IQ2_XXS_R4 (#154)

* iq2_xxs_r4: Zen4

Disapointing gain: 134.7 t/s -> 151.1 t/s for PP-512
TG-128 is better: 3.45 -> 4.61 t/s @ 1 thread

* Minor

* iq2_xxs_r4: NEON

---------

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

6 files changed, 310 insertions, 14 deletions

diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
index b9da9771..ad21dd50 100644
--- a/ggml/src/ggml-common.h
+++ b/ggml/src/ggml-common.h
@@ -397,6 +397,13 @@ typedef struct {
 } block_iq2_xxs;
 static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
 
+typedef struct {
+    ggml_half d[4];
+    uint8_t   sas[QK_K/2];
+    uint8_t   qs[QK_K/2];
+} block_iq2_xxs_r4;
+static_assert(sizeof(block_iq2_xxs_r4) == 4*sizeof(block_iq2_xxs), "wrong iq2_xxs_r4 block size/padding");
+
 // 2.3125 bpw quants
 typedef struct {
     ggml_half d;
diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c
index 7821fa79..23c60182 100644
--- a/ggml/src/ggml-quants.c
+++ b/ggml/src/ggml-quants.c
@@ -15198,6 +15198,7 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
         case GGML_TYPE_IQ4_KSS: break;
         case GGML_TYPE_IQ4_NL_R4: break;
         case GGML_TYPE_IQ4_XS_R4: break;
+        case GGML_TYPE_IQ2_XXS_R4: break;
         case GGML_TYPE_IQ3_XXS_R4: break;
         case GGML_TYPE_Q4_0_R4: break;
         case GGML_TYPE_Q5_0_R4: break;
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index a2d3efdc..27794cd3 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1005,6 +1005,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
         .row_meta_size            = 0,
     },
+    [GGML_TYPE_IQ2_XXS_R4] = {
+        .type_name                = "iq2_xxs_r4",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq2_xxs),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq2_xxs_r4,
+        .from_float               = quantize_row_iq2_xxs_r4,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq2_xxs_r4_ref,
+        .vec_dot                  = vec_dot_iq2_xxs_r4_q8_k,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+        .row_meta_size            = 0,
+    },
     [GGML_TYPE_IQ2_XS] = {
         .type_name                = "iq2_xs",
         .blck_size                = QK_K,
@@ -4211,6 +4224,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_Q6_K_R4:       wtype = GGML_TYPE_Q6_K_R4;  break;
         case GGML_FTYPE_MOSTLY_Q8_K_R8:       wtype = GGML_TYPE_Q8_K_R8;  break;
         case GGML_FTYPE_MOSTLY_IQ2_XXS:       wtype = GGML_TYPE_IQ2_XXS;  break;
+        case GGML_FTYPE_MOSTLY_IQ2_XXS_R4:    wtype = GGML_TYPE_IQ2_XXS_R4;break;
         case GGML_FTYPE_MOSTLY_IQ2_XS:        wtype = GGML_TYPE_IQ2_XS;   break;
         case GGML_FTYPE_MOSTLY_IQ3_XXS:       wtype = GGML_TYPE_IQ3_XXS;  break;
         case GGML_FTYPE_MOSTLY_IQ3_XXS_R4:    wtype = GGML_TYPE_IQ3_XXS_R4;break;
@@ -10753,6 +10767,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_Q6_K_R4:
         case GGML_TYPE_Q8_K_R8:
         case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
@@ -11214,6 +11229,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_Q6_K_R4:
         case GGML_TYPE_Q8_K_R8:
         case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
@@ -11372,6 +11388,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_Q6_K_R4:
         case GGML_TYPE_Q8_K_R8:
         case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
@@ -14576,6 +14593,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_Q6_K_R4:
         case GGML_TYPE_Q8_K_R8:
         case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
@@ -14974,6 +14992,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_Q6_K_R4:
         case GGML_TYPE_Q8_K_R8:
         case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
@@ -15266,6 +15285,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q6_K_R4:
         case GGML_TYPE_Q8_K_R8:
