perf(native): use simd to speedup audio buffer mix (#11717)

Run `cargo bench -p affine_media_capture`

```
test result: ok. 0 passed; 0 failed; 6 ignored; 0 measured; 0 filtered out; finished in 0.00s

     Running benches/mix_audio_samples.rs (target/release/deps/mix_audio_samples-ffbc55dcf90d3468)
audio mix/simd          time:   [98.380 ns 99.339 ns 100.57 ns]
                        change: [−19.199% −16.928% −14.569%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 12 outliers among 100 measurements (12.00%)
  6 (6.00%) high mild
  6 (6.00%) high severe
audio mix/scalar        time:   [123.99 ns 126.11 ns 128.71 ns]
                        change: [+0.2703% +1.2739% +2.5727%] (p = 0.02 < 0.05)
                        Change within noise threshold.
Found 11 outliers among 100 measurements (11.00%)
  4 (4.00%) high mild
  7 (7.00%) high severe
```

.cargo/config.toml

@@ -5,7 +5,7 @@ rustflags = ["-C", "target-feature=+crt-static"]
 [target.'cfg(target_os = "linux")']
 rustflags = ["-C", "link-args=-Wl,--warn-unresolved-symbols"]
 [target.'cfg(target_os = "macos")']
-rustflags = ["-C", "link-args=-all_load", "-C", "link-args=-weak_framework ScreenCaptureKit"]
+rustflags = ["-C", "link-args=-Wl,-undefined,dynamic_lookup,-no_fixup_chains", "-C", "link-args=-all_load", "-C", "link-args=-weak_framework ScreenCaptureKit"]
 # https://sourceware.org/bugzilla/show_bug.cgi?id=21032
 # https://sourceware.org/bugzilla/show_bug.cgi?id=21031
 # https://github.com/rust-lang/rust/issues/134820

Cargo.lock

@@ -79,6 +79,7 @@ dependencies = [
  "block2",
  "core-foundation",
  "coreaudio-rs",
+ "criterion2",
  "dispatch2",
  "libc",
  "napi",

packages/frontend/native/media_capture/Cargo.toml

@@ -6,6 +6,10 @@ version = "0.0.0"
 [lib]
 crate-type = ["cdylib", "rlib"]
 
+[[bench]]
+harness = false
+name    = "mix_audio_samples"
+
 [dependencies]
 napi            = { workspace = true, features = ["napi4"] }
 napi-derive     = { workspace = true, features = ["type-def"] }
@@ -24,5 +28,8 @@ objc2-foundation = { workspace = true }
 screencapturekit = { workspace = true }
 uuid             = { workspace = true, features = ["v4"] }
 
+[dev-dependencies]
+criterion2 = { workspace = true }
+
 [build-dependencies]
 napi-build = { workspace = true }

packages/frontend/native/media_capture/benches/mix_audio_samples.rs

@@ -0,0 +1,55 @@
+#[cfg(target_os = "macos")]
+use std::hint::black_box;
+
+#[cfg(target_os = "macos")]
+use affine_media_capture::macos::audio_buffer::{mix_audio_samples, mix_audio_samples_scalar};
+#[cfg(target_os = "macos")]
+use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion};
+
+#[cfg(target_os = "macos")]
+fn generate_test_samples() -> [f32; 1024] {
+  let mut samples = [0.0; 1024];
+
+  // Generate a simple sine wave with some variation
+  for (i, sample) in samples.iter_mut().enumerate() {
+    let t = i as f32 / 1024.0;
+    // Create a complex waveform with multiple frequencies
+    let value = 0.008 * (2.0 * std::f32::consts::PI * t * 5.0).sin()
+      + 0.004 * (2.0 * std::f32::consts::PI * t * 10.0).cos()
+      + 0.002 * (2.0 * std::f32::consts::PI * t * 20.0).sin();
+    *sample = value;
+  }
+
+  samples
+}
+
+#[cfg(target_os = "macos")]
+fn bench_audio_mix(c: &mut Criterion) {
+  let mut group = c.benchmark_group("audio mix");
+
+  let input = generate_test_samples();
+  let output = generate_test_samples();
+
+  group.bench_function(BenchmarkId::from_parameter("simd"), |b| {
+    b.iter(|| {
+      let mixed = mix_audio_samples(&input, &output);
+      black_box(mixed);
+    });
+  });
+
+  group.bench_function(BenchmarkId::from_parameter("scalar"), |b| {
+    b.iter(|| {
+      let mut mixed = vec![0.0; 1024];
+      mix_audio_samples_scalar(&input, &output, &mut mixed, 0, input.len());
+      black_box(mixed);
+    });
+  });
+}
+
+#[cfg(target_os = "macos")]
+criterion_group!(benches, bench_audio_mix);
+#[cfg(target_os = "macos")]
+criterion_main!(benches);
+
+#[cfg(not(target_os = "macos"))]
+fn main() {}

packages/frontend/native/media_capture/src/macos/audio_buffer.rs

@@ -5,6 +5,141 @@ use objc2::{Encode, Encoding, RefEncode};
 
 use crate::{error::CoreAudioError, utils::process_audio_frame};
 
