#version 450

#include "dequant_head.comp"

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

layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};

void main() {
    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
        const uint ib = gl_WorkGroupID.x * 256 + wgy;
        if (ib >= p.M * p.K / QUANT_K) {
            return;
        }

        const uint tid = gl_LocalInvocationID.x;
        const uint il = tid / 8;
        const uint ir = tid % 8;
        const uint is = 2 * il;
        const uint n = 4;

        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);

        const uint y_idx = ib * QUANT_K + 64 * il + n * ir;
        const uint qs_idx = 32*il + n * ir;

        uint scidx0 = (is < 4) ? is : (is + 4);
        uint scidx1 = (is < 4) ? is : (is - 4);
        uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
        uint scidxshift1 = (is < 4) ? 0 : 2;
        uint mbidx0 = is + 4;
        uint mbidx1 = (is < 4) ? is + 4 : is;
        uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
        uint mbidxshift0 = (is < 4) ? 0 : 4;
        uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
        uint mbidxshift1 = (is < 4) ? 0 : 2;

        uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
        uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));

        const FLOAT_TYPE d1 = dall * sc;
        const FLOAT_TYPE m1 = dmin * mbyte;

        scidx0 = (is < 4) ? is + 1 : (is + 5);
        scidx1 = (is < 4) ? is + 1 : (is - 3);
        scidxmask1 = (is < 4) ? 0x30 : 0xC0;
        scidxshift1 = (is < 4) ? 0 : 2;
        mbidx0 = is + 5;
        mbidx1 = (is < 4) ? is + 5 : is + 1;
        mbidxmask0 = (is < 4) ? 0xF : 0xF0;
        mbidxshift0 = (is < 4) ? 0 : 4;
        mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
        mbidxshift1 = (is < 4) ? 0 : 2;

        sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
        mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));

        const FLOAT_TYPE d2 = dall * sc;
        const FLOAT_TYPE m2 = dmin * mbyte;

        [[unroll]] for (uint l = 0; l < n; ++l) {
            data_b[y_idx + l     ] = D_TYPE(d1 * FLOAT_TYPE(data_a[ib].qs[qs_idx + l] & 0xF) - m1);
            data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[ib].qs[qs_idx + l] >>  4) - m2);
        }
    }
}