Method runInitFiniKernel

rocclr/device/devprogram.cpp:2151–2207

amd::Monitor Program::initFiniLock_(true);

bool Program::runInitFiniKernel(const std::vector<const Kernel*>& kernels) const {
  amd::HostQueue* queue = nullptr;

  for (const auto& kernel : kernels) {
    amd::ScopedLock sl(initFiniLock_);

    if (queue == nullptr) {
      queue = new amd::HostQueue(device_().context(), device_(), 0);
      if (queue == nullptr) {
        LogError("Unable to create queue");
        return false;
      }
      queue->create();
    }

    LogPrintfInfo("%s is marked init/fini", kernel->name().c_str());

    size_t globalWorkOffset[3] = {0};
    size_t globalWorkSize[3] = {1, 1, 1};
    size_t localWorkSize[3] = {1, 1, 1};
    amd::NDRangeContainer ndrange(3, globalWorkOffset, globalWorkSize, localWorkSize);
    amd::Command::EventWaitList waitList;

    auto symbol = owner_.findSymbol(kernel->name().c_str());
    amd::Kernel* k = new amd::Kernel(owner_, *symbol, kernel->name().c_str());
    if (!k) {
      queue->release();
      LogError("Unable to create kernel");
      return false;
    }

    amd::NDRangeKernelCommand* kernelCommand =
        new amd::NDRangeKernelCommand(*queue, waitList, *k, ndrange);
    if (!kernelCommand) {
      LogError("Unale to allocate memory to launch kernel");
      k->release();
      queue->release();
      return false;
    }
    if (CL_SUCCESS != kernelCommand->captureAndValidate()) {
      LogError("Kernel Capture and Validate failed");
      kernelCommand->release();
      k->release();
      queue->release();
      return false;
    }
    kernelCommand->enqueue();
    queue->finish();
    k->release();
    kernelCommand->release();
  }

  if (queue != nullptr) {
    queue->release();
  }
  return true;
}