dtourolle/DarkRoom
1
0
Fork 0
Code Issues 39 Pull Requests Actions Packages Projects Releases 20 Wiki Activity

Assert the frame budget, and commit the numbers behind the FR-DSP-2 verdict

Browse Source 
FR-DSP-3 states a latency requirement and nothing checked it, which makes it a
wish. This adds the check and the measurements it guards.

`docs/frame-budget.md` is the bench's output with the reading of §2's decision
rule attached. The short version: every point-operation chain at every viewport
size, fit and at 1:1, is inside 16 ms at the 99th percentile — the widest is
4.5 ms of GPU at 4K — so FR-DSP-2 should be rewritten rather than implemented.
The measurement did find a stage that misses the budget, and it is the one §2
predicted: clarity's 52-pixel separable kernel costs 34 ms at 4K. Tiles make
that worse rather than better, since a tiled convolution reads a halo per tile;
the fix `local_contrast` already names for itself is a base computed at reduced
resolution.

The test guards the fused path and says so, at length, rather than quietly
excluding the expensive stage and letting the tag imply otherwise (§7). What it
asserts is exactly the claim the recommendation rests on: one dispatch over a
viewport-sized target, at a full chain, is comfortably inside a frame.

Two things the numbers forced:

- The two cases are one `#[test]`. As two they ran on a thread each, contended
  for the same device, and took the 1:1 case from 2.5 ms to 14.9 ms — a
  measurement of the harness that would have flickered either side of the
  budget forever.

- The CPU half of the frame is judged only in an optimised build. Composition
  is real per-frame work on the UI thread and belongs in the budget, but the
  workspace builds its own crates at `opt-level = 0` in dev and `cargo test` is
  a dev build, so measuring it there measures rustc. The GPU half is asserted
  either way.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
dtourolle
2026-08-27 19:00:45 +02:00
co-authored by Claude Opus 5
parent 7235972ca4
commit 13deaa2fbb
3 changed files with 643 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
core/dr-gpu/tests/frame_budget.rs
+338
View File
@@ -0,0 +1,338 @@
//! The frame budget, asserted rather than hoped for.
//!
//! FR-DSP-3 says a slider updates the visible region within one frame budget at
//! proxy resolution. Until this file existed nothing checked it, which made it
//! a wish — `docs/display-and-extension.md` §3 is blunt about that, and §7 is
//! blunt about what tagging an unchecked requirement does to the coverage
//! figure.
//!
//! The measurements this guards are in [`docs/frame-budget.md`], produced by
//! `examples/frame_budget.rs`. This file is the part of them that has to keep
//! being true: it renders the **whole point-operation chain** through the real
//! `render_detailed` for a hundred frames, moving a slider between each, and
//! fails if the 99th percentile leaves the budget.
//!
//! # What it does not cover, said out loud
//!
//! **The neighbourhood stage is deliberately not in the asserted chain.** It is
//! over the budget today — clarity alone is 34 ms at 4K, because its kernel is
//! a fraction of the frame and reaches a 52-pixel radius there — and
//! `docs/frame-budget.md` records that, names the fix (a base computed at
//! reduced resolution) and does not pretend otherwise. Asserting a budget the
//! code does not meet would produce a red suite that everyone learns to ignore;
//! asserting it on a chain that quietly excluded the expensive stage *without
//! saying so* would be the coverage overstatement §7 warns about. So it is
//! excluded, loudly, here.
//!
//! What is asserted is exactly the claim the FR-DSP-2 recommendation rests on:
//! that **one fused dispatch over a viewport-sized target is comfortably inside
//! the budget**, at a full chain, fit and at 1:1. If that stops being true, the
//! recommendation to strike tiled computation from the interactive path stops
//! being supported, and this test is what says so.
//!
//! # Why the percentile and not the mean
//!
//! A drag is judged by its worst frame. Nearest-rank over 100 frames puts the
//! 99th percentile at the second-worst, which is strict enough to catch a
//! stutter and forgiving enough that one scheduler hiccup from an unrelated
//! process does not decide the verdict.
//!
//! # Why the CPU half is asserted only in an optimised build
//!
//! Composing the shader is per-frame work on the UI thread and belongs in the
//! budget — `DevelopSession::render` calls `compose` on every frame, and on a
//! full chain it is milliseconds of string formatting. But the workspace builds
//! its own crates at `opt-level = 0` in dev (see the root `Cargo.toml`), and
//! `cargo test` is a dev build, so that formatting runs unoptimised here and
//! measures rustc rather than the pipeline. The GPU half is unaffected: a
//! shader is compiled by the driver either way.
//!
//! So the GPU half is always asserted, and the composition is folded in only
//! when `debug_assertions` is off. Running `cargo test --release -p dr-gpu`
//! therefore checks strictly more than the default run does, and the numbers
//! printed on failure say which of the two halves was over.
use dr_gpu::{AdjustPass, DemosaicedImage, GpuContext};
use dr_pipeline::descriptor::ParamKind;
use dr_pipeline::ops::exposure;
use dr_pipeline::{Affects, Attribute, CropRect, EditGraph, OpId, ParamId};
use std::time::Instant;
/// 60 Hz. FR-DSP-3 does not name a number; this is the one every interactive
/// application means by "one frame".
const BUDGET_MS: f64 = 16.0;
/// Measured frames per case. Nearest-rank p99 of 100 is the second-worst.
const FRAMES: usize = 100;
/// Discarded before measurement: the first frame at a size allocates a render
/// target and the first frame of a chain compiles a pipeline. Neither recurs
/// during a drag, so neither belongs in a drag's percentile.
const WARMUP: usize = 12;
/// A 24 MP source — a full-frame camera, and large enough that a 1:1 view of it
/// is a genuine zoom rather than a rounding error.
///
/// Smaller than the bench's 60 MP on purpose. The fused pass costs what the
/// *output* costs, so the source size barely moves these numbers, and 24 MP
/// keeps the fixture inside a second even at `opt-level = 0`.
const SOURCE: (u32, u32) = (6000, 4000);
/// The viewport the budget is asserted at: a 16:10 desktop display.
///
/// Not 4K, and the reason is worth stating. At 4K the fused chain still passes
/// with room to spare (4.5 ms of GPU; see `docs/frame-budget.md`), but a test
/// that renders 8.3 M pixels a hundred times twice over is four seconds of
/// suite time to re-establish a conclusion 4.1 M pixels already establishes.
const VIEWPORT: (u32, u32) = (2560, 1600);
fn ctx() -> Option<GpuContext> {
// CI runners and headless machines may have no usable adapter. Skip rather
// than fail, exactly as the rest of this crate's device tests do.
match pollster::block_on(GpuContext::new_headless()) {
Ok(c) => Some(c),
Err(e) => {
eprintln!("skipping: no GPU adapter ({e})");
None
}
}
}
/// TRACES: FR-DSP-3 | FR-DSP-5
/// A slider drag on the full point-operation chain stays inside one frame —
/// fit, and at 1:1.
///
/// The develop view's ordinary case, at a full chain rather than a flattering
/// one: every operation that contributes a fragment to the fused shader is
/// active, and exposure moves between frames exactly as a drag moves it.
///
/// The 1:1 case is the one FR-DSP-5 names and the one FR-DSP-3's asynchronous
/// clause was written for. It is asserted here because the measurement found
/// that clause unnecessary rather than merely unimplemented: a 1:1 view is
/// *cheaper* than a fit view of the same file, since the dispatch is the same
/// size and the reads are contiguous rather than strided. If that ever inverts,
/// the argument for striking the clause weakens, and this is what would notice.
///
/// # One test and not two, deliberately
///
/// The two cases were two `#[test]` functions until the numbers said otherwise.
/// Cargo runs a binary's tests on a thread each, both of these want the same
/// GPU, and contending for it took the 1:1 case from 2.5 ms to 14.9 ms — a
/// measurement of the test harness that would have flickered either side of the
/// budget forever. A timing assertion has to own the device while it runs, and
/// the only way to say that in a test binary is to be the only test in it.
#[test]
fn a_slider_drag_stays_inside_the_frame_budget() {
let Some(ctx) = ctx() else { return };
let source = synthetic_source(&ctx);
let mut fit = full_point_chain();
drag(&ctx, &source, &mut fit, VIEWPORT).assert_inside_budget("proxy resolution, fit", VIEWPORT);
let mut zoomed = full_point_chain();
zoomed.framing_mut().set_view(one_to_one(VIEWPORT));
drag(&ctx, &source, &mut zoomed, VIEWPORT)
.assert_inside_budget("1:1 on a 24 MP source", VIEWPORT);
}
// ---------------------------------------------------------------------------
// The measurement
// ---------------------------------------------------------------------------
struct Run {
/// Per frame: compose, compose the detail chain, hash the invalidation.
cpu_p99: f64,
/// Per frame: submit and wait for the device to go idle.
gpu_p99: f64,
/// `cpu + gpu` summed within each frame, then ranked. Not the sum of the
/// two percentiles above, which would be a frame that never happened.
total_p99: f64,
}
impl Run {
/// Fail if the budget was missed, saying which half missed it.
///
/// In a dev build only the GPU half is judged — see the module
/// documentation for why — and the CPU figure is still printed, because a
/// reader looking at a failure wants both numbers even when only one of
/// them is the verdict.
fn assert_inside_budget(&self, case: &str, viewport: (u32, u32)) {
let judged = if cfg!(debug_assertions) {
self.gpu_p99
} else {
self.total_p99
};
assert!(
judged <= BUDGET_MS,
"{case} at {}x{}: p99 of {FRAMES} frames was {judged:.2} ms, over the \
{BUDGET_MS:.0} ms budget (cpu {:.2} ms, gpu {:.2} ms, total {:.2} ms). \
FR-DSP-3 is what this violates; docs/frame-budget.md holds the \
numbers it used to be.",
viewport.0,
viewport.1,
self.cpu_p99,
self.gpu_p99,
self.total_p99,
);
}
}
/// Render `FRAMES` frames with the exposure slider moving between each.
///
/// The three calls before the dispatch are the three `DevelopSession::render`
/// makes, in the same order, so this is the develop view's frame rather than an
/// idealisation of it.
fn drag(
ctx: &GpuContext,
source: &DemosaicedImage,
graph: &mut EditGraph,
viewport: (u32, u32),
) -> Run {
// Stands for the `VersionId` the app mixes in. Constant because every frame
