There is now support for OA (Observation Architecture) performance counters via the i965 Mesa DRI driver. The OA performance counters are for Haswell and newer.
The Observation Architecture performance counters expose compute metrics, memory reads distribution, and memory writes distribution. Some of the individual counters include the GPU core clocks, the percentage of time the Execution Units are processing, the number of pixels dropped, the number of control shader hardware threads dispatched, the percentage of time when geometry shaders were stalled, the number of VS/CS/FS shader hardware threads dispatched, and more. The new metrics are documented with this commit.
These OA counters are exposed to applications via the OpenGL INTEL_performance_query extension. Just another feature for Mesa 17.1!