+pub const AUDIO_MIX_OUTPUT_WEIGHT: f32 = 0.75;
+
+/// Mix audio samples using scalar operations (no SIMD)
+///
+/// # Arguments
+/// * `input_samples` - Samples from the input stream
+/// * `output_samples` - Samples from the output stream
+/// * `mixed_samples` - Buffer to store the result (must be pre-allocated)
+/// * `start_index` - Starting index in the buffers
+/// * `end_index` - Ending index in the buffers (exclusive)
+pub fn mix_audio_samples_scalar(
+  input_samples: &[f32],
+  output_samples: &[f32],
+  mixed_samples: &mut [f32],
+  start_index: usize,
+  end_index: usize,
+) {
+  // performance downgraded 4x if apply suggestion from this lint rule
+  #[allow(clippy::needless_range_loop)]
+  for sample_index in start_index..end_index {
+    let sample_in = input_samples.get(sample_index).unwrap_or(&0.0);
+    let sample_out = output_samples.get(sample_index).unwrap_or(&0.0);
+    mixed_samples[sample_index] = sample_in + sample_out * AUDIO_MIX_OUTPUT_WEIGHT;
+  }
+}
+
+/// Mix audio samples from input and output streams with specified weights
+/// Uses NEON SIMD acceleration on supported platforms
+///
+/// # Arguments
+/// * `input_samples` - Samples from the input stream
+/// * `output_samples` - Samples from the output stream
+///
+/// # Returns
+/// A vector of mixed audio samples
+pub fn mix_audio_samples(input_samples: &[f32], output_samples: &[f32]) -> Vec<f32> {
+  let mixed_samples_length = input_samples.len();
+  let mut mixed_samples = vec![0.0; mixed_samples_length];
+
+  // For very small arrays, use scalar implementation
+  if mixed_samples_length < 16 {
+    mix_audio_samples_scalar(
+      input_samples,
+      output_samples,
+      &mut mixed_samples,
+      0,
+      mixed_samples_length,
+    );
+    return mixed_samples;
+  }
+
+  #[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
+  unsafe {
+    use std::arch::aarch64::{vdupq_n_f32, vld1q_f32, vmlaq_f32, vst1q_f32};
+
+    let output_weight_vec = vdupq_n_f32(AUDIO_MIX_OUTPUT_WEIGHT);
+    // Process the common length where both arrays have data
+    let common_length = input_samples.len();
+
+    // Main SIMD loop - process each block of 4 samples
+    let input_ptr = input_samples.as_ptr();
+    let output_ptr = output_samples.as_ptr();
+    let result_ptr = mixed_samples.as_mut_ptr();
+
+    let mut offset: usize = 0;
+    let mut remaining_offset: Option<usize> = None;
+
+    // Process 16 samples at a time (4 SIMD vectors)
+    while offset < common_length {
+      // Load 4 vectors of 4 floats each
+      let in_vec1 = vld1q_f32(input_ptr.add(offset));
+      let out_vec1 = vld1q_f32(output_ptr.add(offset));
+      let in_vec2 = vld1q_f32(input_ptr.add(offset + 4));
+      let out_vec2 = vld1q_f32(output_ptr.add(offset + 4));
+      let in_vec3 = vld1q_f32(input_ptr.add(offset + 8));
+      let out_vec3 = vld1q_f32(output_ptr.add(offset + 8));
+      let in_vec4 = vld1q_f32(input_ptr.add(offset + 12));
+      let out_vec4 = vld1q_f32(output_ptr.add(offset + 12));
+
+      // Using fused multiply-add: (a * b) + c in one operation
+      // First multiply input by weight
+      let result1 = vmlaq_f32(in_vec1, out_vec1, output_weight_vec);
+      let result2 = vmlaq_f32(in_vec2, out_vec2, output_weight_vec);
+      let result3 = vmlaq_f32(in_vec3, out_vec3, output_weight_vec);
+
+      let result4 = vmlaq_f32(in_vec4, out_vec4, output_weight_vec);
+
+      // Store results
+      vst1q_f32(result_ptr.add(offset), result1);
+      vst1q_f32(result_ptr.add(offset + 4), result2);
+      vst1q_f32(result_ptr.add(offset + 8), result3);
+      vst1q_f32(result_ptr.add(offset + 12), result4);
+      offset += 16;
+      // accept clippy lint suggestion would downgrade the performance by 15%
+      #[allow(clippy::comparison_chain)]
+      // fast path for aligned length
+      if offset == common_length {
+        break;
+      } else if offset > common_length {
+        remaining_offset = Some(offset - 16);
+      } else {
+        let remaining = common_length - offset;
+        if remaining < 16 {
+          remaining_offset = Some(offset);
+          break;
+        }
+      }
+    }
+
+    if let Some(remaining_offset) = remaining_offset {
+      mix_audio_samples_scalar(
+        input_samples,
+        output_samples,
+        &mut mixed_samples,
+        remaining_offset,
+        common_length,
+      );
+    }
+  }
+
+  #[cfg(not(any(target_arch = "aarch64", target_arch = "arm")))]
+  {
+    // Fallback for non-ARM architectures
+    mix_audio_samples_scalar(
+      input_samples,
+      output_samples,
+      &mut mixed_samples,
+      0,
+      mixed_samples_length,
+    );
+  }
+
+  mixed_samples
+}
+
 /// [Apple's documentation](https://developer.apple.com/documentation/coreaudiotypes/audiobuffer?language=objc)
 #[repr(C)]
 #[derive(Clone, Copy, Debug, PartialEq)]
@@ -50,6 +185,9 @@ unsafe impl RefEncode for AudioBufferList {
 pub struct InputAndOutputAudioBufferList(pub AudioBufferList);
 
