Much faster prompt processing for I-quants (ARM_NEON) (#550)

ikawrakow · Iwan Kawrakow · web-flow · commit ddda4d9e64fa · 2025-06-23T15:50:24.000+02:00
* iq2_xxs

55.8 -&gt; 167.5 t/s. iq2_xxs is at 93.7 t/s

* iq2_xs

46.4 -&gt; 166.6 t/s. iq2_xs_r4 is at 72.3 t/s.

* iq2_s

42.8 t/s -&gt; 166.8 t/s. iq2_s_r4 is at 71.1 t/s.

* iq3_xxs

51.8 t/s -&gt; 165.6 t/s. iq3_xxs_r4 is at 84.6 t/s.

* iq3_s

46.0 t/s -&gt; 162.0 t/s. iq3_s_r4 is at 79.4 t/s

---------

Co-authored-by: Iwan Kawrakow &lt;iwan.kawrakow@gmail.com&gt;
diff --git a/ggml/src/iqk/iqk_gemm_iquants.cpp b/ggml/src/iqk/iqk_gemm_iquants.cpp
@@ -3348,20 +3348,328 @@ static void mul_mat_iq3_s_r4_q8_k(int n, const void * vx, size_t bx, const DataI
     }
 }
 
+inline float convert_to_q8_k_r8(float d0, const int8x16x2_t * qx, const int8_t * scales, uint32_t * block, uint32_t * q8_k) {
+    auto max_i16 = vdupq_n_u16(0);
+    int16x8x4_t q[8];
+    for (int ib32 = 0; ib32 < 8; ++ib32) {
+        auto scale_l = vdup_n_s8(scales[2*ib32+0]);
+        auto scale_h = vdup_n_s8(scales[2*ib32+1]);
+        q[ib32].val[0] = vmull_s8(scale_l, vget_low_s8 (qx[ib32].val[0]));
+        q[ib32].val[1] = vmull_s8(scale_l, vget_high_s8(qx[ib32].val[0]));
+        q[ib32].val[2] = vmull_s8(scale_h, vget_low_s8 (qx[ib32].val[1]));
+        q[ib32].val[3] = vmull_s8(scale_h, vget_high_s8(qx[ib32].val[1]));
+        max_i16 = vmaxq_u16(max_i16, vmaxq_u16(vabsq_s16(q[ib32].val[0]), vabsq_s16(q[ib32].val[1])));
+        max_i16 = vmaxq_u16(max_i16, vmaxq_u16(vabsq_s16(q[ib32].val[2]), vabsq_s16(q[ib32].val[3])));
+    }
+    uint16_t imax = vmaxvq_u16(max_i16);
+    if (!imax) {
+        for (int ib32 = 0; ib32 < 8; ++ib32) for (int l = 0; l < 8; ++l) q8_k[64*ib32 + 8*l] = 0;
+        return 0.f;
+    }
+    float dnew = float(imax) * d0;
+    //auto max_u32 = vmaxq_u32(vmovl_u16(vget_low_u16(max_i16)), vmovl_u16(vget_high_u16(max_i16)));
+    //auto max_f32 = vcvtq_f32_u32(max_u32);
+    //auto dnew = vmaxvq_f32(max_f32) * d0;
+    bool needs_scaling = true;
+    if (dnew <= 1.f) {
+        dnew = 1.f; needs_scaling = false;
+    }
+    auto scale = vdupq_n_f32(1/dnew);
+    for (int ib32 = 0; ib32 < 8; ++ib32) {
+        if (needs_scaling) {
+            for (int l = 0; l < 4; ++l) {
+                auto i1 = vcvtnq_s32_f32(vmulq_f32(scale, vcvtq_f32_s32(vmovl_s16(vget_low_s16 (q[ib32].val[l])))));
+                auto i2 = vcvtnq_s32_f32(vmulq_f32(scale, vcvtq_f32_s32(vmovl_s16(vget_high_s16(q[ib32].val[l])))));
+                q[ib32].val[l] = vcombine_s16(vmovn_s32(i1), vmovn_s32(i2));
+            }
+        }
+        for (int l = 0; l < 2; ++l) {
+            auto s8 = vcombine_s8(vmovn_s16(q[ib32].val[2*l+0]), vmovn_s16(q[ib32].val[2*l+1]));
+            vst1q_s8((int8_t *)block + 16*l, s8);
+        }
+        auto qb = q8_k + 64*ib32;
+        for (int l = 0; l < 8; ++l) {
+            qb[8*l] = block[l];
+        }
+    }
+    return dnew;
+}
+
+void iqk_convert_iq2_xxs_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_iq2_xxs * x8[8];
+
+    block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
+
+    int8_t   ls[16];
+    uint32_t block[8];
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    int8x16x2_t xv[8];
+
+    for (int ix = 0; ix < nrc_x; ix += 8) {
