Faset_Engine/docs/studies/20-p2-gpu-visibility-benchmark-2026-09-23.md

P2 GPU visibility: first measured run (2026-09-23)

The acceptance protocol passed all 15 offscreen Vulkan cases on this source state. Three benchmark repetitions then produced 810 measured frames and zero Vulkan validation errors. GPU visibility reduced CPU render-call time in these synthetic scenes, but its GPU command time was substantially higher than the direct path. The compute MainCull pass accounts for almost all of that GPU cost. This is a Debug build with validation and diagnostic counter readback, so the numbers identify work to profile; they do not establish Release-build throughput.

Reproduction and scope

Field	Value
Source commit	ae537c0a4b28e6033a6db045d3401f8d8e96f2ee
Acceptance executable SHA-256	8eba12dab776ee64a2edcdf743f3bad83b91a991608578ee136c072566ab4751
src/render/renderer.cpp SHA-256	d650718b78cb757d322586aef3ebfc97cb412dd36ac29d3071886f59d045752e
shaders/gpu_scene.slang SHA-256	8fb7460ebcabc87a844fa2774f3bdad1d1f594a8d3b84d0a6881c9e439a35d98
Compiled SPIR-V manifest SHA-256	307a13cf30c8211703224072d7d730035dea2f8eaf0e66f0a70e01737ce72238 (SHA-256 of lexically sorted sha256sum build/linux-debug/shaders/*.spv output)
Build	CMake Debug, Ubuntu Clang 21.1.8, Linux 7.0.0-31-generic x86_64
GPU / driver	NVIDIA GeForce RTX 2080 Ti / 595.84; Vulkan device API 1.4.329
Capture	Headless 320 × 240 RGBA, synchronous image readback every frame
Diagnostics	Vulkan validation and visibility_diagnostics enabled; zero reported validation errors
Workload	Deterministic geometry, no random seed; 1,024 grid instances plus scene-specific occluder; shadows disabled for benchmark objects
Sampling	Sequential direct → GPU frustum → GPU occlusion for each scene; 10 warmup + 30 sampled frames per scene/mode/run; three sequential runs

The three raw, unmodified per-frame captures are run 1, run 2, and run 3. Each contains 270 rows: 30 frames for each of three scenes and three visibility modes. Source, shader and executable hashes were the same before and after all three runs. The CSV frame field resets to zero for each scene/mode. The SPIR-V manifest hash above includes file names and hashes in sorted order.

Values below pool 90 frames per scene/mode. P50 is the median and p95 is linearly interpolated at rank (n - 1) × 0.95 after sorting. The run range is the minimum–maximum of the three individual-run p50 values, making run-to-run variation visible. Times are milliseconds.

Scene	Mode	CPU render p50 / p95	CPU run-p50 range	GPU commands p50 / p95	GPU run-p50 range
Frustum-heavy	Direct	73.60 / 90.94	71.37–81.42	0.241 / 0.268	0.240–0.242
Frustum-heavy	GPU frustum	14.58 / 23.04	13.48–15.23	4.00 / 4.69	3.89–4.12
Frustum-heavy	GPU occlusion	15.01 / 22.14	14.00–15.67	3.99 / 4.60	3.96–4.02
Open	Direct	70.34 / 89.71	68.70–71.51	0.502 / 0.531	0.500–0.503
Open	GPU frustum	29.52 / 37.88	28.98–30.67	18.08 / 19.21	17.87–18.25
Open	GPU occlusion	30.82 / 38.47	29.11–33.36	18.36 / 19.18	18.03–18.63
Occluded	Direct	68.00 / 84.26	67.45–68.64	0.499 / 0.554	0.496–0.500
Occluded	GPU frustum	29.63 / 40.42	28.49–30.71	18.26 / 19.23	18.15–18.36
Occluded	GPU occlusion	30.39 / 37.37	29.16–31.10	18.24 / 19.32	17.86–18.47

cpu_ms spans the synchronous Renderer::render call, including CPU preparation, submission, GPU completion wait, and host image copy; it excludes renderer construction and the ten warmup frames. gpu_ms is the Vulkan timestamp interval around recorded GPU commands, including the image-to-buffer capture copy, but excludes later host mapping and copying. The median isolated host readback_cpu_ms is 0.05–0.06 ms in these runs. Neither column is a Player frame rate, and subtracting them does not isolate CPU culling cost.

Passes, work saved, and memory

The table shows pass-level GPU p50 / p95, in milliseconds. A dash means the pass is absent in that mode. These timestamp spans include work recorded under the named pass and are not interchangeable with the whole-frame GPU interval.

Scene	Mode	MainCull	MainRaster	HZB	PostCull	PostRaster
Frustum-heavy	Direct	—	0.195 / 0.198	—	—	—
Frustum-heavy	GPU frustum	3.917 / 4.624	0.029 / 0.050	—	—	—
Frustum-heavy	GPU occlusion	3.869 / 4.484	0.026 / 0.046	0.033 / 0.040	0.005 / 0.006	0.006 / 0.007
Open	Direct	—	0.455 / 0.464	—	—	—
Open	GPU frustum	17.956 / 19.120	0.046 / 0.084	—	—	—
Open	GPU occlusion	18.230 / 19.036	0.046 / 0.083	0.033 / 0.036	0.005 / 0.005	0.006 / 0.007
Occluded	Direct	—	0.453 / 0.511	—	—	—
Occluded	GPU frustum	18.174 / 19.142	0.047 / 0.085	—	—	—
Occluded	GPU occlusion	18.115 / 19.202	0.018 / 0.019	0.033 / 0.035	0.024 / 0.028	0.006 / 0.007

The visibility counters are stable across the 90 sampled frames of each GPU scene/mode:

Scene	Candidate count	GPU main-visible	Frustum rejected	HZB deferred	Post-visible	Direct / GPU-frustum / GPU-occlusion draw calls
Frustum-heavy	1,024	416	608	0	0	416 / 1 / 2
Open	1,024	1,024	0	0	0	1,024 / 1 / 2
Occluded	1,026	1 in occlusion mode	0	1,025 in occlusion mode	0	1,026 / 1 / 2

The direct path does not produce GPU visibility counters; zeros in its CSV rows mean “not measured,” not zero visible objects. Peak FrameStats::gpu_allocated_bytes was about 7.45 MiB for direct, 5.17 MiB for GPU frustum, and 6.50 MiB for GPU occlusion in this harness. This counter is the renderer's tracked allocations, not whole-process GPU memory.

The frustum fixture removes 608 main-view candidates and cuts main raster time, while the occluded fixture defers 1,025 candidates and cuts its main raster time. Yet the current MainCull cost exceeds those savings by an order of magnitude on this device/configuration. The next performance investigation should isolate compute dispatches, transfer/barrier and indirect-buffer work inside MainCull, then repeat in Release with validation and diagnostic counter readback disabled. The fixed mode order and GPU clock/thermal state may affect absolute values; a randomized or interleaved follow-up would reduce order bias. The existing acceptance tests verify image agreement in memory, but this benchmark does not store reference images or measure visual quality over real game content.