// here is the same photograph.
const PHOTOGRAPH: u64 = 0x0dd_ba11;
let mut adjust = AdjustPass::new(ctx);
let src = source.size();
let (w, h) = viewport;
let mut cpu = Vec::with_capacity(FRAMES);
let mut gpu = Vec::with_capacity(FRAMES);
let mut total = Vec::with_capacity(FRAMES);
for i in 0..WARMUP + FRAMES {
// A hundredth of a stop per frame: what a drag does, and what stops
// `render_detailed` reusing the previous frame's colour result and
// turning this into a measurement of nothing.
graph.set_param(exposure::ID, exposure::EXPOSURE, 0.30 + i as f32 * 0.01);
let t0 = Instant::now();
let shader = graph.compose();
let detail = graph.compose_detail(src, viewport);
let colour_key = graph.invalidation().through(Affects::Colour) ^ PHOTOGRAPH;
let cpu_elapsed = t0.elapsed();
let t1 = Instant::now();
adjust
.render_detailed(source, &shader, w, h, None, &detail, colour_key)
.expect("render");
ctx.device
.poll(wgpu::PollType::wait_indefinitely())
.expect("poll");
let gpu_elapsed = t1.elapsed();
if i >= WARMUP {
cpu.push(cpu_elapsed.as_secs_f64() * 1e3);
gpu.push(gpu_elapsed.as_secs_f64() * 1e3);
total.push((cpu_elapsed + gpu_elapsed).as_secs_f64() * 1e3);
}
}
Run {
cpu_p99: p99(cpu),
gpu_p99: p99(gpu),
total_p99: p99(total),
}
}
/// Nearest-rank 99th percentile.
///
/// Nearest-rank because the samples *are* the population: there is no
/// distribution being estimated, only a hundred frames that either fitted in
/// the budget or did not.
fn p99(mut samples: Vec<f64>) -> f64 {
samples.sort_by(f64::total_cmp);
let n = samples.len();
samples[((0.99 * n as f64).ceil() as usize).clamp(1, n) - 1]
}
// ---------------------------------------------------------------------------
// Fixtures
// ---------------------------------------------------------------------------
/// Every operation that contributes a fragment to the fused shader, active.
///
/// Built from [`EditGraph::capabilities`] rather than from a list of operation
/// names, for the same reason the develop panel is: declaring a new node must
/// not silently shrink what this test calls "the full chain". A quarter of the
/// way from each parameter's default towards whichever end is further from it,
/// which is a plausible setting and — the part that matters — is never the
/// neutral, since a neutral operation contributes nothing at all.
///
/// The film stock is not reachable this way (it is a choice of material, not a
/// slider) and is left off. It is one texture lookup and two curve reads; the
/// bench includes it and it is worth about a millisecond at this size.
fn full_point_chain() -> EditGraph {
let mut graph = EditGraph::default_chain();
let moves: Vec<(OpId, ParamId, f32)> = graph
.capabilities()
.iter()
// The neighbourhood operations. See the module documentation for why
// they are not here.
.filter(|op| !op.attributes.contains(&Attribute::Detail))
.flat_map(|op| {
op.params.iter().map(move |p| {
let value = match p.kind {
ParamKind::Scalar { min, max, .. } => {
let far = if (max - p.default).abs() >= (p.default - min).abs() {
max
} else {
min
};
p.default + (far - p.default) * 0.25
}
ParamKind::Bool => 1.0,
ParamKind::Enum { variants } => {
if variants.len() > 1 {
1.0
} else {
0.0
}
}
};
(op.id, p.id, value)
})
})
.collect();
for (op, param, value) in moves {
graph.set_param(op, param, value);
}
graph
}
/// The view rect that puts one render pixel on one source pixel, centred.
fn one_to_one(render: (u32, u32)) -> CropRect {
let w = render.0 as f32 / SOURCE.0 as f32;
let h = render.1 as f32 / SOURCE.1 as f32;
CropRect {
x: (1.0 - w) * 0.5,
y: (1.0 - h) * 0.5,
width: w,
height: h,
}
}
/// A source with structure at every scale.
///
/// Not flat: a flat frame lets the memory system serve every sample of every
/// pixel from one cache line, which flatters a bandwidth-bound pass by an amount
/// that has nothing to do with photographs.
fn synthetic_source(ctx: &GpuContext) -> DemosaicedImage {
let (w, h) = SOURCE;
let mut rgba = vec![0u8; (w as usize) * (h as usize) * 4];
for y in 0..h as usize {
let row = y * (w as usize) * 4;
for x in 0..w as usize {
let n = (x.wrapping_mul(2_654_435_761) ^ y.wrapping_mul(1_640_531_527)) >> 13;
let dither = (n & 0x1f) as u32;
let gx = (x * 200 / w as usize) as u32;
let gy = (y * 55 / h as usize) as u32;
let px = &mut rgba[row + x * 4..row + x * 4 + 4];
px[0] = (30 + gx + dither).min(255) as u8;
px[1] = (40 + gy + dither).min(255) as u8;
px[2] = (60 + gx / 2 + gy + dither).min(255) as u8;
px[3] = 255;
}
}
DemosaicedImage::from_rgba8(ctx, &rgba, w, h).expect("upload")
}