#version 450

#include "dequant_head.comp"

layout(local_size_x = 64, 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 i = gl_WorkGroupID.x * 256 + wgy;
        if (i >= p.M * p.K / QUANT_K) {
            return;
        }

        const uint tid = gl_LocalInvocationID.x;
        const uint ip = tid / 32;
        const uint il = tid - 32 * ip;
        const uint is = 8 * ip + il / 16;

        const uint y_idx = i * QUANT_K + 128 * ip + il;

        const uint ql_idx = 32 * ip + il;
        const uint8_t qs = data_a[i].qs[32 * ip + il];

        FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
        FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
        data_b[y_idx +  0] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+0] & 0xF) * ((qs >> 0) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+0] >> 4));
        data_b[y_idx + 32] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+2] & 0xF) * ((qs >> 2) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+2] >> 4));
        data_b[y_idx + 64] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+4] & 0xF) * ((qs >> 4) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+4] >> 4));
        data_b[y_idx + 96] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+6] & 0xF) * ((qs >> 6) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+6] >> 4));
    }
}