+        for (int k = 0; k < 8; ++k) x8[k] = (const block_iq2_xxs *)((const char *)vx + (ix + k)*bx);
+        for (int i = 0; i < nb; ++i) {
+            for (int k = 0; k < 8; ++k) {
+                float d = 0.125f * GGML_FP16_TO_FP32(x8[k][i].d);
+                for (int ib32 = 0; ib32 < 8; ++ib32) {
+                    std::memcpy(aux32, x8[k][i].qs + 4*ib32, 2*sizeof(uint32_t));
+                    ls[2*ib32+0] = ls[2*ib32+1] = (2*(aux32[1] >> 28) + 1);
+                    xv[ib32].val[0] = vreinterpretq_s8_u64(uint64x2_t{iq2xxs_grid[aux8[0]], iq2xxs_grid[aux8[1]]});
+                    xv[ib32].val[1] = vreinterpretq_s8_u64(uint64x2_t{iq2xxs_grid[aux8[2]], iq2xxs_grid[aux8[3]]});
+                    apply_signs_2((uint8x16_t *)xv[ib32].val, keven_signs, aux32[1]);
+                }
+                float dnew = convert_to_q8_k_r8(1.f/124, xv, ls, block, (uint32_t *)y[i].qs + k);
+                y[i].d[k] = GGML_FP32_TO_FP16(d*dnew);
+            }
+        }
+        y += nb;
+    }
+}
+
+void iqk_convert_iq2_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_iq2_xs * x8[8];
+
+    block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
+
+    uint32_t block[8];
+
+    int8x16x2_t xv[8];
+
+    union { int8x16_t vec; int8_t val[16]; } helper;
+
+    for (int ix = 0; ix < nrc_x; ix += 8) {
+        for (int k = 0; k < 8; ++k) x8[k] = (const block_iq2_xs *)((const char *)vx + (ix + k)*bx);
+        for (int i = 0; i < nb; ++i) {
+            for (int k = 0; k < 8; ++k) {
+                float d = 0.125f * GGML_FP16_TO_FP32(x8[k][i].d);
+                auto aux = vld1_u8(x8[k][i].scales);
+                auto scales_l = vand_u8(aux, vdup_n_u8(0xf));
+                auto scales_h = vshr_n_u8(aux, 4);
+                auto aux1 = vcombine_u8(vzip1_u8(scales_l, scales_h), vzip2_u8(scales_l, scales_h));
+                helper.vec = vreinterpretq_s8_u8(vorrq_u8(vshlq_n_u8(aux1, 1), vdupq_n_u8(1)));
+                for (int ib32 = 0; ib32 < 8; ++ib32) {
+                    xv[ib32].val[0] = vreinterpretq_s8_u64(uint64x2_t{iq2xs_grid[x8[k][i].qs[4*ib32+0] & 511], iq2xs_grid[x8[k][i].qs[4*ib32+1] & 511]});
+                    xv[ib32].val[1] = vreinterpretq_s8_u64(uint64x2_t{iq2xs_grid[x8[k][i].qs[4*ib32+2] & 511], iq2xs_grid[x8[k][i].qs[4*ib32+3] & 511]});
+                    auto s1 = vreinterpretq_s8_u64(uint64x2_t{keven_signs[x8[k][i].qs[4*ib32+0] >> 9], keven_signs[x8[k][i].qs[4*ib32+1] >> 9]});
+                    auto s2 = vreinterpretq_s8_u64(uint64x2_t{keven_signs[x8[k][i].qs[4*ib32+2] >> 9], keven_signs[x8[k][i].qs[4*ib32+3] >> 9]});
+                    xv[ib32].val[0] = vmulq_s8(xv[ib32].val[0], s1);
+                    xv[ib32].val[1] = vmulq_s8(xv[ib32].val[1], s2);
+                }
+                float dnew = convert_to_q8_k_r8(1.f/124, xv, helper.val, block, (uint32_t *)y[i].qs + k);
+                y[i].d[k] = GGML_FP32_TO_FP16(d*dnew);
+            }
+        }
+        y += nb;
+    }
+}
+
+void iqk_convert_iq2_s_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_iq2_s * x8[8];
+
+    block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
+
+    uint32_t block[8];
+
+    union { int8x16_t vec; int8_t val[16]; } helper;
+    int8x16x2_t xv[8];
+
+    SignHelper sh;
+
+    for (int ix = 0; ix < nrc_x; ix += 8) {
