Trellis quants: faster CPU prompt processing (#482)

* Experimenting with dequant + f32 GEMM

For iq4_kt this results in a massive PP improvement
from PP512 = ~42 t/s to PP512 = 128 t/s.

* Experimenting with dequant + f32 GEMM

iq2_kt: from PP512 = 57.3 t/s to PP512 = 135.0 t/s
iq3_kt: from PP512 = 43.8 t/s to PP512 = 131.4 t/s

* Experimenting with dequant + f16 GEMM on NEON

iq2_kt: PP512 = 79 t/s from 42 t/s
iq3_kt: PP512 = 81 t/s from 35 t/s

Also, found the reason why the f16 implementation for iq4_kt was
not working: it overflows. It works after mltiplying with the row scale
before doing the multiply-adds.

* Experimenting with dequant + f16 GEMM on NEON

iq4_kt: PP512 = 86 t/s from 29 t/s

* Minor

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

5 files changed, 410 insertions, 23 deletions

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 871a0968..5bb75d32 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1618,8 +1618,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_ref           = (ggml_from_float_t)quantize_row_iq4_kt_ref,
         .vec_dot                  = vec_dot_iq4_kt_q8_k,
 #ifdef __ARM_NEON
-        //.vec_dot_type             = GGML_TYPE_F16,
-        .vec_dot_type             = GGML_TYPE_F32,
+        .vec_dot_type             = GGML_TYPE_F16,
 #else
         .vec_dot_type             = GGML_TYPE_F32,
 #endif
diff --git a/ggml/src/iqk/iqk_gemm_ktquants.cpp b/ggml/src/iqk/iqk_gemm_ktquants.cpp
index 38e76e1e..bc7bcf8b 100644
--- a/ggml/src/iqk/iqk_gemm_ktquants.cpp
+++ b/ggml/src/iqk/iqk_gemm_ktquants.cpp
@@ -13,7 +13,7 @@
 
 namespace {
 
-static inline uint32_t trellis_next(uint32_t& val) {
+inline uint32_t trellis_next(uint32_t& val) {
     constexpr uint32_t ka = 89226354;
     constexpr uint32_t kb = 64248484;
     constexpr uint32_t kmask = 0x8fff8fff;
@@ -22,7 +22,7 @@ static inline uint32_t trellis_next(uint32_t& val) {
     return (val & kmask) ^ km32;
 }
 
-static inline float trellis_gen(uint32_t& val, uint32_t* s) {
+inline float trellis_gen(uint32_t& val, uint32_t* s) {
     const ggml_fp16_t * h = (const ggml_fp16_t *)s;
     s[0] = trellis_next(val);
     return GGML_FP16_TO_FP32(h[0]) + GGML_FP16_TO_FP32(h[1]);
@@ -59,7 +59,7 @@ struct Trellis1 {
     }
 };
 
-static inline __m256 trellis_gen8(__m256i i8) {
+inline __m256 trellis_gen8(__m256i i8) {
     // split upper and lower bits of each 32-bit lane into two 8xfloat16 `hlo`, `hhi`
     __m256i low_16_bits_mask = _mm256_set1_epi32(0x0000FFFF);
     __m256i lower_halves_lanes32 = _mm256_and_si256(i8, low_16_bits_mask);
@@ -97,8 +97,47 @@ struct Trellis2 {
     }
 };
 
+void iqk_dequantize_iq2_kt(int n, const void * vx, size_t bx, float * y, size_t stride_y, int nrc_x) {
+    GGML_ASSERT(n%QK_K == 0);
+    const int nb = n/QK_K;
+
+    Trellis1 trellis;
+
+    auto shifts = _mm_set_epi32(0, 0, 4, 0);
+    auto values = _mm_loadu_si128((const __m128i *)iq4k_values);
+
+    union { __m256 vec; float val[8]; } s_helper;
+
+    for (int ix = 0; ix < nrc_x; ++ix) {
+        const float * dptr = (const float *)((const char*)vx + ix*bx);
+        auto d = _mm256_set1_ps(*dptr * 31.75f * 1.05f);
+        const block_iq2_kt * x = (const block_iq2_kt *)(dptr + 1);
+
+        for (int i = 0; i < nb; ++i) {
+            const uint16_t * ql = (const uint16_t *)x[i].ql;
+            auto s8 = _mm_set1_epi32(*(const uint32_t *)x[i].scales);
+            s8 = _mm_and_si128(_mm_srlv_epi32(s8, shifts), _mm_set1_epi8(0xf));
+            s8 = _mm_shuffle_epi8(values, s8);
+            auto s32 = _mm256_cvtepi8_epi32(s8);
+            s_helper.vec = _mm256_mul_ps(d, _mm256_cvtepi32_ps(s32));
+            for (int ib = 0; ib < QK_K/64; ++ib) {
+                auto scale1 = _mm256_set1_ps(s_helper.val[2*ib+0]);
+                auto scale2 = _mm256_set1_ps(s_helper.val[2*ib+1]);
+                for (int j = 0; j < 4; ++j) {
+                    auto xval1 = _mm256_mul_ps(scale1, trellis_gen8(trellis.next8(ql[8*ib+j+0]+4096)));
+                    auto xval2 = _mm256_mul_ps(scale2, trellis_gen8(trellis.next8(ql[8*ib+j+4]+4096)));
+                    _mm256_storeu_ps(y + i*QK_K + 64*ib + 8*j +  0, xval1);
+                    _mm256_storeu_ps(y + i*QK_K + 64*ib + 8*j + 32, xval2);
+                }
+            }
+        }
+
+        y += stride_y;
+    }
+}
+
 template <int nrc_y>
-static void mul_mat_iq2_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
+void mul_mat_iq2_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
     assert(n%QK_K == 0);
     const int nb = n/QK_K;
 