         case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
@@ -15887,6 +15907,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_Q8_K_R8:
         case GGML_TYPE_Q8_KR8:
         case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
@@ -22658,6 +22679,7 @@ void ggml_quantize_init(enum ggml_type type) {
     ggml_critical_section_start();
 
     switch (type) {
+        case GGML_TYPE_IQ2_XXS_R4: iq2xs_init_impl(GGML_TYPE_IQ2_XXS); break;
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ2_S:
@@ -22735,6 +22757,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_Q6_K_R4: result = quantize_q6_k_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q8_K_R8: result = quantize_q8_k_r8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_XXS: result = quantize_iq2_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ2_XXS_R4:result = quantize_iq2_xxs_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_XS:  result = quantize_iq2_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ3_XXS: result = quantize_iq3_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ3_XXS_R4:result = quantize_iq3_xxs_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp
index b6b6da15..0733d4ea 100644
--- a/ggml/src/iqk/iqk_mul_mat.cpp
+++ b/ggml/src/iqk/iqk_mul_mat.cpp
@@ -187,6 +187,7 @@ struct MulMat {
             case GGML_TYPE_IQ4_K_R4:
             case GGML_TYPE_IQ5_K_R4:
             case GGML_TYPE_IQ4_KS_R4:
+            case GGML_TYPE_IQ2_XXS_R4:
             case GGML_TYPE_IQ3_XXS_R4:
             case GGML_TYPE_IQ2_BN_R4: return 4;
             case GGML_TYPE_Q8_K_R8: return 8;
@@ -3214,6 +3215,92 @@ static void mul_mat_iq4_ks_r4_q8_k(int n, const void * vx, size_t bx, const Data
 }
 
 template <int nrc_y>
+static void mul_mat_iq2_xxs_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
+    GGML_ASSERT(nrc_x%4 == 0);
+    Q8<nrc_y, block_q8_K> q8(info);
+    int nbl = n / QK_K;
+#ifndef HAVE_FANCY_SIMD
+    auto smask = _mm256_set1_epi64x(0x8040201008040201);
+    auto sign_shuffle = _mm256_set_epi64x(0x0303030303030303, 0x0202020202020202, 0x0101010101010101, 0x0000000000000000);
+    auto m4 = _mm256_set1_epi8(4);
+    auto m1 = _mm256_set1_epi16(1);
+#endif
+    __m256  acc[nrc_y] = {};
+    __m256i isum[nrc_y] = {};
+    __m256i qx[4];
+    for (int ix = 0; ix < nrc_x; ix += 4) {
+        auto iq2 = (const block_iq2_xxs_r4 *)((const char *)vx + (ix+0)*bx);
+        for (int ibl = 0; ibl < nbl; ++ibl) { // Block of 256
+            auto dl = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq2[ibl].d));
+            auto d4 = _mm256_set_m128(dl, dl);
+            auto qs = iq2[ibl].qs;
+            for (int ib = 0; ib < QK_K/32; ++ib) {
+                qx[0] = _mm256_set_epi64x(iq2xxs_grid[qs[ 3]], iq2xxs_grid[qs[ 2]], iq2xxs_grid[qs[ 1]], iq2xxs_grid[qs[ 0]]);
+                qx[1] = _mm256_set_epi64x(iq2xxs_grid[qs[ 7]], iq2xxs_grid[qs[ 6]], iq2xxs_grid[qs[ 5]], iq2xxs_grid[qs[ 4]]);
+                qx[2] = _mm256_set_epi64x(iq2xxs_grid[qs[11]], iq2xxs_grid[qs[10]], iq2xxs_grid[qs[ 9]], iq2xxs_grid[qs[ 8]]);
+                qx[3] = _mm256_set_epi64x(iq2xxs_grid[qs[15]], iq2xxs_grid[qs[14]], iq2xxs_grid[qs[13]], iq2xxs_grid[qs[12]]);
+                qs += 16;
+                auto sas = _mm_loadu_si128((const __m128i *)iq2[ibl].sas + ib);
+                auto scales = _mm_and_si128(sas, _mm_set1_epi8(1));
+#ifdef HAVE_FANCY_SIMD
+                scales = _mm_dpbusd_epi32(_mm_set1_epi32(1), scales, _mm_set1_epi32(0x10080402));
+#else
+                scales = _mm_maddubs_epi16(scales, _mm_set1_epi32(0x10080402));
+                scales = _mm_add_epi32(_mm_madd_epi16(_mm_set1_epi16(1), scales), _mm_set1_epi32(1));
+#endif
+                auto scales32 = MM256_SET_M128I(scales, scales);
+                auto signs128 = _mm_and_si128(sas, _mm_set1_epi8(-2)); // 0xfe = -2 as signed. Needed to shutup compiler warning.