 impl InputAndOutputAudioBufferList {
+  /// # Safety
+  ///
+  /// The caller must ensure that the input data is a valid AudioBufferList
   pub unsafe fn from_raw(in_input_data: *mut c_void) -> Result<Self, i32> {
     let buffer_list: AudioBufferList = unsafe { *in_input_data.cast() };
     if buffer_list.mNumberBuffers != 2 {
@@ -93,18 +231,157 @@ impl InputAndOutputAudioBufferList {
       return Err(CoreAudioError::ProcessAudioFrameFailed("output"));
     };
 
-    let mixed_samples_length = processed_samples_input
+    // Use the extracted mixing function with the const weights
-      .len()
+    let mixed_samples = mix_audio_samples(&processed_samples_input, &processed_samples_output);
-      .max(processed_samples_output.len());
-
-    let mut mixed_samples = vec![0.0; mixed_samples_length];
-    for (sample_index, mixed_sample) in mixed_samples.iter_mut().enumerate() {
-      let sample_in = processed_samples_input.get(sample_index).unwrap_or(&0.0);
-      let sample_out = processed_samples_output.get(sample_index).unwrap_or(&0.0);
-
-      *mixed_sample = (sample_in * 2.0 + sample_out * 1.5) / 2.0;
-    }
 
     Ok(mixed_samples)
   }
 }
+
+#[cfg(test)]
+mod tests {
+  use super::*;
+
+  #[test]
+  fn test_mix_audio_samples_empty() {
+    let input: Vec<f32> = vec![];
+    let output: Vec<f32> = vec![];
+    let mixed = mix_audio_samples(&input, &output);
+    assert_eq!(mixed.len(), 0);
+  }
+
+  #[test]
+  fn test_mix_audio_samples_equal_length() {
+    let input = vec![0.1, 0.2, 0.3, 0.4, 0.5];
+    let output = vec![0.5, 0.4, 0.3, 0.2, 0.1];
+    let mixed = mix_audio_samples(&input, &output);
+
+    assert_eq!(mixed.len(), 5);
+
+    // Verify calculations: (input + output * 0.75)
+    let expected = [
+      (0.1 + 0.5 * AUDIO_MIX_OUTPUT_WEIGHT),
+      (0.2 + 0.4 * AUDIO_MIX_OUTPUT_WEIGHT),
+      (0.3 + 0.3 * AUDIO_MIX_OUTPUT_WEIGHT),
+      (0.4 + 0.2 * AUDIO_MIX_OUTPUT_WEIGHT),
+      (0.5 + 0.1 * AUDIO_MIX_OUTPUT_WEIGHT),
+    ];
+
+    for i in 0..mixed.len() {
+      assert!(
+        (mixed[i] - expected[i]).abs() < 1e-6,
+        "Mismatch at index {}: expected {}, got {}",
+        i,
+        expected[i],
+        mixed[i]
+      );
+    }
+  }
+
+  #[test]
+  fn test_mix_audio_samples_input_longer() {
+    let input = vec![0.1, 0.2, 0.3, 0.4, 0.5];
+    let output = vec![0.5, 0.4, 0.3];
+    let mixed = mix_audio_samples(&input, &output);
+
+    assert_eq!(mixed.len(), 5);
+
+    // Verify calculations
+    let expected = [
+      (0.1 + 0.5 * AUDIO_MIX_OUTPUT_WEIGHT),
+      (0.2 + 0.4 * AUDIO_MIX_OUTPUT_WEIGHT),
+      (0.3 + 0.3 * AUDIO_MIX_OUTPUT_WEIGHT),
+      (0.4 + 0.0 * AUDIO_MIX_OUTPUT_WEIGHT),
+      (0.5 + 0.0 * AUDIO_MIX_OUTPUT_WEIGHT),
+    ];
+
+    for i in 0..mixed.len() {
+      assert!(
+        (mixed[i] - expected[i]).abs() < 1e-6,
