Skip to content

Performance

github-actions[bot] edited this page Sep 13, 2026 · 10 revisions

Performance

xemu performance work should connect a measured symptom to a specific processing path, then prove that a proposed change improves that path without moving cost elsewhere or weakening correctness.

Principles

  1. Separate layers. Shader binding, stage-module compilation, graphics-pipeline creation, command-buffer submission, GPU execution, and guest-visible cadence are different measurements.
  2. Measure tails. Average FPS can stay flat while p95/p99 or single-frame stalls get worse.
  3. Keep source identity. A result belongs to the exact executable/source state that produced it.
  4. Retain adverse data. A workload-specific win does not erase a regression elsewhere.
  5. Distinguish diagnostics from product runs. Instrumentation can change timing.
  6. Preserve correctness oracles. A faster path is not an optimization if ordering, output, or guest-visible behavior changes incorrectly.

Vulkan performance

See Vulkan Performance for the current Vulkan documentation set.

The three initial pages deliberately cover different cost classes:

Reading performance tables

Where a page reports an "Improvement %", verify the metric direction. Lower frame interval or CPU/GPU duration is good; higher guest progress/cadence can be good. Do not interpret a large percentage of time in a region as equally large removable overhead: the region may include useful work.

Clone this wiki locally