+                signs128 = _mm_xor_si128(signs128, _mm_srli_epi16(signs128, 1));
+#ifdef HAVE_FANCY_SIMD
+                auto mask = (const __mmask32 *)&signs128;
+                for (int iy = 0; iy < nrc_y; ++iy) {
+                    auto y = _mm256_loadu_si256((const __m256i *)q8.y[iy][ibl].qs + ib);
+                    auto sumi1 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), qx[0], _mm256_mask_sub_epi8(y, mask[0], _mm256_setzero_si256(), y));
+                    auto sumi2 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), qx[1], _mm256_mask_sub_epi8(y, mask[1], _mm256_setzero_si256(), y));
+                    auto sumi3 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), qx[2], _mm256_mask_sub_epi8(y, mask[2], _mm256_setzero_si256(), y));
+                    auto sumi4 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), qx[3], _mm256_mask_sub_epi8(y, mask[3], _mm256_setzero_si256(), y));
+                    auto s12 = _mm256_add_epi32(_mm256_unpacklo_epi32(sumi1, sumi2), _mm256_unpackhi_epi32(sumi1, sumi2)); // 0,1, 0,1, 0,1, 0,1
+                    auto s34 = _mm256_add_epi32(_mm256_unpacklo_epi32(sumi3, sumi4), _mm256_unpackhi_epi32(sumi3, sumi4)); // 2,3, 2,3, 2,3, 2,3
+                    auto sumi = _mm256_add_epi32(_mm256_unpacklo_epi64(s12, s34), _mm256_unpackhi_epi64(s12, s34)); // 0,1,2,3, 0,1,2,3
+                    isum[iy] = _mm256_add_epi32(isum[iy], _mm256_mullo_epi32(scales32, sumi));
+                }
+#else
+                auto signs = MM256_SET_M128I(signs128, signs128);
+                auto shuffle = sign_shuffle;
+                auto s1 = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(signs, shuffle), smask), smask), _mm256_set1_epi8(1));
+                shuffle = _mm256_add_epi8(shuffle, m4);
+                auto s2 = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(signs, shuffle), smask), smask), _mm256_set1_epi8(1));
+                shuffle = _mm256_add_epi8(shuffle, m4);
+                auto s3 = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(signs, shuffle), smask), smask), _mm256_set1_epi8(1));
+                shuffle = _mm256_add_epi8(shuffle, m4);
+                auto s4 = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(signs, shuffle), smask), smask), _mm256_set1_epi8(1));
+                for (int iy = 0; iy < nrc_y; ++iy) {
+                    auto y = _mm256_loadu_si256((const __m256i *)q8.y[iy][ibl].qs + ib);
+                    auto sumi1 = _mm256_madd_epi16(m1, _mm256_maddubs_epi16(qx[0], _mm256_sign_epi8(y, s1)));
+                    auto sumi2 = _mm256_madd_epi16(m1, _mm256_maddubs_epi16(qx[1], _mm256_sign_epi8(y, s2)));
+                    auto sumi3 = _mm256_madd_epi16(m1, _mm256_maddubs_epi16(qx[2], _mm256_sign_epi8(y, s3)));
+                    auto sumi4 = _mm256_madd_epi16(m1, _mm256_maddubs_epi16(qx[3], _mm256_sign_epi8(y, s4)));