@@ -159,14 +198,63 @@ static void mul_mat_iq2_kt_F32_T(int n, const void * vx, size_t bx, const DataIn
     }
 }
 
-static inline __m256 abs_ps(__m256 vals) {
+inline __m256 abs_ps(__m256 vals) {
     // Clear sign-bit of all the 32-bit floats in vals
     __m256 sign_bit = _mm256_set1_ps(-0.0f);
     return _mm256_andnot_ps(sign_bit, vals);
 }
 
+void iqk_dequantize_iq3_kt(int n, const void * vx, size_t bx, float * y, size_t stride_y, int nrc_x) {
+    GGML_ASSERT(n%QK_K == 0);
+    const int nb = n/QK_K;
+
+    Trellis1 trellis;
+
+    union { __m256 vec; float val[8]; } s_helper;
+
+    auto shifts = _mm_set_epi32(0, 0, 4, 0);
+
+    __m256i all_signs[4];
+    auto mask1 = _mm256_set1_epi32(0x01);
+    auto mask2 = _mm256_set1_epi32(0x10);
+
+    for (int ix = 0; ix < nrc_x; ++ix) {
+        const float * dptr = (const float *)((const char*)vx + ix*bx);
+        auto d  = _mm256_set1_ps(*dptr * 31.75f * 1.015f);
+        const block_iq3_kt * x = (const block_iq3_kt *)(dptr + 1);
+
+        for (int i = 0; i < nb; ++i) {
+            const uint16_t * ql = (const uint16_t *)x[i].ql;
+            const uint8_t * qh = x[i].qh;
+            auto s8 = _mm_set1_epi32(*(const uint32_t *)x[i].scales);
+            s8 = _mm_and_si128(_mm_srlv_epi32(s8, shifts), _mm_set1_epi8(0xf));
+            auto s32 = _mm256_cvtepi8_epi32(s8);
+            s_helper.vec = _mm256_mul_ps(d, _mm256_cvtepi32_ps(s32));
+            for (int j = 0; j < 4; ++j) all_signs[j] = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i *)(qh + 8*j)));
+            for (int ib = 0; ib < 4; ++ib) {
+                auto scale1 = _mm256_set1_ps(s_helper.val[ib+0]);
+                auto scale2 = _mm256_set1_ps(s_helper.val[ib+4]);
+                for (int j = 0; j < 4; ++j) {
+                    uint32_t val1 = ql[4*ib+j   ] + 4096;
+                    uint32_t val2 = ql[4*ib+j+16] + 4096;
+                    auto sign1 = _mm256_and_si256(_mm256_cmpeq_epi32(_mm256_and_si256(all_signs[j], mask1), mask1), _mm256_set1_epi32(0x80000000));
+                    auto sign2 = _mm256_and_si256(_mm256_cmpeq_epi32(_mm256_and_si256(all_signs[j], mask2), mask2), _mm256_set1_epi32(0x80000000));
+                    all_signs[j] = _mm256_srli_epi32(all_signs[j], 1);
+                    auto x_val1 = abs_ps(trellis_gen8(trellis.next8(val1)));
+                    auto x_val2 = abs_ps(trellis_gen8(trellis.next8(val2)));
+                    x_val1 = _mm256_mul_ps(scale1, _mm256_xor_ps(x_val1, _mm256_castsi256_ps(sign1)));
+                    x_val2 = _mm256_mul_ps(scale2, _mm256_xor_ps(x_val2, _mm256_castsi256_ps(sign2)));
+                    _mm256_storeu_ps(y + i*QK_K+32*ib+8*j    , x_val1);
+                    _mm256_storeu_ps(y + i*QK_K+32*ib+8*j+128, x_val2);
+                }
+            }
+        }
+        y += stride_y;
+    }
+}
+
 template <int nrc_y>
-static void mul_mat_iq3_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
+void mul_mat_iq3_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
     assert(n%QK_K == 0);
     const int nb = n/QK_K;
 