+        for (int k = 0; k < 8; ++k) x8[k] = (const block_iq2_s *)((const char *)vx + (ix + k)*bx);
+        for (int i = 0; i < nb; ++i) {
+            for (int k = 0; k < 8; ++k) {
+                float d = 0.125f * GGML_FP16_TO_FP32(x8[k][i].d);
+                auto aux = vld1_u8(x8[k][i].scales);
+                auto scales_l = vand_u8(aux, vdup_n_u8(0xf));
+                auto scales_h = vshr_n_u8(aux, 4);
+                auto aux1 = vcombine_u8(vzip1_u8(scales_l, scales_h), vzip2_u8(scales_l, scales_h));
+                helper.vec = vreinterpretq_s8_u8(vorrq_u8(vshlq_n_u8(aux1, 1), vdupq_n_u8(1)));
+                for (int j = 0; j < 2; ++j) {
+                    auto qs = x8[k][i].qs + 16*j;
+                    auto qh = x8[k][i].qh + 4*j;
+                    auto signs16 = vld1q_u8(qs + QK_K/8);
+                    sh.init();
+                    DequantizerIQ2S::make4(sh, signs16, qs+0, qh+0, (uint8x16_t *)&xv[4*j+0]);
+                    DequantizerIQ2S::make4(sh, signs16, qs+8, qh+2, (uint8x16_t *)&xv[4*j+2]);
+                }
+                float dnew = convert_to_q8_k_r8(1.f/124, xv, helper.val, block, (uint32_t *)y[i].qs + k);
+                y[i].d[k] = GGML_FP32_TO_FP16(d*dnew);
+            }
+        }
+        y += nb;
+    }
+}
+
+void iqk_convert_iq3_xxs_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_iq3_xxs * x8[8];
+
+    block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
+
+    int8_t ls[16];
+    int8x16x2_t xv[8];
+    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) {
+            for (int k = 0; k < 8; ++k) {
+                float d = 0.25f * GGML_FP16_TO_FP32(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));
+                    ls[2*ib32 + 0] = ls[2*ib32 + 1] = (2*(aux32 >> 28) + 1);
+                    xv[ib32].val[0] = vreinterpretq_s8_u32(uint32x4_t{iq3xxs_grid[qs[0]], iq3xxs_grid[qs[1]], iq3xxs_grid[qs[2]], iq3xxs_grid[qs[3]]});
+                    xv[ib32].val[1] = vreinterpretq_s8_u32(uint32x4_t{iq3xxs_grid[qs[4]], iq3xxs_grid[qs[5]], iq3xxs_grid[qs[6]], iq3xxs_grid[qs[7]]});
+                    apply_signs_2((uint8x16_t *)xv[ib32].val, keven_signs, aux32);
+                    qs += 8;
+                }
+                float dnew = convert_to_q8_k_r8(1.f/124, xv, ls, block, (uint32_t *)y[i].qs + k);
+                y[i].d[k] = GGML_FP32_TO_FP16(d*dnew);
+            }
+        }
+        y += nb;
+    }
+}
+
+//struct DequantizerIQ3S final : public BaseDequantizer<block_iq3_s> {
+//    DequantizerIQ3S(const void * vx, size_t bx, int nrc) : BaseDequantizer(vx, bx, nrc) {}
+//
+//    constexpr static int num_blocks() { return 8; }
+//    constexpr static bool should_scale_quants() { return false; }
+//
+//    template <typename Q8>
+//    inline int32x4x2_t new_block(int i, const Q8& /*q8*/, float32x4_t * /*acc*/) {
+//        d = GGML_FP16_TO_FP32(x[i].d);
+//        uint32_t scales32[2];
+//        std::memcpy(scales32, x[i].scales, 4);
+//        scales32[1] = (((scales32[0] >> 4) & 0x0f0f0f0f) << 1) | 0x01010101;
+//        scales32[0] = ((scales32[0] & 0x0f0f0f0f) << 1) | 0x01010101;
+//        auto scales8 = vld1_u8((const uint8_t *)scales32); // 0, 2, 4, 6, 1, 3, 5, 7
+//        scales8 = vtbl1_u8(scales8, vreinterpret_u8_u64(vdup_n_u64(0x0703060205010400)));
