Metal: cull only frames with 4,096 section draws or more

On Apple's GPUs the passes the test adds cost a fixed share of the frame
(it is stored and reloaded at each), which the hidden sections repay only
at high render distances: at 12 culling cost an M4 Pro 0.05 ms of GPU time
a frame, at 32 it saved 0.25 to 1.6 ms. Other backends keep culling every
frame; HORIZON_GPU_CULL=all culls every frame on Metal too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Ludovic Marie
2026-10-23 13:22:10 +02:00
parent 14be6ca394
commit 660407fa58
2 changed files with 20 additions and 4 deletions
+10 -4
View File
@@ -2146,8 +2146,8 @@ impl Renderer {
.features()
.contains(wgpu::Features::INDIRECT_FIRST_INSTANCE);
// HORIZON_GPU_CULL=0 turns occlusion culling off, `all` culls every
// frame however little it hides, `shader` tests in the shader as on
// Metal (A/B checks).
// frame however few its draws and little it hides, `shader` tests in
// the shader as on Metal (A/B checks).
let cull_env = std::env::var("HORIZON_GPU_CULL");
let cull = cull::Cull::supported(&adapter, &device)
.filter(|_| cull_env.as_deref() != Ok("0"))
@@ -2159,6 +2159,9 @@ impl Renderer {
let mut cull =
cull::Cull::new(&device, config.format.remove_srgb_suffix(), depth_test);
cull.always = cull_env.as_deref() == Ok("all");
if !cull.always && adapter.get_info().backend == wgpu::Backend::Metal {
cull.min_draws = cull::TILED_MIN_DRAWS;
}
cull
});
let depth = Self::depth(&device, &config, cull.as_ref());
@@ -3148,11 +3151,14 @@ impl Renderer {
.inverse()
.transform_point3(Vec3::ZERO);
// Occlusion culling leaves the sections its last result did not see
// to phase 2, after the others, with instance counts of 0.
// to phase 2, after the others, with instance counts of 0. Whether
// the frame has draws enough goes by the last one's.
let mut cull = self
.cull
.as_mut()
.filter(|_| self.cull_on && !frame.title && self.multi_draw)
.filter(|c| {
self.cull_on && !frame.title && self.multi_draw && self.section_draws >= c.min_draws
})
.and_then(|cull| cull.begin(&self.device).then_some(cull));
self.phase2_draws.clear();
self.phase2_pages.clear();
+10
View File
@@ -46,6 +46,13 @@ const RETEST: u32 = 8;
/// more than they save. Else one frame in `PROBE` tests every section.
const MIN_SHARE: f32 = 0.1;
const PROBE: u32 = 64;
/// On Metal, section draws from which a frame culls. Apple's GPUs store and
/// reload the whole frame at each of the passes the test adds, a fixed cost
/// that the sections it hides repay only at high render distances: on an M4
/// Pro at 1080p and render distance 12 (1,000–1,600 draws) culling added
/// about 0.05 ms of GPU time while hiding 15–30% of the solid indices; at
/// 32 (drawn flat, over this many draws) it saved 0.25–1.6 ms.
pub const TILED_MIN_DRAWS: usize = 4096;
/// Boxes over more than this share of the screen count as seen untested.
const BIG: f32 = 0.25;
/// Results read back at once, some frames late so the CPU never waits.
@@ -141,6 +148,8 @@ pub struct Cull {
pub hidden: ([usize; 2], f32),
/// Cull every frame, whatever the hidden share (A/B checks).
pub always: bool,
/// Frames with fewer section draws than this do not cull.
pub min_draws: usize,
}
fn buffer(
@@ -337,6 +346,7 @@ impl Cull {
stamp: 0,
hidden: ([0; 2], 1.0),
always: false,
min_draws: 0,
}
}