Much faster iq3_xxs GEMM via repacking to q8_0_r8 (AVX2) (#516)

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

3 files changed, 76 insertions, 1 deletions

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 9e3c4b90..bf5a8519 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1123,7 +1123,11 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float               = quantize_row_iq3_xxs,
         .from_float_ref           = (ggml_from_float_t)quantize_row_iq3_xxs_ref,
         .vec_dot                  = ggml_vec_dot_iq3_xxs_q8_K,
+#ifdef __AVX2__
+        .vec_dot_type             = GGML_TYPE_Q8_2_X4,
+#else
         .vec_dot_type             = GGML_TYPE_Q8_K,
+#endif
         .nrows                    = 1,
         .row_meta_size            = 0,
     },
diff --git a/ggml/src/iqk/iqk_gemm_iquants.cpp b/ggml/src/iqk/iqk_gemm_iquants.cpp
index 9e2ddc10..e8ed3d78 100644
--- a/ggml/src/iqk/iqk_gemm_iquants.cpp
+++ b/ggml/src/iqk/iqk_gemm_iquants.cpp
@@ -472,6 +472,15 @@ struct DequantizerIQ3XXS final : public BaseDequantizer<block_iq3_xxs> {
         auto scales16 = prepare_scales(i);
         scales[0] = MM256_SET_M128I(scales16, scales16);
     }
+    inline void new_block_f(int i, __m256 * scales) {
+        auto sc16 = prepare_scales(i);
+        auto scf  = _mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(sc16)));
+        auto scf_l = _mm256_castps256_ps128(scf);
+        auto scf_h = _mm256_extractf128_ps(scf, 1);
+        scales[0] = _mm256_set_m128(scf_l, scf_l);
+        scales[1] = _mm256_set_m128(scf_h, scf_h);
+        scales[2] = _mm256_mul_ps(scf, _mm256_set1_ps(-minv));
+    }
     inline float new_block(int i, __m256i * scales, __m256i& mins) {
         auto scales16 = prepare_scales(i);
         mins = scb.shuffle(scales16);
@@ -1771,6 +1780,58 @@ void iqk_convert_iq2_xxs_q8_0_r8(int n, const void * vx, size_t bx, void * vy, i
     }
 }
 
+void iqk_convert_iq3_xxs_q8_0_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_iq3_xxs * x8[8];
+
+    block_q8_0_r8 * y = (block_q8_0_r8 *)vy;
+
+    ggml_half dh[8];
+    uint16_t all_ls[64];
+    EvenSignHelper esh;
+
+    uint32_t block[8];
+    uint32_t aux32;
+
+    for (int ix = 0; ix < nrc_x; ix += 8) {
+        for (int k = 0; k < 8; ++k) x8[k] = (const block_iq3_xxs *)((const char *)vx + (ix + k)*bx);
+        for (int i = 0; i < nb; ++i) {
+            // TODO: simdify
+            for (int k = 0; k < 8; ++k) {
+                dh[k] = x8[k][i].d;
+                auto qs  = x8[k][i].qs;
+                auto sas = qs + QK_K/4;
+                for (int ib32 = 0; ib32 < 8; ++ib32) {
+                    std::memcpy(&aux32, sas + 4*ib32, sizeof(uint32_t));
+                    all_ls[8*ib32 + k] = (2*(aux32 >> 28) + 1);
+                    auto value = _mm256_set_epi32(iq3xxs_grid[qs[7]], iq3xxs_grid[qs[6]], iq3xxs_grid[qs[5]], iq3xxs_grid[qs[4]],
+                                                  iq3xxs_grid[qs[3]], iq3xxs_grid[qs[2]], iq3xxs_grid[qs[1]], iq3xxs_grid[qs[0]]);
+                    esh.sign_value(aux32, value);
+                    _mm256_storeu_si256((__m256i *)block, value);
+                    auto q8 = (uint32_t *)y[ib32].qs;
+                    for (int l = 0; l < 4; ++l) {
+                        q8[8*l + k +  0] = block[l + 0];
+                        q8[8*l + k + 32] = block[l + 4];
+                    }
+                    qs += 8;
+                }
+            }
+            auto vd = _mm256_mul_ps(_mm256_set1_ps(0.25f), _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)dh)));
+            for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+                auto iscales16 = _mm_loadu_si128((const __m128i *)all_ls + ib32);
+                auto iscales32 = _mm256_cvtepi16_epi32(iscales16);
+                auto scales = _mm256_mul_ps(vd, _mm256_cvtepi32_ps(iscales32));
+                _mm_storeu_si128((__m128i *)y[ib32].d, _mm256_cvtps_ph(scales, _MM_FROUND_TO_NEAREST_INT));
+            }
+            y += QK_K/32;
+        }
+    }
+}
+
 template <typename Dequantizer> void set_functions(std::array<mul_mat_t, IQK_MAX_NY>& funcs) {
     funcs[0] = mul_mat_qX_K_q8_K_IQ<Dequantizer, 1>;
     funcs[1] = mul_mat_qX_K_q8_K_IQ<Dequantizer, 2>;
@@ -1791,7 +1852,15 @@ bool iqk_set_kernels_iquants(int ne00, int typeA, int typeB, std::array<mul_mat_
     if (ggml_type(typeA) == GGML_TYPE_IQ2_XXS) {
         if (ggml_type(typeB) == GGML_TYPE_Q8_2_X4) {
             IQK_SET_MUL_MAT_FUNCTIONS_T(mul_mat_qX_K_q8_2_IQ_N, DequantizerIQ2XXS, kernels);
-            //IQK_SET_MUL_MAT_FUNCTIONS(mul_mat_iq2_xxs_q8_2_IQ_N, kernels);
+            func16 = nullptr;
+            return true;
+        }
+        return false;
+    }
+
+    if (ggml_type(typeA) == GGML_TYPE_IQ3_XXS) {
+        if (ggml_type(typeB) == GGML_TYPE_Q8_2_X4) {
+            IQK_SET_MUL_MAT_FUNCTIONS_T(mul_mat_qX_K_q8_2_IQ_N, DequantizerIQ3XXS, kernels);
             func16 = nullptr;
             return true;
         }
@@ -1856,6 +1925,7 @@ bool iqk_convert_iquants_q80_r8(int type, int n, const void * vx, size_t bx, voi
     if (n%QK_K != 0 || nrc_x%8 != 0) return false;
     switch (ggml_type(type)) {
         case GGML_TYPE_IQ2_XXS: iqk_convert_iq2_xxs_q8_0_r8(n, vx, bx, vy, nrc_x); break;
+        case GGML_TYPE_IQ3_XXS: iqk_convert_iq3_xxs_q8_0_r8(n, vx, bx, vy, nrc_x); break;
         default: return false;
     }
     return true;
diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp
index 6f8c0106..16896909 100644
--- a/ggml/src/iqk/iqk_mul_mat.cpp
+++ b/ggml/src/iqk/iqk_mul_mat.cpp
@@ -240,6 +240,7 @@ struct MulMat {
             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;
             case GGML_TYPE_IQ2_XXS: return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
+            case GGML_TYPE_IQ3_XXS: return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
             default: break;
         }
 #else