#version 450

#include "mul_mat_vec_base.comp"

layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;

shared FLOAT_TYPE tmp[32];

void main() {
    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;

    uint a_offset, b_offset, d_offset;
    get_offsets(a_offset, b_offset, d_offset);

    const uint num_blocks_per_row = p.ncols / QUANT_K;
    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;

    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1

    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8

    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7

#if K_QUANTS_PER_ITERATION == 1
    const uint l0 = v_in;                                   // 0...15
    const uint is = 0;
#else
    const uint l0 = 4 * v_in;                               // 0, 4, 8, ..., 28
    const uint is = v_in / 4;
#endif

    const uint ql_offset = 64*v_im + l0;
    const uint qh_offset = 32*v_im + l0;
    const uint s_offset  =  8*v_im + is;
    const uint y_offset = 128*v_im + l0;

    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp

    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
        const uint y_idx   = i * QUANT_K + y_offset;

        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);

#if K_QUANTS_PER_ITERATION == 1
        FLOAT_TYPE sum = FLOAT_TYPE(data_b[b_offset + y_idx +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x03) << 4)) - 32)
                       + FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32)
                       + FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x0c) << 2)) - 32)
                       + FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32)
                       + FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x30) >> 0)) - 32)
                       + FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32)
                       + FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0xc0) >> 2)) - 32)
                       + FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32);
        tmp[16 * ix + tid] += sum;
#else
        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
        [[unroll]] for (int l = 0; l < 4; ++l) {
            sum += FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32)
                 + FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32)
                 + FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32)
                 + FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32);
        }
        tmp[16 * ix + tid] += sum;
#endif
    }

    // sum up partial sums and write back result
    barrier();
    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
        if (tid < s) {
            tmp[tid] += tmp[tid + s];
       }
        barrier();
    }
    if (tid == 0) {
        data_d[d_offset + row] = D_TYPE(tmp[0]);
    }
}