gpui: Take advantage of unified memory on Apple silicon (#45577)

Reapplies #44273

I included metal-rs upgrade so we can get this working on Intel-based
Macs

cc: @JosephTLyons @notpeter @rtfeldman @Anthony-Eid 

Release Notes:

- Reduced memory usage on Apple-silicon Macs by using shared memory
where appropriate

---------

Signed-off-by: Marco Mihai Condrache <52580954+marcocondrache@users.noreply.github.com>

Cargo.lock

@@ -3880,9 +3880,9 @@ dependencies = [
 
 [[package]]
 name = "core-graphics2"
-version = "0.4.1"
+version = "0.5.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7e4583956b9806b69f73fcb23aee05eb3620efc282972f08f6a6db7504f8334d"
+checksum = "4416167a69126e617f8d0a214af0e3c1dbdeffcb100ddf72dcd1a1ac9893c146"
 dependencies = [
  "bitflags 2.9.4",
  "block",
@@ -3914,9 +3914,9 @@ dependencies = [
 
 [[package]]
 name = "core-video"
-version = "0.4.3"
+version = "0.5.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d45e71d5be22206bed53c3c3cb99315fc4c3d31b8963808c6bc4538168c4f8ef"
+checksum = "139679cc63eb9504bdbe37e37874b0247136177655f0008588781e90863afa62"
 dependencies = [
  "block",
  "core-foundation 0.10.0",
@@ -10093,13 +10093,13 @@ dependencies = [
 
 [[package]]
 name = "metal"
-version = "0.29.0"
+version = "0.33.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7ecfd3296f8c56b7c1f6fbac3c71cefa9d78ce009850c45000015f206dc7fa21"
+checksum = "c7047791b5bc903b8cd963014b355f71dc9864a9a0b727057676c1dcae5cbc15"
 dependencies = [
  "bitflags 2.9.4",
  "block",
- "core-graphics-types 0.1.3",
+ "core-graphics-types 0.2.0",
  "foreign-types 0.5.0",
  "log",
  "objc",

Cargo.toml

@@ -507,7 +507,7 @@ cocoa-foundation = "=0.2.0"
 convert_case = "0.8.0"
 core-foundation = "=0.10.0"
 core-foundation-sys = "0.8.6"
-core-video = { version = "0.4.3", features = ["metal"] }
+core-video = { version = "0.5.2", features = ["metal"] }
 cpal = "0.16"
 crash-handler = "0.6"
 criterion = { version = "0.5", features = ["html_reports"] }
@@ -558,7 +558,7 @@ log = { version = "0.4.16", features = ["kv_unstable_serde", "serde"] }
 lsp-types = { git = "https://github.com/zed-industries/lsp-types", rev = "b71ab4eeb27d9758be8092020a46fe33fbca4e33" }
 mach2 = "0.5"
 markup5ever_rcdom = "0.3.0"
-metal = "0.29"
+metal = "0.33"
 minidumper = "0.8"
 moka = { version = "0.12.10", features = ["sync"] }
 naga = { version = "25.0", features = ["wgsl-in"] }

crates/gpui/src/platform/mac/metal_atlas.rs

@@ -15,6 +15,9 @@ pub(crate) struct MetalAtlas(Mutex<MetalAtlasState>);
 impl MetalAtlas {
     pub(crate) fn new(device: Device) -> Self {
         MetalAtlas(Mutex::new(MetalAtlasState {
+            // Shared memory can be used only if CPU and GPU share the same memory space.
+            // https://developer.apple.com/documentation/metal/setting-resource-storage-modes
+            unified_memory: device.has_unified_memory(),
             device: AssertSend(device),
             monochrome_textures: Default::default(),
             polychrome_textures: Default::default(),
@@ -29,6 +32,7 @@ impl MetalAtlas {
 
 struct MetalAtlasState {
     device: AssertSend<Device>,
+    unified_memory: bool,
     monochrome_textures: AtlasTextureList<MetalAtlasTexture>,
     polychrome_textures: AtlasTextureList<MetalAtlasTexture>,
     tiles_by_key: FxHashMap<AtlasKey, AtlasTile>,
@@ -149,6 +153,11 @@ impl MetalAtlasState {
         }
         texture_descriptor.set_pixel_format(pixel_format);
         texture_descriptor.set_usage(usage);
+        texture_descriptor.set_storage_mode(if self.unified_memory {
+            metal::MTLStorageMode::Shared
+        } else {
+            metal::MTLStorageMode::Managed
+        });
         let metal_texture = self.device.new_texture(&texture_descriptor);
 
         let texture_list = match kind {

crates/gpui/src/platform/mac/metal_renderer.rs

@@ -78,12 +78,22 @@ impl InstanceBufferPool {
         self.buffers.clear();
     }
 
