Fine-Grained Resource Sharing for Concurrent GPGPU Kernels