+                    auto s12 = _mm256_add_epi32(_mm256_unpacklo_epi32(sumi1, sumi2), _mm256_unpackhi_epi32(sumi1, sumi2)); // 0,1, 0,1, 0,1, 0,1
+                    auto s34 = _mm256_add_epi32(_mm256_unpacklo_epi32(sumi3, sumi4), _mm256_unpackhi_epi32(sumi3, sumi4)); // 2,3, 2,3, 2,3, 2,3
+                    auto sumi = _mm256_add_epi32(_mm256_unpacklo_epi64(s12, s34), _mm256_unpackhi_epi64(s12, s34)); // 0,1,2,3, 0,1,2,3
+                    isum[iy] = _mm256_add_epi32(isum[iy], _mm256_mullo_epi32(scales32, sumi));
+                }
+#endif
+            }
+            for (int iy = 0; iy < nrc_y; ++iy) {
+                acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(isum[iy]), acc[iy]);
+                isum[iy] = _mm256_setzero_si256();
+            }
+        }
+        for (int iy = 0; iy < nrc_y; ++iy) {
+            auto sum = _mm_add_ps(_mm256_castps256_ps128(acc[iy]), _mm256_extractf128_ps(acc[iy], 1));
+            info.store(ix, iy, _mm_mul_ps(_mm_set1_ps(0.125f), sum));
+            acc[iy] = _mm256_setzero_ps();
+        }
+    }
+}
+
+template <int nrc_y>
 static void mul_mat_iq3_xxs_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
     GGML_ASSERT(nrc_x%4 == 0);
     Q8<nrc_y, block_q8_K> q8(info);
@@ -6702,6 +6789,18 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) {
             mm.funcs[7] = mul_mat_iq4_ks_r4_q8_k<8>;
             expected_typeB = GGML_TYPE_Q8_K32;
             break;
+        case GGML_TYPE_IQ2_XXS_R4:
+            assert (ne00 % QK_K == 0);
+            mm.funcs[0] = mul_mat_iq2_xxs_r4_q8_k<1>;
+            mm.funcs[1] = mul_mat_iq2_xxs_r4_q8_k<2>;
+            mm.funcs[2] = mul_mat_iq2_xxs_r4_q8_k<3>;
+            mm.funcs[3] = mul_mat_iq2_xxs_r4_q8_k<4>;
+            mm.funcs[4] = mul_mat_iq2_xxs_r4_q8_k<5>;
+            mm.funcs[5] = mul_mat_iq2_xxs_r4_q8_k<6>;
+            mm.funcs[6] = mul_mat_iq2_xxs_r4_q8_k<7>;
+            mm.funcs[7] = mul_mat_iq2_xxs_r4_q8_k<8>;
+            expected_typeB = GGML_TYPE_Q8_K;
+            break;
         case GGML_TYPE_IQ3_XXS_R4:
             assert (ne00 % QK_K == 0);
             mm.funcs[0] = mul_mat_iq3_xxs_r4_q8_k<1>;
@@ -9586,6 +9685,56 @@ void mul_mat_iq4_ks_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& i
 }
 
 template <int nrc_y>
+static void mul_mat_iq2_xxs_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
+    GGML_ASSERT(nrc_x%4 == 0);
+    Q8<nrc_y, block_q8_K> q8(info);
+    int nbl = n / QK_K;
+    float32x4_t acc[nrc_y] = {};
+    int32x4_t   isum[nrc_y] = {};
+    int8x16_t   qx[8];
+    SignHelper  sh;
+    for (int ix = 0; ix < nrc_x; ix += 4) {
+        auto iq2 = (const block_iq2_xxs_r4 *)((const char *)vx + (ix+0)*bx);
+        for (int ibl = 0; ibl < nbl; ++ibl) { // Block of 256
+            auto d4 = vcvt_f32_f16(vld1_f16((const float16_t *)iq2[ibl].d));
+            auto qs = iq2[ibl].qs;
+            for (int ib = 0; ib < QK_K/32; ++ib) {