@@ -227,8 +315,57 @@ static void mul_mat_iq3_kt_F32_T(int n, const void * vx, size_t bx, const DataIn
     }
 }
 
+void iqk_dequantize_iq4_kt(int n, const void * vx, size_t bx, float * y, size_t stride_y, int nrc_x) {
+    GGML_ASSERT(n%QK_K == 0);
+    const int nb = n/QK_K;
+    constexpr int kNumGroups = 64;
+
+    Trellis2 trellis;
+
+    union { __m256  vec; float    val[8]; } s_helper;
+    union { __m256i vec; uint32_t val[8]; } o_helper;
+
+    for (int ix = 0; ix < nrc_x; ++ix) {
+        const float * dptr = (const float *)((const char*)vx + ix*bx);
+        auto d = _mm256_set1_ps(dptr[0] * 31.75f * 1.01f);
+        auto dav = _mm256_set1_ps(dptr[1]);
+        const block_iq4_kt * x = (const block_iq4_kt *)(dptr + 2);
+
+        for (int i = 0; i < nb; ++i) {
+            auto vshb = _mm256_loadu_si256((const __m256i *)x[i].qs);
+            const uint32_t * shb = x[i].qs;
+            const uint8_t * ql = (const uint8_t *)(shb + 8);
+            const uint8_t * qh = ql + kNumGroups;
+            auto iscales = _mm256_srli_epi32(_mm256_and_si256(vshb, _mm256_set1_epi32(0xff)), 1);
+            s_helper.vec = _mm256_mul_ps(d, _mm256_cvtepi32_ps(_mm256_sub_epi32(iscales, _mm256_set1_epi32(64))));
+            o_helper.vec = _mm256_add_epi32(_mm256_slli_epi32(_mm256_and_si256(vshb, _mm256_set1_epi32(1)), 15), _mm256_set1_epi32(4096));
+            for (int ib = 0; ib < 4; ++ib) {
+                auto scale1 = _mm256_set1_ps(s_helper.val[ib+0]);
+                auto scale2 = _mm256_set1_ps(s_helper.val[ib+4]);
+                for (int j = 0; j < 4; ++j) {
+                    const uint32_t sh1 = shb[ib+0] >> (8 + 6*j);
+                    const uint32_t sh2 = shb[ib+4] >> (8 + 6*j);
+                    uint32_t val1 = ql[8*ib+2*j+ 0] + ((qh[8*ib+2*j+0] << 8) & 0xf00) + ((sh1 & 7) << 12) + o_helper.val[ib+0];
+                    uint32_t val2 = ql[8*ib+2*j+32] + ((qh[8*ib+2*j+0] << 4) & 0xf00) + ((sh2 & 7) << 12) + o_helper.val[ib+4];
+                    uint32_t val3 = ql[8*ib+2*j+ 1] + ((qh[8*ib+2*j+1] << 8) & 0xf00) + ((sh1 & 56) << 9) + o_helper.val[ib+0];
+                    uint32_t val4 = ql[8*ib+2*j+33] + ((qh[8*ib+2*j+1] << 4) & 0xf00) + ((sh2 & 56) << 9) + o_helper.val[ib+4];
+                    auto x_val1 = _mm256_fmadd_ps(scale1, trellis_gen8(trellis.next8(val1, val3)), dav);
+                    auto x_val2 = _mm256_fmadd_ps(scale2, trellis_gen8(trellis.next8(val2, val4)), dav);
+
+                    _mm256_storeu_ps(y + i*QK_K + 32*ib + 8*j,          x_val1);
+                    _mm256_storeu_ps(y + i*QK_K + 32*ib + 8*j + QK_K/2, x_val2);
+
+                }
+            }
+        }
+
+        y += stride_y;
+
+    }
+}
+
 template <int nrc_y>
-static void mul_mat_iq4_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
+void mul_mat_iq4_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
     assert(n%QK_K == 0);
     const int nb = n/QK_K;
     constexpr int kNumGroups = 64;
@@ -333,6 +470,16 @@ bool iqk_set_kernels_ktquants(int ne00, int typeA, int typeB, std::array<mul_mat
 
 }
 
+bool iqk_dequantize_ktquants(int type, int n, const void * vx, size_t bx, void * y, size_t stride_y, int nrc_x) {
+    switch (type) {
+        case GGML_TYPE_IQ2_KT: iqk_dequantize_iq2_kt(n, vx, bx, (float *)y, stride_y, nrc_x); break;
+        case GGML_TYPE_IQ3_KT: iqk_dequantize_iq3_kt(n, vx, bx, (float *)y, stride_y, nrc_x); break;
+        case GGML_TYPE_IQ4_KT: iqk_dequantize_iq4_kt(n, vx, bx, (float *)y, stride_y, nrc_x); break;
+        default: return false;
+    }
+    return true;
+}
+
 #else // !__x86_64__
 
 namespace {
@@ -403,8 +550,52 @@ struct Trellis1 {
     }
 };
 
