ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-25 08:59:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq3_xs

Browse Source 
178 t/s -> 363 t/s. iq4_xs_r4 is at 275 t/s.
This commit is contained in:
Iwan Kawrakow
2025-06-18 12:25:32 +03:00
parent c6048c478a
commit 222833d373
2 changed files with 94 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

92
ggml/src/iqk/iqk_gemm_kquants.cpp
View File

@@ -2430,6 +2430,97 @@ void iqk_convert_q3_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
}
}
inline float convert_to_q8_k_r8(int k, float d0, const __m256i * qx, const int16_t * scales, uint32_t * block, int8_t * q8_k) {
auto max_i16 = _mm256_setzero_si256();
__m256i qs[16];
for (int ib32 = 0; ib32 < 8; ++ib32) {
qs[2*ib32+0] = _mm256_cvtepi8_epi16(_mm256_castsi256_si128(qx[ib32]));
qs[2*ib32+1] = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(qx[ib32], 1));
qs[2*ib32+0] = _mm256_mullo_epi16(qs[2*ib32+0], _mm256_set1_epi16(scales[2*ib32+0]));
qs[2*ib32+1] = _mm256_mullo_epi16(qs[2*ib32+1], _mm256_set1_epi16(scales[2*ib32+1]));
max_i16 = _mm256_max_epi16(max_i16, _mm256_sign_epi16(qs[2*ib32+0], qs[2*ib32+0]));
max_i16 = _mm256_max_epi16(max_i16, _mm256_sign_epi16(qs[2*ib32+1], qs[2*ib32+1]));
}
auto max_q32 = _mm256_cvtepi16_epi32(_mm_max_epi16(_mm256_castsi256_si128(max_i16), _mm256_extracti128_si256(max_i16, 1)));
auto imax4 = _mm_max_epi32(_mm256_castsi256_si128(max_q32), _mm256_extracti128_si256(max_q32, 1));
auto max4 = _mm_cvtepi32_ps(imax4);
max4 = _mm_max_ps(max4, _mm_movehl_ps(max4, max4));
max4 = _mm_max_ss(max4, _mm_movehdup_ps(max4));
bool needs_scaling = true;
float dnew = _mm_cvtss_f32(max4) * d0;
if (dnew < 1.f) {
dnew = 1.f; needs_scaling = false;
}
auto scale = _mm256_set1_ps(std::abs(dnew) > 1e-9f ? 1/dnew : 0.f);
for (int ib32 = 0; ib32 < 8; ++ib32) {
if (needs_scaling) {
auto i0 = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(qs[2*ib32+0]));
auto i1 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(qs[2*ib32+0], 1));
auto i2 = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(qs[2*ib32+1]));
auto i3 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(qs[2*ib32+1], 1));
i0 = _mm256_cvtps_epi32(_mm256_round_ps(_mm256_mul_ps(scale, _mm256_cvtepi32_ps(i0)), _MM_ROUND_NEAREST));
i1 = _mm256_cvtps_epi32(_mm256_round_ps(_mm256_mul_ps(scale, _mm256_cvtepi32_ps(i1)), _MM_ROUND_NEAREST));
i2 = _mm256_cvtps_epi32(_mm256_round_ps(_mm256_mul_ps(scale, _mm256_cvtepi32_ps(i2)), _MM_ROUND_NEAREST));
i3 = _mm256_cvtps_epi32(_mm256_round_ps(_mm256_mul_ps(scale, _mm256_cvtepi32_ps(i3)), _MM_ROUND_NEAREST));
i0 = _mm256_packs_epi32(i0, i1);
i2 = _mm256_packs_epi32(i2, i3);
i0 = _mm256_packs_epi16(i0, i2);
i0 = _mm256_permutevar8x32_epi32(i0, _mm256_setr_epi32(0, 4, 1, 5, 2, 6, 3, 7));
_mm256_storeu_si256((__m256i *)block, i0);
} else {
// 0, 1, 2, 3, 4, 5, 6, 7, 8, 16, 17, 18, 19, 20, 21, 22, 23, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31
auto i0 = _mm256_packs_epi16(qs[2*ib32+0], qs[2*ib32+1]);
auto i0_l = _mm256_castsi256_si128(i0);
auto i0_h = _mm256_extracti128_si256(i0, 1);
_mm_storeu_si128((__m128i *)block+0, _mm_unpacklo_epi64(i0_l, i0_h));
_mm_storeu_si128((__m128i *)block+1, _mm_unpackhi_epi64(i0_l, i0_h));
}
auto qs = (uint32_t *)q8_k + 64*ib32;
for (int l = 0; l < 8; ++l) {
qs[8*l + k] = block[l];
}
}
return dnew;
}
// TODO: move this to iqk_gemm_iquants
void iqk_convert_iq4_xs_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_iq4_xs * x8[8];
block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
auto values128 = _mm_loadu_si128((const __m128i *)iq4k_values);
auto values = MM256_SET_M128I(values128, values128);
int16_t ls[16];
float dnew[8];
uint32_t block[8];
__m256i xv[8];
for (int ix = 0; ix < nrc_x; ix += 8) {
for (int k = 0; k < 8; ++k) x8[k] = (const block_iq4_xs *)((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);
for (int ib32 = 0; ib32 < 8; ++ib32) {
ls[2*ib32+0] = ls[2*ib32+1] = (((x8[k][i].scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((x8[k][i].scales_h >> 2*ib32) & 3) << 4)) - 32;
auto bits = _mm_loadu_si128((const __m128i *)x8[k][i].qs + ib32);
xv[ib32] = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(bits, 4), bits), _mm256_set1_epi8(0xf));
xv[ib32] = _mm256_shuffle_epi8(values, xv[ib32]);
}
dnew[k] = d * convert_to_q8_k_r8(k, 1.f/127, xv, ls, block, y[i].qs);
}
_mm_storeu_si128((__m128i *)y[i].d, _mm256_cvtps_ph(_mm256_loadu_ps(dnew), _MM_ROUND_NEAREST));
}
y += nb;
}
}
} // namespace
@@ -2521,6 +2612,7 @@ bool iqk_convert_kquants_q8X_r8(int type, int n, const void * vx, size_t bx, voi
case GGML_TYPE_Q4_K: iqk_convert_q4_k_q8_1_r8(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_Q5_K: iqk_convert_q5_k_q8_1_r8(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_Q6_K: iqk_convert_q6_k_q8_0_r8(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ4_XS: iqk_convert_iq4_xs_q8_k_r8(n, vx, bx, vy, nrc_x); break;
default: return false;
}
return true;

3
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -243,6 +243,7 @@ struct MulMat {
case GGML_TYPE_IQ2_XS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ2_S : return nrc_y >= 16 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ3_XXS: return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ4_XS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ3_S : return nrc_y >= 32 ? GGML_TYPE_Q8_K_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 >= 32 ? GGML_TYPE_Q8_K_R8 : type;
@@ -368,7 +369,7 @@ bool iqk_convert_repack(int typeA, int n, const void * vx, size_t bx, void * vy,
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
//case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ4_XS:
//case GGML_TYPE_Q2_K_R4:
//case GGML_TYPE_Q3_K_R4:
//case GGML_TYPE_Q4_K_R4: