ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-10 22:10:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

mxfp4: AVX2 GEMM

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-08-08 16:27:12 +03:00
parent a5e87adfa7
commit 9238445f19
3 changed files with 32 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/ggml.c
View File

@@ -1301,14 +1301,10 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.from_float = quantize_row_iq4_nl,
.from_float_ref = (ggml_from_float_t)quantize_row_iq4_nl_ref,
.vec_dot = ggml_vec_dot_iq4_nl_q8_0,
#if GGML_USE_IQK_MULMAT
#if defined HAVE_FANCY_SIMD
.vec_dot_type = GGML_TYPE_Q8_2_X4,
#else
.vec_dot_type = GGML_TYPE_Q8_0_X4,
#endif
#else
.vec_dot_type = GGML_TYPE_Q8_0,
#endif
.nrows = 1,
.row_meta_size = 0,
@@ -1335,7 +1331,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.from_float = quantize_row_mxfp4,
.from_float_ref = (ggml_from_float_t)quantize_row_mxfp4_ref,
.vec_dot = vec_dot_mxfp4_q8_0_x4,
#if defined HAVE_FANCY_SIMD
#if defined __AVX2__
.vec_dot_type = GGML_TYPE_Q8_2_X4,
#else
.vec_dot_type = GGML_TYPE_Q8_0_X4,