-    pub(crate) fn acquire(&mut self, device: &metal::Device) -> InstanceBuffer {
+    pub(crate) fn acquire(
+        &mut self,
+        device: &metal::Device,
+        unified_memory: bool,
+    ) -> InstanceBuffer {
         let buffer = self.buffers.pop().unwrap_or_else(|| {
-            device.new_buffer(
-                self.buffer_size as u64,
-                MTLResourceOptions::StorageModeManaged,
-            )
+            let options = if unified_memory {
+                MTLResourceOptions::StorageModeShared
+                    // Buffers are write only which can benefit from the combined cache
+                    // https://developer.apple.com/documentation/metal/mtlresourceoptions/cpucachemodewritecombined
+                    | MTLResourceOptions::CPUCacheModeWriteCombined
+            } else {
+                MTLResourceOptions::StorageModeManaged
+            };
+
+            device.new_buffer(self.buffer_size as u64, options)
         });
         InstanceBuffer {
             metal_buffer: buffer,
@@ -101,6 +111,7 @@ impl InstanceBufferPool {
 pub(crate) struct MetalRenderer {
     device: metal::Device,
     layer: metal::MetalLayer,
+    unified_memory: bool,
     presents_with_transaction: bool,
     command_queue: CommandQueue,
     paths_rasterization_pipeline_state: metal::RenderPipelineState,
@@ -186,6 +197,10 @@ impl MetalRenderer {
             output
         }
 
+        // Shared memory can be used only if CPU and GPU share the same memory space.
+        // https://developer.apple.com/documentation/metal/setting-resource-storage-modes
+        let unified_memory = device.has_unified_memory();
+
         let unit_vertices = [
             to_float2_bits(point(0., 0.)),
             to_float2_bits(point(1., 0.)),
@@ -197,7 +212,12 @@ impl MetalRenderer {
         let unit_vertices = device.new_buffer_with_data(
             unit_vertices.as_ptr() as *const c_void,
             mem::size_of_val(&unit_vertices) as u64,
-            MTLResourceOptions::StorageModeManaged,
+            if unified_memory {
+                MTLResourceOptions::StorageModeShared
+                    | MTLResourceOptions::CPUCacheModeWriteCombined
+            } else {
+                MTLResourceOptions::StorageModeManaged
+            },
         );
 
         let paths_rasterization_pipeline_state = build_path_rasterization_pipeline_state(
@@ -275,6 +295,7 @@ impl MetalRenderer {
             device,
             layer,
             presents_with_transaction: false,
+            unified_memory,
             command_queue,
             paths_rasterization_pipeline_state,
             path_sprites_pipeline_state,
@@ -344,14 +365,23 @@ impl MetalRenderer {
         texture_descriptor.set_width(size.width.0 as u64);
         texture_descriptor.set_height(size.height.0 as u64);
         texture_descriptor.set_pixel_format(metal::MTLPixelFormat::BGRA8Unorm);
+        texture_descriptor.set_storage_mode(metal::MTLStorageMode::Private);
         texture_descriptor
             .set_usage(metal::MTLTextureUsage::RenderTarget | metal::MTLTextureUsage::ShaderRead);
         self.path_intermediate_texture = Some(self.device.new_texture(&texture_descriptor));
 
         if self.path_sample_count > 1 {
+            // https://developer.apple.com/documentation/metal/choosing-a-resource-storage-mode-for-apple-gpus
+            // Rendering MSAA textures are done in a single pass, so we can use memory-less storage on Apple Silicon
+            let storage_mode = if self.unified_memory {
+                metal::MTLStorageMode::Memoryless
+            } else {
+                metal::MTLStorageMode::Private
+            };
+
             let mut msaa_descriptor = texture_descriptor;
             msaa_descriptor.set_texture_type(metal::MTLTextureType::D2Multisample);
-            msaa_descriptor.set_storage_mode(metal::MTLStorageMode::Private);
+            msaa_descriptor.set_storage_mode(storage_mode);
             msaa_descriptor.set_sample_count(self.path_sample_count as _);
             self.path_intermediate_msaa_texture = Some(self.device.new_texture(&msaa_descriptor));
         } else {
@@ -385,7 +415,10 @@ impl MetalRenderer {
         };
 
         loop {
-            let mut instance_buffer = self.instance_buffer_pool.lock().acquire(&self.device);
+            let mut instance_buffer = self
+                .instance_buffer_pool
+                .lock()
+                .acquire(&self.device, self.unified_memory);
 
             let command_buffer =
                 self.draw_primitives(scene, &mut instance_buffer, drawable, viewport_size);
@@ -449,7 +482,10 @@ impl MetalRenderer {
             .ok_or_else(|| anyhow::anyhow!("Failed to get drawable for render_to_image"))?;
 
         loop {
-            let mut instance_buffer = self.instance_buffer_pool.lock().acquire(&self.device);
+            let mut instance_buffer = self
+                .instance_buffer_pool
+                .lock()
+                .acquire(&self.device, self.unified_memory);
 
             let command_buffer =
                 self.draw_primitives(scene, &mut instance_buffer, drawable, viewport_size);
@@ -649,10 +685,14 @@ impl MetalRenderer {
 
         command_encoder.end_encoding();
 
-        instance_buffer.metal_buffer.did_modify_range(NSRange {
-            location: 0,
-            length: instance_offset as NSUInteger,
-        });
+        if !self.unified_memory {
+            // Sync the instance buffer to the GPU
+            instance_buffer.metal_buffer.did_modify_range(NSRange {
+                location: 0,
+                length: instance_offset as NSUInteger,
+            });
+        }
+
         Ok(command_buffer.to_owned())
     }