diff --git a/Cargo.lock b/Cargo.lock
index a04cbf4..0619b43 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -851,6 +851,7 @@ dependencies = [
  "ash-window",
  "bitflags",
  "raw-window-handle",
+ "smallvec",
  "winapi 0.3.9",
  "wio",
 ]
@@ -1036,9 +1037,9 @@ checksum = "c111b5bd5695e56cffe5129854aa230b39c93a305372fdbb2668ca2394eea9f8"
 
 [[package]]
 name = "smallvec"
-version = "1.4.2"
+version = "1.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fbee7696b84bbf3d89a1c2eccff0850e3047ed46bfcd2e92c29a2d074d57e252"
+checksum = "fe0f37c9e8f3c5a4a66ad655a93c74daac4ad00c441533bf5c6e7990bb42604e"
 
 [[package]]
 name = "smithay-client-toolkit"
diff --git a/piet-gpu-hal/Cargo.toml b/piet-gpu-hal/Cargo.toml
index abbe98a..763da23 100644
--- a/piet-gpu-hal/Cargo.toml
+++ b/piet-gpu-hal/Cargo.toml
@@ -11,7 +11,9 @@ ash = "0.31"
 ash-window = "0.5"
 raw-window-handle = "0.3"
 bitflags = "1.2.1"
+smallvec = "1.6.1"
 
+[target.'cfg(target_os="windows")'.dependencies]
 winapi = { version = "0.3.9", features = [
     'd3d12', 'd3d12sdklayers', 'd3dcommon', 'd3dcompiler', 'dxgi',
     'dxgi1_2', 'dxgi1_3', 'dxgi1_4', 'dxgidebug', 'dxgiformat', 'dxgitype',
diff --git a/piet-gpu-hal/examples/collatz.rs b/piet-gpu-hal/examples/collatz.rs
index bde4084..5dcce21 100644
--- a/piet-gpu-hal/examples/collatz.rs
+++ b/piet-gpu-hal/examples/collatz.rs
@@ -1,16 +1,16 @@
 use piet_gpu_hal::hub;
-use piet_gpu_hal::vulkan::VkInstance;
-use piet_gpu_hal::{BufferUsage, CmdBuf};
+use piet_gpu_hal::mux::{Instance, ShaderCode};
+use piet_gpu_hal::BufferUsage;
 
 fn main() {
-    let (instance, _) = VkInstance::new(None).unwrap();
+    let (instance, _) = Instance::new(None).unwrap();
     unsafe {
         let device = instance.device(None).unwrap();
         let session = hub::Session::new(device);
         let usage = BufferUsage::MAP_READ | BufferUsage::STORAGE;
         let src = (0..256).map(|x| x + 1).collect::<Vec<u32>>();
         let buffer = session.create_buffer_init(&src, usage).unwrap();
-        let code = include_bytes!("./shader/collatz.spv");
+        let code = ShaderCode::Spv(include_bytes!("./shader/collatz.spv"));
         let pipeline = session.create_simple_compute_pipeline(code, 1).unwrap();
         let descriptor_set = session
             .create_simple_descriptor_set(&pipeline, &[&buffer])
diff --git a/piet-gpu-hal/examples/dx12_toy.rs b/piet-gpu-hal/examples/dx12_toy.rs
index 892ec0c..63e3db4 100644
--- a/piet-gpu-hal/examples/dx12_toy.rs
+++ b/piet-gpu-hal/examples/dx12_toy.rs
@@ -50,8 +50,8 @@ void main(SPIRV_Cross_Input stage_input)
 "#;
 
 fn toy() -> Result<(), Error> {
-    let instance = dx12::Dx12Instance::new()?;
-    let device = instance.device()?;
+    let (instance, _surface) = dx12::Dx12Instance::new(None)?;
+    let device = instance.device(None)?;
     let buf = device.create_buffer(
         1024,
         BufferUsage::MAP_READ
@@ -87,7 +87,7 @@ fn toy() -> Result<(), Error> {
         cmd_buf.host_barrier();
         cmd_buf.finish();
         device.run_cmd_bufs(&[&cmd_buf], &[], &[], Some(&fence))?;
-        device.wait_and_reset(&[fence])?;
+        device.wait_and_reset(&[&fence])?;
         let mut readback: Vec<u32> = vec![0u32; 256];
         device.read_buffer(&buf, readback.as_mut_ptr() as *mut u8, 0, 1024)?;
         println!("{:?}", readback);
diff --git a/piet-gpu-hal/src/dx12.rs b/piet-gpu-hal/src/dx12.rs
index bf0e44e..cc9e77b 100644
--- a/piet-gpu-hal/src/dx12.rs
+++ b/piet-gpu-hal/src/dx12.rs
@@ -9,6 +9,8 @@ use winapi::shared::dxgi1_3;
 use winapi::shared::minwindef::TRUE;
 use winapi::um::d3d12;
 
+use smallvec::SmallVec;
+
 use crate::{BufferUsage, Error, ImageLayout};
 
 use self::wrappers::{
@@ -19,6 +21,12 @@ pub struct Dx12Instance {
     factory: Factory4,
 }
 
+// TODO
+pub struct Dx12Surface;
+
+// TODO
+pub struct Dx12Swapchain;
+
 pub struct Dx12Device {
     device: Device,
     command_allocator: CommandAllocator,
@@ -30,7 +38,7 @@ pub struct Dx12Device {
 #[derive(Clone)]
 pub struct Buffer {
     resource: Resource,
-    size: u64,
+    pub size: u64,
 }
 
 #[derive(Clone)]
@@ -95,7 +103,7 @@ impl Dx12Instance {
     ///
     /// TODO: take a raw window handle.
     /// TODO: can probably be a trait.
-    pub fn new() -> Result<Dx12Instance, Error> {
+    pub fn new(window_handle: Option<&dyn raw_window_handle::HasRawWindowHandle>) -> Result<(Dx12Instance, Option<Dx12Surface>), Error> {
         unsafe {
             #[cfg(debug_assertions)]
             if let Err(e) = wrappers::enable_debug_layer() {
@@ -110,7 +118,7 @@ impl Dx12Instance {
             let factory_flags: u32 = 0;
 
             let factory = Factory4::create(factory_flags)?;
-            Ok(Dx12Instance { factory })
+            Ok((Dx12Instance { factory }, None))
         }
     }
 
@@ -118,7 +126,7 @@ impl Dx12Instance {
     ///
     /// TODO: handle window.
     /// TODO: probably can also be trait'ified.
-    pub fn device(&self) -> Result<Dx12Device, Error> {
+    pub fn device(&self, surface: Option<&Dx12Surface>) -> Result<Dx12Device, Error> {
         unsafe {
             let device = Device::create_device(&self.factory)?;
             let list_type = d3d12::D3D12_COMMAND_LIST_TYPE_DIRECT;
@@ -262,16 +270,15 @@ impl crate::Device for Dx12Device {
     unsafe fn run_cmd_bufs(
         &self,
         cmd_bufs: &[&Self::CmdBuf],
-        wait_semaphores: &[Self::Semaphore],
-        signal_semaphores: &[Self::Semaphore],
+        wait_semaphores: &[&Self::Semaphore],
+        signal_semaphores: &[&Self::Semaphore],
         fence: Option<&Self::Fence>,
     ) -> Result<(), Error> {
         // TODO: handle semaphores
-        // SmallVec?
         let lists = cmd_bufs
             .iter()
             .map(|c| c.0.as_raw_command_list())
-            .collect::<Vec<_>>();
+            .collect::<SmallVec<[_; 4]>>();
         self.command_queue.execute_command_lists(&lists);
         if let Some(fence) = fence {
             let val = fence.val.get() + 1;
@@ -319,7 +326,7 @@ impl crate::Device for Dx12Device {
         Ok(Fence { fence, event, val })
     }
 
