ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-10 22:10:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

Q3_K_R4 (#134)

Browse Source 
* q3_k_r4: Zen4 works, but not as good as it should be

238 t/s, so sloghtly slower than q6_k_r4.

* q3_k_r4: NEON

We get PP-512(LLaMA-3.1-8B) = 106.9 t/s.
This is 1.93X faster than q3_K_S!

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-12-11 11:19:00 +01:00
committed by GitHub
parent e78e47b857
commit 4872f2f57e
10 changed files with 356 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

136
ggml/src/iqk/iqk_quantize.cpp
View File

@@ -4301,3 +4301,139 @@ void vec_dot_q5_k_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t b
GGML_UNUSED(by);
}
//
// ========================================= q3_k_r4
//
void quantize_row_q3_k_r4_ref(const float * x, block_q3_k_r4 * y, int64_t k) {
quantize_q3_k_r4(x, (void *)y, 4, k/4, nullptr);
}
void quantize_row_q3_k_r4(const float * x, void * y, int64_t k) {
quantize_q3_k_r4(x, y, 4, k/4, nullptr);
}
namespace {
inline void convert_q3_k(const block_q3_K& x, uint8_t * L, uint8_t * Ld) {
constexpr uint32_t kmask1 = 0x03030303;
constexpr uint32_t kmask2 = 0x0f0f0f0f;
uint32_t aux[4];
memcpy(aux, x.scales, 12);
uint32_t tmp = aux[2];
aux[2] = ((aux[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
aux[3] = ((aux[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
aux[0] = (aux[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
aux[1] = (aux[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
std::memcpy(Ld, aux, 16);
const uint8_t * q = x.qs;
const uint8_t * hm = x.hmask;
uint8_t m = 1;
for (int n = 0; n < QK_K; n += 128) {
int shift = 0;
for (int j = 0; j < 4; ++j) {
for (int l = 0; l < 32; ++l) {
*L++ = ((q[l] >> shift) & 3) + ((hm[l] & m) ? 4 : 0);
}
shift += 2;
m <<= 1;
}
q += 32;
}
}
}
static void repack_q3_k(int nrows, int n_per_row, const block_q3_K * x, block_q3_k_r4 * y) {
GGML_ASSERT(nrows%4 == 0);
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
const block_q3_K * x4[4];
uint8_t L[QK_K], Ld[QK_K/16];
for (int row = 0; row < nrows; row += 4) {
for (int k = 0; k < 4; ++k) x4[k] = x + nblock*k;
for (int ibl = 0; ibl < nblock; ++ibl) {
std::memset(y[ibl].scales_l, 0, QK_K/8);
std::memset(y[ibl].scales_h, 0, QK_K/16);
for (int k = 0; k < 4; ++k) {
y[ibl].d[k] = x4[k][ibl].d;
convert_q3_k(x4[k][ibl], L, Ld);
for (int ib = 0; ib < QK_K/32; ++ib) {
int is = 8*ib+k;
y[ibl].scales_l[is%32] |= (Ld[2*ib+0] & 0xf) << 4*(is/32);
y[ibl].scales_h[is%16] |= (Ld[2*ib+0] >> 4) << 2*(is/16);
is += 4;
y[ibl].scales_l[is%32] |= (Ld[2*ib+1] & 0xf) << 4*(is/32);
y[ibl].scales_h[is%16] |= (Ld[2*ib+1] >> 4) << 2*(is/16);
for (int i = 0; i < 4; ++i) {
y[ibl].qs[32*ib+4*k+i+ 0] = ((L[32*ib+i+ 0] & 0x3) << 0) | ((L[32*ib+i+ 4] & 0x3) << 2) | ((L[32*ib+i+ 8] & 0x3) << 4) | ((L[32*ib+i+12] & 0x3) << 6);
