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

iq1_m

Browse Source 
19 t/s -> 163 t/s.
This commit is contained in:
Iwan Kawrakow
2025-06-24 14:05:30 +02:00
parent 3c5b788b70
commit e5e5acfdda
2 changed files with 59 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
ggml/src/iqk/iqk_gemm_1bit.cpp
View File

@@ -2606,6 +2606,63 @@ void iqk_convert_iq1_s_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
}
}
void iqk_convert_iq1_m_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_iq1_m * x8[8];
block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
int8_t ls[16];
uint32_t block[8];
int8x16x2_t qx[8];
uint32x4x2_t mask = {uint32x4_t{0x00000008, 0x00000008, 0x00000080, 0x00000080}, uint32x4_t {0x00080000, 0x00080000, 0x00800000, 0x00800000}};
for (int ix = 0; ix < nrc_x; ix += 8) {
for (int k = 0; k < 8; ++k) x8[k] = (const block_iq1_m *)((const char *)vx + (ix + k)*bx);
for (int i = 0; i < nb; ++i) {
for (int k = 0; k < 8; ++k) {
const uint16_t * sc = (const uint16_t *)x8[k][i].scales;
iq1m_scale_t scale;
scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
float d = 0.125f * GGML_FP16_TO_FP32(scale.f16);
auto qs = x8[k][i].qs;
auto qh = x8[k][i].qh;
int8x16x2_t value;
for (int ib32 = 0; ib32 < 8; ++ib32) {
ls[2*ib32 + 0] = (2*((sc[ib32/2] >> (6*(ib32%2)+0)) & 0x7) + 1);
ls[2*ib32 + 1] = (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1);
//value = _mm256_set_epi64x(iq1s_grid[qs[3] | ((qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | ((qh[1] << 8) & 0x700)],
// iq1s_grid[qs[1] | ((qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | ((qh[0] << 8) & 0x700)]);
value.val[0] = vreinterpretq_s8_u64(uint64x2_t{iq1s_grid[qs[0] | ((qh[0] << 8) & 0x700)], iq1s_grid[qs[1] | ((qh[0] << 4) & 0x700)]});
value.val[1] = vreinterpretq_s8_u64(uint64x2_t{iq1s_grid[qs[2] | ((qh[1] << 8) & 0x700)], iq1s_grid[qs[3] | ((qh[1] << 4) & 0x700)]});
value.val[0] = vshlq_n_s8(vaddq_s8(value.val[0], vdupq_n_s8(1)), 3);
value.val[1] = vshlq_n_s8(vaddq_s8(value.val[1], vdupq_n_s8(1)), 3);
auto aux = vdupq_n_u32(qh[0] | qh[1] << 16);
uint32x4x2_t delta_mask{ vceqq_u32(vandq_u32(aux, mask.val[0]), mask.val[0]), vceqq_u32(vandq_u32(aux, mask.val[1]), mask.val[1]) };
uint8x16x2_t delta{ vaddq_s8(vdupq_n_s8(7), vandq_s8(vdupq_n_s8(2), vreinterpretq_s8_u32(delta_mask.val[0]))),
vaddq_s8(vdupq_n_s8(7), vandq_s8(vdupq_n_s8(2), vreinterpretq_s8_u32(delta_mask.val[1]))) };
qx[ib32].val[0] = vsubq_s8(value.val[0], delta.val[0]);
qx[ib32].val[1] = vsubq_s8(value.val[1], delta.val[1]);
qs += 4;
qh += 2;
}
float dnew = convert_to_q8_k_r8(1.f/126, qx, ls, block, (uint32_t *)y[i].qs + k);
y[i].d[k] = GGML_FP32_TO_FP16(d*dnew);
}
}
y += nb;
}
}
}
bool iqk_set_kernels_1bit(int ne00, int typeA, int typeB, std::array<mul_mat_t, IQK_MAX_NY>& funcs, mul_mat_t& func16) {
@@ -2668,7 +2725,7 @@ bool iqk_convert_1bit_q80_r8(int type, int n, const void * vx, size_t bx, void *
if (n%QK_K != 0 || nrc_x%8 != 0) return false;
switch (ggml_type(type)) {
case GGML_TYPE_IQ1_S: iqk_convert_iq1_s_q8_k_r8(n, vx, bx, vy, nrc_x); break;
//case GGML_TYPE_IQ1_M: iqk_convert_iq1_m_q8_k_r8(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ1_M: iqk_convert_iq1_m_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

@@ -277,6 +277,7 @@ struct MulMat {
case GGML_TYPE_Q5_K : return nrc_y >= 32 ? GGML_TYPE_Q8_1 : type;
case GGML_TYPE_Q6_K : return nrc_y >= 64 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ1_S : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ1_M : return nrc_y >= 8 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ2_XXS: return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ2_XS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ2_S : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;