llama.cpp/kompute-shaders/op_getrows.comp

void main() {
    const uint i = gl_WorkGroupID.x;
    const int r = inB[i + pcs.inBOff];

    int z = 0;
    for (uint ind = gl_LocalInvocationID.x; ind < pcs.ne00/16; ind += gl_WorkGroupSize.x) {
        const uint inIndex = (r * pcs.nb01 + pcs.inAOff) + ind/NL * SIZE_OF_BLOCK;
        const mat4 result = dequantize_block(inIndex, ind%NL);
        for (uint j = 0; j < 4; ++j) {
            for (uint k = 0; k < 4; ++k) {
                const uint outIndex = i * pcs.nb1/BYTES_FOR_TYPE + pcs.outOff + z;
                out_[outIndex] = result[j][k];
                ++z;
            }
        }
    }
}