cbuffer cbuffer_u : register(b0) {
  uint4 u[8];
};
RWByteAddressBuffer s : register(u1);

float4x2 u_load(uint offset) {
  const uint scalar_offset = ((offset + 0u)) / 4;
  uint4 ubo_load = u[scalar_offset / 4];
  const uint scalar_offset_1 = ((offset + 8u)) / 4;
  uint4 ubo_load_1 = u[scalar_offset_1 / 4];
  const uint scalar_offset_2 = ((offset + 16u)) / 4;
  uint4 ubo_load_2 = u[scalar_offset_2 / 4];
  const uint scalar_offset_3 = ((offset + 24u)) / 4;
  uint4 ubo_load_3 = u[scalar_offset_3 / 4];
  return float4x2(asfloat(((scalar_offset & 2) ? ubo_load.zw : ubo_load.xy)), asfloat(((scalar_offset_1 & 2) ? ubo_load_1.zw : ubo_load_1.xy)), asfloat(((scalar_offset_2 & 2) ? ubo_load_2.zw : ubo_load_2.xy)), asfloat(((scalar_offset_3 & 2) ? ubo_load_3.zw : ubo_load_3.xy)));
}

[numthreads(1, 1, 1)]
void f() {
  float2x4 t = transpose(u_load(64u));
  float l = length(asfloat(u[0].zw).yx);
  float a = abs(asfloat(u[0].zw).yx.x);
  s.Store(0u, asuint(((t[0].x + float(l)) + float(a))));
  return;
}