Various Metal performance optimizations

Makefile

@@ -36,7 +36,7 @@ else
 endif
 
 
-.PHONY: all check test reftests travis-sdl2
+.PHONY: all check quad test reftests travis-sdl2
 
 all: check test
 
@@ -62,6 +62,9 @@ reftests-ci:
 	cd src/warden && cargo test --features "gl"
 	cd src/warden && cargo run --features "gl" -- ci #TODO: "gl-headless"
 
+quad:
+	cd examples && cargo run --bin quad --features ${FEATURES_HAL}
+
 travis-sdl2:
 	#TODO
 	#if [ -e $(SDL2_CONFIG) ]; then exit 1; fi

examples/quad/main.rs

@@ -30,8 +30,8 @@ use hal::pso::{PipelineStage, ShaderStageFlags, Specialization};
 use hal::queue::Submission;
 
 use std::fs;
-use std::io::Cursor;
-use std::io::Read;
+use std::io::{Cursor, Read};
+
 
 const ENTRY_NAME: &str = "main";
 
@@ -480,7 +480,7 @@ fn main() {
             cmd_buffer.set_viewports(0, &[viewport.clone()]);
             cmd_buffer.set_scissors(0, &[viewport.rect]);
             cmd_buffer.bind_graphics_pipeline(&pipeline);
-            cmd_buffer.bind_vertex_buffers(0, pso::VertexBufferSet(vec![(&vertex_buffer, 0)]));
+            cmd_buffer.bind_vertex_buffers(0, Some((&vertex_buffer, 0)));
             cmd_buffer.bind_graphics_descriptor_sets(&pipeline_layout, 0, Some(&desc_set), &[]); //TODO
 
             {

src/backend/dx11/src/lib.rs

@@ -866,9 +866,14 @@ impl hal::command::RawCommandBuffer<Backend> for CommandBuffer {
         }
     }
 
-    fn bind_vertex_buffers(&mut self, first_binding: u32, vbs: pso::VertexBufferSet<Backend>) {
-        let (buffers, offsets): (Vec<*mut d3d11::ID3D11Buffer>, Vec<u32>) = vbs.0.iter()
-            .map(|(buf, offset)| (buf.internal.raw, *offset as u32))
+    fn bind_vertex_buffers<I, T>(&mut self, first_binding: u32, buffers: I)
+    where
+        I: IntoIterator<Item = (T, buffer::Offset)>,
+        T: Borrow<Buffer>,
+    {
+        let (buffers, offsets): (Vec<*mut d3d11::ID3D11Buffer>, Vec<u32>) = buffers
+            .into_iter()
+            .map(|(buf, offset)| (buf.borrow().internal.raw, offset as u32))
             .unzip();
 
         // TODO: strides

src/backend/dx12/src/command.rs

@@ -1634,15 +1634,21 @@ impl com::RawCommandBuffer<Backend> for CommandBuffer {
         }
     }
 
-    fn bind_vertex_buffers(&mut self, first_binding: u32, vbs: pso::VertexBufferSet<Backend>) {
+    fn bind_vertex_buffers<I, T>(&mut self, first_binding: u32, buffers: I)
+    where
+        I: IntoIterator<Item = (T, buffer::Offset)>,
+        T: Borrow<n::Buffer>,
+    {
         // Only cache the vertex buffer views as we don't know the stride (PSO).
         assert!(first_binding as usize <= MAX_VERTEX_BUFFERS);
-        for (&(buffer, offset), view) in vbs.0.iter()
+        for (&(buffer, offset), view) in buffers
+            .into_iter()
             .zip(self.vertex_buffer_views[first_binding as _..].iter_mut())
         {
-            let base = unsafe { (*buffer.resource).GetGPUVirtualAddress() };
+            let b = buffer.borrow();
+            let base = unsafe { (*b.resource).GetGPUVirtualAddress() };
             view.BufferLocation = base + offset as u64;
-            view.SizeInBytes = buffer.size_in_bytes - offset as u32;
+            view.SizeInBytes = b.size_in_bytes - offset as u32;
         }
     }
 

src/backend/empty/src/lib.rs

@@ -499,7 +499,11 @@ impl command::RawCommandBuffer<Backend> for RawCommandBuffer {
         unimplemented!()
     }
 