-    unsafe fn wait_and_reset(&self, fences: &[Self::Fence]) -> Result<(), Error> {
+    unsafe fn wait_and_reset(&self, fences: &[&Self::Fence]) -> Result<(), Error> {
         for fence in fences {
             // TODO: probably handle errors here.
             let _status = fence.event.wait(winapi::um::winbase::INFINITE);
@@ -327,7 +334,7 @@ impl crate::Device for Dx12Device {
         Ok(())
     }
 
-    unsafe fn get_fence_status(&self, fence: Self::Fence) -> Result<bool, Error> {
+    unsafe fn get_fence_status(&self, fence: &Self::Fence) -> Result<bool, Error> {
         let fence_val = fence.fence.get_value();
         Ok(fence_val == fence.val.get())
     }
diff --git a/piet-gpu-hal/src/hub.rs b/piet-gpu-hal/src/hub.rs
index 3238376..80c03cf 100644
--- a/piet-gpu-hal/src/hub.rs
+++ b/piet-gpu-hal/src/hub.rs
@@ -1,35 +1,28 @@
-//! A convenience layer on top of raw hal.
+//! A somewhat higher level GPU abstraction.
 //!
-//! This layer takes care of some lifetime and synchronization bookkeeping.
-//! It is likely that it will also take care of compile time and runtime
-//! negotiation of backends (Vulkan, DX12), but right now it's Vulkan-only.
+//! This layer is on top of the lower-level layer that multiplexes different
+//! back-ends. It handles details such as managing staging buffers for creating
+//! buffers with initial content, deferring dropping of resources until command
+//! submission is complete, and a bit more. These conveniences might expand
+//! even more in time.
 
 use std::convert::TryInto;
 use std::sync::{Arc, Mutex, Weak};
 
-use crate::vulkan;
-use crate::CmdBuf as CmdBufTrait;
-use crate::DescriptorSetBuilder as DescriptorSetBuilderTrait;
-use crate::PipelineBuilder as PipelineBuilderTrait;
-use crate::{BufferUsage, Device, Error, GpuInfo, SamplerParams};
+use smallvec::SmallVec;
 
-pub type Semaphore = <vulkan::VkDevice as Device>::Semaphore;
-pub type Pipeline = <vulkan::VkDevice as Device>::Pipeline;
-pub type DescriptorSet = <vulkan::VkDevice as Device>::DescriptorSet;
-pub type QueryPool = <vulkan::VkDevice as Device>::QueryPool;
-pub type Sampler = <vulkan::VkDevice as Device>::Sampler;
+use crate::mux;
 
-type Fence = <vulkan::VkDevice as Device>::Fence;
+use crate::{BufferUsage, Error, GpuInfo, SamplerParams};
 
-type VkImage = <vulkan::VkDevice as Device>::Image;
-type VkBuffer = <vulkan::VkDevice as Device>::Buffer;
+pub use crate::mux::{DescriptorSet, Fence, Pipeline, QueryPool, Sampler, Semaphore, ShaderCode};
 
 #[derive(Clone)]
 pub struct Session(Arc<SessionInner>);
 
 struct SessionInner {
-    device: vulkan::VkDevice,
-    cmd_buf_pool: Mutex<Vec<(vulkan::CmdBuf, Fence)>>,
+    device: mux::Device,
+    cmd_buf_pool: Mutex<Vec<(mux::CmdBuf, Fence)>>,
     /// Command buffers that are still pending (so resources can't be freed).
     pending: Mutex<Vec<SubmittedCmdBufInner>>,
     /// A command buffer that is used for copying from staging buffers.
@@ -38,7 +31,7 @@ struct SessionInner {
 }
 
 pub struct CmdBuf {
-    cmd_buf: vulkan::CmdBuf,
+    cmd_buf: mux::CmdBuf,
     fence: Fence,
     resources: Vec<RetainResource>,
     session: Weak<SessionInner>,
@@ -50,7 +43,7 @@ pub struct SubmittedCmdBuf(Option<SubmittedCmdBufInner>, Weak<SessionInner>);
 struct SubmittedCmdBufInner {
     // It's inconsistent, cmd_buf is unpacked, staging_cmd_buf isn't. Probably
     // better to chose one or the other.
-    cmd_buf: vulkan::CmdBuf,
+    cmd_buf: mux::CmdBuf,
     fence: Fence,
     resources: Vec<RetainResource>,
     staging_cmd_buf: Option<CmdBuf>,
@@ -60,7 +53,7 @@ struct SubmittedCmdBufInner {
 pub struct Image(Arc<ImageInner>);
 
 struct ImageInner {
-    image: VkImage,
+    image: mux::Image,
     session: Weak<SessionInner>,
 }
 
@@ -68,13 +61,13 @@ struct ImageInner {
 pub struct Buffer(Arc<BufferInner>);
 
 struct BufferInner {
-    buffer: VkBuffer,
+    buffer: mux::Buffer,
     session: Weak<SessionInner>,
 }
 
-pub struct PipelineBuilder(vulkan::PipelineBuilder);
+pub struct PipelineBuilder(mux::PipelineBuilder);
 
-pub struct DescriptorSetBuilder(vulkan::DescriptorSetBuilder);
+pub struct DescriptorSetBuilder(mux::DescriptorSetBuilder);
 
 /// Data types that can be stored in a GPU buffer.
 pub unsafe trait PlainData {}
@@ -97,7 +90,7 @@ pub enum RetainResource {
 }
 
 impl Session {
-    pub fn new(device: vulkan::VkDevice) -> Session {
+    pub fn new(device: mux::Device) -> Session {
         let gpu_info = device.query_gpu_info();
         Session(Arc::new(SessionInner {
             device,
@@ -130,10 +123,10 @@ impl Session {
         unsafe {
             let mut i = 0;
             while i < pending.len() {
-                if let Ok(true) = self.0.device.get_fence_status(pending[i].fence) {
+                if let Ok(true) = self.0.device.get_fence_status(&pending[i].fence) {
                     let item = pending.swap_remove(i);
                     // TODO: wait is superfluous, can just reset
-                    let _ = self.0.device.wait_and_reset(&[item.fence]);
+                    let _ = self.0.device.wait_and_reset(&[&item.fence]);
                     let mut pool = self.0.cmd_buf_pool.lock().unwrap();
                     pool.push((item.cmd_buf, item.fence));
                     std::mem::drop(item.resources);
@@ -151,8 +144,8 @@ impl Session {
     pub unsafe fn run_cmd_buf(
         &self,
         cmd_buf: CmdBuf,
-        wait_semaphores: &[Semaphore],
-        signal_semaphores: &[Semaphore],
+        wait_semaphores: &[&Semaphore],
+        signal_semaphores: &[&Semaphore],
     ) -> Result<SubmittedCmdBuf, Error> {
         // Again, SmallVec here?
         let mut cmd_bufs = Vec::with_capacity(2);
@@ -222,7 +215,7 @@ impl Session {
         let create_buf = self.create_buffer(size, create_usage)?;
         self.0
             .device
-            .write_buffer(&create_buf.vk_buffer(), contents, 0, size)?;
+            .write_buffer(&create_buf.mux_buffer(), contents, 0, size)?;
         if use_staging_buffer {
             let buf = self.create_buffer(size, usage | BufferUsage::COPY_DST)?;
             let mut staging_cmd_buf = self.0.staging_cmd_buf.lock().unwrap();
@@ -232,8 +225,8 @@ impl Session {
                 *staging_cmd_buf = Some(cmd_buf);
             }
             let staging_cmd_buf = staging_cmd_buf.as_mut().unwrap();
-            // This will ensure the staging buffer is deallocated. It would be nice to
-            staging_cmd_buf.copy_buffer(create_buf.vk_buffer(), buf.vk_buffer());
+            // This will ensure the staging buffer is deallocated.
+            staging_cmd_buf.copy_buffer(create_buf.mux_buffer(), buf.mux_buffer());
             staging_cmd_buf.add_resource(create_buf);
             Ok(buf)
         } else {
@@ -256,9 +249,9 @@ impl Session {
     /// This creates a pipeline that operates on some buffers and images.
     ///
     /// The descriptor set layout is just some number of storage buffers and storage images (this might change).
-    pub unsafe fn create_simple_compute_pipeline(
+    pub unsafe fn create_simple_compute_pipeline<'a>(
         &self,
-        code: &[u8],
+        code: ShaderCode<'a>,
         n_buffers: u32,
     ) -> Result<Pipeline, Error> {
         self.pipeline_builder()
@@ -295,7 +288,8 @@ impl Session {
     }
 
     pub unsafe fn create_sampler(&self, params: SamplerParams) -> Result<Sampler, Error> {
-        self.0.device.create_sampler(params)
+        todo!()
+        //self.0.device.create_sampler(params)
     }
 
     pub fn gpu_info(&self) -> &GpuInfo {
@@ -322,7 +316,7 @@ impl SubmittedCmdBuf {
         let item = self.0.take().unwrap();
         if let Some(session) = Weak::upgrade(&self.1) {
             unsafe {
-                session.device.wait_and_reset(&[item.fence])?;
+                session.device.wait_and_reset(&[&item.fence])?;
             }
             session
                 .cmd_buf_pool
@@ -366,10 +360,9 @@ impl Drop for ImageInner {
     }
 }
 
-/// For now, we deref, but for runtime backend switching we'll need to wrap
-/// all methods.
+// Probably migrate from deref here to wrapping all methods.
 impl std::ops::Deref for CmdBuf {
-    type Target = vulkan::CmdBuf;
+    type Target = mux::CmdBuf;
     fn deref(&self) -> &Self::Target {
         &self.cmd_buf
     }
@@ -382,13 +375,13 @@ impl std::ops::DerefMut for CmdBuf {
 }
 
 impl Image {
-    pub fn vk_image(&self) -> &vulkan::Image {
+    pub fn mux_image(&self) -> &mux::Image {
         &self.0.image
     }
 }
 
 impl Buffer {
-    pub fn vk_buffer(&self) -> &vulkan::Buffer {
+    pub fn mux_buffer(&self) -> &mux::Buffer {
         &self.0.buffer
     }
 
@@ -405,7 +398,7 @@ impl Buffer {
         Ok(())
     }
     pub unsafe fn read<T: PlainData>(&self, result: &mut Vec<T>) -> Result<(), Error> {
-        let size = self.vk_buffer().size;
+        let size = self.mux_buffer().size();
         let len = size as usize / std::mem::size_of::<T>();
         if len > result.len() {
             result.reserve(len - result.len());
@@ -440,10 +433,10 @@ impl PipelineBuilder {
         self
     }
 
-    pub unsafe fn create_compute_pipeline(
+    pub unsafe fn create_compute_pipeline<'a>(
         self,
         session: &Session,
-        code: &[u8],
+        code: ShaderCode<'a>,
     ) -> Result<Pipeline, Error> {
         self.0.create_compute_pipeline(&session.0.device, code)
     }
@@ -451,23 +444,29 @@ impl PipelineBuilder {
 
 impl DescriptorSetBuilder {
     pub fn add_buffers<'a>(mut self, buffers: impl IntoRefs<'a, Buffer>) -> Self {
-        let vk_buffers = buffers
+        let mux_buffers = buffers
             .into_refs()
-            .map(|b| b.vk_buffer())
-            .collect::<Vec<_>>();
-        self.0.add_buffers(&vk_buffers);
+            .map(|b| b.mux_buffer())
+            .collect::<SmallVec<[_; 8]>>();
+        self.0.add_buffers(&mux_buffers);
         self
     }
 
     pub fn add_images<'a>(mut self, images: impl IntoRefs<'a, Image>) -> Self {
-        let vk_images = images.into_refs().map(|i| i.vk_image()).collect::<Vec<_>>();
-        self.0.add_images(&vk_images);
+        let mux_images = images
+            .into_refs()
+            .map(|i| i.mux_image())
+            .collect::<Vec<_>>();
+        self.0.add_images(&mux_images);
         self
     }
 
     pub fn add_textures<'a>(mut self, images: impl IntoRefs<'a, Image>) -> Self {
-        let vk_images = images.into_refs().map(|i| i.vk_image()).collect::<Vec<_>>();
-        self.0.add_textures(&vk_images);
+        let mux_images = images
+            .into_refs()
+            .map(|i| i.mux_image())
+            .collect::<Vec<_>>();
+        self.0.add_textures(&mux_images);
         self
     }
 
diff --git a/piet-gpu-hal/src/lib.rs b/piet-gpu-hal/src/lib.rs
index 2fc7a8d..9a16686 100644
--- a/piet-gpu-hal/src/lib.rs
+++ b/piet-gpu-hal/src/lib.rs
@@ -6,9 +6,20 @@ use bitflags::bitflags;
 
 pub mod hub;
 
-#[cfg(target_os = "windows")]
-pub mod dx12;
-pub mod vulkan;
+#[macro_use]
+mod macros;
+
+// TODO make this not pub
+pub mod mux;
+
+mux_cfg! {
+    #[cfg(vk)]
+    pub mod vulkan;
+}
+mux_cfg! {
+    #[cfg(dx12)]
+    pub mod dx12;
+}
 
 /// This isn't great but is expedient.
 pub type Error = Box<dyn std::error::Error>;
@@ -178,8 +189,8 @@ pub trait Device: Sized {
     unsafe fn run_cmd_bufs(
         &self,
         cmd_buf: &[&Self::CmdBuf],
-        wait_semaphores: &[Self::Semaphore],
-        signal_semaphores: &[Self::Semaphore],
+        wait_semaphores: &[&Self::Semaphore],
+        signal_semaphores: &[&Self::Semaphore],
         fence: Option<&Self::Fence>,
     ) -> Result<(), Error>;
 
@@ -217,8 +228,8 @@ pub trait Device: Sized {
 
     unsafe fn create_semaphore(&self) -> Result<Self::Semaphore, Error>;
     unsafe fn create_fence(&self, signaled: bool) -> Result<Self::Fence, Error>;
-    unsafe fn wait_and_reset(&self, fences: &[Self::Fence]) -> Result<(), Error>;
-    unsafe fn get_fence_status(&self, fence: Self::Fence) -> Result<bool, Error>;
+    unsafe fn wait_and_reset(&self, fences: &[&Self::Fence]) -> Result<(), Error>;
+    unsafe fn get_fence_status(&self, fence: &Self::Fence) -> Result<bool, Error>;
 
     unsafe fn create_sampler(&self, params: SamplerParams) -> Result<Self::Sampler, Error>;
 }
diff --git a/piet-gpu-hal/src/macros.rs b/piet-gpu-hal/src/macros.rs
new file mode 100644
index 0000000..f1d7019
--- /dev/null
+++ b/piet-gpu-hal/src/macros.rs
@@ -0,0 +1,104 @@
+// Copyright 2021 The piet-gpu authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// Also licensed under MIT license, at your choice.
+
+//! Macros, mostly to automate backend selection tedium.
+
+#[macro_export]
+macro_rules! mux_cfg {
+    ( #[cfg(vk)] $($tokens:tt)* ) => {
+        #[cfg(not(target_os="macos"))] $( $tokens )*
+    };
+
+    ( #[cfg(dx12)] $($tokens:tt)* ) => {
+        #[cfg(target_os="windows")] $( $tokens )*
+    };
+}
+
+#[macro_export]
+macro_rules! mux_enum {
+    ( $(#[$outer:meta])* $v:vis enum $name:ident {
+        Vk($vk:ty),
+        Dx12($dx12:ty),
+    } ) => {
+        $(#[$outer])* $v enum $name {
+            #[cfg(not(target_os="macos"))]
+            Vk($vk),
+            #[cfg(target_os="windows")]
+            Dx12($dx12),
+        }
+
+        impl $name {
+            $crate::mux_cfg! {
+                #[cfg(vk)]
+                #[allow(unused)]
+                fn vk(&self) -> &$vk {
+                    match self {
+                        $name::Vk(x) => x,
+                        _ => panic!("downcast error")
+                    }
+                }
+            }
+
+            $crate::mux_cfg! {
+                #[cfg(dx12)]
+                #[allow(unused)]
+                fn dx12(&self) -> &$dx12 {
+                    match self {
+                        $name::Dx12(x) => x,
+                        _ => panic!("downcast error")
+                    }
+                }
+            }
+        }
+    };
+}
+
+macro_rules! mux_device_enum {
+    ( $(#[$outer:meta])* $assoc_type: ident) => {
+        $crate::mux_enum! {
+            $(#[$outer])*
+            pub enum $assoc_type {
+                Vk(<$crate::vulkan::VkDevice as $crate::Device>::$assoc_type),
+                Dx12(<$crate::dx12::Dx12Device as $crate::Device>::$assoc_type),
+            }
+        }
+    }
+}
+
+#[macro_export]
+macro_rules! mux_match {
+    ( $e:expr ;
+        $vkname:ident::Vk($vkvar:ident) => $vkblock: block
+        $dx12name:ident::Dx12($dx12var:ident) => $dx12block: block
+    ) => {
+        match $e {
+            #[cfg(not(target_os="macos"))]
+            $vkname::Vk($vkvar) => $vkblock
+            #[cfg(target_os="windows")]
+            $dx12name::Dx12($dx12var) => $dx12block
+        }
+    };
+
+    ( $e:expr ;
+        $vkname:ident::Vk($vkvar:ident) => $vkblock: expr,
+        $dx12name:ident::Dx12($dx12var:ident) => $dx12block: expr,
+    ) => {
+        $crate::mux_match! { $e;
+            $vkname::Vk($vkvar) => { $vkblock }
+            $dx12name::Dx12($dx12var) => { $dx12block }
+        }
+    };
+}
diff --git a/piet-gpu-hal/src/mux.rs b/piet-gpu-hal/src/mux.rs
new file mode 100644
index 0000000..70525c4
--- /dev/null
+++ b/piet-gpu-hal/src/mux.rs
@@ -0,0 +1,601 @@
+// Copyright 2021 The piet-gpu authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// Also licensed under MIT license, at your choice.
+
+//! A multiplexer module that selects a back-end at runtime.
+
+use smallvec::SmallVec;
+
+mux_cfg! {
+    #[cfg(vk)]
+    use crate::vulkan;    
+}
+mux_cfg! {
+    #[cfg(dx12)]
+    use crate::dx12;
+}
+use crate::CmdBuf as CmdBufTrait;
+use crate::DescriptorSetBuilder as DescriptorSetBuilderTrait;
+use crate::Device as DeviceTrait;
+use crate::PipelineBuilder as PipelineBuilderTrait;
+use crate::{BufferUsage, Error, GpuInfo, ImageLayout};
+
+mux_enum! {
+    /// An instance, selected from multiple backends.
+    pub enum Instance {
+        Vk(vulkan::VkInstance),
+        Dx12(dx12::Dx12Instance),
+    }
+}
+
+mux_enum! {
+    /// A device, selected from multiple backends.
+    pub enum Device {
+        Vk(vulkan::VkDevice),
+        Dx12(dx12::Dx12Device),
+    }
+}
+
+mux_enum! {
+    /// A surface, which can apply to one of multiple backends.
+    pub enum Surface {
+        Vk(vulkan::VkSurface),
+        Dx12(dx12::Dx12Surface),
+    }
+}
+
+mux_enum! {
+    /// A surface, which can apply to one of multiple backends.
+    pub enum Swapchain {
+        Vk(vulkan::VkSwapchain),
+        Dx12(dx12::Dx12Swapchain),
+    }
+}
+
+mux_device_enum! { Buffer }
+mux_device_enum! { Image }
+mux_device_enum! { Fence }
+mux_device_enum! { Semaphore }
+mux_device_enum! { PipelineBuilder }
+mux_device_enum! { Pipeline }
+mux_device_enum! { DescriptorSetBuilder }
+mux_device_enum! { DescriptorSet }
+mux_device_enum! { CmdBuf }
+mux_device_enum! { QueryPool }
+mux_device_enum! { Sampler }
+
+/// The code for a shader, either as source or intermediate representation.
+pub enum ShaderCode<'a> {
+    Spv(&'a [u8]),
+    Hlsl(&'a str),
+}
+
+impl Instance {
+    pub fn new(
+        window_handle: Option<&dyn raw_window_handle::HasRawWindowHandle>,
+    ) -> Result<(Instance, Option<Surface>), Error> {
+        mux_cfg! {
+            #[cfg(vk)]
+            {
+                let result = vulkan::VkInstance::new(window_handle);
+                if let Ok((instance, surface)) = result {
+                    return Ok((Instance::Vk(instance), surface.map(Surface::Vk)));
+                }
+            }
+        }
+        mux_cfg! {
+            #[cfg(dx12)]
+            {
+                let result = dx12::Dx12Instance::new(window_handle);
+                if let Ok((instance, surface)) = result {
+                    return Ok((Instance::Dx12(instance), surface.map(Surface::Dx12)));
+                }
+            }
+        }
+        // TODO plumb creation errors through.
+        Err("No suitable instances found".into())
+    }
+
+    pub unsafe fn device(&self, surface: Option<&Surface>) -> Result<Device, Error> {
+        mux_match! { self;
+            Instance::Vk(i) => i.device(surface.map(Surface::vk)).map(Device::Vk),
+            Instance::Dx12(i) => i.device(surface.map(Surface::dx12)).map(Device::Dx12),
+        }
+    }
+
+    pub unsafe fn swapchain(
+        &self,
+        width: usize,
+        height: usize,
+        device: &Device,
+        surface: &Surface,
+    ) -> Result<Swapchain, Error> {
+        mux_match! { self;
+            Instance::Vk(i) => i
+                .swapchain(width, height, device.vk(), surface.vk())
+                .map(Swapchain::Vk),
+            Instance::Dx12(_i) => todo!(),
+        }
+    }
+}
+
+// This is basically re-exporting the backend device trait, and we could do that,
+// but not doing so lets us diverge more easily (at the moment, the divergence is
+// missing functionality).
+impl Device {
+    pub fn query_gpu_info(&self) -> GpuInfo {
+        mux_match! { self;
+            Device::Vk(d) => d.query_gpu_info(),
+            Device::Dx12(d) => d.query_gpu_info(),
+        }
+    }
+
+    pub fn create_buffer(&self, size: u64, usage: BufferUsage) -> Result<Buffer, Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.create_buffer(size, usage).map(Buffer::Vk),
+            Device::Dx12(d) => d.create_buffer(size, usage).map(Buffer::Dx12),
+        }
+    }
+
+    pub unsafe fn destroy_buffer(&self, buffer: &Buffer) -> Result<(), Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.destroy_buffer(buffer.vk()),
+            Device::Dx12(d) => d.destroy_buffer(buffer.dx12()),
