static const uint3 gl_WorkGroupSize = uint3(1u, 1u, 1u);

RWByteAddressBuffer _11 : register(u0);

void comp_main()
{
    _11.Store(0, asuint(asfloat(_11.Load(0)) * asfloat(_11.Load(96))));
    _11.Store4(16, asuint(asfloat(_11.Load4(16)) * asfloat(_11.Load4(112))));
    float4x4 _35 = asfloat(uint4x4(_11.Load4(128), _11.Load4(144), _11.Load4(160), _11.Load4(176)));
    float4x4 _37 = asfloat(uint4x4(_11.Load4(32), _11.Load4(48), _11.Load4(64), _11.Load4(80)));
    float4x4 _38 = mul(_35, _37);
    _11.Store4(32, asuint(_38[0]));
    _11.Store4(48, asuint(_38[1]));
    _11.Store4(64, asuint(_38[2]));
    _11.Store4(80, asuint(_38[3]));
}

[numthreads(1, 1, 1)]
void main()
{
    comp_main();
}