+void iqk_dequantize_iq2_kt(int n, const void * vx, size_t bx, float16_t * y, size_t stride_y, int nrc_x) {
+    GGML_ASSERT(n%QK_K == 0);
+    const int nb = n/QK_K;
+
+    Trellis1 trellis;
+
+    auto values = vld1q_s8(iq4k_values);
+
+    union { float16x8_t vec; float16_t val[8]; } s_helper;
+
+    for (int ix = 0; ix < nrc_x; ++ix) {
+        const float * dptr = (const float *)((const char*)vx + ix*bx);
+        const float d = *dptr * 31.75f * 1.05f;
+        auto vd = vdupq_n_f32(d);
+        const block_iq2_kt * x = (const block_iq2_kt *)(dptr + 1);
+
+        for (int i = 0; i < nb; ++i) {
+            const uint16_t * ql = (const uint16_t *)x[i].ql;
+            auto u32 = *(const uint32_t *)x[i].scales;
+            auto s8_u32 = uint32x2_t{u32, u32 >> 4};
+            s8_u32 = vand_u8(s8_u32, vdup_n_u32(0x0f0f0f0f));
+            auto s8 = vqtbl1_s8(values, vreinterpret_u8_u32(s8_u32));
+            auto s16 = vmovl_s8(s8);
+            auto s32l = vmovl_s16(vget_low_s16 (s16));
+            auto s32h = vmovl_s16(vget_high_s16(s16));
+            auto f32l = vmulq_f32(vd, vcvtq_f32_s32(s32l));
+            auto f32h = vmulq_f32(vd, vcvtq_f32_s32(s32h));
+            s_helper.vec = vcombine_f16(vcvt_f16_f32(f32l), vcvt_f16_f32(f32h));
+            for (int ib = 0; ib < QK_K/64; ++ib) {
+                auto scale1 = vdupq_n_f16(s_helper.val[2*ib+0]);
+                auto scale2 = vdupq_n_f16(s_helper.val[2*ib+1]);
+                for (int j = 0; j < 4; ++j) {
+                    auto xval1 = vmulq_f16(scale1, trellis.gen8(ql[8*ib+j+0]+4096));
+                    auto xval2 = vmulq_f16(scale2, trellis.gen8(ql[8*ib+j+4]+4096));
+                    vst1q_f16(y + i*QK_K + 64*ib + 8*j +  0, xval1);
+                    vst1q_f16(y + i*QK_K + 64*ib + 8*j + 32, xval2);
+                }
+            }
+        }
+
+        y += stride_y;
+    }
+}
+
 template <int nrc_y>
-static void mul_mat_iq2_kt_F16_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
+void mul_mat_iq2_kt_F16_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
     assert(n%QK_K == 0);
     const int nb = n/QK_K;
 
@@ -466,8 +657,61 @@ static void mul_mat_iq2_kt_F16_T(int n, const void * vx, size_t bx, const DataIn
     }
 }
 
+void iqk_dequantize_iq3_kt(int n, const void * vx, size_t bx, float16_t * y, size_t stride_y, int nrc_x) {
+    GGML_ASSERT(n%QK_K == 0);
+    const int nb = n/QK_K;
+
+    Trellis1 trellis;
+
+    union { float16x8_t vec; float16_t val[8]; } s_helper;
+
+    uint16x8_t all_signs[4];
+    auto mask1 = vdupq_n_u16(0x01);
+    auto mask2 = vdupq_n_u16(0x10);
+
+    for (int ix = 0; ix < nrc_x; ++ix) {
+        const float * dptr = (const float *)((const char*)vx + ix*bx);
+        const float d = *dptr * 31.75f * 1.015f;
+        auto vd = vdupq_n_f32(d);
+        const block_iq3_kt * x = (const block_iq3_kt *)(dptr + 1);
+
+        for (int i = 0; i < nb; ++i) {
+            const uint16_t * ql = (const uint16_t *)x[i].ql;
+            const uint8_t * qh = x[i].qh;
+            auto u32 = *(const uint32_t *)x[i].scales;
+            auto s8_u32 = uint32x2_t{u32, u32 >> 4};
+            s8_u32 = vand_u8(s8_u32, vdup_n_u32(0x0f0f0f0f));
+            auto s16 = vmovl_s8(vreinterpret_s8_u32(s8_u32));
+            auto s32l = vmovl_s16(vget_low_s16 (s16));
+            auto s32h = vmovl_s16(vget_high_s16(s16));
+            auto f32l = vmulq_f32(vd, vcvtq_f32_s32(s32l));
+            auto f32h = vmulq_f32(vd, vcvtq_f32_s32(s32h));
+            s_helper.vec = vcombine_f16(vcvt_f16_f32(f32l), vcvt_f16_f32(f32h));
+            for (int j = 0; j < 4; ++j) all_signs[j] = vmovl_u8(vld1_u8(qh + 8*j));
+            for (int ib = 0; ib < 4; ++ib) {
+                auto scale1 = vdupq_n_f16(s_helper.val[ib+0]);
+                auto scale2 = vdupq_n_f16(s_helper.val[ib+4]);
+                for (int j = 0; j < 4; ++j) {
+                    uint32_t val1 = ql[4*ib+j   ] + 4096;
+                    uint32_t val2 = ql[4*ib+j+16] + 4096;
+                    auto sign1 = vshlq_n_u16(vandq_u16(all_signs[j], mask1), 15);
+                    auto sign2 = vshlq_n_u16(vandq_u16(all_signs[j], mask2), 11);
+                    all_signs[j] = vshrq_n_u16(all_signs[j], 1);
+                    auto x_val1 = vabsq_f16(trellis.gen8(val1));
+                    auto x_val2 = vabsq_f16(trellis.gen8(val2));
+                    x_val1 = vmulq_f16(scale1, vreinterpretq_f16_u16(vorrq_u16(vreinterpretq_u16_f16(x_val1), sign1)));
+                    x_val2 = vmulq_f16(scale2, vreinterpretq_f16_u16(vorrq_u16(vreinterpretq_u16_f16(x_val2), sign2)));
+                    vst1q_f16(y + i*QK_K+32*ib+8*j    , x_val1);
+                    vst1q_f16(y + i*QK_K+32*ib+8*j+128, x_val2);
+                }
+            }
+        }
+        y += stride_y;
+    }
+}
+
 template <int nrc_y>
-static void mul_mat_iq3_kt_F16_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
+void mul_mat_iq3_kt_F16_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
     assert(n%QK_K == 0);
     const int nb = n/QK_K;
 