+                auto sas = vld1q_u8(iq2[ibl].sas + 16*ib);
+                auto scale_bits = vandq_u8(sas, vdupq_n_u8(1));
+                auto scales = ggml_vdotq_s32(vdupq_n_s32(1), scale_bits, vreinterpretq_s8_u32(vdupq_n_u32(0x10080402)));
+                auto signs128 = vandq_u8(sas, vdupq_n_u8(254));
+                signs128 = veorq_u8(signs128, vshrq_n_u8(signs128, 1));
+                sh.init();
+                for (int i = 0; i < 8; ++i) {
+                    qx[i] = vreinterpretq_s8_u64(uint64x2_t{iq2xxs_grid[qs[2*i+0]], iq2xxs_grid[qs[2*i+1]]});
+                    sh.apply_signs_1((uint8x16_t *)qx+i, signs128);
+                }
+                for (int iy = 0; iy < nrc_y; ++iy) {
+                    auto y = vld1q_s8_x2(q8.y[iy][ibl].qs + 32*ib);
+                    auto sumi1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), qx[0], y.val[0]), qx[1], y.val[1]);
+                    auto sumi2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), qx[2], y.val[0]), qx[3], y.val[1]);
+                    auto sumi3 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), qx[4], y.val[0]), qx[5], y.val[1]);
+                    auto sumi4 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), qx[6], y.val[0]), qx[7], y.val[1]);
+                    auto sumi12 = vpaddq_s32(sumi1, sumi2);
+                    auto sumi34 = vpaddq_s32(sumi3, sumi4);
+                    auto sumi = vpaddq_s32(sumi12, sumi34);
+                    isum[iy] = vmlaq_s32(isum[iy], scales, sumi);
+                }
+                qs += 16;
+            }
+            for (int iy = 0; iy < nrc_y; ++iy) {
+                acc[iy] = vfmaq_f32(acc[iy], vmulq_f32(d4, vdupq_n_f32(q8.scale(iy, ibl))), vcvtq_f32_s32(isum[iy]));
+                isum[iy] = vdupq_n_s32(0);
+            }
+        }
+        for (int iy = 0; iy < nrc_y; ++iy) {
+            info.store(ix, iy, vmulq_f32(vdupq_n_f32(0.125f), acc[iy]));
+            acc[iy] = vdupq_n_f32(0.f);
+        }
+    }
+}
+
+template <int nrc_y>
 static void mul_mat_iq3_xxs_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
     GGML_ASSERT(nrc_x%4 == 0);
     Q8<nrc_y, block_q8_K> q8(info);
@@ -10932,6 +11081,10 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& m, int /*Ny*/) {
             SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq4_ks_r4_q8_k);
             expected_Btype = GGML_TYPE_Q8_K;
             break;
+        case GGML_TYPE_IQ2_XXS_R4:
+            SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq2_xxs_r4_q8_k);
+            expected_Btype = GGML_TYPE_Q8_K;
+            break;
         case GGML_TYPE_IQ3_XXS_R4:
             SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq3_xxs_r4_q8_k);
             expected_Btype = GGML_TYPE_Q8_K;
diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp
index 8d2439ed..4c49836e 100644
--- a/ggml/src/iqk/iqk_quantize.cpp
+++ b/ggml/src/iqk/iqk_quantize.cpp