+        }
+    }
+
+    pub unsafe fn create_image2d(&self, width: u32, height: u32) -> Result<Image, Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.create_image2d(width, height).map(Image::Vk),
+            Device::Dx12(d) => d.create_image2d(width, height).map(Image::Dx12),
+        }
+    }
+
+    pub unsafe fn destroy_image(&self, image: &Image) -> Result<(), Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.destroy_image(image.vk()),
+            Device::Dx12(d) => d.destroy_image(image.dx12()),
+        }
+    }
+
+    pub unsafe fn create_fence(&self, signaled: bool) -> Result<Fence, Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.create_fence(signaled).map(Fence::Vk),
+            Device::Dx12(d) => d.create_fence(signaled).map(Fence::Dx12),
+        }
+    }
+
+    pub unsafe fn wait_and_reset(&self, fences: &[&Fence]) -> Result<(), Error> {
+        mux_match! { self;
+            Device::Vk(d) => {
+                let fences = fences
+                    .iter()
+                    .copied()
+                    .map(Fence::vk)
+                    .collect::<SmallVec<[_; 4]>>();
+                d.wait_and_reset(&*fences)
+            }
+            Device::Dx12(d) => {
+                let fences = fences
+                    .iter()
+                    .copied()
+                    .map(Fence::dx12)
+                    .collect::<SmallVec<[_; 4]>>();
+                d.wait_and_reset(&*fences)
+            }
+        }
+    }
+
+    pub unsafe fn get_fence_status(&self, fence: &Fence) -> Result<bool, Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.get_fence_status(fence.vk()),
+            Device::Dx12(d) => d.get_fence_status(fence.dx12()),
+        }
+    }
+
+    pub unsafe fn create_semaphore(&self) -> Result<Semaphore, Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.create_semaphore().map(Semaphore::Vk),
+            Device::Dx12(d) => d.create_semaphore().map(Semaphore::Dx12),
+        }
+    }
+
+    pub unsafe fn pipeline_builder(&self) -> PipelineBuilder {
+        mux_match! { self;
+            Device::Vk(d) => PipelineBuilder::Vk(d.pipeline_builder()),
+            Device::Dx12(d) => PipelineBuilder::Dx12(d.pipeline_builder()),
+        }
+    }
+
+    pub unsafe fn descriptor_set_builder(&self) -> DescriptorSetBuilder {
+        mux_match! { self;
+            Device::Vk(d) => DescriptorSetBuilder::Vk(d.descriptor_set_builder()),
+            Device::Dx12(d) => DescriptorSetBuilder::Dx12(d.descriptor_set_builder()),
+        }
+    }
+
+    pub fn create_cmd_buf(&self) -> Result<CmdBuf, Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.create_cmd_buf().map(CmdBuf::Vk),
+            Device::Dx12(d) => d.create_cmd_buf().map(CmdBuf::Dx12),
+        }
+    }
+
+    pub fn create_query_pool(&self, n_queries: u32) -> Result<QueryPool, Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.create_query_pool(n_queries).map(QueryPool::Vk),
+            Device::Dx12(d) => d.create_query_pool(n_queries).map(QueryPool::Dx12),
+        }
+    }
+
+    pub unsafe fn fetch_query_pool(&self, pool: &QueryPool) -> Result<Vec<f64>, Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.fetch_query_pool(pool.vk()),
+            Device::Dx12(d) => d.fetch_query_pool(pool.dx12()),
+        }
+    }
+
+    pub unsafe fn run_cmd_bufs(
+        &self,
+        cmd_bufs: &[&CmdBuf],
+        wait_semaphores: &[&Semaphore],
+        signal_semaphores: &[&Semaphore],
+        fence: Option<&Fence>,
+    ) -> Result<(), Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.run_cmd_bufs(
+                &cmd_bufs
+                    .iter()
+                    .map(|c| c.vk())
+                    .collect::<SmallVec<[_; 4]>>(),
+                &wait_semaphores
+                    .iter()
+                    .copied()
+                    .map(Semaphore::vk)
+                    .collect::<SmallVec<[_; 4]>>(),
+                &signal_semaphores
+                    .iter()
+                    .copied()
+                    .map(Semaphore::vk)
+                    .collect::<SmallVec<[_; 4]>>(),
+                fence.map(Fence::vk),
+            ),
+            Device::Dx12(d) => d.run_cmd_bufs(
+                &cmd_bufs
+                    .iter()
+                    .map(|c| c.dx12())
+                    .collect::<SmallVec<[_; 4]>>(),
+                &wait_semaphores
+                    .iter()
+                    .copied()
+                    .map(Semaphore::dx12)
+                    .collect::<SmallVec<[_; 4]>>(),
+                &signal_semaphores
+                    .iter()
+                    .copied()
+                    .map(Semaphore::dx12)
+                    .collect::<SmallVec<[_; 4]>>(),
+                fence.map(Fence::dx12),
+            ),
+        }
+    }
+
+    pub unsafe fn read_buffer(
+        &self,
+        buffer: &Buffer,
+        dst: *mut u8,
+        offset: u64,
+        size: u64,
+    ) -> Result<(), Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.read_buffer(buffer.vk(), dst, offset, size),
+            Device::Dx12(d) => d.read_buffer(buffer.dx12(), dst, offset, size),
+        }
+    }
+
+    pub unsafe fn write_buffer(
+        &self,
+        buffer: &Buffer,
+        contents: *const u8,
+        offset: u64,
+        size: u64,
+    ) -> Result<(), Error> {
+        mux_match! { self;
+            Device::Vk(d) => d.write_buffer(buffer.vk(), contents, offset, size),
+            Device::Dx12(d) => d.write_buffer(buffer.dx12(), contents, offset, size),
+        }
+    }
+}
+
+impl PipelineBuilder {
+    pub fn add_buffers(&mut self, n_buffers: u32) {
+        mux_match! { self;
+            PipelineBuilder::Vk(x) => x.add_buffers(n_buffers),
+            PipelineBuilder::Dx12(x) => x.add_buffers(n_buffers),
+        }
+    }
+
+    pub fn add_images(&mut self, n_buffers: u32) {
+        mux_match! { self;
+            PipelineBuilder::Vk(x) => x.add_images(n_buffers),
+            PipelineBuilder::Dx12(x) => x.add_images(n_buffers),
+        }
+    }
+
+    pub fn add_textures(&mut self, n_buffers: u32) {
+        mux_match! { self;
+            PipelineBuilder::Vk(x) => x.add_textures(n_buffers),
+            PipelineBuilder::Dx12(x) => x.add_textures(n_buffers),
+        }
+    }
+
+    pub unsafe fn create_compute_pipeline<'a>(
+        self,
+        device: &Device,
+        code: ShaderCode<'a>,
+    ) -> Result<Pipeline, Error> {
+        mux_match! { self;
+            PipelineBuilder::Vk(x) => {
+                let shader_code = match code {
+                    ShaderCode::Spv(spv) => spv,
+                    // Panic or return "incompatible shader" error here?
+                    _ => panic!("Vulkan backend requires shader code in SPIR-V format"),
+                };
+                x.create_compute_pipeline(device.vk(), shader_code)
+                    .map(Pipeline::Vk)
+            }
+            PipelineBuilder::Dx12(x) => {
+                let shader_code = match code {
+                    ShaderCode::Hlsl(hlsl) => hlsl,
+                    // Panic or return "incompatible shader" error here?
+                    _ => panic!("DX12 backend requires shader code in HLSL format"),
+                };
+                x.create_compute_pipeline(device.dx12(), shader_code)
+                    .map(Pipeline::Dx12)
+            }
+        }