-    fn bind_vertex_buffers(&mut self, _: u32, _: pso::VertexBufferSet<Backend>) {
+    fn bind_vertex_buffers<I, T>(&mut self, _: u32, _: I)
+    where
+        I: IntoIterator<Item = (T, buffer::Offset)>,
+        T: Borrow<()>,
+    {
         unimplemented!()
     }
 

src/backend/gl/src/command.rs

@@ -698,17 +698,20 @@ impl command::RawCommandBuffer<Backend> for RawCommandBuffer {
         self.push_cmd(Command::BindIndexBuffer(ibv.buffer.raw));
     }
 
-    fn bind_vertex_buffers(&mut self, _first_binding: u32, vbs: hal::pso::VertexBufferSet<Backend>) {
-        if vbs.0.len() == 0 {
-            return
-        }
-
-        let needed_length = vbs.0.iter().map(|vb| vb.1).max().unwrap() + 1;
-
-        self.cache.vertex_buffers.resize(needed_length as usize, 0);
-
-        for vb in vbs.0 {
-            self.cache.vertex_buffers[vb.1 as usize] = vb.0.raw;
+    fn bind_vertex_buffers<I, T>(&mut self, first_binding: u32, buffers: I)
+    where
+        I: IntoIterator<Item = (T, buffer::Offset)>,
+        T: Borrow<n::Buffer>,
+    {
+        for (i, (buffer, offset)) in buffers.into_iter().enumerate() {
+            let index = first_binding as usize + i;
+            if self.cache.vertex_buffers.len() <= index {
+                self.cache.vertex_buffers.resize(index+1, 0);
+            }
+            self.cache.vertex_buffers[index] = buffer.borrow().raw;
+            if offset != 0 {
+                error!("Vertex buffer offset {} is not supported", offset);
+            }
         }
     }
 

src/backend/metal/src/command.rs

@@ -2233,13 +2233,26 @@ impl com::RawCommandBuffer<Backend> for CommandBuffer {
         });
     }
 
-    fn bind_vertex_buffers(&mut self, first_binding: u32, buffer_set: pso::VertexBufferSet<Backend>) {
-        while self.state.vertex_buffers.len() < first_binding as usize + buffer_set.0.len() {
-            self.state.vertex_buffers.push(None);
-        }
-        for (i, &(buffer, offset)) in buffer_set.0.iter().enumerate() {
-            let buffer_ptr = BufferPtr(buffer.raw.as_ptr());
-            self.state.vertex_buffers[first_binding as usize + i] = Some((buffer_ptr, buffer.range.start + offset));
+
+    fn bind_vertex_buffers<I, T>(&mut self, first_binding: u32, buffers: I)
+    where
+        I: IntoIterator<Item = (T, buffer::Offset)>,
+        T: Borrow<native::Buffer>,
+    {
+        if self.state.vertex_buffers.len() <= first_binding as usize {
+            self.state.vertex_buffers.resize(first_binding as usize + 1, None);
+        }
+        for (i, (buffer, offset)) in buffers.into_iter().enumerate() {
+            let b = buffer.borrow();
+            let buffer_ptr = BufferPtr(b.raw.as_ptr());
+            let index = first_binding as usize + i;
+            let value = Some((buffer_ptr, b.range.start + offset));
+            if index >= self.state.vertex_buffers.len() {
+                debug_assert_eq!(index, self.state.vertex_buffers.len());
+                self.state.vertex_buffers.push(value);
+            } else {
+                self.state.vertex_buffers[index] = value;
+            }
         }
 
         let mask = self.state.set_vertex_buffers();

src/backend/vulkan/src/command.rs

@@ -483,11 +483,15 @@ impl com::RawCommandBuffer<Backend> for CommandBuffer {
         }
     }
 
-    fn bind_vertex_buffers(&mut self, first_binding: u32, vbs: pso::VertexBufferSet<Backend>) {
-        let buffers: SmallVec<[vk::Buffer; 16]> =
-            vbs.0.iter().map(|&(ref buffer, _)| buffer.raw).collect();
-        let offsets: SmallVec<[vk::DeviceSize; 16]> =
-            vbs.0.iter().map(|&(_, offset)| offset as u64).collect();
+    fn bind_vertex_buffers<I, T>(&mut self, first_binding: u32, buffers: I)
+    where
+        I: IntoIterator<Item = (T, buffer::Offset)>,
+        T: Borrow<n::Buffer>,
+    {
+        let (buffers, offsets): (SmallVec<[vk::Buffer; 16]>, SmallVec<[vk::DeviceSize; 16]>) = buffers
+            .into_iter()
+            .map(|(buffer, offset)| (buffer.borrow().raw, offset))
+            .unzip();
 
         unsafe {
             self.device.0.cmd_bind_vertex_buffers(

src/hal/src/command/graphics.rs

@@ -3,8 +3,7 @@ use std::borrow::Borrow;
 use std::ops::Range;
 
 use Backend;
-use {image, pso};
-use buffer::IndexBufferView;
+use {buffer, image, pso};
 use query::{Query, QueryControl, QueryId};
 use queue::capability::{Graphics, GraphicsOrCompute, Supports};
 use super::{
@@ -181,13 +180,17 @@ impl<'a, B: Backend, C: Supports<Graphics>, S: Shot, L: Level> CommandBuffer<'a,
     }
 