@@ -5304,6 +5304,122 @@ struct Repack {
 }
 
 //
+// ========================================= iq2_xxs_r4
+//
+
+void quantize_row_iq2_xxs_r4_ref(const float * x, block_iq2_xxs_r4 * y, int64_t k) {
+    quantize_iq2_xxs_r4(x, (void *)y, 4, k/4, nullptr);
+}
+
+void quantize_row_iq2_xxs_r4(const float * x, void * y, int64_t k) {
+    quantize_iq2_xxs_r4(x, y, 4, k/4, nullptr);
+}
+
+namespace {
+inline uint8_t scrambled_sign(uint8_t s) {
+    static const uint8_t k_table[128] = {
+        0x00, 0x7f, 0x7e, 0x01, 0x7c, 0x03, 0x02, 0x7d, 0x78, 0x07, 0x06, 0x79, 0x04, 0x7b, 0x7a, 0x05,
+        0x70, 0x0f, 0x0e, 0x71, 0x0c, 0x73, 0x72, 0x0d, 0x08, 0x77, 0x76, 0x09, 0x74, 0x0b, 0x0a, 0x75,
+        0x60, 0x1f, 0x1e, 0x61, 0x1c, 0x63, 0x62, 0x1d, 0x18, 0x67, 0x66, 0x19, 0x64, 0x1b, 0x1a, 0x65,
+        0x10, 0x6f, 0x6e, 0x11, 0x6c, 0x13, 0x12, 0x6d, 0x68, 0x17, 0x16, 0x69, 0x14, 0x6b, 0x6a, 0x15,
+        0x40, 0x3f, 0x3e, 0x41, 0x3c, 0x43, 0x42, 0x3d, 0x38, 0x47, 0x46, 0x39, 0x44, 0x3b, 0x3a, 0x45,
+        0x30, 0x4f, 0x4e, 0x31, 0x4c, 0x33, 0x32, 0x4d, 0x48, 0x37, 0x36, 0x49, 0x34, 0x4b, 0x4a, 0x35,
+        0x20, 0x5f, 0x5e, 0x21, 0x5c, 0x23, 0x22, 0x5d, 0x58, 0x27, 0x26, 0x59, 0x24, 0x5b, 0x5a, 0x25,
+        0x50, 0x2f, 0x2e, 0x51, 0x2c, 0x53, 0x52, 0x2d, 0x28, 0x57, 0x56, 0x29, 0x54, 0x2b, 0x2a, 0x55,
+    };
+    return k_table[s];
+}
+}
+
+static void repack_iq2_xxs(int nrows, int n_per_row, const block_iq2_xxs * x, block_iq2_xxs_r4 * y) {
+    GGML_ASSERT(nrows%4 == 0);
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    int nblock = n_per_row/QK_K;
+    const block_iq2_xxs * x4[4];
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+    for (int row = 0; row < nrows; row += 4) {
+        for (int k = 0; k < 4; ++k) x4[k] = x + nblock*k;
+        for (int ibl = 0; ibl < nblock; ++ibl) {
+            auto ysas = (uint32_t *)y[ibl].sas;
+            for (int k = 0; k < 4; ++k) {
+                y[ibl].d[k] = x4[k][ibl].d;
+                for (int ib = 0; ib < QK_K/32; ++ib) {
+                    std::memcpy(aux32, x4[k][ibl].qs + 4*ib, 2*sizeof(uint32_t));
+                    for (int i = 0; i < 4; ++i) {
+                        y[ibl].qs[16*ib+4*k+i] = aux8[i];
+                    }
+                    uint8_t scale = aux32[1] >> 28;
+                    uint8_t s1 = (scrambled_sign((aux32[1] >>  0) & 127) << 1) | ((scale >> 0) & 1);
+                    uint8_t s2 = (scrambled_sign((aux32[1] >>  7) & 127) << 1) | ((scale >> 1) & 1);
+                    uint8_t s3 = (scrambled_sign((aux32[1] >> 14) & 127) << 1) | ((scale >> 2) & 1);
+                    uint8_t s4 = (scrambled_sign((aux32[1] >> 21) & 127) << 1) | ((scale >> 3) & 1);
+                    aux32[1] = uint32_t(s1) | (uint32_t(s2) << 8) | (uint32_t(s3) << 16) | (uint32_t(s4) << 24);
+                    ysas[4*ib+k] = aux32[1];
+                }
+            }
+        }
+        x += 4*nblock;