@@ -527,8 +771,63 @@ static void mul_mat_iq3_kt_F16_T(int n, const void * vx, size_t bx, const DataIn
     }
 }
 
+void iqk_dequantize_iq4_kt(int n, const void * vx, size_t bx, float16_t * y, size_t stride_y, int nrc_x) {
+    GGML_ASSERT(n%QK_K == 0);
+    const int nb = n/QK_K;
+    constexpr int kNumGroups = 64;
+
+    Trellis1 trellis;
+
+    union { float16x8_t vec; float16_t val[8]; } s_helper;
+    union { uint16x8_t  vec; uint16_t  val[8]; } o_helper;
+
+    for (int ix = 0; ix < nrc_x; ++ix) {
+        const float * dptr = (const float *)((const char*)vx + ix*bx);
+        auto d = dptr[0] * 31.75f * 1.01f;
+        //auto dav = dptr[1];
+        // Something goes wrong when we add the average. Why?
+        //auto vav = std::abs(dav) > 0.00006103515625f ? vdupq_n_f16(GGML_FP32_TO_FP16(dav)) : vdupq_n_f16(0);
+        auto vd = vdupq_n_f32(d);
+        const block_iq4_kt * x = (const block_iq4_kt *)(dptr + 2);
+
+        for (int i = 0; i < nb; ++i) {
+            const uint32_t * shb = x[i].qs;
+            auto vshb = vld1q_u32_x2(shb);
+            auto vshb16 = vcombine_u16(vmovn_u32(vandq_u32(vshb.val[0], vdupq_n_u32(0xff))), vmovn_u32(vandq_u32(vshb.val[1], vdupq_n_u32(0xff))));
+            const uint8_t * ql = (const uint8_t *)(shb + 8);
+            const uint8_t * qh = ql + kNumGroups;
+            auto iscales = vsubq_s16(vreinterpretq_s16_u16(vshrq_n_u16(vshb16, 1)), vdupq_n_s16(64));
+            auto s32l = vmovl_s16(vget_low_s16(iscales));
+            auto s32h = vmovl_s16(vget_high_s16(iscales));
+            auto f32l = vmulq_f32(vd, vcvtq_f32_s32(s32l));
+            auto f32h = vmulq_f32(vd, vcvtq_f32_s32(s32h));
+            s_helper.vec = vcombine_f16(vcvt_f16_f32(f32l), vcvt_f16_f32(f32h));
+            o_helper.vec = vaddq_u16(vshlq_n_u16(vandq_u16(vshb16, vdupq_n_u16(1)), 15), vdupq_n_u16(4096));
+            for (int ib = 0; ib < 4; ++ib) {
+                auto scale1 = vdupq_n_f16(s_helper.val[ib+0]);
+                auto scale2 = vdupq_n_f16(s_helper.val[ib+4]);
+                for (int j = 0; j < 4; ++j) {
+                    const uint32_t sh1 = shb[ib+0] >> (8 + 6*j);
+                    const uint32_t sh2 = shb[ib+4] >> (8 + 6*j);
+                    uint32_t val1 = ql[8*ib+2*j+ 0] + ((qh[8*ib+2*j+0] << 8) & 0xf00) + ((sh1 & 7) << 12) + o_helper.val[ib+0];
+                    uint32_t val2 = ql[8*ib+2*j+32] + ((qh[8*ib+2*j+0] << 4) & 0xf00) + ((sh2 & 7) << 12) + o_helper.val[ib+4];
+                    uint32_t val3 = ql[8*ib+2*j+ 1] + ((qh[8*ib+2*j+1] << 8) & 0xf00) + ((sh1 & 56) << 9) + o_helper.val[ib+0];
+                    uint32_t val4 = ql[8*ib+2*j+33] + ((qh[8*ib+2*j+1] << 4) & 0xf00) + ((sh2 & 56) << 9) + o_helper.val[ib+4];
+                    //auto x_val1 = vfmaq_f16(vav, scale1, trellis.gen8(val1, val3));
+                    //auto x_val2 = vfmaq_f16(vav, scale2, trellis.gen8(val2, val4));
+                    auto x_val1 = vmulq_f16(scale1, trellis.gen8(val1, val3));
+                    auto x_val2 = vmulq_f16(scale2, trellis.gen8(val2, val4));
+                    vst1q_f16(y + i*QK_K+32*ib+8*j+  0, x_val1);
+                    vst1q_f16(y + i*QK_K+32*ib+8*j+128, x_val2);
+                }
+            }
+        }
+        y += stride_y;
+    }
+}
+
 template <int nrc_y>
-static void mul_mat_iq4_kt_F16_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
+void mul_mat_iq4_kt_F16_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
     assert(n%QK_K == 0);
     const int nb = n/QK_K;
     constexpr int kNumGroups = 64;
@@ -548,8 +847,6 @@ static void mul_mat_iq4_kt_F16_T(int n, const void * vx, size_t bx, const DataIn
         auto sum = vdupq_n_f16(0);
         for (int i = 0; i < n/8; ++i) sum = vaddq_f16(sum, vld1q_f16(y[iy] + 8*i));
         auto sum32 = vaddq_f32(vcvt_f32_f16(vget_low_f16(sum)), vcvt_f32_f16(vget_high_f16(sum)));
-        //auto sum32 = vdupq_n_f32(0);
-        //for (int i = 0; i < n/4; ++i) sum32 = vaddq_f32(sum32, vcvt_f32_f16(vld1_f16(y[iy] + 4*i)));
         row_sum[iy] = vaddvq_f32(sum32);
     }
 
