#if defined(INTELFPGA_CL)
#define UTL_ENABLE_CHANNELS _Pragma("OPENCL EXTENSION cl_intel_channels : enable")
#define UTL_READ_CH(ch) read_channel_intel(ch)
#define UTL_WRITE_CH(ch, dat) write_channel_intel(ch, dat)
#else
#define UTL_ENABLE_CHANNELS _Pragma("OPENCL_EXTENSION cl_altera_channels : enable")
#define UTL_READ_CH(ch) read_channel_altera(ch)
#define UTL_WRITE_CH(ch, dat) write_channel_altera(ch, dat)
#endif

UTL_ENABLE_CHANNELS

#define DATA_SIZE 4
#define VEC_SIZE 4

typedef struct {
    char d[DATA_SIZE];
} data_t;

typedef struct {
    data_t v[VEC_SIZE];
} data_ch_t;

typedef struct {
    char v[VEC_SIZE];
} ch_t;

channel ch_t ch_1 __attribute((depth(0)));
channel data_ch_t ch_2 __attribute((depth(0)));
channel data_ch_t ch_3 __attribute((depth(0)));
channel ch_t ch_4 __attribute((depth(0)));

__kernel
__attribute((task))
void k1(__global const ch_t* restrict data1,
        __global const data_ch_t* restrict data2,
        __global const data_t* restrict data3,
        const int l1,
        const int l2) {
    data_t buf_d3[2][2];
    data_ch_t buf_d2[8];

    #pragma loop_coalesce 2
    for (int i = 0; i < l1; ++i) {
        for (int j = 0; j < l2; ++j) {
            buf_d3[0][j] = data3[j];
            buf_d2[j] = data2[j];

            UTL_WRITE_CH(ch_1, data1[j]);
            UTL_WRITE_CH(ch_2, buf_d2[i]);

            data_ch_t dch;
            dch.v[0] = buf_d3[0][j];
            UTL_WRITE_CH(ch_3, dch);
        }
    }
}

__kernel
__attribute((task))
void k2(const int l1, const int l2) {
    for (int i = 0; i < l1; ++i) {
        ch_t t;
        for (int j = 0; j < l2; ++j) {
            const data_ch_t d3 = UTL_READ_CH(ch_3);
            const data_ch_t d2 = UTL_READ_CH(ch_2);
            t.v[0] += d2.v[0].d[0] * d3.v[0].d[0];

            const ch_t d1 = UTL_READ_CH(ch_1);
            t.v[0] += d1.v[0];
        }

        UTL_WRITE_CH(ch_4, t);
    }
}

__kernel
__attribute((task))
void k3(__global char* restrict output, const int l1) {
    for (int j = 0; j < l1; ++j) {
        ch_t dat = UTL_READ_CH(ch_4);
        output[j] = dat.v[0];
    }
}