ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-22 15:39:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq6_k

Browse Source 
44 t/s -> 164.3 t/s. There is no iq6_k_r4
This commit is contained in:
Iwan Kawrakow
2025-06-22 19:31:45 +02:00
parent eeea8af04f
commit 425bbe7609
2 changed files with 49 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
ggml/src/iqk/iqk_gemm_iqk_quants.cpp
View File

@@ -4137,6 +4137,53 @@ void iqk_convert_iq5_ks_q8_k_r8(int n, const void * vx, size_t bx, void * vy, in
}
}
void iqk_convert_iq6_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_iq6_k * x8[8];
block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
int8x16x4_t values[2];
values[0] = vld1q_s8_x4(iq6nl_values);
values[1] = vld1q_s8_x4(iq6nl_values+64);
int8x16x2_t xv[8];
uint32_t block[8];
auto ml = vdupq_n_u8(0x0f);
auto mh = vdupq_n_u8(0x30);
for (int ix = 0; ix < nrc_x; ix += 8) {
for (int k = 0; k < 8; ++k) x8[k] = (const block_iq6_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) {
auto hbits = vld1q_u8_x2(x8[k][i].qh+i128*32);
auto bits = vld1q_u8_x4(x8[k][i].qs+i128*64);
xv[4*i128+0].val[0] = vqtbl4q_s8(values[(extra >> 0) & 1], vorrq_u8(vandq_u8(bits.val[0], ml), vandq_u8(vshlq_n_u8(hbits.val[0], 4), mh)));
xv[4*i128+0].val[1] = vqtbl4q_s8(values[(extra >> 1) & 1], vorrq_u8(vandq_u8(bits.val[1], ml), vandq_u8(vshlq_n_u8(hbits.val[1], 4), mh)));
xv[4*i128+1].val[0] = vqtbl4q_s8(values[(extra >> 2) & 1], vorrq_u8(vshrq_n_u8(bits.val[0], 4), vandq_u8(vshlq_n_u8(hbits.val[0], 2), mh)));
xv[4*i128+1].val[1] = vqtbl4q_s8(values[(extra >> 3) & 1], vorrq_u8(vshrq_n_u8(bits.val[1], 4), vandq_u8(vshlq_n_u8(hbits.val[1], 2), mh)));
xv[4*i128+2].val[0] = vqtbl4q_s8(values[(extra >> 4) & 1], vorrq_u8(vandq_u8(bits.val[2], ml), vandq_u8(hbits.val[0], mh)));
xv[4*i128+2].val[1] = vqtbl4q_s8(values[(extra >> 5) & 1], vorrq_u8(vandq_u8(bits.val[3], ml), vandq_u8(hbits.val[1], mh)));
xv[4*i128+3].val[0] = vqtbl4q_s8(values[(extra >> 6) & 1], vorrq_u8(vshrq_n_u8(bits.val[2], 4), vandq_u8(vshrq_n_u8(hbits.val[0], 2), mh)));
xv[4*i128+3].val[1] = vqtbl4q_s8(values[(extra >> 7) & 1], vorrq_u8(vshrq_n_u8(bits.val[3], 4), vandq_u8(vshrq_n_u8(hbits.val[1], 2), mh)));
extra >>= 8;
}
float dnew = convert_to_q8_k_r8(1.f/127, xv, x8[k][i].scales, block, (uint32_t *)y[i].qs + k);
y[i].d[k] = GGML_FP32_TO_FP16(d*dnew);
}
}
y += nb;
}
}
}
bool iqk_convert_iqk_quants_q80_r8(int type, int n, const void * vx, size_t bx, void * vy, int nrc_x) {
@@ -4149,7 +4196,7 @@ bool iqk_convert_iqk_quants_q80_r8(int type, int n, const void * vx, size_t bx,
// case GGML_TYPE_IQ4_K : iqk_convert_iq4_k_q8_k_r8 (n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ5_KS : iqk_convert_iq5_ks_q8_k_r8(n, vx, bx, vy, nrc_x); break;
// case GGML_TYPE_IQ5_K : iqk_convert_iq5_k_q8_k_r8 (n, vx, bx, vy, nrc_x); break;
// case GGML_TYPE_IQ6_K : iqk_convert_iq6_k_q8_k_r8 (n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ6_K : iqk_convert_iq6_k_q8_k_r8 (n, vx, bx, vy, nrc_x); break;
default: return false;
}
return true;

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_IQ6_K : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
default: break;
}
#endif