     /// Identical to the `RawCommandBuffer` method of the same name.
-    pub fn bind_index_buffer(&mut self, ibv: IndexBufferView<B>) {
+    pub fn bind_index_buffer(&mut self, ibv: buffer::IndexBufferView<B>) {
         self.raw.bind_index_buffer(ibv)
     }
 
     /// Identical to the `RawCommandBuffer` method of the same name.
-    pub fn bind_vertex_buffers(&mut self, first_binding: u32, vbs: pso::VertexBufferSet<B>) {
-        self.raw.bind_vertex_buffers(first_binding, vbs)
+    pub fn bind_vertex_buffers<I, T>(&mut self, first_binding: u32, buffers: I)
+    where
+        I: IntoIterator<Item = (T, buffer::Offset)>,
+        T: Borrow<B::Buffer>,
+    {
+        self.raw.bind_vertex_buffers(first_binding, buffers)
     }
 
     /// Identical to the `RawCommandBuffer` method of the same name.

src/hal/src/command/raw.rs

@@ -206,7 +206,10 @@ pub trait RawCommandBuffer<B: Backend>: Clone + Any + Send + Sync {
     ///
     /// Each buffer passed corresponds to the vertex input binding with the same index,
     /// starting from an offset index `first_binding`.
-    fn bind_vertex_buffers(&mut self, first_binding: u32, pso::VertexBufferSet<B>);
+    fn bind_vertex_buffers<I, T>(&mut self, first_binding: u32, buffers: I)
+    where
+        I: IntoIterator<Item = (T, buffer::Offset)>,
+        T: Borrow<B::Buffer>;
 
     /// Set the viewport parameters for the rasterizer.
     ///

src/hal/src/command/render_pass.rs

@@ -65,8 +65,12 @@ impl<'a, B: Backend> RenderSubpassCommon<'a, B> {
     }
 
     ///
-    pub fn bind_vertex_buffers(&mut self, first_binding: u32, vbs: pso::VertexBufferSet<B>) {
-        self.0.bind_vertex_buffers(first_binding, vbs);
+    pub fn bind_vertex_buffers<I, T>(&mut self, first_binding: u32, buffers: I)
+    where
+        I: IntoIterator<Item = (T, buffer::Offset)>,
+        T: Borrow<B::Buffer>,
+    {
+        self.0.bind_vertex_buffers(first_binding, buffers);
     }
 
     ///

src/hal/src/pso/input_assembler.rs

@@ -2,8 +2,7 @@
 //! The input assembler collects raw vertex and index data.
 
 use format;
-use buffer::Offset;
-use {Backend, Primitive};
+use {Primitive};
 
 /// Shader binding location.
 pub type Location = u32;
@@ -88,17 +87,3 @@ impl InputAssemblerDesc {
         }
     }
 }
-
-/// A complete set of vertex buffers to be used for vertex import in PSO.
-#[derive(Clone, Debug)]
-pub struct VertexBufferSet<'a, B: Backend>(
-    /// Array of buffer handles with offsets in them
-    pub Vec<(&'a B::Buffer, Offset)>,
-);
-
-impl<'a, B: Backend> VertexBufferSet<'a, B> {
-    /// Create an empty set
-    pub fn new() -> Self {
-        VertexBufferSet(Vec::new())
-    }
-}

src/warden/src/gpu.rs

@@ -945,10 +945,8 @@ impl<B: hal::Backend> Scene<B, hal::General> {
                                                 .expect(&format!("Missing vertex buffer: {}", name))
                                                 .handle;
                                             (buf, offset)
-                                        })
-                                        .collect::<Vec<_>>();
-                                    let set = pso::VertexBufferSet(buffers_raw);
-                                    encoder.bind_vertex_buffers(0, set);
+                                        });
+                                    encoder.bind_vertex_buffers(0, buffers_raw);
                                 }
                                 Dc::BindPipeline(ref name) => {
                                     let pso = resources.graphics_pipelines