y[ibl].qs[32*ib+4*k+i+16] = ((L[32*ib+i+16] & 0x3) << 0) | ((L[32*ib+i+20] & 0x3) << 2) | ((L[32*ib+i+24] & 0x3) << 4) | ((L[32*ib+i+28] & 0x3) << 6);
y[ibl].qh[16*ib+4*k+i+ 0] = ((L[32*ib+i+ 0] >> 2) << 0) | ((L[32*ib+i+ 4] >> 2) << 1) | ((L[32*ib+i+ 8] >> 2) << 2) | ((L[32*ib+i+12] >> 2) << 3)
| ((L[32*ib+i+16] >> 2) << 4) | ((L[32*ib+i+20] >> 2) << 5) | ((L[32*ib+i+24] >> 2) << 6) | ((L[32*ib+i+28] >> 2) << 7);
}
}
}
}
x += 4*nblock;
y += nblock;
}
}
size_t quantize_q3_k_r4(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) {
GGML_ASSERT(nrows%4 == 0);
GGML_ASSERT(n_per_row%QK_K == 0);
char * qcur = (char *)dst;
auto row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row);
std::vector<char> qtmp(4*row_size);
for (int row = 0; row < nrows; row += 4) {
quantize_q3_K(src, (void *)qtmp.data(), 4, n_per_row, imatrix);
repack_q3_k(4, n_per_row, (const block_q3_K *)qtmp.data(), (block_q3_k_r4 *)qcur);
qcur += 4*row_size;
src += 4*n_per_row;
}
return nrows*row_size;
}
void dequantize_row_q3_k_r4(const block_q3_k_r4 * x, float * y, int64_t k) {
auto n_per_row = k/4;
float * y4[4] = {y, y + n_per_row, y + 2*n_per_row, y + 3*n_per_row};
int nblock = n_per_row/QK_K;
for (int ibl = 0; ibl < nblock; ++ibl) {
for (int k = 0; k < 4; ++k) {
const float d = GGML_FP16_TO_FP32(x[ibl].d[k]);
auto ql = x[ibl].qs;
auto qh = x[ibl].qh;
for (int ib = 0; ib < QK_K/32; ++ib) {
int is = 8*ib + k;
float dl1 = d * ((((x[ibl].scales_l[is%32] >> 4*(is/32)) & 0xf) | (((x[ibl].scales_h[is%16] >> 2*(is/16)) & 0x03) << 4)) - 32);
is += 4;
float dl2 = d * ((((x[ibl].scales_l[is%32] >> 4*(is/32)) & 0xf) | (((x[ibl].scales_h[is%16] >> 2*(is/16)) & 0x03) << 4)) - 32);
for (int i = 0; i < 4; ++i) {
y4[k][QK_K*ibl+32*ib+i+ 0] = dl1 * ((((ql[4*k+i+ 0] >> 0) & 3) | ((qh[4*k+i] << 2) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+ 4] = dl1 * ((((ql[4*k+i+ 0] >> 2) & 3) | ((qh[4*k+i] << 1) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+ 8] = dl1 * ((((ql[4*k+i+ 0] >> 4) & 3) | ((qh[4*k+i] << 0) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+12] = dl1 * ((((ql[4*k+i+ 0] >> 6) & 3) | ((qh[4*k+i] >> 1) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+16] = dl2 * ((((ql[4*k+i+16] >> 0) & 3) | ((qh[4*k+i] >> 2) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+20] = dl2 * ((((ql[4*k+i+16] >> 2) & 3) | ((qh[4*k+i] >> 3) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+24] = dl2 * ((((ql[4*k+i+16] >> 4) & 3) | ((qh[4*k+i] >> 4) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+28] = dl2 * ((((ql[4*k+i+16] >> 6) & 3) | ((qh[4*k+i] >> 5) & 4)) - 4);
}
ql += 32;
qh += 16;
}
}
}
}
void vec_dot_q3_k_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) {
#if GGML_USE_IQK_MULMAT
if (iqk_mul_mat(1, 1, n, GGML_TYPE_Q3_K_R4, vx, 0, GGML_TYPE_Q8_K, vy, 0, s, 0, 0, 1)) {
return;
}
#endif
GGML_ASSERT(n%QK4_NL == 0);
GGML_ASSERT(nrc == 1);
GGML_UNUSED(bs);
GGML_UNUSED(bx);
GGML_UNUSED(by);
}