On Thu, Mar 19, 2015 at 03:13:15PM +0000, Chris Wilson wrote:
Pretty much an igt that compared the speed of just querying the hw counter vs querying with a regular vblank interrupt would be ideal for measuring the impact here.
ickle@crystalwell:/usr/src/intel-gpu-tools$ sudo ./tests/kms_vblank IGT-Version: 1.10-gc15fb9e (x86_64) (Linux: 4.0.0-rc4+ x86_64) Time to query current counter (idle): 2.217µs Subtest query-idle: SUCCESS (1.006s) Time to query current counter (busy): 0.244µs Subtest query-busy: SUCCESS (1.201s
- 28.04% kms_vblank [kernel.kallsyms] [k] gen6_read32 - gen6_read32 + 65.27% gm45_get_vblank_counter + 17.50% ironlake_disable_display_irq + 16.53% ironlake_enable_display_irq + 21.57% kms_vblank [kernel.kallsyms] [k] hsw_unclaimed_reg_detect.isra + 10.69% kms_vblank [kernel.kallsyms] [k] _raw_spin_lock_irqsave + 9.91% kms_vblank [kernel.kallsyms] [k] __intel_get_crtc_scanline + 6.66% kms_vblank [kernel.kallsyms] [k] _raw_spin_unlock_irqrestore + 1.34% kms_vblank [kernel.kallsyms] [k] i915_get_crtc_scanoutpos + 1.29% kms_vblank [kernel.kallsyms] [k] drm_wait_vblank + 1.27% kms_vblank [kernel.kallsyms] [k] hsw_unclaimed_reg_debug.isra. + 1.17% kms_vblank [kernel.kallsyms] [k] copy_user_enhanced_fast_strin + 1.12% kms_vblank [kernel.kallsyms] [k] drm_ioctl + 1.05% kms_vblank [kernel.kallsyms] [k] vblank_disable_and_save