+    }
+}
+
+impl DescriptorSetBuilder {
+    pub fn add_buffers(&mut self, buffers: &[&Buffer]) {
+        mux_match! { self;
+            DescriptorSetBuilder::Vk(x) => x.add_buffers(
+                &buffers
+                    .iter()
+                    .copied()
+                    .map(Buffer::vk)
+                    .collect::<SmallVec<[_; 8]>>(),
+            ),
+            DescriptorSetBuilder::Dx12(x) => x.add_buffers(
+                &buffers
+                    .iter()
+                    .copied()
+                    .map(Buffer::dx12)
+                    .collect::<SmallVec<[_; 8]>>(),
+            ),
+        }
+    }
+
+    pub fn add_images(&mut self, images: &[&Image]) {
+        mux_match! { self;
+            DescriptorSetBuilder::Vk(x) => x.add_images(
+                &images
+                    .iter()
+                    .copied()
+                    .map(Image::vk)
+                    .collect::<SmallVec<[_; 8]>>(),
+            ),
+            DescriptorSetBuilder::Dx12(x) => x.add_images(
+                &images
+                    .iter()
+                    .copied()
+                    .map(Image::dx12)
+                    .collect::<SmallVec<[_; 8]>>(),
+            ),
+        }
+    }
+
+    pub fn add_textures(&mut self, images: &[&Image]) {
+        mux_match! { self;
+            DescriptorSetBuilder::Vk(x) => x.add_textures(
+                &images
+                    .iter()
+                    .copied()
+                    .map(Image::vk)
+                    .collect::<SmallVec<[_; 8]>>(),
+            ),
+            DescriptorSetBuilder::Dx12(x) => x.add_textures(
+                &images
+                    .iter()
+                    .copied()
+                    .map(Image::dx12)
+                    .collect::<SmallVec<[_; 8]>>(),
+            ),
+        }
+    }
+
+    pub unsafe fn build(
+        self,
+        device: &Device,
+        pipeline: &Pipeline,
+    ) -> Result<DescriptorSet, Error> {
+        mux_match! { self;
+            DescriptorSetBuilder::Vk(x) =>
+                x.build(device.vk(), pipeline.vk()).map(DescriptorSet::Vk),
+            DescriptorSetBuilder::Dx12(x) => x
+                .build(device.dx12(), pipeline.dx12())
+                .map(DescriptorSet::Dx12),
+        }
+    }
+}
+
+impl CmdBuf {
+    pub unsafe fn begin(&mut self) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.begin(),
+            CmdBuf::Dx12(c) => c.begin(),
+        }
+    }
+
+    pub unsafe fn finish(&mut self) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.finish(),
+            CmdBuf::Dx12(c) => c.finish(),
+        }
+    }
+
+    pub unsafe fn dispatch(
+        &mut self,
+        pipeline: &Pipeline,
+        descriptor_set: &DescriptorSet,
+        size: (u32, u32, u32),
+    ) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.dispatch(pipeline.vk(), descriptor_set.vk(), size),
+            CmdBuf::Dx12(c) => c.dispatch(pipeline.dx12(), descriptor_set.dx12(), size),
+        }
+    }
+
+    pub unsafe fn memory_barrier(&mut self) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.memory_barrier(),
+            CmdBuf::Dx12(c) => c.memory_barrier(),
+        }
+    }
+
+    pub unsafe fn host_barrier(&mut self) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.host_barrier(),
+            CmdBuf::Dx12(c) => c.host_barrier(),
+        }
+    }
+
+    pub unsafe fn image_barrier(
+        &mut self,
+        image: &Image,
+        src_layout: ImageLayout,
+        dst_layout: ImageLayout,
+    ) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.image_barrier(image.vk(), src_layout, dst_layout),
+            CmdBuf::Dx12(c) => c.image_barrier(image.dx12(), src_layout, dst_layout),
+        }
+    }
+
+    pub unsafe fn clear_buffer(&mut self, buffer: &Buffer, size: Option<u64>) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.clear_buffer(buffer.vk(), size),
+            CmdBuf::Dx12(c) => c.clear_buffer(buffer.dx12(), size),
+        }
+    }
+
+    pub unsafe fn copy_buffer(&mut self, src: &Buffer, dst: &Buffer) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.copy_buffer(src.vk(), dst.vk()),
+            CmdBuf::Dx12(c) => c.copy_buffer(src.dx12(), dst.dx12()),
+        }
+    }
+
+    pub unsafe fn copy_image_to_buffer(&mut self, src: &Image, dst: &Buffer) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.copy_image_to_buffer(src.vk(), dst.vk()),
+            CmdBuf::Dx12(c) => c.copy_image_to_buffer(src.dx12(), dst.dx12()),
+        }
+    }
+
+    pub unsafe fn copy_buffer_to_image(&mut self, src: &Buffer, dst: &Image) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.copy_buffer_to_image(src.vk(), dst.vk()),
+            CmdBuf::Dx12(c) => c.copy_buffer_to_image(src.dx12(), dst.dx12()),
+        }
+    }
+
+    pub unsafe fn blit_image(&mut self, src: &Image, dst: &Image) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.blit_image(src.vk(), dst.vk()),
+            CmdBuf::Dx12(c) => c.blit_image(src.dx12(), dst.dx12()),
+        }
+    }
+
+    pub unsafe fn reset_query_pool(&mut self, pool: &QueryPool) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.reset_query_pool(pool.vk()),
+            CmdBuf::Dx12(c) => c.reset_query_pool(pool.dx12()),
+        }
+    }
+
+    pub unsafe fn write_timestamp(&mut self, pool: &QueryPool, query: u32) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.write_timestamp(pool.vk(), query),
+            CmdBuf::Dx12(c) => c.write_timestamp(pool.dx12(), query),
+        }
+    }
+
+    pub unsafe fn finish_timestamps(&mut self, pool: &QueryPool) {
+        mux_match! { self;
+            CmdBuf::Vk(c) => c.finish_timestamps(pool.vk()),
+            CmdBuf::Dx12(c) => c.finish_timestamps(pool.dx12()),
+        }
+    }
+}
+
+impl Buffer {
+    pub fn size(&self) -> u64 {
+        mux_match! { self;
+            Buffer::Vk(b) => b.size,
+            Buffer::Dx12(b) => b.size,
+        }
+    }
+}
+
+impl Swapchain {
+    pub unsafe fn next(&mut self) -> Result<(usize, Semaphore), Error> {
+        mux_match! { self;
+            Swapchain::Vk(s) => {
+                let (idx, sem) = s.next()?;
+                Ok((idx, Semaphore::Vk(sem)))
+            }
+            Swapchain::Dx12(_s) => {
+                todo!()
+            }
+        }
+    }
+
+    pub unsafe fn image(&self, idx: usize) -> Image {
+        mux_match! { self;
+            Swapchain::Vk(s) => Image::Vk(s.image(idx)),
+            Swapchain::Dx12(_s) => todo!(),
+        }
+    }
+
+    pub unsafe fn present(
+        &self,
+        image_idx: usize,
+        semaphores: &[&Semaphore],
+    ) -> Result<bool, Error> {
+        mux_match! { self;
+            Swapchain::Vk(s) => s.present(
+                image_idx,
+                &semaphores
+                    .iter()
+                    .copied()
+                    .map(Semaphore::vk)
+                    .collect::<SmallVec<[_; 4]>>(),
+            ),
+            Swapchain::Dx12(_s) => todo!(),
+        }
+    }
+}
diff --git a/piet-gpu-hal/src/vulkan.rs b/piet-gpu-hal/src/vulkan.rs
index b93b0c0..619c6bd 100644
--- a/piet-gpu-hal/src/vulkan.rs
+++ b/piet-gpu-hal/src/vulkan.rs
@@ -10,6 +10,8 @@ use ash::extensions::{ext::DebugUtils, khr};
 use ash::version::{DeviceV1_0, EntryV1_0, InstanceV1_0, InstanceV1_1};
 use ash::{vk, Device, Entry, Instance};
 