+//        auto scales16 = vreinterpretq_s16_u16(vmovl_u8(scales8));
+//        int32x4x2_t scales;
+//        scales.val[0] = vmovl_s16(vget_low_s16(scales16));
+//        scales.val[1] = vmovl_s16(vget_high_s16(scales16));
+//        return scales;
+//    }
+//
+//    static inline void make2(SignHelper& sh, const uint8x16_t& signs16, const uint16x8_t& idx_l, uint8_t qh,
+//            const int8x16_t& hshift, uint8x16_t * b) {
+//        auto vindex = vorrq_u16(idx_l, vandq_u16(vshlq_u16(vdupq_n_u16(qh), hshift), vdupq_n_u16(256)));
+//        const uint16_t * idx = (const uint16_t *)&vindex;
+//        b[0] = vreinterpretq_u8_u32(uint32x4_t{iq3s_grid[idx[0]], iq3s_grid[idx[1]], iq3s_grid[idx[2]], iq3s_grid[idx[3]]});
+//        b[1] = vreinterpretq_u8_u32(uint32x4_t{iq3s_grid[idx[4]], iq3s_grid[idx[5]], iq3s_grid[idx[6]], iq3s_grid[idx[7]]});
+//        sh.apply_signs_1(b+0, signs16);
+//        sh.apply_signs_1(b+1, signs16);
+//    }
+//    static inline void make4(SignHelper& sh, const uint8x16_t& signs16, const uint8_t * qs, const uint8_t * qh,
+//            const int8x16_t& hshift, uint8x16_t * b) {
+//        auto idx_l = vld1q_u8(qs);
+//        make2(sh, signs16, vmovl_u8(vget_low_u8 (idx_l)), qh[0], hshift, b+0);
+//        make2(sh, signs16, vmovl_u8(vget_high_u8(idx_l)), qh[1], hshift, b+2);
+//    }
+//
+//    inline void prepare(int i, int j) {
+//
+//        static const int16_t k_shift[8] = {8, 7, 6, 5, 4, 3, 2, 1};
+//        const auto hshift  = vld1q_s16(k_shift);
+//
+//        const auto * qs = x[i].qs + 32*j;
+//        const auto * qh = x[i].qh + 4*j;
+//        const auto signs16 = vld1q_u8(x[i].signs + 16*j);
+//
+//        sh.init();
+//        make4(sh, signs16, qs+ 0, qh+0, hshift, bits.b1.val);
+//        make4(sh, signs16, qs+16, qh+2, hshift, bits.b2.val);
+//    }
+//
+//    SimpleBits bits;
+//    SignHelper sh;
+//    uint32x4x2_t gas;
+//
+//};
+
+void iqk_convert_iq3_s_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_iq3_s * x8[8];
+
+    block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
+
+    int8_t ls[16];
+    SignHelper sh;
+
+    uint32_t block[8];
+    int8x16x2_t xv[8];
+
+    static const int16_t k_shift[8] = {8, 7, 6, 5, 4, 3, 2, 1};
+    const auto hshift  = vld1q_s16(k_shift);
+
+    for (int ix = 0; ix < nrc_x; ix += 8) {
+        for (int k = 0; k < 8; ++k) x8[k] = (const block_iq3_s *)((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 j = 0; j < 2; ++j) {
+                    const auto * qs = x8[k][i].qs + 32*j;
+                    const auto * qh = x8[k][i].qh + 4*j;
+                    const auto signs16 = vld1q_u8(x8[k][i].signs + 16*j);
+                    sh.init();
+                    DequantizerIQ3S::make4(sh, signs16, qs+ 0, qh+0, hshift, (uint8x16_t *)&xv[4*j+0]);
+                    DequantizerIQ3S::make4(sh, signs16, qs+16, qh+2, hshift, (uint8x16_t *)&xv[4*j+2]);
+                }
+                for (int ib32 = 0; ib32 < 8; ++ib32) {
+                    ls[2*ib32 + 0] = ls[2*ib32 + 1] = (2*((x8[k][i].scales[ib32/2] >> 4*(ib32%2)) & 0xf) + 1);
+                }
+                float dnew = convert_to_q8_k_r8(1.f/127, xv, ls, block, (uint32_t *)y[i].qs + k);
+                y[i].d[k] = GGML_FP32_TO_FP16(d*dnew);
+            }
+        }
+        y += nb;
+    }
+}
+
+
 }
 
