Metal encoder & pass timestamp support

.deny.toml

@@ -27,6 +27,7 @@ allow = [
 [sources]
 allow-git = [
 	"https://github.com/grovesNL/glow",
+	"https://github.com/gfx-rs/metal-rs",
 ]
 unknown-registry = "deny"
 unknown-git = "deny"

CHANGELOG.md

@@ -88,6 +88,10 @@ By @Valaphee in [#3402](https://github.com/gfx-rs/wgpu/pull/3402)
 ### Documentation
 - Use WGSL for VertexFormat example types. By @ScanMountGoat in [#4305](https://github.com/gfx-rs/wgpu/pull/4035)
 
+#### Metal
+
+- Support for timestamp queries on encoders and passes. By @wumpf in [#4008](https://github.com/gfx-rs/wgpu/pull/4008)
+
 ### Bug Fixes
 
 #### General

Cargo.toml

@@ -158,6 +158,8 @@ termcolor = "1.2.0"
 #glow = { path = "../glow" }
 #d3d12 = { path = "../d3d12-rs" }
 #metal = { path = "../metal-rs" }
+#metal = { path = "../metal-rs" }
+metal = { git = "https://github.com/gfx-rs/metal-rs/", rev = "d24f1a4" } # More timer support via https://github.com/gfx-rs/metal-rs/pull/280
 #web-sys = { path = "../wasm-bindgen/crates/web-sys" }
 #js-sys = { path = "../wasm-bindgen/crates/js-sys" }
 #wasm-bindgen = { path = "../wasm-bindgen" }

examples/timestamp-queries/src/main.rs

@@ -47,6 +47,7 @@ impl QueryResults {
     // * compute end
     const NUM_QUERIES: u64 = 8;
 
+    #[allow(clippy::redundant_closure)] // False positive
     fn from_raw_results(timestamps: Vec<u64>, timestamps_inside_passes: bool) -> Self {
         assert_eq!(timestamps.len(), Self::NUM_QUERIES as usize);
 
@@ -60,9 +61,9 @@ impl QueryResults {
         let mut encoder_timestamps = [0, 0];
         encoder_timestamps[0] = get_next_slot();
         let render_start_end_timestamps = [get_next_slot(), get_next_slot()];
-        let render_inside_timestamp = timestamps_inside_passes.then_some(get_next_slot());
+        let render_inside_timestamp = timestamps_inside_passes.then(|| get_next_slot());
         let compute_start_end_timestamps = [get_next_slot(), get_next_slot()];
-        let compute_inside_timestamp = timestamps_inside_passes.then_some(get_next_slot());
+        let compute_inside_timestamp = timestamps_inside_passes.then(|| get_next_slot());
         encoder_timestamps[1] = get_next_slot();
 
         QueryResults {
@@ -79,8 +80,8 @@ impl QueryResults {
         let elapsed_us = |start, end: u64| end.wrapping_sub(start) as f64 * period as f64 / 1000.0;
 
         println!(
-            "Elapsed time render + compute: {:.2} μs",
-            elapsed_us(self.encoder_timestamps[0], self.encoder_timestamps[1])
+            "Elapsed time before render until after compute: {:.2} μs",
+            elapsed_us(self.encoder_timestamps[0], self.encoder_timestamps[1]),
         );
         println!(
             "Elapsed time render pass: {:.2} μs",
@@ -464,13 +465,10 @@ mod tests {
             render_start_end_timestamps[1].wrapping_sub(render_start_end_timestamps[0]);
         let compute_delta =
             compute_start_end_timestamps[1].wrapping_sub(compute_start_end_timestamps[0]);
+        let encoder_delta = encoder_timestamps[1].wrapping_sub(encoder_timestamps[0]);
 