@@ -557,6 +854,7 @@ static void mul_mat_iq4_kt_F16_T(int n, const void * vx, size_t bx, const DataIn
         const float * dptr = (const float *)((const char*)vx + ix*bx);
         auto d = dptr[0] * 31.75f * 1.01f;
         auto dav = dptr[1];
+        auto vd = vdupq_n_f32(d);
         const block_iq4_kt * x = (const block_iq4_kt *)(dptr + 2);
 
         for (int iy = 0; iy < k_acc; ++iy) accd[iy] = vdupq_n_f16(0);
@@ -568,7 +866,12 @@ static void mul_mat_iq4_kt_F16_T(int n, const void * vx, size_t bx, const DataIn
             const uint8_t * ql = (const uint8_t *)(shb + 8);
             const uint8_t * qh = ql + kNumGroups;
             auto iscales = vsubq_s16(vreinterpretq_s16_u16(vshrq_n_u16(vshb16, 1)), vdupq_n_s16(64));
-            s_helper.vec = vcvtq_f16_s16(iscales);
+            auto s32l = vmovl_s16(vget_low_s16(iscales));
+            auto s32h = vmovl_s16(vget_high_s16(iscales));
+            auto f32l = vmulq_f32(vd, vcvtq_f32_s32(s32l));
+            auto f32h = vmulq_f32(vd, vcvtq_f32_s32(s32h));
+            s_helper.vec = vcombine_f16(vcvt_f16_f32(f32l), vcvt_f16_f32(f32h));
+            //s_helper.vec = vcvtq_f16_s16(iscales);
             o_helper.vec = vaddq_u16(vshlq_n_u16(vandq_u16(vshb16, vdupq_n_u16(1)), 15), vdupq_n_u16(4096));
             for (int ib = 0; ib < 4; ++ib) {
                 auto scale1 = vdupq_n_f16(s_helper.val[ib+0]);
@@ -602,18 +905,18 @@ static void mul_mat_iq4_kt_F16_T(int n, const void * vx, size_t bx, const DataIn
         if constexpr (nrc_y == 1) {
             auto sum16 = vaddq_f16(accd[0], accd[1]);
             auto sum = vaddq_f32(vcvt_f32_f16(vget_low_f16(sum16)), vcvt_f32_f16(vget_high_f16(sum16)));
-            info.store(ix, 0, d*vaddvq_f32(sum) + dav*row_sum[0]);
+            info.store(ix, 0, vaddvq_f32(sum) + dav*row_sum[0]);
         } else {
             for (int iy = 0; iy < nrc_y; ++iy) {
                 auto sum = vaddq_f32(vcvt_f32_f16(vget_low_f16(accd[iy])), vcvt_f32_f16(vget_high_f16(accd[iy])));
-                info.store(ix, iy, d*vaddvq_f32(sum) + dav*row_sum[iy]);
+                info.store(ix, iy, vaddvq_f32(sum) + dav*row_sum[iy]);
             }
         }
     }
 }
 
