Add a basic metal example with capture (#2324)

* Add some tracing.

* Get the trace to work.

candle-core/Cargo.toml

@@ -48,3 +48,7 @@ metal = ["dep:metal", "dep:candle-metal-kernels"]
 [[bench]]
 name = "bench_main"
 harness = false
+
+[[example]]
+name = "metal_basics"
+required-features = ["metal"]

candle-core/examples/metal_basics.rs

@@ -0,0 +1,28 @@
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+use anyhow::Result;
+use candle_core::{Device, Tensor};
+
+fn main() -> Result<()> {
+    // This requires the code to be run with MTL_CAPTURE_ENABLED=1
+    let device = Device::new_metal(0)?;
+    let metal_device = match &device {
+        Device::Metal(m) => m,
+        _ => anyhow::bail!("unexpected device"),
+    };
+    metal_device.capture("/tmp/candle.gputrace")?;
+    // This first synchronize ensures that a new command buffer gets created after setting up the
+    // capture scope.
+    device.synchronize()?;
+    let x = Tensor::randn(0f32, 1.0, (128, 128), &device)?;
+    let x1 = x.add(&x)?;
+    println!("{x1:?}");
+    // This second synchronize ensures that the command buffer gets commited before the end of the
+    // capture scope.
+    device.synchronize()?;
+    Ok(())
+}

candle-core/src/metal_backend/device.rs

@@ -273,7 +273,13 @@ impl MetalDevice {
         let descriptor = metal::CaptureDescriptor::new();
         descriptor.set_destination(metal::MTLCaptureDestination::GpuTraceDocument);
         descriptor.set_capture_device(self);
+        // The [set_output_url] call requires an absolute path so we convert it if needed.
+        if path.as_ref().is_absolute() {
             descriptor.set_output_url(path);
+        } else {
+            let path = std::env::current_dir()?.join(path);
+            descriptor.set_output_url(path);
+        }
 
         capture
             .start_capture(&descriptor)