// This file has been autogenerated by generate_cu_files.py, do not edit manually.

#include "../mmq.cuh"

template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_k_r4(
    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {

#ifdef INT8_MMA_AVAILABLE
    int   * x_qs = (int   *)  x_tile;
    float * x_df = (float *) (x_qs + WARP_SIZE*2);
#else
    constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16;
    int   * x_qs = (int   *)  x_tile;
    float * x_df = (float *) (x_qs + txs.qs);
#endif // INT8_MMA_AVAILABLE

    const int kqsx = threadIdx.x/4;  // 0...7 -> block of 32

    uint32_t aux32[4];
    const uint8_t * aux8 = (const uint8_t *)aux32;
#pragma unroll
    for (int i0 = 0; i0 < mmq_y; i0 += 4*nwarps) {
        int i = i0 + 4*threadIdx.y + threadIdx.x%4;

        if (need_check) {
            i = min(i, i_max);
        }
        int i4 = i/4;
        int ir = i%4;

        const block_iq4_k_r4 * bxi = (const block_iq4_k_r4 *)(x + 4*i4*stride) + kbx0;

        const float d = __half2float(bxi->d[ir]);

    #pragma unroll
        for (int l = 0; l < 2; ++l) {

            auto values_l = iq4k_table + ((bxi->extra[ir+4*l] << (8 - kqsx)) & 0x100);

            const int ql1 = get_int_b4(bxi->qs, 16*kqsx + ir + 4*l + 0);
            const int ql2 = get_int_b4(bxi->qs, 16*kqsx + ir + 4*l + 8);
            aux32[0] = (ql1 >> 0) & 0x0f0f0f0f;
            aux32[1] = (ql1 >> 4) & 0x0f0f0f0f;
            aux32[2] = (ql2 >> 0) & 0x0f0f0f0f;
            aux32[3] = (ql2 >> 4) & 0x0f0f0f0f;

#ifdef INT8_MMA_AVAILABLE
            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + 4*l + 0] = int_from_table_x(aux8+ 0, values_l);
            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + 4*l + 2] = int_from_table_x(aux8+ 4, values_l);
            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + 4*l + 1] = int_from_table_x(aux8+ 8, values_l);
            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + 4*l + 3] = int_from_table_x(aux8+12, values_l);
#else
            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + 4*l + 0] = int_from_table_x(aux8+ 0, values_l);
            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + 4*l + 2] = int_from_table_x(aux8+ 4, values_l);
            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + 4*l + 1] = int_from_table_x(aux8+ 8, values_l);
            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + 4*l + 3] = int_from_table_x(aux8+12, values_l);
#endif // INT8_MMA_AVAILABLE

        }

        int is = 8*kqsx + ir;
        float dl1 = d * ((((bxi->scales_l[is%32] >> 4*(is/32)) & 0xf) | (((bxi->scales_h[is%16] >> 2*(is/16)) & 3) << 4)) - 32);
        is += 4;
        float dl2 = d * ((((bxi->scales_l[is%32] >> 4*(is/32)) & 0xf) | (((bxi->scales_h[is%16] >> 2*(is/16)) & 3) << 4)) - 32);

#ifdef INT8_MMA_AVAILABLE
        x_df[i*MMQ_MMA_TILE_X_K_Q3_K               + 2*kqsx+0] = dl1;
        x_df[i*MMQ_MMA_TILE_X_K_Q3_K               + 2*kqsx+1] = dl2;
#else
        x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = dl1;
        x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = dl2;
#endif // INT8_MMA_AVAILABLE
    }
}

template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ4_K_R4> {
    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq4_k_r4<mmq_y, nwarps, need_check>;
    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_16_q8_1_mma<mmq_x, mmq_y, nwarps>;
    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
};

DECL_MMQ_CASE(GGML_TYPE_IQ4_K_R4);