 template <int nrc_y>
-static void mul_mat_iq4_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
+void mul_mat_iq4_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
     assert(n%QK_K == 0);
     const int nb = n/QK_K;
     constexpr int kNumGroups = 64;
@@ -693,11 +996,11 @@ static void mul_mat_iq4_kt_F32_T(int n, const void * vx, size_t bx, const DataIn
 
 bool iqk_set_kernels_ktquants(int ne00, int typeA, int typeB, std::array<mul_mat_t, IQK_MAX_NY>& kernels, mul_mat_t& func16) {
 
-    if (ne00%QK_K == 0 && ggml_type(typeB) == GGML_TYPE_F32 && ggml_type(typeA) == GGML_TYPE_IQ4_KT) {
-        IQK_SET_MUL_MAT_FUNCTIONS(mul_mat_iq4_kt_F32_T, kernels);
-        func16 = nullptr;
-        return true;
-    }
+    //if (ne00%QK_K == 0 && ggml_type(typeB) == GGML_TYPE_F32 && ggml_type(typeA) == GGML_TYPE_IQ4_KT) {
+    //    IQK_SET_MUL_MAT_FUNCTIONS(mul_mat_iq4_kt_F32_T, kernels);
+    //    func16 = nullptr;
+    //    return true;
+    //}
 
     if (ne00%QK_K != 0 || ggml_type(typeB) != GGML_TYPE_F16) {
         return false;
@@ -722,6 +1025,17 @@ bool iqk_set_kernels_ktquants(int ne00, int typeA, int typeB, std::array<mul_mat
     return true;
 }
 
+bool iqk_dequantize_ktquants(int type, int n, const void * vx, size_t bx, void * y, size_t stride_y, int nrc_x) {
+    switch (type) {
+        case GGML_TYPE_IQ2_KT: iqk_dequantize_iq2_kt(n, vx, bx, (float16_t *)y, stride_y, nrc_x); break;
+        case GGML_TYPE_IQ3_KT: iqk_dequantize_iq3_kt(n, vx, bx, (float16_t *)y, stride_y, nrc_x); break;
+        case GGML_TYPE_IQ4_KT: iqk_dequantize_iq4_kt(n, vx, bx, (float16_t *)y, stride_y, nrc_x); break;
+        default: return false;
+    }
+
+    return true;
+}
+
 #endif
 
 #endif
diff --git a/ggml/src/iqk/iqk_gemm_ktquants.h b/ggml/src/iqk/iqk_gemm_ktquants.h
index b1e84d63..304e0235 100644
--- a/ggml/src/iqk/iqk_gemm_ktquants.h
+++ b/ggml/src/iqk/iqk_gemm_ktquants.h
@@ -8,4 +8,6 @@
 
 bool iqk_set_kernels_ktquants(int ne00, int typeA, int typeB, std::array<mul_mat_t, IQK_MAX_NY>& kernels, mul_mat_t& func16);
 
+bool iqk_dequantize_ktquants(int type, int n, const void * vx, size_t bx, void * vy, size_t stride_y, int nrc_x);
+
 #endif
diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp
index 2de6c933..ea8a60d4 100644
--- a/ggml/src/iqk/iqk_mul_mat.cpp
+++ b/ggml/src/iqk/iqk_mul_mat.cpp
@@ -233,6 +233,24 @@ struct MulMat {
         }
     }
     static bool prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny);
+    static inline ggml_type is_dequant_better(ggml_type type, int nrc_y) {
+#ifdef __AVX2__
+        switch (type) {
+            case GGML_TYPE_IQ2_KT: return nrc_y >= 32 ? GGML_TYPE_F32 : type;
+            case GGML_TYPE_IQ3_KT: return nrc_y >= 32 ? GGML_TYPE_F32 : type;
+            case GGML_TYPE_IQ4_KT: return nrc_y >= 32 ? GGML_TYPE_F32 : type;
+            default: break;
+        }
+#else
+        switch (type) {
+            case GGML_TYPE_IQ2_KT: return nrc_y >= 32 ? GGML_TYPE_F16 : type;
+            case GGML_TYPE_IQ3_KT: return nrc_y >= 32 ? GGML_TYPE_F16 : type;
+            case GGML_TYPE_IQ4_KT: return nrc_y >= 32 ? GGML_TYPE_F16 : type;
+            default: break;
+        }
+#endif
+        return type;
+    }
     static inline int num_rows(ggml_type type) {
 #ifdef HAVE_FANCY_SIMD
         switch (type) {
@@ -312,6 +330,49 @@ extern "C" IQK_API bool iqk_mul_mat(long Nx, long Ny, long ne00,
         float * C, long stride_C, int ith, int nth) {
 