+        y += nblock;
+    }
+}
+
+size_t quantize_iq2_xxs_r4(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) {
+    GGML_ASSERT(nrows%4 == 0);
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    char * qcur = (char *)dst;
+    auto row_size = ggml_row_size(GGML_TYPE_IQ2_XXS, n_per_row);
+    std::vector<char> qtmp(4*row_size);
+    for (int row = 0; row < nrows; row += 4) {
+        quantize_iq2_xxs(src, (void *)qtmp.data(), 4, n_per_row, imatrix);
+        repack_iq2_xxs(4, n_per_row, (const block_iq2_xxs *)qtmp.data(), (block_iq2_xxs_r4 *)qcur);
+        qcur += 4*row_size;
+        src += 4*n_per_row;
+    }
+    return nrows*row_size;
+}
+
+void dequantize_row_iq2_xxs_r4(const block_iq2_xxs_r4 * x, float * y, int64_t k) {
+    auto n_per_row = k/4;
+    float * y4[4] = {y, y + n_per_row, y + 2*n_per_row, y + 3*n_per_row};
+    int nblock = n_per_row/QK_K;
+    uint32_t s32;
+    const uint8_t * s8 = (const uint8_t *)&s32;
+    for (int ibl = 0; ibl < nblock; ++ibl) {
+        const uint32_t * sas = (const uint32_t *)x[ibl].sas;
+        for (int k = 0; k < 4; ++k) {
+            const float d = 0.125f*GGML_FP16_TO_FP32(x[ibl].d[k]);
+            for (int ib = 0; ib < QK_K/32; ++ib) {
+                uint32_t aux32 = sas[4*ib+k];
+                s32 = aux32 & 0x01010101;
+                uint8_t scale = s8[0] | (s8[1] << 1) | (s8[2] << 2) | (s8[3] << 3);
+                float dl = d*(2*scale+1);
+                aux32 &= 0xfefefefe;
+                aux32 ^= (aux32 >> 1);
+                for (int i = 0; i < 4; ++i) {
+                    auto val = (const int8_t *)(iq2xxs_grid + x[ibl].qs[16*ib+4*k+i]);
+                    for (int j = 0; j < 8; ++j) y4[k][QK_K*ibl+32*ib+8*i+j] = dl * val[j] * (aux32 & (1 << j) ? -1 : 1);
+                    aux32 >>= 8;
+                }
+            }
+        }
+    }
+}
+
+void vec_dot_iq2_xxs_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) {
+#if GGML_USE_IQK_MULMAT
+    if (iqk_mul_mat(1, 1, n, GGML_TYPE_IQ2_XXS_R4, vx, 0, GGML_TYPE_Q8_K, vy, 0, s, 0, 0, 1)) {
+        return;
+    }
+#endif
+    GGML_ASSERT(n%QK4_NL == 0);
+    GGML_ASSERT(nrc == 1);
+    GGML_UNUSED(bs);
+    GGML_UNUSED(bx);
+    GGML_UNUSED(by);
+}
+
+//
 // ========================================= iq3_xxs_r4
 //
 
@@ -5321,16 +5437,6 @@ namespace {
 static void repack_iq3_xxs(int nrows, int n_per_row, const block_iq3_xxs * x, block_iq3_xxs_r4 * y) {
     GGML_ASSERT(nrows%4 == 0);
     GGML_ASSERT(n_per_row%QK_K == 0);
-    static uint8_t k_table[128] = {
-        0x00, 0x7f, 0x7e, 0x01, 0x7c, 0x03, 0x02, 0x7d, 0x78, 0x07, 0x06, 0x79, 0x04, 0x7b, 0x7a, 0x05,
-        0x70, 0x0f, 0x0e, 0x71, 0x0c, 0x73, 0x72, 0x0d, 0x08, 0x77, 0x76, 0x09, 0x74, 0x0b, 0x0a, 0x75,
-        0x60, 0x1f, 0x1e, 0x61, 0x1c, 0x63, 0x62, 0x1d, 0x18, 0x67, 0x66, 0x19, 0x64, 0x1b, 0x1a, 0x65,
-        0x10, 0x6f, 0x6e, 0x11, 0x6c, 0x13, 0x12, 0x6d, 0x68, 0x17, 0x16, 0x69, 0x14, 0x6b, 0x6a, 0x15,
