ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-24 15:14:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq3_ks: NEON convert to q8_k_r8

Browse Source 
This gives us PP-512 = 164 t/s.
This commit is contained in:
Iwan Kawrakow
2025-07-01 11:29:27 +02:00
parent 8fb48218cb
commit cb09eb5c19
2 changed files with 67 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

66
ggml/src/iqk/iqk_gemm_iqk_quants.cpp
View File

@@ -4409,6 +4409,71 @@ void iqk_convert_iq2_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
}
}
void iqk_convert_iq3_ks_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_iq3_ks * x8[8];
block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
int8x16x2_t values;
{
auto v1 = vld1_s8(iq3nl_values+0);
auto v2 = vld1_s8(iq3nl_values+8);
values.val[0] = vcombine_s8(v1, v1);
values.val[1] = vcombine_s8(v2, v2);
}
ggml_half dh[8];
int8x16x2_t xv[8];
uint32_t block[8];
int8_t ls[16];
auto ml = vdupq_n_u8(0x03);
auto mh = vdupq_n_u8(0x04);
for (int ix = 0; ix < nrc_x; ix += 8) {
for (int k = 0; k < 8; ++k) {
auto dptr = (const ggml_half *)((const char *)vx + (ix+k)*bx);
dh[k] = dptr[0];
x8[k] = (const block_iq3_ks *)(dptr + 1);
}
for (int i = 0; i < nb; ++i) {
for (int k = 0; k < 8; ++k) {
auto extra = x8[k][i].extra;
auto extra_v = extra >> 8;
auto hbits = vld1q_u8_x2(x8[k][i].qh);
for (int i128 = 0; i128 < 2; ++i128) {
ls[8*i128+0] = ls[8*i128+1] = int8_t(((x8[k][i].scales[0] >> 4*i128) & 0xf) | ((extra << 4) & 0x10)) - 16;
ls[8*i128+2] = ls[8*i128+3] = int8_t(((x8[k][i].scales[1] >> 4*i128) & 0xf) | ((extra << 3) & 0x10)) - 16;
ls[8*i128+4] = ls[8*i128+5] = int8_t(((x8[k][i].scales[2] >> 4*i128) & 0xf) | ((extra << 2) & 0x10)) - 16;
ls[8*i128+6] = ls[8*i128+7] = int8_t(((x8[k][i].scales[3] >> 4*i128) & 0xf) | ((extra << 1) & 0x10)) - 16;
auto bits = vld1q_u8_x2(x8[k][i].qs+32*i128);
xv[4*i128+0].val[0] = vqtbl1q_s8(values.val[extra_v & 1], vorrq_u8(vandq_u8(bits.val[0], ml), vandq_u8(vshlq_n_u8(hbits.val[0], 2), mh)));
xv[4*i128+0].val[1] = vqtbl1q_s8(values.val[extra_v & 1], vorrq_u8(vandq_u8(bits.val[1], ml), vandq_u8(vshlq_n_u8(hbits.val[1], 2), mh))); extra_v >>= 1;
xv[4*i128+1].val[0] = vqtbl1q_s8(values.val[extra_v & 1], vorrq_u8(vandq_u8(vshrq_n_u8(bits.val[0], 2), ml), vandq_u8(vshlq_n_u8(hbits.val[0], 1), mh)));
xv[4*i128+1].val[1] = vqtbl1q_s8(values.val[extra_v & 1], vorrq_u8(vandq_u8(vshrq_n_u8(bits.val[1], 2), ml), vandq_u8(vshlq_n_u8(hbits.val[1], 1), mh))); extra_v >>= 1;
xv[4*i128+2].val[0] = vqtbl1q_s8(values.val[extra_v & 1], vorrq_u8(vandq_u8(vshrq_n_u8(bits.val[0], 4), ml), vandq_u8(hbits.val[0], mh)));
xv[4*i128+2].val[1] = vqtbl1q_s8(values.val[extra_v & 1], vorrq_u8(vandq_u8(vshrq_n_u8(bits.val[1], 4), ml), vandq_u8(hbits.val[1], mh))); extra_v >>= 1;
xv[4*i128+3].val[0] = vqtbl1q_s8(values.val[extra_v & 1], vorrq_u8(vshrq_n_u8(bits.val[0], 6), vandq_u8(vshrq_n_u8(hbits.val[0], 1), mh)));
xv[4*i128+3].val[1] = vqtbl1q_s8(values.val[extra_v & 1], vorrq_u8(vshrq_n_u8(bits.val[1], 6), vandq_u8(vshrq_n_u8(hbits.val[1], 1), mh))); extra_v >>= 1;
hbits.val[0] = vshrq_n_u8(hbits.val[0], 4);
hbits.val[1] = vshrq_n_u8(hbits.val[1], 4);
extra >>= 4;
}
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(GGML_FP16_TO_FP32(dh[k])*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);
@@ -4618,6 +4683,7 @@ bool iqk_convert_iqk_quants_q80_r8(int type, int n, const void * vx, size_t bx,
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_IQ3_KS : iqk_convert_iq3_ks_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

@@ -296,6 +296,7 @@ struct MulMat {
case GGML_TYPE_IQ3_KT : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ4_KT : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ2_KS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ3_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;