     MulMat mm;
+
+    auto etypeA = ggml_type(typeA);
+    if (auto dequant_type = MulMat::is_dequant_better(etypeA, Ny); dequant_type != etypeA) {
+        if (!MulMat::prepare(dequant_type, typeB, ne00, mm, Ny)) {
+            return false;
+        }
+
+        constexpr int k_x_step = 32;
+
+        auto num_rows = MulMat::num_rows(ggml_type(dequant_type));
+        GGML_ASSERT(Nx%num_rows == 0);
+        auto nrc_x = (Nx/num_rows + nth - 1)/nth;
+        auto first_x = ith*nrc_x;
+        if (first_x + nrc_x > Nx/num_rows) nrc_x = Nx/num_rows - first_x;
+        first_x *= num_rows;
+        nrc_x   *= num_rows;
+
+        auto type_size = ggml_type_size(dequant_type);
+
+        thread_local std::vector<char> f;
+
+        size_t row_size_qx = ne00*type_size;
+        size_t row_size_qy = strideB;
+
+        //printf("Dequant mul mat %s x %s: ne00 = %d, row_size = %d\n", ggml_type_name(dequant_type), ggml_type_name(ggml_type(typeB)), (int)ne00, (int)row_size_qx);
+
+        DataInfo info{C + first_x, (const char *)B, (size_t)stride_C, row_size_qy, 0, 1, nullptr, 0};
+
+        for (int ix = 0; ix < nrc_x; ix += k_x_step) {
+            auto this_info = info;
+            this_info.s += ix;
+            int this_nrc_x = ix + k_x_step <= nrc_x ? k_x_step : nrc_x - ix;
+            if (f.size() < row_size_qx*this_nrc_x) f.resize(row_size_qx*this_nrc_x);
+            if (!iqk_dequantize_ktquants(typeA, ne00, (const char *)A + (first_x + ix)*strideA, strideA, f.data(), ne00, this_nrc_x)) {
+                GGML_ABORT("Fatal error");
+            }
+            mm.mul_mat_NxM(ne00, f.data(), row_size_qx, this_info, this_nrc_x, Ny);
+        }
+
+        return true;
+
+    }
+
     if (!MulMat::prepare(typeA, typeB, ne00, mm, Ny)) {
         return false;
     }
diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp
index c1f7a8e4..ee85344d 100644
--- a/ggml/src/iqk/iqk_quantize.cpp
+++ b/ggml/src/iqk/iqk_quantize.cpp
@@ -14,6 +14,8 @@
 #include "iqk_quantize.h"
 #include "iqk_config.h"
 
+#include "iqk_gemm_ktquants.h"
+
 #include <vector>
 #include <utility>
 #include <cstdint>
@@ -8241,6 +8243,9 @@ size_t quantize_iq2_kt(const float * src, void * dst, int64_t nrows, int64_t n_p
 
 void dequantize_row_iq2_kt(const block_iq2_kt * x, float * y, int64_t k) {
     assert(k % QuantizerIQ2KT::kSuperBlockSize == 0);
+#ifdef __AVX2__
+    if (iqk_dequantize_ktquants(GGML_TYPE_IQ2_KT, k, x, 0, y, 0, 1)) return;
+#endif
     const int nb = k / QuantizerIQ2KT::kSuperBlockSize;
     const float * dptr = (const float *)x;
     const float d = *dptr * QuantizerIQ2KT::kScale;
@@ -8494,6 +8499,9 @@ size_t quantize_iq3_kt(const float * src, void * dst, int64_t nrows, int64_t n_p
 }
 
 void dequantize_row_iq3_kt(const block_iq3_kt * x, float * y, int64_t k) {
+#ifdef __AVX2__
+    if (iqk_dequantize_ktquants(GGML_TYPE_IQ3_KT, k, x, 0, y, 0, 1)) return;
+#endif
     using Q = QuantizerIQ3KT;
     constexpr int kNumGroups = Q::kSuperBlockSize/Q::kGroupSize;
     assert(k % Q::kSuperBlockSize == 0);
@@ -8750,6 +8758,9 @@ size_t quantize_iq4_kt(const float * src, void * dst, int64_t nrows, int64_t n_p
 }
 
 void dequantize_row_iq4_kt(const block_iq4_kt * x, float * y, int64_t k) {
+#ifdef __AVX2__
+    if (iqk_dequantize_ktquants(GGML_TYPE_IQ4_KT, k, x, 0, y, 0, 1)) return;
+#endif
     using Q = QuantizerIQ4KT;
     assert(k % Q::kSuperBlockSize == 0);
     constexpr int kNumGroups = Q::kSuperBlockSize/Q::kGroupSize;