46
ggml/src/iqk/iqk_gemm_legacy_quants.cpp
View File

@@ -105,6 +105,21 @@ struct ScaleHelperQ_0 {
template <typename Q> inline float prepare1(float d, const Q * y) const { return d*prepare1(y); }
};
struct ScaleHelperQ_0_MXFP4 {
float scales[4];
template <typename Q>
inline __m128 prepare4(const Q * y) {
for (int j = 0; j < 4; ++j) scales[j] = GGML_E8M0_TO_FP32_HALF(y[j].e);
return _mm_loadu_ps(scales);
}
template <typename Q>
inline __m128 prepare4(__m128 other_scales, const Q * y) {
return _mm_mul_ps(other_scales, prepare4<Q>(y));
}
template <typename Q> inline float prepare1(const Q * y) const { return GGML_E8M0_TO_FP32_HALF(y->e); }
template <typename Q> inline float prepare1(float d, const Q * y) const { return d*prepare1(y); }
};
template <int min_value>
struct ScaleHelperQ_0_1 {
ggml_half scales8[4];
@@ -582,11 +597,7 @@ static inline __m256i load_mxfp4_values_256() {
struct MXFP4_Dequantizer {
Dequantizer4bit b4;
#ifdef HAVE_FANCY_SIMD
const __m256i values = load_unsigned_mxfp4_values_256();
#else
const __m256i values = load_mxfp4_values_256();
#endif
inline __m256i dequant(const block_mxfp4 * x) const {
return _mm256_shuffle_epi8(values, b4.dequant(x->qs));
}
@@ -712,20 +723,20 @@ struct Q4_0_1_Unpacker final : public Q_Unpacker<block_q4_0, ScaleHelperQ_0_1<8>
using Sum4T = Sum4q4<block_q8_2, block_q8_2_x4>;
inline static int block_size() { return QK4_0; }
};
struct MXFP4_Unpacker final : public Q_Unpacker<block_mxfp4, ScaleHelperQ_0_1_MXFP4<12>, MXFP4_Dequantizer> {
MXFP4_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {}
using Sum4T = Sum4TypeQ82;
inline static int block_size() { return QK4_NL; }
};
#ifdef HAVE_FANCY_SIMD
struct IQ4_NL_Unpacker final : public Q_Unpacker<block_iq4_nl, ScaleHelperQ_0_1<128>, IQ4_NL_Dequantizer> {
IQ4_NL_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {}
using Sum4T = Sum4TypeQ82;
inline static int block_size() { return QK4_NL; }
};
struct IQ4_MXFP4_Unpacker final : public Q_Unpacker<block_mxfp4, ScaleHelperQ_0_1_MXFP4<12>, MXFP4_Dequantizer> {
IQ4_MXFP4_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {}
using Sum4T = Sum4TypeQ82;
inline static int block_size() { return QK4_NL; }
};
#else
struct IQ4_MXFP4_Unpacker final : public Q_Unpacker<block_mxfp4, ScaleHelperQ_0, MXFP4_Dequantizer> {
IQ4_MXFP4_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {}
struct IQ4_NL_Unpacker final : public Q_Unpacker<block_iq4_nl, ScaleHelperQ_0, IQ4_NL0_Dequantizer> {
IQ4_NL_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {}
using Sum4T = Sum4TypeQ80;
inline static int block_size() { return QK4_NL; }
};
@@ -1867,7 +1878,7 @@ template <typename Dequantizer> void set_functions(std::array<mul_mat_t, IQK_MAX
else if constexpr (std::is_same_v<Dequantizer, Q4_1_Unpacker> || std::is_same_v<Dequantizer, Q5_1_Unpacker>) {
IQK_SET_MUL_MAT_FUNCTIONS_T(mul_mat_qX_1_q8_2_T, Dequantizer, funcs)
}
else if constexpr (std::is_same_v<Dequantizer, IQ4_NL_Unpacker> || std::is_same_v<Dequantizer, IQ4_MXFP4_Unpacker>) {
else if constexpr (std::is_same_v<Dequantizer, IQ4_NL_Unpacker>) {
#ifdef HAVE_FANCY_SIMD
IQK_SET_MUL_MAT_FUNCTIONS_T(mul_mat_qX_1_q8_2_T, Dequantizer, funcs)
#else
@@ -1875,7 +1886,8 @@ template <typename Dequantizer> void set_functions(std::array<mul_mat_t, IQK_MAX
#endif
}
else if constexpr (std::is_same_v<Dequantizer, Q8_0_1_Unpacker> || std::is_same_v<Dequantizer, Q4_0_1_Unpacker> ||
std::is_same_v<Dequantizer, Q5_0_1_Unpacker> || std::is_same_v<Dequantizer, Q6_0_1_Unpacker>) {
std::is_same_v<Dequantizer, Q5_0_1_Unpacker> || std::is_same_v<Dequantizer, Q6_0_1_Unpacker> ||
std::is_same_v<Dequantizer, MXFP4_Unpacker>) {
IQK_SET_MUL_MAT_FUNCTIONS_T(mul_mat_qX_1_q8_2_T, Dequantizer, funcs)
}
}
@@ -1936,10 +1948,10 @@ bool iqk_set_kernels_legacy_quants(int ne00, int typeA, int typeB, std::array<mu
#endif
break;
case GGML_TYPE_MXFP4:
set_functions<IQ4_MXFP4_Unpacker>(kernels);
#ifndef HAVE_FANCY_SIMD
expected_typeB = GGML_TYPE_Q8_0_X4;
#endif
set_functions<MXFP4_Unpacker>(kernels);
//#ifndef HAVE_FANCY_SIMD
// expected_typeB = GGML_TYPE_Q8_0_X4;
//#endif
break;
case GGML_TYPE_Q4_0_R8:
IQK_SET_MUL_MAT_FUNCTIONS(mul_mat_q4_0_r8_q8_2, kernels)

2
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -265,7 +265,9 @@ struct MulMat {
case GGML_TYPE_Q5_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_Q5_1 : return nrc_y >= 32 ? GGML_TYPE_Q8_1 : type;
case GGML_TYPE_Q6_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
#ifdef HAVE_FANCY_SIMD
case GGML_TYPE_IQ4_NL : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
#endif
case GGML_TYPE_MXFP4 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_Q8_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ1_KT : return nrc_y >= 16 ? GGML_TYPE_Q8_0_R8 : type;