SKIP: FAILED

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

struct tint_symbol_1 {
  uint local_invocation_index : SV_GroupIndex;
};

matrix<float16_t, 4, 2> u_load_1(uint offset) {
  const uint scalar_offset = ((offset + 0u)) / 4;
  uint ubo_load = u[scalar_offset / 4][scalar_offset % 4];
  const uint scalar_offset_1 = ((offset + 4u)) / 4;
  uint ubo_load_1 = u[scalar_offset_1 / 4][scalar_offset_1 % 4];
  const uint scalar_offset_2 = ((offset + 8u)) / 4;
  uint ubo_load_2 = u[scalar_offset_2 / 4][scalar_offset_2 % 4];
  const uint scalar_offset_3 = ((offset + 12u)) / 4;
  uint ubo_load_3 = u[scalar_offset_3 / 4][scalar_offset_3 % 4];
  return matrix<float16_t, 4, 2>(vector<float16_t, 2>(float16_t(f16tof32(ubo_load & 0xFFFF)), float16_t(f16tof32(ubo_load >> 16))), vector<float16_t, 2>(float16_t(f16tof32(ubo_load_1 & 0xFFFF)), float16_t(f16tof32(ubo_load_1 >> 16))), vector<float16_t, 2>(float16_t(f16tof32(ubo_load_2 & 0xFFFF)), float16_t(f16tof32(ubo_load_2 >> 16))), vector<float16_t, 2>(float16_t(f16tof32(ubo_load_3 & 0xFFFF)), float16_t(f16tof32(ubo_load_3 >> 16))));
}

typedef matrix<float16_t, 4, 2> u_load_ret[4];
u_load_ret u_load(uint offset) {
  matrix<float16_t, 4, 2> arr[4] = (matrix<float16_t, 4, 2>[4])0;
  {
    for(uint i_1 = 0u; (i_1 < 4u); i_1 = (i_1 + 1u)) {
      arr[i_1] = u_load_1((offset + (i_1 * 16u)));
    }
  }
  return arr;
}

void f_inner(uint local_invocation_index) {
  {
    for(uint idx = local_invocation_index; (idx < 4u); idx = (idx + 1u)) {
      const uint i = idx;
      w[i] = matrix<float16_t, 4, 2>((float16_t(0.0h)).xx, (float16_t(0.0h)).xx, (float16_t(0.0h)).xx, (float16_t(0.0h)).xx);
    }
  }
  GroupMemoryBarrierWithGroupSync();
  w = u_load(0u);
  w[1] = u_load_1(32u);
  uint ubo_load_4 = u[0].y;
  w[1][0] = vector<float16_t, 2>(float16_t(f16tof32(ubo_load_4 & 0xFFFF)), float16_t(f16tof32(ubo_load_4 >> 16))).yx;
  w[1][0].x = float16_t(f16tof32(((u[0].y) & 0xFFFF)));
}

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