+        "Mismatch at index {}: expected {}, got {}",
+        i,
+        expected[i],
+        mixed[i]
+      );
+    }
+  }
+
+  #[test]
+  fn test_mix_audio_samples_custom_weights() {
+    // Note: We're using the constant weights so we can't really test custom values
+    // directly
+    let input = vec![0.1, 0.2, 0.3];
+    let output = vec![0.5, 0.4, 0.3];
+    let mixed = mix_audio_samples(&input, &output);
+
+    // Calculate expected values based on the constants
+    let expected = [
+      (0.1 + 0.5 * AUDIO_MIX_OUTPUT_WEIGHT),
+      (0.2 + 0.4 * AUDIO_MIX_OUTPUT_WEIGHT),
+      (0.3 + 0.3 * AUDIO_MIX_OUTPUT_WEIGHT),
+    ];
+
+    for i in 0..mixed.len() {
+      assert!(
+        (mixed[i] - expected[i]).abs() < 1e-6,
+        "Mismatch at index {}: expected {}, got {}",
+        i,
+        expected[i],
+        mixed[i]
+      );
+    }
+  }
+
+  #[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
+  #[test]
+  fn test_simd_implementation_used() {
+    const BUFFER_SIZES: [usize; 4] = [100, 127, 128, 512];
+    for size in BUFFER_SIZES {
+      // Create arrays large enough to trigger SIMD path
+      let input: Vec<f32> = (0..size).map(|i| i as f32 * 0.01).collect();
+      let output: Vec<f32> = (0..size).map(|i| (size - i) as f32 * 0.01).collect();
+
+      // Mix with standard weights
+      let mixed = mix_audio_samples(&input, &output);
+
+      // Compute the same mix using scalar implementation for comparison
+      let mut expected = vec![0.0; input.len()];
+      mix_audio_samples_scalar(&input, &output, &mut expected, 0, input.len());
+
+      // Verify results match between SIMD and scalar implementations
+      for i in 0..mixed.len() {
+        assert!(
+          (mixed[i] - expected[i]).abs() < 1e-6,
+          "SIMD and scalar implementations should produce identical results at index {}",
+          i
+        );
+      }
+    }
+  }
+
+  #[test]
+  fn test_small_vector_uses_scalar() {
+    // Create small arrays that should use scalar path even with SIMD available
+    let input = vec![0.1, 0.2, 0.3];
+    let output = vec![0.5, 0.4, 0.3];
+
+    // Mix with standard weights
+    let mixed = mix_audio_samples(&input, &output);
+
+    // Calculate expected values manually
+    let expected = [
+      (0.1 + 0.5 * AUDIO_MIX_OUTPUT_WEIGHT),
+      (0.2 + 0.4 * AUDIO_MIX_OUTPUT_WEIGHT),
+      (0.3 + 0.3 * AUDIO_MIX_OUTPUT_WEIGHT),
+    ];
+
+    // Verify results
+    for i in 0..mixed.len() {
+      assert!(
+        (mixed[i] - expected[i]).abs() < 1e-6,
+        "Small vector mixing should be correct at index {}",
+        i
+      );
+    }
+  }
+}

packages/frontend/native/media_capture/src/macos/mod.rs

@@ -1,4 +1,4 @@
-pub(crate) mod audio_buffer;
+pub mod audio_buffer;
 pub mod audio_stream_basic_desc;
 pub mod av_audio_file;
 pub mod av_audio_format;