Function CheckMask

tensorflow/core/common_runtime/gpu/gpu_debug_allocator.cc:43–63

int64* after_mask = NewMask(0xcdcdcdcdcdcdcdcd);

bool CheckMask(se::StreamExecutor* exec, void* ptr, int64* mask) {
  se::DeviceMemory<int64> gpu_ptr{se::DeviceMemoryBase{ptr, MASK_BYTES}};
  int64 tmp[MASK_WORDS];

  Status result = exec->SynchronousMemcpyD2H(gpu_ptr, MASK_BYTES, tmp);
  if (!result.ok()) {
    LOG(FATAL) << "Could not copy debug mask, " << result;
  }

  bool ok = true;
  for (int i = 0; i < MASK_WORDS; ++i) {
    ok &= (mask[i] == tmp[i]);
    if (!ok) {
      LOG(ERROR) << "i=" << i
                 << " mask=" << reinterpret_cast<const void*>(mask[i])
                 << " field=" << reinterpret_cast<const void*>(tmp[i]);
    }
  }

  return ok;
}

void InitMask(se::StreamExecutor* exec, void* ptr, int64* mask) {
  se::DeviceMemory<int64> gpu_ptr{se::DeviceMemoryBase{ptr, MASK_BYTES}};