Can GPGPU Programming Be Liberated from the Data-Parallel Bottleneck?