8(-32);
//auto mx = vdupq_n_u8(0xaa);
int nbl = n / QK_K;
int8x16_t qx[8];
float32x4x2_t scales;
float32x4_t acc[nrc_y] = {};
for (int ix = 0; ix < nrc_x; ix += 4) {
const block_q6_k_r4 * iq6 = (const block_q6_k_r4 *)((const char *)vx + ix*bx);
for (int ibl = 0; ibl < nbl; ++ibl) {
auto d4 = vcvt_f32_f16(vld1_f16((const float16_t *)iq6[ibl].d));
if constexpr (nrc_y == 1) {
d4 = vmulq_f32(d4, vdupq_n_f32(q8.scale(0, ibl)));
}
for (int is = 0; is < 2; ++is) {
for (int ib = 0; ib < 4; ++ib) {
auto lbits = vld1q_u8_x4(iq6[ibl].ql + 256*is + 64*ib);
auto hbits = vld1q_u8_x2(iq6[ibl].qh + 128*is + 32*ib);
auto iscales = vmovl_s8(vld1_s8(iq6[ibl].scales + 32*is + 8*ib));
scales.val[0] = vmulq_f32(d4, vcvtq_f32_s32(vmovl_s16(vget_low_s16(iscales))));
scales.val[1] = vmulq_f32(d4, vcvtq_f32_s32(vmovl_s16(vget_high_s16(iscales))));
//hbits.val[0] = veorq_u8(hbits.val[0], mx);
//hbits.val[1] = veorq_u8(hbits.val[1], mx);
//qx[0] = vorrq_u8(vandq_u8(vshlq_n_u8(hbits.val[0], 6), m3), vandq_u8(vshlq_n_u8(lbits.val[0], 4), mf));
prepare_q4_k_quants(mf, lbits, qx);
qx[0] = vaddq_s8(m32, vorrq_u8(qx[0], vandq_u8(m3, vshlq_n_u8(hbits.val[0], 4))));
qx[1] = vaddq_s8(m32, vorrq_u8(qx[1], vandq_u8(m3, vshlq_n_u8(hbits.val[1], 4))));
qx[2] = vaddq_s8(m32, vorrq_u8(qx[2], vandq_u8(m3, hbits.val[0])));
qx[3] = vaddq_s8(m32, vorrq_u8(qx[3], vandq_u8(m3, hbits.val[1])));
qx[4] = vaddq_s8(m32, vorrq_u8(qx[4], vandq_u8(m3, vshlq_n_u8(hbits.val[0], 2))));
qx[5] = vaddq_s8(m32, vorrq_u8(qx[5], vandq_u8(m3, vshlq_n_u8(hbits.val[1], 2))));
qx[6] = vaddq_s8(m32, vorrq_u8(qx[6], vandq_u8(m3, vshrq_n_u8(hbits.val[0], 2))));
qx[7] = vaddq_s8(m32, vorrq_u8(qx[7], vandq_u8(m3, vshrq_n_u8(hbits.val[1], 2))));
for (int iy = 0; iy < nrc_y; ++iy) {
auto y = vld1q_s8_x2(q8.y[iy][ibl].qs+128*is+32*ib);
auto sumi = interleaved_dotq_b16(qx, y);
if constexpr (nrc_y == 1) {
acc[iy] = vfmaq_f32(acc[iy], scales.val[0], vcvtq_f32_s32(sumi.val[0]));
acc[iy] = vfmaq_f32(acc[iy], scales.val[1], vcvtq_f32_s32(sumi.val[1]));
} else {
auto d8 = vdupq_n_f32(q8.scale(iy, ibl));
acc[iy] = vfmaq_f32(acc[iy], vmulq_f32(scales.val[0], d8), vcvtq_f32_s32(sumi.val[0]));
acc[iy] = vfmaq_f32(acc[iy], vmulq_f32(scales.val[1], d8), vcvtq_f32_s32(sumi.val[1]));
}
}
}
}
}
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_r4_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);
@@ -8502,6 +8575,10 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& m, int /*Ny*/) {
SET_MUL_MAT_FUNCTIONS(m, mul_mat_q4_k_r4_q8_k);
expected_Btype = GGML_TYPE_Q8_K32;
break;
case GGML_TYPE_Q6_K_R4:
SET_MUL_MAT_FUNCTIONS(m, mul_mat_q6_k_r4_q8_k);
expected_Btype = GGML_TYPE_Q8_K;
break;
case GGML_TYPE_Q4_0_R4:
SET_MUL_MAT_FUNCTIONS_T(m, mul_mat_qx_r4_q8_0, Q4_0_R4_Dequantizer);
expected_Btype = GGML_TYPE_Q8_0;