ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-24 07:04:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq2_k

Browse Source 
47.4 t/s -> 167 t/s. iq2_k_r4 is at 113.3 t/s.
This commit is contained in:
Iwan Kawrakow
2025-06-23 11:15:25 +02:00
parent c15688ca3a
commit ec4b4536ea
2 changed files with 61 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

61
ggml/src/iqk/iqk_gemm_iqk_quants.cpp
View File

@@ -4137,6 +4137,65 @@ void iqk_convert_iq5_ks_q8_k_r8(int n, const void * vx, size_t bx, void * vy, in
}
}
void iqk_convert_iq2_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) {
GGML_ASSERT(n%QK_K == 0);
GGML_ASSERT(nrc_x%8 == 0);
int nb = n/QK_K;
const block_iq2_k * x8[8];
block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
int8x16x2_t values;
{
auto v1 = vld1_s8(iq2nl_values);
auto v2 = vreinterpret_s8_u64(vshr_n_u64(vreinterpret_u64_s8(v1), 32));
values.val[0] = vcombine_s8(v1, v1);
values.val[1] = vcombine_s8(v2, v2);
}
int8x16x2_t xv[8];
uint32_t block[8];
int8_t ls[16];
auto ml = vdupq_n_u8(0x03);
for (int ix = 0; ix < nrc_x; ix += 8) {
for (int k = 0; k < 8; ++k) x8[k] = (const block_iq2_k *)((const char *)vx + (ix+k)*bx);
for (int i = 0; i < nb; ++i) {
for (int k = 0; k < 8; ++k) {
float d = GGML_FP16_TO_FP32(x8[k][i].d);
auto extra = x8[k][i].extra;
for (int i128 = 0; i128 < 2; ++i128) {
ls[8*i128+0] = ((x8[k][i].scales[4*i128+0] & 0xf) - 8);
ls[8*i128+1] = ((x8[k][i].scales[4*i128+0] >> 4) - 8);
ls[8*i128+2] = ((x8[k][i].scales[4*i128+1] & 0xf) - 8);
ls[8*i128+3] = ((x8[k][i].scales[4*i128+1] >> 4) - 8);
ls[8*i128+4] = ((x8[k][i].scales[4*i128+2] & 0xf) - 8);
ls[8*i128+5] = ((x8[k][i].scales[4*i128+2] >> 4) - 8);
ls[8*i128+6] = ((x8[k][i].scales[4*i128+3] & 0xf) - 8);
ls[8*i128+7] = ((x8[k][i].scales[4*i128+3] >> 4) - 8);
auto bits = vld1q_u8_x2(x8[k][i].qs+32*i128);
xv[4*i128+0].val[0] = vqtbl1q_s8(values.val[extra & 1], vandq_u8(bits.val[0], ml)); extra >>= 1;
xv[4*i128+0].val[1] = vqtbl1q_s8(values.val[extra & 1], vandq_u8(bits.val[1], ml)); extra >>= 1;
xv[4*i128+1].val[0] = vqtbl1q_s8(values.val[extra & 1], vandq_u8(vshrq_n_u8(bits.val[0], 2), ml)); extra >>= 1;
xv[4*i128+1].val[1] = vqtbl1q_s8(values.val[extra & 1], vandq_u8(vshrq_n_u8(bits.val[1], 2), ml)); extra >>= 1;
xv[4*i128+2].val[0] = vqtbl1q_s8(values.val[extra & 1], vandq_u8(vshrq_n_u8(bits.val[0], 4), ml)); extra >>= 1;
xv[4*i128+2].val[1] = vqtbl1q_s8(values.val[extra & 1], vandq_u8(vshrq_n_u8(bits.val[1], 4), ml)); extra >>= 1;
xv[4*i128+3].val[0] = vqtbl1q_s8(values.val[extra & 1], vshrq_n_u8(bits.val[0], 6)); extra >>= 1;
xv[4*i128+3].val[1] = vqtbl1q_s8(values.val[extra & 1], vshrq_n_u8(bits.val[1], 6)); extra >>= 1;
}
float dnew = convert_to_q8_k_r8(1.f/127, xv, ls, block, (uint32_t *)y[i].qs + k);
y[i].d[k] = GGML_FP32_TO_FP16(d*dnew);
}
}
y += nb;
}
}
void iqk_convert_iq3_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) {
GGML_ASSERT(n%QK_K == 0);
GGML_ASSERT(nrc_x%8 == 0);
@@ -4345,7 +4404,7 @@ bool iqk_convert_iqk_quants_q80_r8(int type, int n, const void * vx, size_t bx,
if (n%QK_K != 0 || nrc_x%8 != 0) return false;
switch (ggml_type(type)) {
case GGML_TYPE_IQ2_KS : iqk_convert_iq2_ks_q8_k_r8(n, vx, bx, vy, nrc_x); break;
// case GGML_TYPE_IQ2_K : iqk_convert_iq2_k_q8_k_r8 (n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ2_K : iqk_convert_iq2_k_q8_k_r8 (n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ3_K : iqk_convert_iq3_k_q8_k_r8 (n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ4_KS : iqk_convert_iq4_ks_q8_k_r8(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ4_K : iqk_convert_iq4_k_q8_k_r8 (n, vx, bx, vy, nrc_x); break;

1
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -284,6 +284,7 @@ struct MulMat {
case GGML_TYPE_IQ2_KS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ4_KS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ5_KS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ2_K : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ3_K : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ4_K : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ5_K : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;