 bool iqk_convert_iquants_q80_r8([[maybe_unused]] int type, int n, [[maybe_unused]] const void * vx, [[maybe_unused]] size_t bx, [[maybe_unused]] void * vy, int nrc_x) {
     if (n%QK_K != 0 || nrc_x%8 != 0) return false;
-    return false;
-    //switch (ggml_type(type)) {
-    //    case GGML_TYPE_IQ2_XXS: iqk_convert_iq2_xxs_q8_k_r8(n, vx, bx, vy, nrc_x); break;
-    //    case GGML_TYPE_IQ2_XS : iqk_convert_iq2_xs_q8_k_r8 (n, vx, bx, vy, nrc_x); break;
-    //    case GGML_TYPE_IQ2_S  : iqk_convert_iq2_s_q8_k_r8  (n, vx, bx, vy, nrc_x); break;
-    //    case GGML_TYPE_IQ3_XXS: iqk_convert_iq3_xxs_q8_k_r8(n, vx, bx, vy, nrc_x); break;
-    //    case GGML_TYPE_IQ3_S  : iqk_convert_iq3_s_q8_k_r8  (n, vx, bx, vy, nrc_x); break;
-    //    default: return false;
-    //}
-    //return true;
+    switch (ggml_type(type)) {
+        case GGML_TYPE_IQ2_XXS: iqk_convert_iq2_xxs_q8_k_r8(n, vx, bx, vy, nrc_x); break;
+        case GGML_TYPE_IQ2_XS : iqk_convert_iq2_xs_q8_k_r8 (n, vx, bx, vy, nrc_x); break;
+        case GGML_TYPE_IQ2_S  : iqk_convert_iq2_s_q8_k_r8  (n, vx, bx, vy, nrc_x); break;
+        case GGML_TYPE_IQ3_XXS: iqk_convert_iq3_xxs_q8_k_r8(n, vx, bx, vy, nrc_x); break;
+        case GGML_TYPE_IQ3_S  : iqk_convert_iq3_s_q8_k_r8  (n, vx, bx, vy, nrc_x); break;
+        default: return false;
+    }
+    return true;
 }
 
 bool iqk_set_kernels_iquants(int ne00, int typeA, int typeB, std::array<mul_mat_t, IQK_MAX_NY>& kernels, mul_mat_t& func16) {
diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp
@@ -271,6 +271,11 @@ struct MulMat {
         }
 #else
         switch (type) {
+            case GGML_TYPE_IQ2_XXS: return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
+            case GGML_TYPE_IQ2_XS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
+            case GGML_TYPE_IQ2_S  : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
+            case GGML_TYPE_IQ3_XXS: 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_Q4_0   : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
             case GGML_TYPE_Q4_1   : return nrc_y >= 32 ? GGML_TYPE_Q8_1    : type;
             case GGML_TYPE_Q5_0   : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