-        // TODO: Metal encoder timestamps aren't implemented yet.
-        if ctx.adapter.get_info().backend != wgpu::Backend::Metal {
-            let encoder_delta = encoder_timestamps[1].wrapping_sub(encoder_timestamps[0]);
-            assert!(encoder_delta > 0);
-            assert!(encoder_delta >= render_delta + compute_delta);
-        }
+        assert!(encoder_delta > 0);
+        assert!(encoder_delta >= render_delta + compute_delta);
 
         if let Some(render_inside_timestamp) = render_inside_timestamp {
             assert!(render_inside_timestamp >= render_start_end_timestamps[0]);

wgpu-hal/src/metal/adapter.rs

@@ -5,6 +5,8 @@ use wgt::{AstcBlock, AstcChannel};
 
 use std::{sync::Arc, thread};
 
+use super::TimestampQuerySupport;
+
 const MAX_COMMAND_BUFFERS: u64 = 2048;
 
 unsafe impl Send for super::Adapter {}
@@ -536,6 +538,26 @@ impl super::PrivateCapabilities {
             MTLReadWriteTextureTier::TierNone
         };
 
+        let mut timestamp_query_support = TimestampQuerySupport::empty();
+        if version.at_least((11, 0), (14, 0), os_is_mac)
+            && device.supports_counter_sampling(metal::MTLCounterSamplingPoint::AtStageBoundary)
+        {
+            // If we don't support at stage boundary, don't support anything else.
+            timestamp_query_support.insert(TimestampQuerySupport::STAGE_BOUNDARIES);
+
+            if device.supports_counter_sampling(metal::MTLCounterSamplingPoint::AtDrawBoundary) {
+                timestamp_query_support.insert(TimestampQuerySupport::ON_RENDER_ENCODER);
+            }
+            if device.supports_counter_sampling(metal::MTLCounterSamplingPoint::AtDispatchBoundary)
+            {
+                timestamp_query_support.insert(TimestampQuerySupport::ON_COMPUTE_ENCODER);
+            }
+            if device.supports_counter_sampling(metal::MTLCounterSamplingPoint::AtBlitBoundary) {
+                timestamp_query_support.insert(TimestampQuerySupport::ON_BLIT_ENCODER);
+            }
+            // `TimestampQuerySupport::INSIDE_WGPU_PASSES` emerges from the other flags.
+        }
+
         Self {
             family_check,
             msl_version: if os_is_xr || version.at_least((12, 0), (15, 0), os_is_mac) {
@@ -773,13 +795,7 @@ impl super::PrivateCapabilities {
             } else {
                 None
             },
-            support_timestamp_query: version.at_least((11, 0), (14, 0), os_is_mac)
-                && device
-                    .supports_counter_sampling(metal::MTLCounterSamplingPoint::AtStageBoundary),
-            support_timestamp_query_in_passes: version.at_least((11, 0), (14, 0), os_is_mac)
-                && device.supports_counter_sampling(metal::MTLCounterSamplingPoint::AtDrawBoundary)
-                && device
-                    .supports_counter_sampling(metal::MTLCounterSamplingPoint::AtDispatchBoundary),
+            timestamp_query_support,
         }
     }
 
@@ -807,12 +823,16 @@ impl super::PrivateCapabilities {
             | F::DEPTH32FLOAT_STENCIL8
             | F::MULTI_DRAW_INDIRECT;
 
-        features.set(F::TIMESTAMP_QUERY, self.support_timestamp_query);
-        // TODO: Not yet implemented.
-        // features.set(
-        //     F::TIMESTAMP_QUERY_INSIDE_PASSES,
-        //     self.support_timestamp_query_in_passes,
-        // );
+        features.set(
+            F::TIMESTAMP_QUERY,
+            self.timestamp_query_support
+                .contains(TimestampQuerySupport::STAGE_BOUNDARIES),
+        );
+        features.set(
+            F::TIMESTAMP_QUERY_INSIDE_PASSES,
+            self.timestamp_query_support
+                .contains(TimestampQuerySupport::INSIDE_WGPU_PASSES),
+        );
         features.set(F::TEXTURE_COMPRESSION_ASTC, self.format_astc);
         features.set(F::TEXTURE_COMPRESSION_ASTC_HDR, self.format_astc_hdr);
         features.set(F::TEXTURE_COMPRESSION_BC, self.format_bc);