Method Invoke

tensorflow/lite/delegates/gpu/gl_delegate.cc:276–334

  }

  Status Invoke(TfLiteContext* context) {
    const EGLContext egl_context_at_delegate_init = env_->context().context();
    const EGLContext egl_context_at_delegate_invoke = eglGetCurrentContext();
    if (egl_context_at_delegate_init != egl_context_at_delegate_invoke) {
      return FailedPreconditionError(
          "Delegate should run on the same thread where it was initialized.");
    }

    // Push input data from a tensor to GPU.
    for (ValueId id : inputs_) {
      const ValueRef& ref = tensors_[id];
      auto external_object = bhwc_objects_.FindBuffer(ref.tensor_index);
      if (external_object) {
        // Use input from GPU.
        // Conversion is needed only when external object is not phwc4.
        if (!IsPHWC4(tensors_[id].shape)) {
          RETURN_IF_ERROR(bhwc_to_phwc4_.Convert(
              ref.shape, *external_object, command_queue_.get(),
              phwc4_objects_.FindBuffer(id)));
        }
      } else {
        // Copy from CPU to GPU
        TfLiteTensor& tensor = context->tensors[ref.tensor_index];
        RETURN_IF_ERROR(CopyToBufferHandle(id, &tensor));
      }
    }

    // Run inference.
    RETURN_IF_ERROR(inference_context_->Reset());
    RETURN_IF_ERROR(inference_context_->Execute());

    // Push output data from GPU to a tensor.
    bool finished_gpu_processing = false;
    for (ValueId id : outputs_) {
      const ValueRef& ref = tensors_[id];
      auto external_object = bhwc_objects_.FindBuffer(ref.tensor_index);
      if (external_object) {
        // Convert data from PHWC4 to BHWC and leave it in GPU object.
        // Conversion is needed only when external object is not phwc4.
        if (!IsPHWC4(tensors_[id].shape)) {
          RETURN_IF_ERROR(
              phwc4_to_bhwc_.Convert(ref.shape, *phwc4_objects_.FindBuffer(id),
                                     command_queue_.get(), external_object));
        }
      } else {
        // Wait until all GPU command are completed. This call leads to a lower
        // processing latency because a buffer reading below will not stall if
        // data is not yet ready.
        if (!finished_gpu_processing) {
          RETURN_IF_ERROR(command_queue_->WaitForCompletion());
          finished_gpu_processing = true;
        }
        // Copy from GPU to CPU.
        TfLiteTensor& tensor = context->tensors[ref.tensor_index];
        RETURN_IF_ERROR(CopyFromBufferHandle(id, &tensor));
      }
    }
    return OkStatus();