-        0x40, 0x3f, 0x3e, 0x41, 0x3c, 0x43, 0x42, 0x3d, 0x38, 0x47, 0x46, 0x39, 0x44, 0x3b, 0x3a, 0x45,
-        0x30, 0x4f, 0x4e, 0x31, 0x4c, 0x33, 0x32, 0x4d, 0x48, 0x37, 0x36, 0x49, 0x34, 0x4b, 0x4a, 0x35,
-        0x20, 0x5f, 0x5e, 0x21, 0x5c, 0x23, 0x22, 0x5d, 0x58, 0x27, 0x26, 0x59, 0x24, 0x5b, 0x5a, 0x25,
-        0x50, 0x2f, 0x2e, 0x51, 0x2c, 0x53, 0x52, 0x2d, 0x28, 0x57, 0x56, 0x29, 0x54, 0x2b, 0x2a, 0x55,
-    };
     int nblock = n_per_row/QK_K;
     const block_iq3_xxs * x4[4];
     uint32_t aux32;
@@ -5347,10 +5453,10 @@ static void repack_iq3_xxs(int nrows, int n_per_row, const block_iq3_xxs * x, bl
                     }
                     std::memcpy(&aux32, xsas + 4*ib, 4);
                     uint8_t scale = aux32 >> 28;
-                    uint8_t s1 = (k_table[(aux32 >>  0) & 127] << 1) | ((scale >> 0) & 1);
-                    uint8_t s2 = (k_table[(aux32 >>  7) & 127] << 1) | ((scale >> 1) & 1);
-                    uint8_t s3 = (k_table[(aux32 >> 14) & 127] << 1) | ((scale >> 2) & 1);
-                    uint8_t s4 = (k_table[(aux32 >> 21) & 127] << 1) | ((scale >> 3) & 1);
+                    uint8_t s1 = (scrambled_sign((aux32 >>  0) & 127) << 1) | ((scale >> 0) & 1);
+                    uint8_t s2 = (scrambled_sign((aux32 >>  7) & 127) << 1) | ((scale >> 1) & 1);
+                    uint8_t s3 = (scrambled_sign((aux32 >> 14) & 127) << 1) | ((scale >> 2) & 1);
+                    uint8_t s4 = (scrambled_sign((aux32 >> 21) & 127) << 1) | ((scale >> 3) & 1);
                     aux32 = uint32_t(s1) | (uint32_t(s2) << 8) | (uint32_t(s3) << 16) | (uint32_t(s4) << 24);
                     ysas[4*ib+k] = aux32;
                 }
diff --git a/ggml/src/iqk/iqk_quantize.h b/ggml/src/iqk/iqk_quantize.h
index af15f1dd..18fc0773 100644
--- a/ggml/src/iqk/iqk_quantize.h
+++ b/ggml/src/iqk/iqk_quantize.h
@@ -169,6 +169,12 @@ size_t quantize_iq4_ks_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT
 void   dequantize_row_iq4_ks_r4(const block_iq4_ks_r4  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 void   vec_dot_iq4_ks_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
+void   quantize_row_iq2_xxs_r4_ref(const float * GGML_RESTRICT x, block_iq2_xxs_r4  * GGML_RESTRICT y, int64_t k);
+void   quantize_row_iq2_xxs_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+size_t quantize_iq2_xxs_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+void   dequantize_row_iq2_xxs_r4(const block_iq2_xxs_r4  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void   vec_dot_iq2_xxs_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
 void   quantize_row_iq3_xxs_r4_ref(const float * GGML_RESTRICT x, block_iq3_xxs_r4  * GGML_RESTRICT y, int64_t k);
 void   quantize_row_iq3_xxs_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 size_t quantize_iq3_xxs_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);