+use smallvec::SmallVec;
+
 use crate::{
     BufferUsage, Device as DeviceTrait, Error, GpuInfo, ImageLayout, SamplerParams, SubgroupSize,
 };
@@ -617,16 +619,21 @@ impl crate::Device for VkDevice {
         Ok(device.create_semaphore(&vk::SemaphoreCreateInfo::default(), None)?)
     }
 
-    unsafe fn wait_and_reset(&self, fences: &[Self::Fence]) -> Result<(), Error> {
+    unsafe fn wait_and_reset(&self, fences: &[&Self::Fence]) -> Result<(), Error> {
         let device = &self.device.device;
-        device.wait_for_fences(fences, true, !0)?;
-        device.reset_fences(fences)?;
+        let fences = fences
+            .iter()
+            .copied()
+            .copied()
+            .collect::<SmallVec<[_; 4]>>();
+        device.wait_for_fences(&fences, true, !0)?;
+        device.reset_fences(&fences)?;
         Ok(())
     }
 
-    unsafe fn get_fence_status(&self, fence: Self::Fence) -> Result<bool, Error> {
+    unsafe fn get_fence_status(&self, fence: &Self::Fence) -> Result<bool, Error> {
         let device = &self.device.device;
-        Ok(device.get_fence_status(fence)?)
+        Ok(device.get_fence_status(*fence)?)
     }
 
     unsafe fn pipeline_builder(&self) -> PipelineBuilder {
@@ -709,8 +716,8 @@ impl crate::Device for VkDevice {
     unsafe fn run_cmd_bufs(
         &self,
         cmd_bufs: &[&CmdBuf],
-        wait_semaphores: &[Self::Semaphore],
-        signal_semaphores: &[Self::Semaphore],
+        wait_semaphores: &[&Self::Semaphore],
+        signal_semaphores: &[&Self::Semaphore],
         fence: Option<&Self::Fence>,
     ) -> Result<(), Error> {
         let device = &self.device.device;
@@ -722,16 +729,28 @@ impl crate::Device for VkDevice {
         let wait_stages = wait_semaphores
             .iter()
             .map(|_| vk::PipelineStageFlags::ALL_COMMANDS)
-            .collect::<Vec<_>>();
-        // Use SmallVec or similar here to reduce allocation?
-        let cmd_bufs = cmd_bufs.iter().map(|c| c.cmd_buf).collect::<Vec<_>>();
+            .collect::<SmallVec<[_; 4]>>();
+        let cmd_bufs = cmd_bufs
+            .iter()
+            .map(|c| c.cmd_buf)
+            .collect::<SmallVec<[_; 4]>>();
+        let wait_semaphores = wait_semaphores
+            .iter()
+            .copied()
+            .copied()
+            .collect::<SmallVec<[_; 2]>>();
+        let signal_semaphores = signal_semaphores
+            .iter()
+            .copied()
+            .copied()
+            .collect::<SmallVec<[_; 2]>>();
         device.queue_submit(
             self.queue,
             &[vk::SubmitInfo::builder()
                 .command_buffers(&cmd_bufs)
-                .wait_semaphores(wait_semaphores)
+                .wait_semaphores(&wait_semaphores)
                 .wait_dst_stage_mask(&wait_stages)
-                .signal_semaphores(signal_semaphores)
+                .signal_semaphores(&signal_semaphores)
                 .build()],
             fence,
         )?;
@@ -1287,14 +1306,15 @@ impl VkSwapchain {
     pub unsafe fn present(
         &self,
         image_idx: usize,
-        semaphores: &[vk::Semaphore],
+        semaphores: &[&vk::Semaphore],
     ) -> Result<bool, Error> {
+        let semaphores = semaphores.iter().copied().copied().collect::<SmallVec<[_; 4]>>();
         Ok(self.swapchain_fn.queue_present(
             self.present_queue,
             &vk::PresentInfoKHR::builder()
                 .swapchains(&[self.swapchain])
                 .image_indices(&[image_idx as u32])
-                .wait_semaphores(semaphores)
+                .wait_semaphores(&semaphores)
                 .build(),
         )?)
     }
diff --git a/piet-gpu/bin/android.rs b/piet-gpu/bin/android.rs
index 3bf3ca6..63ee91a 100644
--- a/piet-gpu/bin/android.rs
+++ b/piet-gpu/bin/android.rs
@@ -12,7 +12,7 @@ use ndk::native_window::NativeWindow;
 use ndk_glue::Event;
 
 use piet_gpu_hal::hub;
-use piet_gpu_hal::vulkan::{QueryPool, VkInstance, VkSurface, VkSwapchain};
+use piet_gpu_hal::mux::{QueryPool, Instance, Surface, Swapchain};
 use piet_gpu_hal::{CmdBuf, Error, ImageLayout};
 
 use piet_gpu::{render_scene, PietGpuRenderContext, Renderer};
@@ -30,7 +30,7 @@ struct MyHandle {
 struct GfxState {
     session: hub::Session,
     renderer: Renderer,
-    swapchain: VkSwapchain,
+    swapchain: Swapchain,
     current_frame: usize,
     last_frame_idx: usize,
     submitted: Option<hub::SubmittedCmdBuf>,
@@ -52,7 +52,7 @@ fn my_main() -> Result<(), Error> {
                     let window = ndk_glue::native_window();
                     if let Some(window) = &*window {
                         let handle = get_handle(window);
-                        let (instance, surface) = VkInstance::new(Some(&handle))?;
+                        let (instance, surface) = Instance::new(Some(&handle))?;
                         gfx_state = Some(GfxState::new(&instance, surface.as_ref())?);
                     } else {
                         println!("native window is sadly none");
@@ -90,7 +90,7 @@ unsafe impl HasRawWindowHandle for MyHandle {
 }
 
 impl GfxState {
-    fn new(instance: &VkInstance, surface: Option<&VkSurface>) -> Result<GfxState, Error> {
+    fn new(instance: &Instance, surface: Option<&Surface>) -> Result<GfxState, Error> {
         unsafe {
             let device = instance.device(surface)?;
             let mut swapchain =
@@ -151,7 +151,7 @@ impl GfxState {
 
             // Image -> Swapchain
             cmd_buf.image_barrier(&swap_image, ImageLayout::Undefined, ImageLayout::BlitDst);
-            cmd_buf.blit_image(self.renderer.image_dev.vk_image(), &swap_image);
+            cmd_buf.blit_image(self.renderer.image_dev.mux_image(), &swap_image);
             cmd_buf.image_barrier(&swap_image, ImageLayout::BlitDst, ImageLayout::Present);
             cmd_buf.finish();
 
@@ -159,15 +159,15 @@ impl GfxState {
                 self.session
                     .run_cmd_buf(
                         cmd_buf,
-                        &[acquisition_semaphore],
-                        &[self.present_semaphores[frame_idx]],
+                        &[&acquisition_semaphore],
+                        &[&self.present_semaphores[frame_idx]],
                     )
                     .unwrap(),
             );
             self.last_frame_idx = frame_idx;
 
             self.swapchain
-                .present(image_idx, &[self.present_semaphores[frame_idx]])
+                .present(image_idx, &[&self.present_semaphores[frame_idx]])
                 .unwrap();
 
             self.current_frame += 1;
diff --git a/piet-gpu/bin/cli.rs b/piet-gpu/bin/cli.rs
index 69e0fc4..a1d766d 100644
--- a/piet-gpu/bin/cli.rs
+++ b/piet-gpu/bin/cli.rs
@@ -5,8 +5,8 @@ use std::path::Path;
 use clap::{App, Arg};
 
 use piet_gpu_hal::hub;
-use piet_gpu_hal::vulkan::VkInstance;
-use piet_gpu_hal::{BufferUsage, CmdBuf, Error};
+use piet_gpu_hal::mux::Instance;
+use piet_gpu_hal::{BufferUsage, Error};
 
 use piet_gpu::{render_scene, render_svg, PietGpuRenderContext, Renderer, HEIGHT, WIDTH};
 
@@ -225,7 +225,7 @@ fn main() -> Result<(), Error> {
                 .takes_value(true),
         )
         .get_matches();
-    let (instance, _) = VkInstance::new(None)?;
+    let (instance, _) = Instance::new(None)?;
     unsafe {
         let device = instance.device(None)?;
         let session = hub::Session::new(device);
@@ -258,7 +258,7 @@ fn main() -> Result<(), Error> {
 
         cmd_buf.begin();
         renderer.record(&mut cmd_buf, &query_pool);
-        cmd_buf.copy_image_to_buffer(renderer.image_dev.vk_image(), image_buf.vk_buffer());
+        cmd_buf.copy_image_to_buffer(renderer.image_dev.mux_image(), image_buf.mux_buffer());
         cmd_buf.host_barrier();
         cmd_buf.finish();
         let start = std::time::Instant::now();
diff --git a/piet-gpu/bin/winit.rs b/piet-gpu/bin/winit.rs
index 5eb5be6..af60a3a 100644
--- a/piet-gpu/bin/winit.rs
+++ b/piet-gpu/bin/winit.rs
@@ -1,6 +1,6 @@
 use piet_gpu_hal::hub;
-use piet_gpu_hal::vulkan::VkInstance;
-use piet_gpu_hal::{CmdBuf, Error, ImageLayout};
+use piet_gpu_hal::mux::Instance;
+use piet_gpu_hal::{Error, ImageLayout};
 
 use piet_gpu::{render_scene, PietGpuRenderContext, Renderer, HEIGHT, WIDTH};
 
