Renderer: Metal's flat draws and occlusion culling switch on by sections in view
Both switched on past 4,096 section draws, measured when a section had 2.2 to 2.5 draws. A section's draw count now depends on the sides it shows the eye, so the bars count sections in view: 1,600, where 4,096 draws fell. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
@@ -2159,7 +2159,7 @@ impl Renderer {
|
||||
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.min_sections = cull::TILED_MIN_SECTIONS;
|
||||
}
|
||||
cull
|
||||
});
|
||||
@@ -3131,7 +3131,7 @@ impl Renderer {
|
||||
.cull
|
||||
.as_mut()
|
||||
.filter(|c| {
|
||||
self.cull_on && !frame.title && self.multi_draw && self.section_draws >= c.min_draws
|
||||
self.cull_on && !frame.title && self.multi_draw && visible.len() >= c.min_sections
|
||||
})
|
||||
.and_then(|cull| cull.begin(&self.device).then_some(cull));
|
||||
self.phase2_draws.clear();
|
||||
@@ -3533,7 +3533,9 @@ impl Renderer {
|
||||
}
|
||||
let flat_frame = match &mut self.flat {
|
||||
Some(flat)
|
||||
if self.flat_on && !self.draws.is_empty() && self.draws.len() >= flat.min_draws =>
|
||||
if self.flat_on
|
||||
&& !self.draws.is_empty()
|
||||
&& self.drawn_sections >= flat.min_sections =>
|
||||
{
|
||||
let sections = &self.sections;
|
||||
let lods = &self.lods;
|
||||
|
||||
@@ -46,13 +46,15 @@ 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;
|
||||
/// On Metal, sections in view 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 (over 4,096 draws) it saved 0.25–1.6 ms. Those draws were
|
||||
/// 2.2 to 2.5 per section; a section's draw count now depends on the sides
|
||||
/// it shows the eye (`mesher::Sides`), so the bar counts sections.
|
||||
pub const TILED_MIN_SECTIONS: usize = 1600;
|
||||
/// 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.
|
||||
@@ -148,8 +150,8 @@ 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,
|
||||
/// Frames with fewer sections in view than this do not cull.
|
||||
pub min_sections: usize,
|
||||
}
|
||||
|
||||
fn buffer(
|
||||
@@ -348,7 +350,7 @@ impl Cull {
|
||||
stamp: 0,
|
||||
hidden: ([0; 2], 1.0),
|
||||
always: false,
|
||||
min_draws: 0,
|
||||
min_sections: 0,
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@@ -96,17 +96,19 @@ pub struct Flat {
|
||||
pub calls: Vec<Call>,
|
||||
pub phase1_calls: usize,
|
||||
pub solid_calls: usize,
|
||||
/// Frames with fewer section draws keep the per-section draws.
|
||||
pub min_draws: usize,
|
||||
/// Frames with fewer sections in view keep the per-section draws.
|
||||
pub min_sections: usize,
|
||||
range_align: u64,
|
||||
}
|
||||
|
||||
/// Section draws from which a frame is drawn flat. Below it the index list
|
||||
/// costs the GPU more than the draw calls cost the CPU: on an M4 Pro at
|
||||
/// render distance 16 (2,900 draws) the copy and the list add about 0.4 ms
|
||||
/// of GPU time for 0.5–0.8 ms less CPU time; at 32 (9,300 draws) the CPU
|
||||
/// saves 1.3 ms.
|
||||
const MIN_DRAWS: usize = 4096;
|
||||
/// Sections in view (and distant tiles) from which a frame is drawn flat.
|
||||
/// Below it the index list costs the GPU more than the draw calls cost the
|
||||
/// CPU: on an M4 Pro at render distance 16 (2,900 draws) the copy and the
|
||||
/// list add about 0.4 ms of GPU time for 0.5–0.8 ms less CPU time; at 32
|
||||
/// (9,300 draws) the CPU saves 1.3 ms. The bar was 4,096 draws when they
|
||||
/// were 2.2 to 2.5 per section; a section's draw count now depends on the
|
||||
/// sides it shows the eye (`mesher::Sides`), so it counts sections.
|
||||
const MIN_SECTIONS: usize = 1600;
|
||||
|
||||
/// Whether flat draws apply: the backend emulates multi-draw, and vertex
|
||||
/// shaders can read storage buffers large enough for an index page.
|
||||
@@ -337,10 +339,10 @@ impl Flat {
|
||||
calls: Vec::new(),
|
||||
phase1_calls: 0,
|
||||
solid_calls: 0,
|
||||
min_draws: if std::env::var("HORIZON_FLAT_DRAWS").as_deref() == Ok("1") {
|
||||
min_sections: if std::env::var("HORIZON_FLAT_DRAWS").as_deref() == Ok("1") {
|
||||
0
|
||||
} else {
|
||||
MIN_DRAWS
|
||||
MIN_SECTIONS
|
||||
},
|
||||
range_align: (u64::from(device.limits().min_storage_buffer_offset_alignment) / 16)
|
||||
.max(1),
|
||||
|
||||
Reference in New Issue
Block a user