Skip to content

Speed baseline for 0.6.5 (S0): Model M and Model L harness rows - #217

Draft
phantoslayer21 wants to merge 1 commit into
mainfrom
param/speed-baseline-065
Draft

phantoslayer21 wants to merge 1 commit into
mainfrom
param/speed-baseline-065

Conversation

@phantoslayer21

Copy link
Copy Markdown
Owner

Speed baseline for 0.6.5 (release item S0), first part: the Model M and Model L rows. The Model K rows and the Ray Reconstruction (preset F) rows follow in this record from their own measurements; the file says so in its model_k and rr_preset_f sections.

What it adds

  • records/speed-baseline-065.json: Model M and Model L on the native D3D12 harness at 1080p Quality (1280x720 -> 1920x1080), 1440p Quality (1706x960 -> 2560x1440) and 1440p Performance (1280x720 -> 2560x1440).
    • Measured at main 25407f2, on an AMD Radeon RX 9070 XT with driver 32.0.32015.7029.
    • Synthetic models and synthetic frames (clean/l/runner/l_frame_check.py --prepare): no game, no real weights.
    • 5 rounds per shape, Model M and Model L interleaved in each round, every run a fresh process.
    • Per shape and model: the frame span, the per-step and per-kernel split, launches per frame, the kernel selection, and the staged-route-only kernels (below).
  • clean/l/runner/l_d3d12_profile.py:
    • the 1440p Performance shape;
    • the post step now counts Model L's own reconstruction (l_post_*), the direct-route helpers (l_direct_*) and color_divide. This is the step map of the earlier shape record (clean/l/records/l-vs-m-shapes-main.json); before, those kernels fell into "other".

Timers (stated in the record)

  • frame_span_ms: whole-frame D3D12 timestamps with no per-kernel timestamps (wg_rt_frame_timed). Median of 30 frames per run, then the median over the runs.
  • dispatch_sum_ms, steps_ms, kernels_ms: per-kernel D3D12 timestamps (WG_D3D12_PROFILE). Mean over 100 frames per run, then the median over the runs. These timestamps perturb scheduling, so read them as attribution, not isolated costs.

Frame span (ms)

Shape Model M Model L
1080p Quality 1.366 2.276
1440p Quality 2.273 3.833
1440p Performance 1.596 2.783

Staged route vs the game's route

The harness passes buffers, so it runs the staged route. On that route, color_divide (the pre-exposure division of the input colour) and scale_rgb (the multiply of the output) run as their own kernels. On the game's direct route the kernels fold the pre-exposure in, so the in-game frame does not run them.

For Model M they cost 0.106 / 0.209 / 0.170 ms at the three shapes (staged_route_only_ms). The speed items S1 (Model M) and S2 (Model L) are therefore judged on the game-route kernels: dispatch_sum_without_staged_route_only_ms and the per-kernel times. The record states this in speed_targets_basis.

How the build was made

  • The runtime library, the HIP code objects and the synthetic models were built from the head's sources.
  • The 71 D3D12 kernels of clean/dx12/kernel_sets.json (model_m + model_l) were reused from an earlier build of d5b0648. Its kernel sources (clean/dx12/kernels, hlsl, gelu_lut, f16lut) differ from 25407f2 only in frame-generation kernels.
  • Check: three kernels rebuilt at 25407f2 (swin_block_c96_pm_rev_nc, l_post_recon_p12, lev/input_encoder_q20) are byte-identical to the reused DXILs.

Tested

  • 30 harness runs (2 models x 3 shapes x 5 rounds), all completed.
  • public_gates.py: the must-pass gates pass. The LOOK lines are all in files this PR does not touch.
  • Leak scan of the changed files: 0 hits.

Not tested

  • Exactness: no kernel or runtime code changes. The tool change only regroups kernels into steps.
  • Model K and RR rows: to come.
  • In game: not re-measured. The record repeats the existing Cyberpunk 2077 1080p Quality reference (Model M 1.174 ms, Model L 2.250 ms, from clean/l/records/cp2077-model-l-timing-2026-11-14.json).

Refs #89

…407f2

records/speed-baseline-065.json: Model M and Model L on the native D3D12 harness at 1080p Quality,
1440p Quality and 1440p Performance (synthetic models and frames, 5 interleaved rounds, each run in
its own GPU slot): unprofiled frame span, per-step and per-kernel split, and the staged-route-only
kernels (color_divide, scale_rgb) the game's direct route does not run. Model K and RR preset F
rows follow.

l_d3d12_profile.py: the 1440p Performance shape; Model L's own reconstruction, the direct-route
helpers and color_divide counted in the post step (the step map of the earlier shape record).

Signed-off-by: phantoslayer21 <156617979+phantoslayer21@users.noreply.github.com>
@phantoslayer21
phantoslayer21 marked this pull request as ready for review November 14, 2026 23:50
@phantoslayer21
phantoslayer21 marked this pull request as draft November 15, 2026 00:39
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant