cbuffer cbuffer_u : register(b0, space0) {
  uint4 u[1];
};
groupshared matrix<float16_t, 2, 4> w;

struct tint_symbol_1 {
  uint local_invocation_index : SV_GroupIndex;
};

matrix<float16_t, 2, 4> tint_symbol_2(uint4 buffer[1], uint offset) {
  const uint scalar_offset = ((offset + 0u)) / 4;
  uint4 ubo_load_1 = buffer[scalar_offset / 4];
  uint2 ubo_load = ((scalar_offset & 2) ? ubo_load_1.zw : ubo_load_1.xy);
  vector<float16_t, 2> ubo_load_xz = vector<float16_t, 2>(f16tof32(ubo_load & 0xFFFF));
  vector<float16_t, 2> ubo_load_yw = vector<float16_t, 2>(f16tof32(ubo_load >> 16));
  const uint scalar_offset_1 = ((offset + 8u)) / 4;
  uint4 ubo_load_3 = buffer[scalar_offset_1 / 4];
  uint2 ubo_load_2 = ((scalar_offset_1 & 2) ? ubo_load_3.zw : ubo_load_3.xy);
  vector<float16_t, 2> ubo_load_2_xz = vector<float16_t, 2>(f16tof32(ubo_load_2 & 0xFFFF));
  vector<float16_t, 2> ubo_load_2_yw = vector<float16_t, 2>(f16tof32(ubo_load_2 >> 16));
  return matrix<float16_t, 2, 4>(vector<float16_t, 4>(ubo_load_xz[0], ubo_load_yw[0], ubo_load_xz[1], ubo_load_yw[1]), vector<float16_t, 4>(ubo_load_2_xz[0], ubo_load_2_yw[0], ubo_load_2_xz[1], ubo_load_2_yw[1]));
}

void f_inner(uint local_invocation_index) {
  {
    w = matrix<float16_t, 2, 4>((float16_t(0.0h)).xxxx, (float16_t(0.0h)).xxxx);
  }
  GroupMemoryBarrierWithGroupSync();
  w = tint_symbol_2(u, 0u);
  uint2 ubo_load_4 = u[0].xy;
  vector<float16_t, 2> ubo_load_4_xz = vector<float16_t, 2>(f16tof32(ubo_load_4 & 0xFFFF));
  vector<float16_t, 2> ubo_load_4_yw = vector<float16_t, 2>(f16tof32(ubo_load_4 >> 16));
  w[1] = vector<float16_t, 4>(ubo_load_4_xz[0], ubo_load_4_yw[0], ubo_load_4_xz[1], ubo_load_4_yw[1]);
  uint2 ubo_load_5 = u[0].xy;
  vector<float16_t, 2> ubo_load_5_xz = vector<float16_t, 2>(f16tof32(ubo_load_5 & 0xFFFF));
  vector<float16_t, 2> ubo_load_5_yw = vector<float16_t, 2>(f16tof32(ubo_load_5 >> 16));
  w[1] = vector<float16_t, 4>(ubo_load_5_xz[0], ubo_load_5_yw[0], ubo_load_5_xz[1], ubo_load_5_yw[1]).ywxz;
  w[0][1] = float16_t(f16tof32(((u[0].z) & 0xFFFF)));
}

[numthreads(1, 1, 1)]
void f(tint_symbol_1 tint_symbol) {
  f_inner(tint_symbol.local_invocation_index);
  return;
}