@@ -22,7 +22,7 @@ fn main() -> Result<(), Error> {
         .with_resizable(false) // currently not supported
         .build(&event_loop)?;
 
-    let (instance, surface) = VkInstance::new(Some(&window))?;
+    let (instance, surface) = Instance::new(Some(&window))?;
     unsafe {
         let device = instance.device(surface.as_ref())?;
         let mut swapchain =
@@ -103,21 +103,21 @@ fn main() -> Result<(), Error> {
                         ImageLayout::Undefined,
                         ImageLayout::BlitDst,
                     );
-                    cmd_buf.blit_image(renderer.image_dev.vk_image(), &swap_image);
+                    cmd_buf.blit_image(renderer.image_dev.mux_image(), &swap_image);
                     cmd_buf.image_barrier(&swap_image, ImageLayout::BlitDst, ImageLayout::Present);
                     cmd_buf.finish();
 
                     submitted = Some(session
                         .run_cmd_buf(
                             cmd_buf,
-                            &[acquisition_semaphore],
-                            &[present_semaphores[frame_idx]],
+                            &[&acquisition_semaphore],
+                            &[&present_semaphores[frame_idx]],
                         )
                         .unwrap());
                     last_frame_idx = frame_idx;
 
                     swapchain
-                        .present(image_idx, &[present_semaphores[frame_idx]])
+                        .present(image_idx, &[&present_semaphores[frame_idx]])
                         .unwrap();
 
                     current_frame += 1;
diff --git a/piet-gpu/src/lib.rs b/piet-gpu/src/lib.rs
index 3c87f68..2b8ed0f 100644
--- a/piet-gpu/src/lib.rs
+++ b/piet-gpu/src/lib.rs
@@ -13,7 +13,8 @@ use piet::{Color, ImageFormat, RenderContext};
 use piet_gpu_types::encoder::Encode;
 
 use piet_gpu_hal::hub;
-use piet_gpu_hal::{BufferUsage, CmdBuf, Error, ImageLayout};
+use piet_gpu_hal::hub::ShaderCode;
+use piet_gpu_hal::{BufferUsage, Error, ImageLayout};
 
 use pico_svg::PicoSvg;
 
@@ -292,35 +293,35 @@ impl Renderer {
         let memory_buf_dev = session.create_buffer(128 * 1024 * 1024, dev)?;
         memory_buf_host.write(&[alloc as u32, 0 /* Overflow flag */])?;
 
-        let el_code = include_bytes!("../shader/elements.spv");
+        let el_code = ShaderCode::Spv(include_bytes!("../shader/elements.spv"));
         let el_pipeline = session.create_simple_compute_pipeline(el_code, 4)?;
         let el_ds = session.create_simple_descriptor_set(
             &el_pipeline,
             &[&memory_buf_dev, &config_buf, &scene_buf, &state_buf],
         )?;
 
-        let tile_alloc_code = include_bytes!("../shader/tile_alloc.spv");
+        let tile_alloc_code = ShaderCode::Spv(include_bytes!("../shader/tile_alloc.spv"));
         let tile_pipeline = session.create_simple_compute_pipeline(tile_alloc_code, 2)?;
         let tile_ds = session
             .create_simple_descriptor_set(&tile_pipeline, &[&memory_buf_dev, &config_buf])?;
 
-        let path_alloc_code = include_bytes!("../shader/path_coarse.spv");
+        let path_alloc_code = ShaderCode::Spv(include_bytes!("../shader/path_coarse.spv"));
         let path_pipeline = session.create_simple_compute_pipeline(path_alloc_code, 2)?;
         let path_ds = session
             .create_simple_descriptor_set(&path_pipeline, &[&memory_buf_dev, &config_buf])?;
 
-        let backdrop_alloc_code = include_bytes!("../shader/backdrop.spv");
+        let backdrop_alloc_code = ShaderCode::Spv(include_bytes!("../shader/backdrop.spv"));
         let backdrop_pipeline = session.create_simple_compute_pipeline(backdrop_alloc_code, 2)?;
         let backdrop_ds = session
             .create_simple_descriptor_set(&backdrop_pipeline, &[&memory_buf_dev, &config_buf])?;
 
         // TODO: constants
-        let bin_code = include_bytes!("../shader/binning.spv");
+        let bin_code = ShaderCode::Spv(include_bytes!("../shader/binning.spv"));
         let bin_pipeline = session.create_simple_compute_pipeline(bin_code, 2)?;
         let bin_ds =
             session.create_simple_descriptor_set(&bin_pipeline, &[&memory_buf_dev, &config_buf])?;
 
-        let coarse_code = include_bytes!("../shader/coarse.spv");
+        let coarse_code = ShaderCode::Spv(include_bytes!("../shader/coarse.spv"));
         let coarse_pipeline = session.create_simple_compute_pipeline(coarse_code, 2)?;
         let coarse_ds = session
             .create_simple_descriptor_set(&coarse_pipeline, &[&memory_buf_dev, &config_buf])?;
@@ -328,10 +329,10 @@ impl Renderer {
         let bg_image = Self::make_test_bg_image(&session);
 
         let k4_code = if session.gpu_info().has_descriptor_indexing {
-            &include_bytes!("../shader/kernel4_idx.spv")[..]
+            ShaderCode::Spv(include_bytes!("../shader/kernel4_idx.spv"))
         } else {
             println!("doing non-indexed k4");
-            &include_bytes!("../shader/kernel4.spv")[..]
+            ShaderCode::Spv(include_bytes!("../shader/kernel4.spv"))
         };
         // This is an arbitrary limit on the number of textures that can be referenced by
         // the fine rasterizer. To set it for real, we probably want to pay attention both
@@ -386,13 +387,13 @@ impl Renderer {
 
     pub unsafe fn record(&self, cmd_buf: &mut hub::CmdBuf, query_pool: &hub::QueryPool) {
         cmd_buf.copy_buffer(
-            self.memory_buf_host.vk_buffer(),
-            self.memory_buf_dev.vk_buffer(),
+            self.memory_buf_host.mux_buffer(),
+            self.memory_buf_dev.mux_buffer(),
         );
-        cmd_buf.clear_buffer(self.state_buf.vk_buffer(), None);
+        cmd_buf.clear_buffer(self.state_buf.mux_buffer(), None);
         cmd_buf.memory_barrier();
         cmd_buf.image_barrier(
-            self.image_dev.vk_image(),
+            self.image_dev.mux_image(),
             ImageLayout::Undefined,
             ImageLayout::General,
         );
@@ -451,7 +452,7 @@ impl Renderer {
         cmd_buf.write_timestamp(&query_pool, 7);
         cmd_buf.memory_barrier();
         cmd_buf.image_barrier(
-            self.image_dev.vk_image(),
+            self.image_dev.mux_image(),
             ImageLayout::General,
             ImageLayout::BlitSrc,
         );
@@ -475,12 +476,16 @@ impl Renderer {
             let mut cmd_buf = session.cmd_buf()?;
             cmd_buf.begin();
             cmd_buf.image_barrier(
-                image.vk_image(),
+                image.mux_image(),
                 ImageLayout::Undefined,
                 ImageLayout::BlitDst,
             );
-            cmd_buf.copy_buffer_to_image(buffer.vk_buffer(), image.vk_image());
-            cmd_buf.image_barrier(image.vk_image(), ImageLayout::BlitDst, ImageLayout::General);
+            cmd_buf.copy_buffer_to_image(buffer.mux_buffer(), image.mux_image());
+            cmd_buf.image_barrier(
+                image.mux_image(),
+                ImageLayout::BlitDst,
+                ImageLayout::General,
+            );
             cmd_buf.finish();
             // Make sure not to drop the buffer and image until the command buffer completes.
             cmd_buf.add_resource(&buffer);